# ASGCT 2026 — AI/ML AAV Engineering: Master Knowledge Document Generated: 2026-08-06 20:33 Source: /Users/rtscheliessnig/workspace-active/parvotec/Machine learning for Rupert/ASGCT2026 --- ## Index - [AAV Engineering III](#aav-engineering-iii) - [AAV Engineering IV](#aav-engineering-iv) - [AAV Trafficking](#aav-trafficking) - [Lir AAV LLM](#lir-aav-llm) - [ShapeTX AAV5engineering](#shapetx-aav5engineering) - [TuningReceptorInteractions Caltech](#tuningreceptorinteractions-caltech) - [ASGCT recap Georg](#asgct-recap-georg) --- ## AAV Engineering III {#aav-engineering-iii} **File:** `AAV_Engineering_III.mp4` **Type:** Video transcript (Whisper medium) ### Pre-analysis AAV Engineering III session. Expected: ML-guided capsid engineering, directed evolution surrogates, multi-trait optimisation approaches. ### Full text / transcript Introduce our first speaker Gang Wang Which will be sharing about the directed evolution in non-human primates which identifies the AAV capsids With the novel motive enabling efficient muscle targeting and reduced liver liver tropism Just before Gang starts just a reminder that no flash photography is a allowed and If you are Haven't connected to the best in this wall room the password is ASGCT 2026 and without any further ado welcome Gang Thank you for the introduction Good morning, everyone I'm Gang Wang from beyond gene therapeutics. It's my honor to be here today I will introduce our capsid and aneurium platform and we identified a muscle targeting motif by this by this platform and This motif can be effectively tuned for optimal liberty targeting All the these findings come from our platform Dynolib we have Capsid libraries we have Calibrator calibrated Barcoded libraries and Invaluable validation system combined with the capture the signal recovery system Signal recovery system this pipeline allows us to find that the clinical lead Parasites our Capacity the library we have two distinct kind of library one is a random peptide insertion and We are eight of AV non capsaicin another one is rationally designed peptide derived from cell surface binding protein and We use do DNA packaging system this test You the very high viral titer of library and this also achieves very good high fidelity packaging with very low backbone contamination and excellent consistency of phenotype and genotype Calibrator is our calibrated barcoded library we use AV non as internal standards with the different doses and then we We create Center the curve based on the dosage of AV non and the in vivo Tissue Expression and then all variants are mapped to these curve to calculate normalized for the change against the AV non In camel we labeled the different variants with distinct Epitope attacks for multi-level capsid bench barking in a single animal this is our Signal recovery platform we named a capture we use magnetic beads Conjected with virus specific probe to recover a viral MRI and These my third facilitate the RT preparation from large amount of NHP tissues This is our integrated dynamic live we integrated multi-library multi-root and the multi-tissue screening or in a single Single NHP model the aim is to maximize the data output per NHP This strategy is very productive From intravenous delivery we identified mild cap series targeting to muscle and the lead candidate targeting to deep brain From intra sequel delivery we identified the thin cap series Which can transduce almost all an HP cooking cells now we move to our discovery of Muscle Targeted targeting motif and this is the first round Invaluable screening as this early stage stage we didn't found the linear reliable linear motif, but we Found a very strong arginine Arginine accumulation across all clusters In the second round of screening we as the selection pressure increase our R6 Arginine at the sixth position Accumulated dramatically and And we then analysis the round of the R6 anchor revealed that dominant PRT motif across all the Muscles and the heart and then we reclassifying Fix the R and G and we confirmed that it's a PRT motif And then we verified this motif in different species and we select four PRT variants and Packages them individually label them with different IP protects and mix them together and then the meter to animal by intravenous injection and we tested them in Resus markup As the rest and the two species of mice and all the variants a PRT variants show a high a higher muscle inspirations a b9 and the best one is be why oh therapy By or 87 and then we further Before they're validated the be why oh a seven in a risk markup And I to the dose of five times the tend to the power of for 12 We see per KG be why oh a seven out forms every now across DNA MRI and protein levels and you know all tested muscles Since we have already Verified the PRT motif then we Did the optimization of this PRT motif and the structure modeling by Alpha fold reveals that there is Distinct protrusion introduced by as a our TPRG Insertion of be why oh it's a win we hypothesis Optimized the Terminal N terminal over the PRG motif could be maximized muscle potency of the PRG so we established a next round of HP screening by by NK saturation library and The correlation analysis of the NK library screening revealed Revealed that there is a very strong positive correlation between muscle inspiration and the liver expression first we identified the Consensus sequence for Muscle tropism it is s a x y PRG and then we focus on the life of the lifter population as shown in the right color with high muscle expression and low liver expression compared to the background We found a distinct Hissing in accumulation in the position in the second position and The fourth position we named it the H switch And then we developed our next generation my cap my cap is from white live direct inclusion and structure informed design using H switch and Also the AI guide is carried which is the panned our Candidate finally we Finally we got 32 unique next generation Variants and we tested them using our calibrator platform The result shows most of the next generation variants show a higher muscle expression than BY 087 and the H switch The H switch works very well in these variants When the double H incorporation Achieves very high Very high muscle liver ratio Around the 32 times and 23 times It's very it's much hair than the BY 087 It is only 3.8 fold and the single H incorporation is also very effective It reached the ratio of 17 fold And then we selected the two kind of variants the blue one is the muscle potency and the right one is the liberty targeting group and For the for the right group is much the published a mile away and This is the BY 438 resulting in calibrator and the linear we can show we can see the linear regulation of AV 9 and Also, we can see the muscle expression hair for the BY 488 much higher than And with liver like license price in the liver and then we make How to head comparison between BY 438 and AV 9 and We can see the BY 483 expressed much hair than much hair MRI in In muscles and liver and only 40% MRI expression in liver compared to AV 9 the protein I have standing conformed protein expression consistent with MRI expression. This is a last presented slide we Did cross species validation in San Amogus for myocap variants and H switch works very well in San Amogus Especially for the double H incorporation the as shown in red color the auto perform Published Control myocap and the MVG7 for conclusion and then of leave but then only way is an end-to-end pipeline for capsaid discovery and the PRT motif for so as a strong muscle targeting motif while flanking H residue actually as Precise switch to abolish liver of targeting the myocap series like BY 438 Demonstrate Expansion of muscle muscle Transduction and hepatic escape in HP serving as ideal vectors for muscle related in therapy. Thank you Thank you Thank you, do we have any questions for time? We have one minute I'm sorry. I may have missed it. Did you mention the receptor that you believe this is binding to? No, we are trying to find as a receptor Unfortunately until now we did not find it I Am nice talk you mentioned that you use dose calibration curve in your libraries I'm curious how you use that data on the other end and what you've learned about dose response of various like AV variants You mean the mechanism of our calibrator Yeah, like how you use the calibrator data? we have algorithm calibrated algorithm and We Use the change fold change of fold to represent the Potency of each variant we we get to the the Expression tissue is the prision data put them my put them into the center curve and then we will calculate the fold change compared the center the curve that means the increase fold against the AV nine Thank you. Thank you. Thank you Thank you with that I would like to invite our next speaker Benjamin and he is joint collaboration of the Therapeutics and the Department of Cellular and Molecular Medicine in Belgium. So Thank you, it's It's nice being here So today I did I want to tell you a little bit about what we have been doing with Tavira over the last year And we've been working on a novel capsic platform, which we which we call the Dave platform some of my disclosures So I think I'm preaching to the choir here when I say that AV development this is kind of hindered by the specificity and the potency of AV vectors and also the manufacturing can be sometimes always quite complex. So at Tavira we we wondered Okay, can we make this more streamlined and more more more directed? By zip coding AVs to send them directly to specific tissues in a rational way And the way we did that is through what we call our Dave platform stands for decorated AVs What we basically did is we started with a parental AV serotype and in our case We checked AV to an AV nine, but we went ahead with AV nine and we did some minimal vector engineering What we did was we added in the surface exposed loops We added a recognition tag for an enzyme called sort ace a and what happens is this sort days can recognize this stack And then when it does so it can attach targeting groups to the caps it and these targeting groups then afford specificity and potency of the vector So as you see here this tag is in all the all the viral proteins of the caps So here a little animation how it works you see the tag is in in the pink loops The sort is recognizes it and a targeting group in this case I'm just showing a VHH domain, but can be anything comes in recognizes it is Covalently attached to the caps So how do we make it so basically we just make the AV first and we go to her up and downstream manufacturing then we do the conjugation in What we sometimes refer to as a shake and bake reaction you just incubate it for a given amount of time at room temperature After which you purify out the the enzyme and the non-reactive targeting groups And that affords a very strong platform modularity as I said we can do with AV to with AV nine we can target Several surface exposed loops some of them work better than the other and we can target anything from small molecules to to VHH domains on those caps it's and that makes that we have a predictable and reproducible product so we have confirmed the structure of the start these conjugated Molecules using mass spec we can quantify them using Western blot and Eliza essays And when we do transaction assays we see in vitro that it is nice and reproducible So the first thing we did For just for the heck of it. I guess was conjugating biotin to it and what you can see here on the left hand side That is a non conjugated AV vector you see the VP one to three bands And then when we do the conjugation with biotin if we just stain for the viral protein we have basically the C terminal fragment so that is downstream of the Of the recognition side that you see here at the bottom And that's the seat the sort is cleaved fragments so that confirms the conjugation But then when we also look at the biotin staining that is attached to the N terminal Fragment of the viral protein you see that we conjugate both VP one VP two and VP three of the caps So that's all really good, but we don't care too much about biotin So we want to do cell targeting and again in the early days of this platform when we just wanted to validate this We used to her to binding VHH and what you can see is in her to negative cells We barely have any transduction. That's because we do the specific engineering that makes the caps it basically not transduce so well in In any cell type but then when we conjugate the her to VHH we have specific targeting Again her to not super interesting for us. What we wanted to do is we wanted to reach the CNS and We did that with obviously the transfer interceptor what so many others are doing as well Trying to piggyback on that receptor shuttling across the endothelial cells to get into the brain prank in my And That's what you can see here We started off with a mouse transfer receptor in in mouse transfer receptor over expressing cells but crucially We also went into human transfer receptor positive cells and you can see that compared to in a v9 Our constructs have at least in vitro an 18 fold improved transduction efficiency, of course Again, we don't want to transduce HD of our positive cells. We want to cross the BBB And so we did an in vivo by distribution study So what we did is just with a GFP reporter gene and human transfer in knocking human transfer receptor knocking mice At those of 5e 11 VG per mouse. We injected this IV and we looked at the bio distribution So on the left here you see a v9 That's kind of typical of what we would see at this dose But then when we took our lead CNS Dave caps it you see that we have nice and widespread brain distribution Including the deeper brain areas and if we zoom in a little bit here in the top row You can see a v9 with a new end counter stain for neurons and in the bottom row you can see our Dave caps it and again, you see that Almost all brain regions are efficiently transduced of course We want to quantify this so here on the left hand side You see the VG copies in the brain and you see that we have two logs improvement in brain targeting And then when we look at the neuronal transduction You see that we have anywhere depending on the brain region between 20 and 80 percent of the neurons transduced which I think is is Quite important and strongly improved relative to for instance a v9 Very importantly the liver of course if we go systemic With many caps it's out there. You see they end up in the liver associated with liver talks Our Dave caps it's are again to log Detargeted from from the liver on the VG level And that's something that is inherent to our platform So we see the same effect if we don't conjugate So this is really due to how we do the engineering in the caps and basically these hundredfold improvement in Brain targeting and hundredfold reduction in liver targeting opens up that therapeutic window with four logs So that's that's that's on CNS But of course this is platform technology. We don't only want to target CNS. We want to go more broadly as well There's a large need for caps is that target kidney. There's a number of Large number of genetic kidney diseases with devastating outcomes and very limited treatment options So we said, okay can we make Vectors AV vectors that more strongly targets cell types in the in the kidney And we focused on a little bit of podocytes. I'm not showing you that work today But also we focus on tubular epithelial cells With a vision of treating tubulopathies down the line And What you can see here, this is just in vitro data. We're still working on the in vivo work But you see that again here compared to an AV nine in cell culture We have that 20 fold improvement in in cell targeting With our ligand tree here, so we're still improving on that. We think that That we haven't reached the ceiling. We can do some engineering there on both the caps it and the ligand level to Improve that and at the same time Tavira is planning to to put this in an in vivo study So to summarize the Dave platform that we have developed Gives a a modular design for for designing potents, but also very specific AV caps it's it's something this post post-manufactured conjugation technologies And I think we will hear some more about it in this session That that has a strong potential to to rationally engineer AV caps It's a very powerful technology To to rationally engineer AV caps it's also in a human targeted context Our platform can do that by enzymatic conjugation With an inherent liberty targeting We have strong proof of concept for cns targeting and then In vitro proof of concept in kidney targeting which we will further work on Of course, this is work that we haven't done alone. We developed the Dave platform was initially developed at At the university of leuven which we whom we have a collaboration and With that I would like to Thank you for your attention and take any question you have Thank you benjamin and so floor is open for questions Great talk real simple system Are you disclosing just how many of the subunits could look are either? Contain your targeting loop and then after conjugation contain an actual conjugation event We're not disclosing it, but I don't want to disappoint you. So i'll say it's not 60 and it's not zero. So we know that No, the reason is I say if if we go too high and that's what others have seen as well Then you end up with a particle. There's not really an av It's an av covered in a number of molecules and probably it doesn't traffic as well in the cell So we see that we can go quite low And it depends a bit on the system and on the ligand that you attach and so on and so forth But the exact number is currently not disclosed Hi, my name is liu jang song from ohio, sorry Okay The retina is part of the cns Did you look and see if you get any targeting in the retina? Sorry come again the retina is part of the central nervous system. We did not look at the retina. No Okay, thank you, uh, really wonderful talk very impressive data so i'm just curious you mentioned that the conjugation was uh, Specifically to the vr1 vr4 and h i'm just curious is the conjugation how can you like specifically? Conjugate to a specific loop or it is actually not specific So we do the we add the tag By which that so the sort is recognized attack. It's five amino acids and we add that either to loop four or to loop eight Okay, so the tag is in the genome av genome packaging. Yes. The tag is in is in the genetic sequence of the caps. Thank you very much Hi great talk So I have a quick question about um how your retargeting impacts the secondary interaction with avr Have you looked at that and you know if you're forcing av through different like Uptake pathways once you've initially bound through transferring receptor or whatever we have not looked at that No, it's something that has come up in the past In one of the earlier versions of the capsid that we had we saw very funky results And then we were thinking somewhat the the interaction with avr or other receptors might and especially the trafficking as a whole Might be impacted and then we did some reengineering solve the problem and didn't look back We haven't looked at it with this specific version of the caps. Thank you very much Uh does your hftr tfr targeted capsid also recognize the center receptor? um Yes, although um, and that's the case with with with any, uh Capsid we know that there's a gap between the affinity for instance for sino and human so it does recognize So we are pretty comfortable going into nhp studies With the caveat that of course It's a different system. So but it does recognize yes. Thanks Thank you great talk just a quick question Why do you think the in cns like deep brain structures such as thalamus being targeted by the conjugated? It's highly vascularized Since we have a little bit of time I have my own questions so How do you remove the? Unbound or the if there is something in there in the system Not the yeah, so so after the conjugation the Purification process. Yes. So after the conjugation, which is just an incubation We do a purification system currently. We were using small scales. So we're using Something like a dialysis or a size exclusion But of course at larger scales we we intend to do it with a tff process for instance or an extra affinity captures that Okay. Thank you engine Great talk Question on the synomologous macaques. Do you have data? No, not yet. That's in the making So we have data on binding how the how the ligand itself binds to the to the sign receptor but no envy for data on that Yeah So I have two questions first one about the stability of the capsid after these conjugations You have any idea what the stability of the capsid is? Any idea about that and also The surface modifications could introduce like new immunogenic epitopes. So are you planning to do like any immunogenic studies? Yeah, so the first one was the um The stability so what we have seen is sometimes inserting these loops destabilizes the The capsid prior to conjugation. So those are variants that we discarded So we assume that the quaternary structure still holds also because we don't fully conjugate the the whole particle So there's only a number of sites that are conjugated But we haven't measured it directly the second thing is the epitopes the Immune The So basically what we add is is just these five amino acids, which is lp et g We don't assume we assume they're not immunogenic and they they haven't been shown and within other systems to be immunogenic and then we just have the the Nanobody for instance or the targeting group that is there. So there's there's a There's a lot of data on the surface that we can't get to So we assume that the surface is not immunogenic The nanobody for instance or the targeting group that is there so there's the sort is is fully out It's not attached to the to the capsid anymore, but we will measure that in our nhp studies for sure Okay, one last question Uh, you show a series of different av conjugates. Are they different? Nanobodies or just different formulations same nanobody They are different ligands They are different ligands You're referring to the For example for the For the neural lines you For instance here ligand one two and three they are different ligands Oh, okay Thank you Benjamin Moving on to our next question Moving on to our third speaker We welcome luizia She will be talking about the uncoupling capsid structure functional performance and clinical translation of aav virans Well, thank you and as always I want to thank the organizers for the opportunity to present our work My name is lucian song. I've been presenting on behalf of the team So I will be talking about Uncoupling capsid structure functional performance and the kinetic translation of av virans very long title So here are my disclosures So To engineer capsid we have so many different strategies that could be used right so we could do Site-directed mutagenesis we could do shuffling we could do A lot of things error prompt pcai machine learning director evolution But whatever strategies we are using usually we share a common goal It's very similar that we want our capsid to be Detargeted from the liver and we want the capsid to cross a certain Biologic barrier better or we wanted the capsid to target to transduce our target cells or tissues better at higher efficiency This is the common goal, but today I don't want to share another Exciting capsid engineering success story. Instead. I wanted to share some of our observations on a few candidates that In our capsid engineering process and hopefully that could provide you a little bit additional perspective Or considerations when we think about capsid engineering and the kinetic translation So I'd like to Hear this slide So here this slide introduces you introduce all the Capsids that I'll be discussing throughout this presentation So firstly AV9 Obviously is our parent capsid and This is our benchmark control and the AV9 02 03 04 and These are the in-house engineered capsid variants So these variants were engineered either including some mutations on the VR1 region or it has the Peptide insertion at the VRH region or a combination of both and all these peptide candidates are Identified in our previous NHP directory evolution libraries So our Structure specialist Structure specialist analyze the structure Like everyone do so as you can see the VP3 unit structure is Usually highly conserved and the VR8 structure actually is very flexible where you insert the peptide So the structure is a little bit difficult to predict accurately Accurately so our structure specialist also analyze the surface Potential charges those charges so if we Basically the red Stands for their positive charge and the blue stands for the negative charge and the white or gray It stands for the neutral charge. So So If we subtract the AV9 from the variants you can see There are some differences Become more obvious that the variants introduced on the surface of the capsid And if we take a closer look you can see at the trifold at the straight fold axis There will be show some different pattern across different candidates and one of the thing is interesting on the Bottom right middle panel for the AV03 you can see there is a strong positive patch at the straight fold so one When we work on the capsid is very naturally to Speculate or imagine maybe this could do something so well, maybe this kind of feature could affect their AV life cycle or some type of 12% something but now I will show you how How does each of these capsid perform in the animal models? so just for your information that We actually have another talk. We are happy right after this session where Dr. Judith Smusky will present our recent Clinical data if you are interested in so here i'm just showing the animal data so first the AV901 This is a very busy slide I want to draw your attention to the this bar graph and you can see hopefully The second of second bar represent the liver so at the DNA level you can see This capsid appears to be a liver-detoxicated Capsid and you can see all the bars which represent for the brain different Part of the brain and the spinal cord then you can tell they are they are actually Having more copies compared to AV9 which indicating that this capsid is crossing BBB better and we also of course examined the RNA protein and it appears that we have a Liver-detoxicated Capsid and it crossed the BBB better and it has a A very kind of good transduction profile for the motor neurons So the next one i'm showing you here is the AV902 Which you can see is also liver-detoxicated And it Transduces the spinal cord brain well, but it doesn't really cross the BBB significantly better than the AV9 and even a little bit worse or similar so we So we can we think this capsid is liver-detoxicated but it doesn't cross BBB more efficiently But it does transduce the neuron at similar comparable level compared to AV9 so The very last two AV903 and 904 AV903 works very quickly From the DNA level you can tell it is not liver-detoxicated capsid at all But it does cross the BBB better more efficiently so So just to summarize what I just showed you in the animal animals so It appears that for these four AV9 variants we have AV1 901 and the 02 appear to be liver-detoxicated And we have all these capsid except the AV902 And the AV902 appear to not better than AV9 in crossing the BBB But all other three cross the BBB better and they all transduce the neuron pretty well But the next major question is how would this translate to Seriputic effect when we put it on the animal models so here I'm going to show you the Performance in the disease model here. We are using the very famous Delta 7 SMA mouse model which is originally Established by Dr. Burgess at the Ohio State University and this is the mouse model the key mouse model that Successfully supports the geosmar development The eventual approver so this actually is a kind of Reproduct a model that could kind of predict that the translatability for the capsid for the CNS disease applications So we packaged all these AV9 variants into a CCE-AAV format And we package the same human SMM1 gene as the Zorgensma So then we put these vectors into the animal model and here I'm going to show you just one data set which is the survival data Of course we test all the other stuff like the body weight the motor functions everything but here just the survival data and here you can see that the dotted line represent the non-traded mouse and the black line represent the AV9 which is similar to Zorgensma and then here is the AV901 which actually from the Biodistribution expression data we can we know that this is a liver-detargeted Capsid and it crossed the BBB Efficiently much more efficient compared to AV9 and actually to transduce our target cell Which is the motor neuron cell pretty well, but it didn't rescue the phenotypes in the SMMA motor Compared to AV9 doesn't rescue better than AV9 and the AV902 Which is also a liver-detargeted capsid but it didn't cross the BBB more efficiently compared to AV9 It actually works very well on the SMMA motor The third and the fourth Capsid if you still remember the big table the three and the four They are very similar to each other in terms of the biodistribution or expression profile but they are eventually eventually They are therapeutic effect on this model is totally different the AV903 is Similar or actually worse compared to AV9 but the AV904 They are just works perfectly even at lower dose like I forgot the exact dose But it could be up to a log lower dose compared to AV9 can still rescue the phenotypes So this is a very interesting because this AV this two peptides As I said, they are all NHP library-directed evolution their top candidates for targeting the CNS They are supposed to cross in the BBB better and they are very similar profile in all what our pre-clinic evaluation But eventually they are performed totally different in the SMMA most motors and the one more Interesting is actually three and the four that contains the same mutations as AV902 Which in AV902 these mutations in the VA1 actually resulted in liver detacting But when we combine the top candidates the peptide candidates come together with this We are one mutations that they're not they're not liver detacting anymore. So they are talking to each other Even one is at a VR8 one is at VR1 Here is the data and the summary it ends up that we have Expected or unexpected we have two capsules that works better than AV9 and the two capsules is actually work worse than AV9 So I think It is obviously that there are some like biology behind this Phenomena that we may need to pay a little more attention or considerations when we talk about capsules engineering and I would say our observation may Demonstrate that the importance of the deriving translation capsules works the best transducing capsules We should always consider everything together the structure the performance and we should choose the appropriate animal disease model to evaluate the really potential for the kinetic translation so with that I would like to Conclude the talk by thank all my team members for their tremendous support and all their contributions Specifically I want to thank Sophia She carried out all the structural analysis for the team and John Leung basically led all the UV-ray experiments Li Ming takes care of all the production aspect and I thank Dr. Jid Seng for the presentation Thank you Dr. Jid Seng for letting me to lead the project and I thank everyone. Thank you for your attention I'm happy to take any questions Any questions for Li? Thank you very much was an interesting talk My name is Adrian from Genathon So when you look at your capsid biodistribution and then the actual effect it had in the Disease models, what cells do you think actually need to be targeted for curative effect? For SMA we majorly want the motor neuron to be transduced in the spinal cord and also because based on the SM1 Protein actually in humans is not just the motor neuron has the SM1 expression. So we think other cells organs may need also need certain levels of Transduction Majorly we want the motor neuron to be transduced and do you think the motor neuron transduction can happen outside of the CNS? So if they're peripherally really transduces enough since you had non BBB crossing variants that still had a curative effect What? Do you think the motor neurons need to be transduced in the CNS? Transduce peripherally and then the AV retrograde the transports to the nucleus I'm still don't think I get the question hundred percent, but I think The motor neuron in the spinal cord. Yeah, we do want to target one high transduction But for the peripheral tissues we wanna so no capsid that you really are talking in the CNS like Absolutely like hundred percent right you always have distribution in the peripheral tissues I'm asking for the peripheral neurons because the motor Neurons, okay Gadex coffin gene therapeutics optimized. I'm curious. I may have missed it, but do you have any data in human cells yet or any? tropism We are building on the AV actually seeing human in humans Yes, we are going to publish very soon for certain capsid not not all capsid of velocity Can you share whether these ones were found in that atlas? This one is still ongoing. We haven't connected any data yet Yeah on human not yet Thank you Interesting talk I'm Travis from Genentech You know BBB transcytosis and liver transduction can both be affected by a VR binding I was curious if you looked at AVR binding with your capsids. We did a structure analysis But we just to be honest we didn't do this structure cryo-EM No, we didn't and we didn't do the affinity test No, but we do what everyone do is to their The structure prediction right the binding I say by the computer it is consistent and inconsistent in a way consistent isn't we don't find any This of variant of the interact with AVR PKD one domain but for the KPK D2 domain For the AV nine or one it actually shows stronger binding Compared to AV nine, but it is liver detected So which is a little bit contradict to what everyone including myself thought it would happen and for the nine AV nine or two it is Consistent with what way current it thought it would be so it is has less biting affinity and AV nine or three and four I'm a little bit forgetting but it is so seems to be stronger which consistent because nine or three is not even targeting And it has stronger biting but again it Even nine one is higher biting but they were detecting so it's hard to explain So For that I would invite our next speaker from Sanofi Sore of Chaudhary and he will be talking about the metagenomics a AV caches with deep better scale mining Welcome Well, thank you Smriti and I like to start by thanking the organizers for this opportunity to talk about our work Here's my disclosure slide and my title slide About 50% of the human population is ineligible for a gene therapy trial and That is largely because of neutralizing antibodies that humans have Actually, can we switch to the non presenter mode just the slides Thank you Now the reason human populations have neutralizing antibodies is largely because we have exposure to natural AV capsids Throughout life, but also because humans have evolved with capsids of primate origin and as you guys know Every single capsid we have in a clinical trial is either of primate origin or has been neutralized Of primate origin or has been engineered from primate origin. So it's a massive problem We hypothesize that if we could find capsids from Non primate origin we'd be able to get around this problem Now, you're not the first ones to do it. There are landmark papers that have done it before but one of the things that we Thought was relatively unexplored is these is the metagenomic space or the extra genomic space So to aid discovery from that we build this custom-built Capsid discovery pipeline called matcap. I know you've up to your neck in acronyms. This is ours The challenge with finding or mining metagenomes Capsids from metagenomes is just the scale at which you have to find it's it's literally a proverbial needle in a haystack We have about four if you just go to NCBI you have about four point two million SRAs or sequence read archives that you have to search through and that represents petabytes of data So we built this three layer screen where we first did a more crude screen To just identify SRAs that had power viral reads so out of these four point two million we landed about twelve thousand of them and then we Essentially assembled these context de novo by by by overlapping these reads through synotaker two and then these context were annotated and we Identified novel AV capsids from that batch of things Using this method we found capsids throughout the globe We found a lot of capsids from the US which is not surprising because there is a lot of studies in Metagenomic studies in wastewater from the US, but we also found a lot of capsids in in in Southeast Asia We found capsids in a variety of architectures. We found linear capsids we found in complete Genomes with just the capsid and no rep we found capsids that looked like they could be circular In their genomes not entirely clear why? but as a summation we found About a hundred and thirty nine novel capsid so just to set the scale this about doubles the universe of known AV capsid variance What is particularly interesting in this cladogram that you see is the is the Divergent madcap sequence or divergent madcap sequence or the sequences that we actually could not find a close natural relative to there wasn't really a Base capsid or a parental capsid that we could see that it had evolved away from and that's what we're going to talk about Throughout this talk because obviously that's where we want to focus on given the context of what we were searching for In this in this data set we are looking at percentage neutralization I like to point you out sort of the two big buckets of capsids that we were able to find there's this blue bucket Which is mildly neutralizable they get neutralized especially as you keep increasing the concentration of human IVIG Which contains pre-existing neutralizing antibodies throughout the globe? They are much better compared to the black lines there Which are the natural isolates like AV1 AV2 and AV9 and then you have what we think is the most promising one Which are the pink squiggly lines the pink capsids here? Which are these non neutralizing capsids that do not seem to hit 50% utilization at any Concentration of IVIG that we dose there's literally no no Nab response that we could record for these which is very surprising we have we've not seen that We have not seen that previously This is just the same data graphed out differently to show you the gulf between the natural isolates The neutralizable mat caps and the completely non neutralizable mat caps We decided to focus on two of these Pink capsids two of the non neutralized on mat caps MC 46 and MC 55 again more acronyms I'll Talk to you about in a second of why we decided to focus on it But the first thing to show is that because these are you know caps as we haven't seen before We confirm that these are able to package these are able to make what we consider classic virions But very interestingly when we dose them into a non human primate these stood out Significantly compared to AV9 you can see AV9 on the bottom right of the curve high neutralization but low brain penetrance and you can see MC 55 and especially MC 46 on the on the left of that If you look at where they are on the cladogram In in relation to other capsids what you're seeing here in the in the in the purple Manhattan skyscrapers is the Extend to which they can be neutralized and you can clearly point out with those two boxes that MC 45 MC 46 is completely non neutralizable whereas MC 55 has a degree of neutralization But is much less compared to its neighbors and natural natural isolates Here is a representative brain map showing the brain penetrance of MC 46 and MC 55 compared to AV9 What you're seeing on the left is DNA fitness or DNA read counts what you're seeing on the right of this Brain heat map is the RNA and you can see that We think that MC 46 and MC 55 show widespread transduction of the non human primate brain after intravenous delivery We Wanted to next confirm how good or how good at avoiding nap neutralization of these capsids So we basically compared them to their closest natural isolate I'm just showing you this MSA just to show you that when we say closes is not very close is actually quite far But their closest natural isolates and we benchmarked it to AV1 and AV9 And what you're seeing here is basically a person neutralization in about 15 individual human donors that we source and what you can see here the more yellow you get the more Zero positive these donors are for a specific Capsid and what you can see is that your natural AV capsids a Majority of donors are zero positive for them as you go to MC 55, which is the last but one Column you can see that few donors are zero positive and as you get to MC 46 We did not see a single donor out of the 15 that was zero positive We've actually done this exact study with about a hundred non human primates right now And we found two that maybe could be zero positive We also confirmed that if you take serum from these donors You can still are you are still able to get full transduction with MC 46 and MC 55 Now the next question is if that is true is This a paradigm for redosing What we confirmed is that MC 46 transduction is unaffected by prior a redosing so on the left We are showing you this passive transfer model where rodents are pre dosed with human IV IG and then they are challenged with AV And what you can see is that as expected AV nine completely loses all Transduction if in a in a mouse model in a passive transfer mouse model, which is rescued by I desk as Expected on the other hand MC 46 loses none of its transduction And a pretty similar assay on the right where we are showing that in Rodents that are those with AV one or AV RS 32 33 pretty diverse capsids You see exactly the same phenomena where gene transfer is not inhibited Using this when challenged with the seed of these mice. There is no inhibition of gene transfer at all with MC 46 and MC 55 So Your holy grail capsid has all three of these components augmented Last year we showed you some data at ASDCT on apex Which is our generative AI modeling where we take a known Capsid and known natural isolate and we evolve the sequence space around it with with generative AI modeling to find a retinotropic Capsid that is both manufacturable and Has high degree of retina tropism This year we're talking about mad gas where instead of going this way We're going this way and discovering novel a B's by mining this vast amount of available metagenomic space Why are we excited about this work we are excited about this work Because of the discovery potential right so we have hit upon this what we think is this incredibly black box unexplored mother lode of you know this rich vein of nav knife capsids that have an exciting properties that we don't know about We are excited to find out more about the biology of this Why do these capsids have these capsids evolved to avoid? NAB recognition how How is MC 46 doing what it's doing in the non-human primate CNS and why it hasn't likely seen a primate? Is it is it does it have deep brain penetrance? Or is it limited to basically existing or transducing the blood-brain barrier does it actually cross it are these novel capsids? Responsive to AVR or is there a different pathway by which they go through what is the natural receptor for these? Capsids so much so much interesting biology that we are we are we are we are digging into and every day is an exciting day We're excited for the impact that it can have on patients very clearly this Satisfied a fun satisfies a very fundamental ask from the capsule engineering community Can you give us a capsule that is completely not nice? And we are already working with with with some teams on exploring this in the pre-clinical space If if this is of interest to any of you, please, please feel free to reach out and we'll have a chat on this We are completely open about sharing these we are completely we really want people to be using it if it's of interest And lastly Speaking as a capsid biologist. I Love this for the platform value that it brings like it's a it's the value for us is now This is a chassis for us to evolve capsid with desired properties based off on we've we've essentially switched from your AV nine and like to these madcap capsids for our chassis almost immediately because I think I think This is how we're gonna our lab or our team is gonna start doing capsule discovery So it takes a village It's not just us. I like to highlight three individuals test our Grossa Eugenia Lashanko and Alex is AC who are the key contributors for this and are the first authors on a manuscript that is is under review And I like to have a heartfelt tank for the tanks for Christian Mueller Was a global head of JMU for investing time and resource into something especially in the early days when it's blue sky research It's sometimes hard to do that with corporate KPIs and term sheets and all of that. So With that, thank you, and I'm happy to take questions Thank you, Sarah Hi, nice talk. I'm Dan stone from Fred Hutch I'm also the associate chief of my Advances which will get to why I'm saying that second so many of us in the room have submitted AV sequences and capsids to SRA So when you are panning this stuff if say I had a novel capsid that I'd submitted the SRA to and I haven't published it Would that be captured in your novel? AV that you're pulling out So really are you discovering things that people have other discovered already discovered or is this stuff that's coming from say? A non AV related study that happens to have an AV in there the part one The latter In fact, I can tell you that MC 46 comes from a completely unrelated study nothing to do with AV Generally more to do with a post covid Virology screen in a seafood market in Southeast Asia so nothing to do with a V. Okay. Great. Thank you And then part two so at molecular therapy advances, we publish a lot of novel AV capsids and we see a lot of Biotechs wanting to publish capsids, but they call them AV X be Etc. Since you are harvesting this from public available database Are you willing to put these capsid sequences back into the databases where you got them from? Absolutely, that's that's I mean I Speak for Sanofi here that we are incredibly open. We've published like apex and other platforms. We have published This is as I said in press. We will absolutely have a full disclosure including not just the sequences Which is obvious but obviously also the code by which you discovered it in case there are things that we didn't discover There are other people find interesting. So we are entirely transparent. That's fantastic And I commend you and I recommend that all other companies that submit to any of the molecular therapy journals have a similar Love it policy. Thank you very much Great talk Archana from University of Louisville So anything that we make novel also can trigger the immunogenicity profile. Have you looked into how this? You know while the transduction is great, but how does it trigger the innate and the innate immunity of these have you looked into the immunity profiles of these animals that you have been investigating Great question. We do We've not seen anything adverse the animals the non-human primates seem to be seem to be able to tolerate it With with the specific set of trans genes with the caveat that you know We have a specific set of trans genes that we're interested in and so on and so forth With that caveat we have seen no adverse effects. We've seen some very interesting phenomena in terms of Which we still have to confirm in terms of it seems, you know Whether it can truth, you know, it can be completely redosed as a second dose by itself I think that's that study is ongoing but in terms of phenomena that might be interesting to the clinical community these behave Pretty much like other capsules. Okay. Thank you In the interest of time maybe we can have the questions later I see there are many more questions for you. So maybe you can talk to all the people I would love to please please please feel free to catch me. Thank you Moving on to our next speaker and I want to emphasize here that he is a travel awardee and Here we welcome John Zachary for his talk on the structural investigation of naturally occurring human anti-av2 antibodies Antigenic profile of a V2 and welcome John. All right. Good morning. Thank you All right So as she said, my name is John and I am a PhD student at the University of Florida So I don't have any disclosures at this time I'm gonna go ahead and launch into a portion of my dissertation work that we're working on Currently so as we just heard The use of a V's in the clinic uses a large variety of native occurring capsids and those capsids have pre-existing neutralizing antibodies circulating throughout the population In fact as of last January about 80% of the clinical trials that had occurred or were occurring at the time used Native type capsids and as we've seen in the past few days the proportion of engineered capsids can be expected to increase but the resistance through neutralizing antibodies may not be affected by some of these changes and Depending on the serotype to investigate you can have anywhere from five to ninety five percent Seropositivity for this neutralizing antibodies in your population In response to this our lab recently and a few years ago did a study using Av9 antibodies that were sourced from Zolgensma treated patients that were neutralizing and Then use high-resolution cryo-em data in order to characterize where those exact antibodies bound on the surface and from that we determined The about three-quarters of them bound to the surface at the two-fold portion of the capsid and for those of you that aren't overly Familiar with the exact structure. I'll show that in a moment and Then from this data we then developed a antibody escape capable variant that was able to escape 17 of those 21 antibodies that we investigated maintained its tropism and biodistributions and maintained transduction in the presence of sera both from these patients that these antibodies were isolated from and From patients whose antibodies weren't included in this portion of the study Recently There's been a in 2022 following the kovat lockdown There was an outbreak of pediatric hepatitis and during the ideological studies for these patients were found to also be co-infected with a AV2 This was not necessarily on inspect unexpected as many of the viruses that were found To be part of this outbreak where are the common helper viruses for AAVs And using access to our collaborators We were able to generate samples of the antibodies from these patients that are specific to this native type infection for AAV2 and so the goal of this work then is to clone out those and generate high resolution cryo-am data again in order to identify those specific epitopes on the surface of the capsid and Compare that for now to the surface to the epitopes found in a AV9 and other mouse derived anti AAV2 antibodies And so beginning with the work for my collaborators they were able to isolate out these demonstrate both binding and neutralization in vitro Which is important for this type of work and moving on from there Using high resolution cryo-am I've characterized the binding of 10 of these so far go ahead and break that down for you The AAV capsid is an icosahedron and so it has a two-fold three-fold and five-fold symmetry axes and these give rise to different important Structural features on the surface of the capsid and for the clusters of antibodies that I have Five of these a AV2 antibodies bound around the three-fold Which is about fifty percent of what of these? Two bound to the two-fold and three bound to the five-fold and then doing the same breakdown for a AV9 Just as a remind us as I said three-quarters of these 21 bound to the two-folds. That's about 16 if I remember correctly and then two bound to the three-fold and three bound to the five-fold now This is some very interesting piece of information, but structurally why is this happening and so Alright here we go next slide alright, so just to reorient you for a moment now We're looking down at a section of the capsid that I've been pulled out and look Isolated and so the three folds they're represented by the triangles and the two fold by this oval shape And I want to draw your attention first to these red squares on the top the top portions of the panel here are Hydrophobicity matrix X on the surface of the capsid and those two squares on the right are highlighting Two residues that are found within the two-fold pocket on the surface of the capsid and these residues were bound By 100% for the one on the furthest right and 50% of the antibodies that bounded the two-fold and then looking on to the left box we see that there's this residue here This is a tryptophan that's found in the center of the three-fold protrusions here And this was bound by both of those three-fold binding antibodies Conversely when we look at a AV2 you can see that those Hydrophobic spots that are represented by those orange colors are not present and a AV2 that bright box has been changed is Changed from a tyrosine to an asparagine no longer has this hydrophobic interaction surface And then the surface of the AV2 caps it actually at the three-fold protrusion Extends out over the two-fold pocket and presents itself and obscures the interaction that would occur to the That would obscures the interaction to that other Residue and the tryptophan also is not present in the AV2 surface At the bottom panel and now we're looking at an electrostatic Representation of the caps at surface draw your attention to the pink squares, okay And this pocket highlighted on the right for the AV9 It's an intensely negatively charged portion of the capsid which in general antibodies tend to favor though It's not entirely 100% of the time negative regions on antigenic surfaces and this pocket is obscured in AAV2 when we look at the left pink box, I'm gonna draw your attention to the greater region of the AV2 surface Which highlights trying to highlight that there's just in general a larger distribution of charges across the AV2 surface outside of that pocket So there's easier points to contact for AV2 versus that deep pocket for AV9 and then when we make this We make this a similar comparison to mouse monoclonal antibodies that were generated towards AAV2 in the past There are two so far That I want to point out here with fab a20 which is a five-fold binder and fab C37 B, which is a three-fold binder and We look at a stereographic representation of the binding footprints of these antibodies We see that in the case of anti AAV2 mouse antibodies They're fairly prototypic of what we see for humans and this is different from what we saw for AAV9 where anti AAV9 antibodies for mice were not highly similar to anti to human antibodies for AAV9 This is true for all except for the two two-fold binders which for now There's not a mouse monoclonal antibody it's representative of these but in the past work has been used 820 as a prototypic escape for human IV IG and so this kind of highlights some of the structural Underlying reasons for that Capability moving forward we taking this on to develop an escape variant as we did in the past And I'm working on that currently So just as a summary talking about the antigenic profile of AAV2 seems to favor the threefold a bit more then for AAV9 and While this is one of our studies they're very there are fairly few studies that fully characterize the antigenic regions of AAVs in general and This represents right now the largest study of anti-human anti AAV2 antibodies and Is the only study that's conducted on naturally occurring AAV2 infections as opposed to a clinical application of a therapy and While I've highlighted some different structural reasons for why these antibodies could interact differently with the surface There is also the difference in the context of the infection for AAV9 Presumably there was no helper virus present for these patients when they were treated and we treated at a much higher dose Than they would have been if they were exposed to AAV9 naturally such as the case for these AAV2 patients and For the future we expanding this study with another patient another set of antibodies And then developing of this vector of an escape capable vector for these pre-existing antibodies And I want to thank my lab my PI dr. Rob McKenna Mario who will be giving a talk at 330 today about AAV Trafficking and being some seeing some really interesting cool data coming out of our lab and our collaborators at University of Sydney Especially dr. Grant who will be giving a talk tomorrow about more of this data concerning the AAV2 antibody and immune response from these same patients that were used for this study but in a different context tomorrow morning Thank you Thank You John This one right there One AV2 a natural infection and presence of helper virus and AV9 with a With this olgenzma is how these two you know because they're mapping to do different regions But in most humans actually have quite cross reactive antibodies And I'm curious if you could talk about the differences that you might have noticed in other studies with cross reactive antibodies Normal humans here. How well where those map to so in terms of cross reactivity for the antibodies We presented we did do and let me see if I have it in here. I Don't have the cross reactivity studies for the a AV2 antibodies, but we did those by native dot immuno, but Should be No, it's not on here, sorry Several of these were cross reactive and we found that that kind of correlate cross reactivity Generally correlated with some of the more conservative regions of the caps of the five-fold region especially which is highly conserved on the surface of the capsid tends to be more cross reactive than others that was true for Three of these for the a AV9 Possibly more and then and that data is included in the paper in those papers there and then for the a AV2 I have One's highly cross reactive and that one also binds to the five-fold surface. Thank you so much Great talk. I may have missed this detail at the moment So this might be a really quick question for you are these all neutralizing antibodies. Yes Okay, do you have any reason to believe why there might be a different? Landscape in terms of where they're binding for non neutralizers versus neutralizers, or you think this would be a common for both classes So I've not looked at where non neutralizing antibodies would necessarily bind. I don't know if we have That data at least in internally Haven't seen any in the literature, but I would imagine that for non neutralizing antibodies They would probably bind away from where you would expect receptor sites and thereby are unable to in inhibit Transduction directly on their own but they still fully capable of recruiting other elements of the immune system complement, etc in order to clear clear on virus I'm Seaver research. Thank you for presenting this it was very interesting and I understood your Assumption that the differences you see might be associated with a way of transduction But have you considered that difference in post translational modifications as part of the AV manufacturing process may contribute to these differences That is an interesting question. I haven't looked into it but It's possible. I don't really know have the knowledge to say much more right now There are some deamidation sites that have been shown to increase in unigenicity and are prone to occur in Storage we can talk later, of course We have a little bit of time I have a question so Do you have any specific amino acid residues that are involved in on this? contact point for the fabs We have identified those and for the sake of my dissertation I'm holding on to them until I get it all Worked out, but I will be publishing that and that will be available Is there any more questions for John if not, thank you John So With that I would like to invite our next speaker Simon Pecora and He will be talking about a modular Capsid engineering platform to confirm neutralizing antibody evasion to CNS and muscle targeted AVs Welcome. Thank you. Thanks for the opportunity to present our work today So There are several ways we can deliver genes to the CNS You can go in triparenchymole you can deliver into the CSF for go intravenous and because some brain diseases Impact the entire brain. We believe that there is a need to deliver Cargos in as many neurons of the CNS as possible and we believe that the only practical way of achieving that is to use the vasculature So the third one and go IV But there are a couple of challenges in achieving CNS It's systemic gene delivery to the CNS One is that naturally occurring stereotypes do not cross the blood-brain barrier efficiently And the second one is that they are susceptible to neutralizing antibodies resulting from pre-exposure to the virus So in the next few minutes, I'm going to tell you how all app tackled both challenges Sequentially and we'll start with the BBB crossing So To cross the BBB my colleagues Ken Shan Helen Wong and Jean lean reprogrammed a v9 to bind to the human transferrin receptor and This led to the discovery of a second generation human transferrin receptor binding capsid that we call BHTF for 1b to AKATF for 1 cap X. So if you want more information about this capsid I strongly recommend that you go to talk to my colleague Ken Shan is right here He'll be presenting this afternoon his poster number 3037 on the HDF for 1b to characterization I'm just going to share a little piece of data here. This second generation capsid is capable of efficient Transduction of the brain in human transferrin receptor knocking animals as seen on the right of that slide here It outperforms the first generation volume BHTF for 1 as well as a v9 and it mainly transduces neurons and What's interesting to notice here is like we work in that type of study as a dose of 2e 10 VG per kilogram 2e12 sorry Which is about 50 fold lower than Zolgensma So lowering the dose is great because it reduces the risk of toxicity But at the same time if you lower the dose then you have more neutralizing antibodies per capsid and the neutralizing antibody Problem becomes even bigger So that's that brings us to a second challenge So how do we took the height? How did we take BHTF for 1v2 and and further engineer it to evade neutralizing antibodies? So I want to start with Model of BHTF for 1 v2 capsid it's an alpha 4 3 model and in blue is represented the region that was engineered To bind to the human transferrin receptor So as you can see, it's a very small region of the capsid and the rest of the capsid actually looks like a v9 So to make it evade neutralizing antibodies We kind of decided that we had to disrupt as many anti v9 antibody epitopes as possible possibly with as few modification as we could and And we looked in nature and found what strategy that was pretty appealing that Consists in modulating surface charge So we have seen that during kovat with the omicron variants that actually was shown to evade the great majority of antibody based Drugs that were developed to neutralize the original strain and this was attributed to a few mutation in the protein spike that created a positively charged patch So we the question became can we apply this strategy to be HDF for 1 v2 in order to make it evade Neutralizing antibodies while retaining its BBB crossing phenotype To do that. We started from the parental capsid v2 We identified surface exposed target residues that we mutated to either positively charge neutral or negatively charged residue We took the resulting library and we screened it for production fitness for in vitro antibody evasion as well as blood-brain barrier crossing in human transfer and receptor knocking animals We did that in two rounds modified a total of seven capsid regions and characterized selected variants individually And to increase the probability of identifying capsid scaffolds that would be compatible with other loop eight modifications We also added three different peptides crossing the BBB in mice using three different receptors So ebbi 28 and 9 p3 1 to the original library So first we can walk through the NGS data Here's the production fitness profile of the library as you can see Av9 and v2 have like pretty similar scores and we selected variants within that blue regions that produced within a reasonable distance of a v9 We looked at brain transduction in tf4c knocking animals and here each dot represents a variant with four codon replicates and you're looking at the Transduction score in one mice versus the other as expected Bhtf1 v2 are perform AV9 and we selected variants within that green region Last we look at antibody evasion in HEC 293 cells And here I'm showing you the Transduction transduction score in the absence of IVIG on the x-axis versus presence of IVIG on the y-axis and We selected variants that had like a high score in the presence of neutralizing antibodies So now when you take the intersection of these three regions you find a limited set of variants Including one that we called bhtf1 v2e for evasion and that we characterized more in detail So the first thing we did was to look at the surface charge profile of v2e And what we could observe is that it was pretty different from the parental capsid as expected Considering the strategy that was used in that project So now what is the impact of those modifications on tropism? We looked at that in human transfer and receptor knock-in animals You're looking at a comparative analysis of v2e v2 and AV9 in human transfer knock-in mice And you're looking at the transduction data on the left and VG by distribution data on the right and as you can see is that The BBB crossing phenotype of v2 is maintained in v2e in spite of the differences in surface charge Let's look at antibody evasion now in vitro So we use we did that with 73 human sera from individual donors and we used AV9 AV5 And bhtf1 v2s controls so here I'm plotting all the neutralization curves that were measured for sera neutralizing AV9 color coded by their IC 50 red meaning high Neutralization and blue meaning low neutralization So as you can see the profile are pretty similar for 5 9 and v2 But you see a lot of curves that get like right shifted for v2e suggesting some level of antibody evasion And when you take that and measure the IC 50 for all these human sera They are all represented as as columns in that heat map You see that the IC 50 values go down across the board for v2e compared to the other three stereotypes with actually 40% of the human sera that neutralize bhtf1 v2 that exhibit IC 50 below limit of detection of 104 for bhtf1 v2e So that's good news. It means like we're potentially able to expand the patient pool for CNS gene therapy But we wanted to validate these results first in vivo So we we worked in on the model where we passively immunized Human transfer and receptor knocking animals with IGG pre fight from human sera a day later We co-injected AV9 v2 and v2e IV and at day 21 we looked at function in the brain by RTD PCR So let's look at the human sera first from which we purified IGG for the passive immunization You can see first that they have different neutralization potencies with C being more neutralizing than B and A overall and Then you can also see that that bhtf1 v2e Evades this sera to a greater extent than the other two because it's IC 50 value. It's actually lower when measured in vitro So now let's look at the sera of mice that were passively immunized with these IGG So again, you can see that the mice that receive the compound C or have a serum That's more neutralizing to all three capsids compared to B and A and you also see a dose response at the 30 milligram Dose versus 10 milligram IGG per animal and again bhtf1 v2e show IC 50 below our limit of detection of 1 to 20 in that particular assay So now let's look at function. This is the function in the absence of IGG so as you can see v2 and v2e as expected outperform AV9 And now when you look at all the conditions and you normalize that score to the selling control you can see that you have a greater rescue of your brain signal in animals treated with bhtf1 v2e as opposed to v2 and AV9 with a full change ranging from 3 to 40 fold in serum C and B It wasn't neutralizing enough that we could really observe anything in vivo So that enabled us to conclude that this stereotype was also able capable of antibody evasion in vivo We looked at production titers in suspension and adherent cells and in the absence of optimization We were able to produce v2e within two fold of AV9 And we were able to also show that it was compatible with porous AVX AV pure AV9 for a capture affinity chromatography but not porous AV9 Last we wanted to see whether Some of our antibody evading scaffolds were compatible with muscle targeting peptides So we identified five potential candidates that we combined with two muscle peptide My oav4a as well as one that we call rgd1 and We characterized the resulting variants and so we were able to show that One was compatible with both and we followed up on this one with the rgd1 peptide only And if you look at antibody neutralization in vitro The second generation that we call rgd1 gen 2 was able to evade pre-existing Antibodies in about 32% of the human sera that neutralized rgd1 and this year we further evolved it and we went up to around 56% and last we looked at function in c57 animals and We saw no major difference in muscle between rgd1 rgd1 gen 2 and gen 3 So to finish we this is actually a GIF of the surface charge modulation between rgd rgd1 gen 2 and gen 3 and this brings us to a conclusion And we applied a charge alteration with the genesis to CNS and muscle tropic capsids The resulting variants maintain their enhanced tropism they were able to produce within twofold of AV9 and The evaded 40 to 56 percent of the human sera that neutralized AV9 So this method is potentially applicable to any tissue specific capsids and with this I would like to thank all my lab mates and colleagues and collaborators from Deverement lab and that at mini care lab at the Broad especially Ken Shan Jean Lynn and Jensen in Johnston who took care of all the NAB assay as well as Pam Brower and John Harvey for the manufacturing and I'll be happy to take any questions Thank you, Simon and we have questions over here Excellent talk So he's I think you said you looked in seven different areas How many edits did the best capsid end up having? How many mutations? Yes More than you can count on let's say between 10 and 50. Okay. Great. Thank you 50 Thanks for your talk I have a conceptual question on the Results of your library that looks like you had the same approach that modifying the charge surface for all your library on you Showcase some kind of positive candidate. Did you try to extract the negative one on to redock the neutralizing? Antibodies to try to identify which position was actually involving in that kind of conceptual Evading of that when I'm you plan to finally make this data accessible You're talking about the dual production fitness distribution The way you modify your your variant you consider on your top right of your graphics the selection of your top variant That is evading but you may have several Avis vector that have a modification of the charge surface, but without evading Yeah the system so you can envision to redock actually the neutralizing antibody itself to see what is the difference? functional difference between what works to what doesn't work finally at the level of your Pure capsid to antibodies binding. Yeah. Yeah, we initiated some work to In that direction, but I think is not as mature as you would expect. Thank you So I don't see any more questions Beautiful talk Seymour. Thanks. I love that you tested both AVX and the AV9 porous resin And you you saw a major difference there, but what about things like anion? Exchange columns and other downstream purificate. How do you think that this charge modifications will affect things like that? Yeah TBD we we haven't looked into that Very great work. Thank you Thank you Simone so with that we have last presentation from Adrian Sevi and he will be talking about one AAV capsid From co-ed and he's from co-ed therapeutics and it's all yours Thank you for the introduction and first of all, thanks the organizer for giving me the opportunity to present We are redefining targeting gene therapy at co-therapeutics And today won't talk about capsid engineering. We are doing bio conjugation Why it's mainly to address the key challenges of IV injection The first one is the targeting how do we bring the vector to the tissue of interest? Bio conjugation sorrows more efficient conjugated capsid because we are looking for known mechanisms How do we do it? We first rationally design ligands that interact with human receptor with a known mechanisms And we have a very modular platform that is based on chemical conjugation creating a covalent bond It's irreversible and does not require any catalyst This is very modular We can really fine-tune the coupling level to perfectly match the requirements of the targeted receptors And as we are completely decoupling the tropisms from the capsid we can Integrate this platform to any existing AVs wild-type engineers and so on At Cove we have many different kind of ligands from small molecules to peptides to VHH each of them is rationally designed Target a specific receptor and also to allow cross reactivity among species and mostly between human and NHPs To simplify it you can see it like antibody drug conjugate, but apply to AVs So once we perform our bio conjugation we are creating a new generation of targeted vectors and at Cove we have three main focuses in ophthalmology in CNS and for muscle diseases I don't know if you had the chance to attempt earlier this morning to our new supracoridal capsids But if you want do not hesitate to be in touch and I also strongly recommend you Attempt later today to the poster 3026 from my colleague Lavanya that will present our ligand conjugation is enabling cross species crossing through human CFR1 receptor If we talk about the CNS and we have a look to our AV9 BBB crossings through TFR1 workflow we first Design the ligands and we can do that because we are doing bio conjugation. We are not interested by the capsid It's really the ligands that we are working on it We are really designed them to be specific for human receptor But also to work in NHPs and thanks to the bio conjugation and to our chemistry we can really fine-tune the coupling level We can clearly increase how many ligands we put at the capsid surface and it's actually manner a control process We then screens our conjugated AVs in vitro in the example with two cell lines the first one a human expressing the Transfusion one receptor cell line and our two conjugated AV9 is always a ligand one or two at the same Transduction efficiency sense of publish TFR1 benchmark But if we switch to a sign or expressing TFR1 cell line or two conjugated AV9 is there in blue or green Maintains their transaction efficiency whereas the published TFR1 benchmark does not We then move to in vivo screening And you know the AV9 is a poor brain transducer when injected by IV But if you take the exact same AV9 And you conjugate it with our TFR1 ligands in knocking humanized mice you can see that the transaction efficiency is greatly Improved with the world brain transduction If we look at the RNA expression We have a fortyfold Difference in term expression compared to AV9 In purple and we are perfectly matching the publish benchmark In term of brain transduction, but we also have a second effect We have a very strong liver detargeting with our conjugated AV9 with more than 15 fold decrease in liver expression Our platform is something that is very versatile so we can move from The brain to the muscles and if we take the exact same AV9 capsule backbone and we this time conjugated with with a peptide that has been designed to interact with integrin alpha V beta 6 we obtain a conjugated AV9 targeting these integrins and This time we have a three hundred fold increase in the gastrocnemius muscles and the two hundred fold increase In the quadriceps both reaching the publish beta 6 benchmark again So the alligator platform is something that is very versatile it's very tunable and it's highly scalable We can do that because again, we are doing bio conjugations we are not that interested by the capsid only by the ligands and We wanted to go beyond in term of addressing the IV challenges. The first one was the targeting I think that we can fix it easily now, but we wanted to go really beyond and address two other challenges The first one was the immune evasion and to do that. We created one AV capsid family. It's kind of proof of concept It has been designed to completely Decorrelate the tropisms from the capsid we wanted a kind of complete neutral shell and it has been engineered from low Seroprevalence several types in humans to address the immune Response problems and we also engineer them to completely obtain a liver the targeting so we had a Neutral shell without any interaction on its own and thanks to the alligator platform We can strongly retarget them to tissue of interest We first validated our one AV capsid by IV injections Quite a high dose for 10 to the 13 VG per Keg with luciferous transgins Under the CAC promoter and we compare it to the AV9 in both male and female wild type mice as expected For the AV9 groups we saw very strong expression in the liver. Whereas the expression was barely detectable without one AV We wanted to know where went the vectors So three weeks but injection at sacrifice we look at the vector copy number in different tissue and In the liver we have eight hundred four lower copy number compared to the AV9 with an astonishing 0.03 vector copy per cell this I believe is truly worthy targeting Our non conjugated one AV does not cross the BBB on its own We have a thousand fold lower expression Not expression vector copy number in the brain and it was quite the same in every peripheral Peripheral organs that we have looked at So we know we have this neutral shell. What can we do with it? And what it's happening if we conjugated it with a TFR1 ligands We first confirm an unprecedented Liver the targeting with a thousand fold lower expression in the liver You can see the AV9 expressions the liver in purple and our conjugated TFR1 when AV capsid is here with a thousand fold In the heart we have 400 fold But this conjugated one AV is perfectly able to cross the BBB and to express In the brain when we compare to the AV9 on the left and the TFR1 conjugated when AV on the right I think it's quite clear that this vector is able to cross the BBB and to express in the deep brain regions And if we zoom in We know that the AV9 does not have any neural tropisms whereas our conjugated one AV has a very strong neural tropism We continue with Beta6 integrin targeting peptides in wild type mice this time We confirm again a very strong liver the targeting with a 500 fold compared to AV9 So non conjugated when AV is in red and the conjugated is in blue In both muscles quadriceps and gastrocnemius We have not been able to detect any expression of the non conjugated vectors But once it's conjugated with a beta6 targeting peptides this time We are between five and tenfold higher in term of expression compared to AV9 Because we are doing bio conjugation and we are working on the ligands We have been able to create this first generation of one AV capsid it's a very first capsid With no tropisms on its own. It's quite inert It has an unprecedented liver of the targeting. It's highly manufacturable. It's very similar to AV9 in term of yield It's based on low seroprevalent serotypes, but thanks to the alligator platform and the chemistry We can really target the vector on demand on the only in the tissue of interest In conclusion as the alligator platform allows us to have cross reactive ligands among species with a very strong liver Strong tissue targeting we went from AV9 to benchmark expression level in both CNS and in muscles And we believe that the combination of the bio conjugation the alligator platform and the one AV We will pave the way of a certain generation of vectors that is safer and more potent Thank you very much And I would like to thank all cove members for this amazing works and mainly the bio conjugation team which is here Because without the girls nothing will have been possible. Thank you I Have a interesting work, thank you I am a bit concerned about the generalization claim of this platform Given that the transfer receptor as far as I know is indigene by symbol linear peptide So how do we feel? the generalization to more complex Receptor requirements like things that requires conformational agility or things that require Imparation dependent judgment or things that require like in the clustering How you How you feel about this? I'm not completely sure to be honest to a fully understood your question Very simply. How do you think these would generalize to other? Receptors that have more complex requirements to get in judging that can be either more complex ligand some has conformational requirements or clustering of more ligands or I get it Is it again we are decorrelating those the capsid we are first screening the ligands to match the Receptors we know that our ligands are reactive with a receptor of interest and then we conjugate them to the capsule Hi Demetri from ring therapeutics very nice talk and very cool platform I'm curious the with the tfr data you showed since you mentioned that there's not kind of an intrinsic tropism of the capsid How much more? Is the tfr doing after the BBB crossing to give you the actual transduction of the neurons like how are you actually targeted neurons is there? Additional ligand in or is it just a increased concentration in the brain that then is leading to a transducer It's a very good question. I can't answer. It's a very preliminary work, but we are looking into it. Thank you Great dog you from their therapeutics I'm sorry if I missed this in the talk But what was the method of finding the kind of neutral base capsid is it one that exists or did you have to do? a lot of editing to make it neutral We first generated a rational library. It was a very tiny one. It was only ten capsids So yeah, very rational Alright, okay. Great. Thanks very much And no question for mana. Thank you a Capsid with zero tropism is fascinating in and of itself But where is that blockade? Is it just not entering any cells? Is it not trafficking? Does it and is it just coming out in the urine because it's like not Where's the block? Who was thinking that you can help me on that when we look on the vector distribution? The copy numbers we have not been able to find it anywhere to be honest We also perform a kinetic study in the sera To try to find if we were just circulating longer, but no it's the same as a v9 And yeah, we are thinking about the urine's but we did not collect them. Yeah Great work. Thank you. Thank you enough Sorry sort of related to that have you looked at interactions of the base capsid with things like AVR to some going So Have you looked into the shelf life physical stability aggregation profile of this it's ongoing work, but The designs that we have performed this capsid should not change those apparent alouan properties So without disclosing the serotype we we are confident that everything will go as well Specificity of your conjugation method sorry can you repeat can you talk a little bit about whether your conjugation method is site specific? We are conjugating a natural amino acid at the capsid surface Any more questions If not then Conclude the session here and If you have time, please provide the feedback for the session It is really helpful for the ASGC for organization and thank you all for coming --- ## AAV Engineering IV {#aav-engineering-iv} **File:** `AAV_Engineering_IV.mp4` **Type:** Video transcript (Whisper medium) ### Pre-analysis AAV Engineering IV session. Expected: continuation of engineering methods, possibly structure-guided or LLM-based capsid design. ### Full text / transcript Welcome to the AAV capsid engineering 4 session. I'm Saurav Chaudhary. Before we begin, I have a couple of quick reminders which you probably heard before. As a reminder, the use of flash photography and filming is prohibited during the session. I'm supposed to tell you that there's Wi-Fi and there's an app, but if you have not used it, that's why you're not going to use it now. So you should be good. And your feedback is incredibly useful to us. I'll circle back on that topic at the end. So without further ado, it's my pleasure to welcome Tyler Moyer from Voyager Therapeutics. He's going to talk to us about the directed evolution of muscular and neuromuscular capsid variants in both mice and non-human primates. Tyler, take it away. All right. Thank you for that introduction and thank you to the organizers for allowing me to share this work with you all that we've been working on for the past few years. What I'm going to talk about today is actually probably a bit of a different direction for Voyager Therapeutics. If you're familiar with our work, you know of our work more in the CNS and gene therapy space for neurodegenerative disorders. Today I'm going to talk about some work that we've done developing novel capsids for muscular and neuromuscular indications. So at Voyager, we've developed the tracer platform, which is an RNA-based AAV capsid evolution platform that allows us to evolve and engineer capsids with tissue-specific targeting for maximal gene delivery to a given tissue of interest. And of course, as I mentioned, we first applied the tracer platform to engineer transformative BBB penetrant AAV vectors for CNS-specific disorders. And on that note, I hope you all got the opportunity to hear our Chief Scientific Officer Todd Carter present this past Wednesday morning about the development of VY 1706. And that is our gene therapy product for Alzheimer's disease. It's a town knockdown. We're really excited about the three-month GLP talks data that he presented, and we're hoping to take that into the clinic later this year. So please stay tuned. But of course, outside of the CNS and beyond the BBB, there is need across many therapeutic areas for better, safer, stronger AAV capsids for gene therapy programs. And that's what I'm going to talk about here momentarily. So given that this is Capsid Engineering 4, by now I presume many of you are familiar with the overall Capsid Engineering platform overview. So for us, instead of processing brain tissue, we can easily just choose another tissue of interest. And now we've begun working in skeletal muscle and cardiac muscle. So we did this over the past few years. We launched a number of campaigns to identify muscle-specific capsids. And here I'm showing the results of one of our campaigns where we're looking at the of a number of pulled variants in Cinnamulgus macaque muscle and mouse muscle relative to AAV9. And we were able to uncover a host of variants that showed enrichment in the muscles in both of those species. And in fact, in the monkeys, some that were improved upon a known monkey benchmark muscle capsid. So we chose a collection of these to begin evaluating. And we were able to quite easily evaluate a lot of them in mice. So that's where I'll start. So relative to AAV9, we examined a bunch of capsids. And when we looked at whole body in vivo imaging, we had a number of capsids that had substantial muscle targeting globally in muscles relative to AAV9. And they were on the order or in some cases better than the benchmark capsid that we evaluated. So we chose one of these capsids. We denoted V-CAP290 and we'll take a closer look at that one. So when we looked at V-CAP290's ability to deliver transgene here to the muscles, what we saw was just really a substantial increase over AAV9 in all muscle tissues that we examined. And again, sort of at the level or better than a benchmark capsid. And that was confirmed when we did more bioanalytics on it. In mice, V-CAP290 had about a hundredfold improvement in expression over AAV9. In heart tissue, it was about tenfold improved. And we want to note that in liver, we actually did see a mild detargeting in terms of biodistribution to the liver. Now when we first looked at these muscle capsids, we were, at least these images, these histology images, we were really struck and it looked at this global view like we were really hitting every single muscle fiber at this dose. And we confirmed that when we zoomed in a bit. In skeletal muscle, at this dose of 2.5 E13 VGs per Kg, which therapeutically would say is a low dose compared to what's being used in the clinic, we were hitting every muscle fiber. So we decided to take that even lower and we did a fivefold reduction in dose down to 5 E12 VGs per Kg. And again, we were able to hit 100% of skeletal muscle fibers, although we are seeing more variation on a fiber by fiber basis. So at this point, I'm going to switch over to our evaluation of V-CAP290 in cinnamalgas macaques and we switched trans genes here and we go into what we presented a few years ago as what we call sort of our multi-tag system. This is our ability to take a number of capsids of interest. In this case, we're going to look at AAV9, a primate benchmark muscle capsid and V-CAP290 and we put them head to head in the exact same macaques. Each capsid is dosed pretty low at 4 E12 VGs per Kg and each is packaged with a histone trans gene, each with a unique epitope tag so that we can differentiate the capsids based on in histology or bioanalytics. And another note before I show the data is that with a histone trans gene, we aren't going to see that fiber-wide staining like we saw in the mouse with the cytosolic protein. You're just going to see targeting in the nuclei that are squeezed up against the cell membrane there in the muscle fibers. But nevertheless, we can still quantify the targeting to the muscle tissue and in the macaque quadriceps, we were able to see about 30% of nuclei being targeted by V-CAP290 as opposed to about 6% in AAV9. And in terms of expression, we saw about 25-fold increase in trans gene expression relative to AAV9, both of those metrics exceeding the benchmark capsid that we evaluated. When we look at heart, we saw a similar thing where V-CAP290 was hitting about 20% of nuclei in the heart at this dose of 4E12, and that led to about a 13-fold increase in expression relative to AAV9. We also evaluated some peripheral tissues. So in the DRGs, we noted a mild detargeting with V-CAP290 relative to AAV9, but in the liver, we actually did see an increase in targeting. We really saw that as an opportunity to continue to develop V-CAP290 and push it into a second generation of evolution. So we performed an evolution scheme with V-CAP290, and we really focused choosing capsids to move forward with that had increased expression in muscle tissue in monkeys, which I'm showing here, but also mouse as well, but detargeting of those peripheral tissues like the liver. We're well underway with evaluating a number of these. We have some exciting data thus far, and we hope to share that in the future. But I'm going to switch gears just a little bit and talk about neuromuscular capsids. The inspiration for our work here is that we took a look, and we have done a lot of work in the CNS space, and we've developed some great capsids that we're excited about, like V-CAP Gen 2. I just described our work that we are now developing muscle capsids like V-CAP290. But of course, in between these two ideas, there's a therapeutic need for neuromuscular disorders, where you want capsids that have broad CNS transduction as well as enhanced muscle tropism. So we set out to evolve or engineer capsids that are capable of this, and I'm going to share results of two separate capsids today. These are V-CAP336 and 337. So V-CAP336 is really just going to be our first sort of proof of concept in mouse, and then 337 is the capsid that we evaluate in primates. So I'll start with V-CAP336. So we're going to compare it in the benchmark muscle-mouse capsid as well as V-CAP Gen 2, our own muscle capsid or brain capsid. And we can see that the benchmark capsid, when you look at skeletal muscle or heart tissue, you get really strong targeting, not so much in the motor cortex, there in the brain. V-CAP Gen 2, of course, has the exact opposite phenotype, where we don't see much targeting in muscles, but certainly in the motor cortex, we see quite a strong tropism. And of course, the question is, do we get a best of both worlds capsid where we combine those phenotypes, and in mice, that's what we saw. With V-CAP336, we were able to successfully increase the targeting in the skeletal muscle and heart, certainly relative to the baseline of where the brain capsid was. And in the cortex, we still have a really strong targeting, you can see there in the histology, although we note that in terms of expression in the brain, we did take a little bit of a hit. But nevertheless, I think this validates our approach that we are, in fact, getting a best of both worlds capsid, that we can target both of these tissues simultaneously with one product. And interestingly, in the liver, we noted that V-CAP336 took on the liver phenotype more so of our CNS capsids, which are quite liver-detargeted. So switching over to monkeys, we're going to revisit that multi-tag system that I described, where we do a head-to-head evaluation in the same primates of various capsids. So here we're using AAV9, a benchmark muscle capsid, and our neuromuscular capsid, V-CAP337. So in terms of the quadriceps, we saw an increase with V-CAP337 relative to the benchmark in terms of its ability to target the cells. We're seeing an increase in the amount of nuclei targeted. Although we do note that we saw a little bit of a reduction there in the expression in the quadriceps. But in the heart, we were able to see an increase in both of those metrics. 337 targeted a higher amount of nuclei in the heart, as well as increases expression about tenfold relative to that of AAV9. But I think the brain is actually where we see our most exciting data. So V-CAP337 in the motor cortex of these primates had upwards of 50% of nuclei being targeted in the motor cortex. And that led to about, that was measured at about a 250-fold increase in AAV9 over, or with V-CAP337. Now if we compare that to V-CAP Gen2, and I'll make the stipulation here that V-CAP Gen2 is taken, is data taken from some historical data that we have and we presented years ago here at ASGCT. But in terms of the number of nuclei targeted and the amount of expression that we're seeing, we're more than doubling in either case, which we think is really exciting. In the liver, we see the same thing that we saw in mouse, in that V-CAP 337 is largely reduced and more similar to our CNS crossing capsids than they are the muscle capsids. I will make the disclaimer, a disclaimer disclosure here, that the brain targeting mechanism of both 336 and 337 is in fact ALPL, that is the mechanism by which they are getting into the BBB. And we are continuing to evolve these where we have next generation variants beyond 336 and 337 that we're currently evaluating. So to just give you a summary of what I described here, we've developed muscle capsids with V-CAP290, has about a hundredfold increase in expression over AAV9 in mouse skeletal muscle, tenfold in cardiac muscle. In primates, we're looking at 25 fold over AAV9 in the skeletal muscle and about 13 fold in heart. At low doses, we can see that we're targeting a hundred percent of muscle fibers in the mice. And we've moved on to developing neuromuscular capsids that we think have sort of the best of both worlds there with V-CAP336 and 337. And those are liver-detargeted, more similar to our brain capsids. So with that, I want to thank all my colleagues at Voyager, current and former. As you all know, these projects take a lot of brain, a lot of muscle. So I just want to give a shout out to them. And with whatever time is remaining, I'm happy to take a question or two. I think we have time for one good question. I tell you, thanks. Really exciting. I was wondering, does 336 work in NHP and does 337 work in mice? We did not test them in the other species. So I just can't say for certain. I think to some degree, they both would work in the other species, but to what degree, I just can't say because we didn't do that experiment. Moving forward, the next generation of variants that we are evaluating, we're same capsid, both animals. Thanks. Thank you, Tyrone. Our next speaker is Nolan Graham from UC San Diego. He'll be talking about receptor-guided avitropism engineering via MATCH. Hello, everyone. My name is Nolan Graham. I'm a Ph.D. student in the lab of Dr. Prashant Malhi at UC San Diego. Today I'll be talking about my research, which is titled Receptor-Guided Avotropism Engineering via MATCH. Several of the main issues with AV gene therapy, including high cost and toxicity, stand to be overcome through the development of more specific, efficient vectors. Classically, these are identified through large-scale screens, which are resource-intensive and generally limited by the intolerance of the AV capsid to larger peptide insertions. So previously, we've incorporated unnatural amino acids and peptide tags into the AV capsid. Based on this, we developed the MATCH system, which stands for Modulation of Avotropism through Conjugation to Homing Proteins. Specifically, we wanted to redirect the tropism of a couple of different AVs to desired targets by targeting ligands to the capsid via SPI tags by Ketrochemistry. So to start, we went into AV-DJ and inserted a SPI tag into one of the protruding loop regions on the VP. This was flanked by flexible GS linkers, and then we incorporated this into a fully SPI tag-containing capsid. These capsids reacted overnight at room temperature with SPI catcher, and upon doing so, we saw a shift in the size of the VP as a result of that covalent SPI catcher attachment. We took these to TEM to ensure that we were still forming fully formed capsids, and then upon SPI catcher binding, they maintained their integrity with a slight topology change, and upon going to DLS, we saw that there was a slight shift in the mean diameter of these capsids. So we wanted to preempt an issue at this point that we foresaw, where if we're covering the entire AV capsid with SPI catcher, we wouldn't be allowing the binding of secondary receptors such as AVR, which are required for the AV to effectively transduce and produce. So to do this, we put together a panel of mosaic AAVs with various ratios of AAV-DJ to AAV-SPI tag, with many of them primarily composed of AAV-DJ. In HeLa cells, when we transduced in vitro, we saw that the fully SPI tag-containing capsids lost nearly all of the transduction ability that native AAV-DJ has, and this is likely due to the insertion of the SPI tag peptide immediately in the heparin sulfate proteoglycan binding domain of AAV-DJ. Interestingly, in the mosaic capsids, we also saw a drop in transduction ability upon SPI catcher binding, which is likely due to that steric occlusion of binding sites that I mentioned before. So we took our 5 to 1 and 11 to 1 mosaic capsids and decided to start targeting. So first, we went into resting PBMCs, which are classically a challenging target for AAV transduction, and attached either a CD3 or CD28 SCFEs to the outside of the capsid, and saw that the CD3 targeting SCFEs, particularly SCFE1, were able to transduce a sizable portion of the resting PBMCs without prior activation. We also wanted to go for the brain, and to do so, attached a myrin transferrin receptor binding SCFE to the outside of the capsid, and injected these retroorbitally in mice, and saw that decorated capsids were able to transduce the brain much more effectively than AAV-DJ by itself. So we wanted to bring this into a more popular wild type AAV to be more therapeutically relevant, and to do so, we decided to make match AAV9. AAV9 is a wild type AAV that's already used to target the CNS with engineered capsids, and it uses a different primary receptor from AAV-DJ, and that it uses galactose to bind, as opposed to heparin sulfate-protected glycan. So to do so, we inserted our SPI tag into both loop 1 and loop 2, which are protruding loop regions on the VP, and flanked these by either 1x or 2x flexible linkers, and upon forming full capsids with these SPI tag-containing VPs, we saw that both L1 and L2, loop 2x contained, were able to readily bind SPI-catching. Additionally, when we incorporated these into mosaic AAVs at a ratio of 11 to 1, they produced capsids with titers similar to that of wild type AAV9. So we went back to our T cell targeting and used our CD3 SCFE1, and saw that the AAV9 was able to bind, and it was able to transmit the transferrin. Upon transduction, we were able to transduce over 50% of the resting PBMC population. Additionally, when we went to murine transferrin, we were able to gain transduction ability of the brain almost 100-fold higher than native AAV9. Based on this success, we wanted to bring this into a more therapeutically relevant space using human transferrin receptor targeting, and to do so, we put together a panel from the literature of SCFEs that are targeting human transferrin receptor, and used HEC cells that had either received a Safe Harbor AAVS1 locus knockout or a transferrin receptor knockout via LNT CRISPR, and saw in our AAVS1 knockout cells that we were able to transduce these cells much more efficiently than AAV9, and also with efficacy comparable to that of the VIHDFR1, which is a recently engineered human transferrin targeting AAV. This improvement was attenuated in the cells that received a transferrin receptor knockout, supporting the concept that we're using the transferrin receptor to do transduce. So we took this in vivo with SCFV9 and SCFV12, and used mice that expressed the apical domain of the human transferrin receptor, as opposed to the murine, and saw that particularly SCFV12 yielded DNA level and mRNA level transduction comparable to that of VIHDFR1. We looked at the protein level next, using IHC, and saw that we were getting brain-wide distribution of our transgene, again comparable to that of VIHDFR1. So this was pretty promising, but before we proceeded, we wanted to ensure that this was more applicable for many other labs. So to do so, up until this point, everything had been produced either separately, with separate protein production from AAV production, as a kind of modular system, where you could retarget AAVs, the same AAV to different targets, depending on what you conjugate. But at this point, we wanted to ensure that we could make everything work together in a triple-transfection style approach. So we have our standard triple-transfection plasmids, which I'm sure many of you are familiar with. And this also includes our spy tag rep cap, which we've used up until this point in the mosaics, and then also our spy catcher ligand. So when we did this with the CD3 targeting SCFV, we got titers similar to the wild type AAV9, once again. And when we took this in vitro, with our L1 and L2, 5 to 1 and 11 to 1 mosaics, we were able to transduce a large portion of the T cell population, or PBMC population. So in conclusion, the Match AAV platform overcomes a major challenge in the AAV engineering by allowing us to put full-size proteins on the outside of the AAV. Through a series of small-scale binder screens, we identified T cell and brain cell targeting motifs that were pulled straight from the literature. And one of these in CD3 targeted vectors, we were able to transduce a majority of the PBMC population, while MTFR1 and HTFR1 targeted vectors transduced the brain effectively, with much higher transduction ability than AAV9 and comparable transduction to that of popular engineered AAVs. And finally, the mix and match vector productions enabled one-pot generation of these targeted vectors in a method that many labs will be familiar with using the standard triple transfection protocols. So with that, I'd like to thank my mentor, Dr. Mali, as well as my other co-authors in bold, as well as the rest of the lab. This work has recently been accepted for publication, so be sure to follow up. Thank you. Thanks, Nolan. The floor is open for questions. Are there any? Very quick questions. Do you have the characterization of your, it's called match AAV? Yes, how many copies of the antibody decorated of your AAV? So we're not able to directly quantify that yet, like how many SCFs per AAV. However, based on the kind of stochastic nature of AAV assembly, we hope that it's relatively similar to what we transfect at. How about the full partial analytics? Like, do you see a lot of full vector genome being packaged with your novel AAV? We haven't looked at that yet, but that's a great question. Thank you. Hey, that's really cool work. So I was curious, was there a difference in the amount of spike catcher that conjugated when you did it in vitro as opposed to when you did it in the cell? Transduction in vitro versus transduction in the cell? No, the amount of spike catcher that bound to the spy tag virus when you did it in vitro as opposed to when you assembled it within the cell. Oh, yes. So we use a large excess, massive excess for everything, all the conjugation reactions in vitro. But when we transfect, it's about a quarter. There's a hundredfold excess of protein when we do it outside the cell, but in the cell, we just transfect at a certain ratio, which is about a quarter of the total DNA transfected. And the second question was, so you used spike catcher three, but you used spy tag one. Is there a reason you didn't use spy tag three? Yes. So spy tag three is a little bit longer than spy tag one. So spy tag one is only 13 times. And based on the cross reactivity of spy catcher three and spy tag one, we elected to ensure that our AAVs would package using a smaller peptide insertion. Okay. And the last question is, so when you did your different ratios, your 11 is to one, is that just based on bulk ratios of how you estimated how many spy tags on each capsule? Yes. So we transfect these plasmids at a molar ratio of 11 to one or five to one. So yeah. The fact that they packaged similarly indicates that hopefully they will be at that appropriate ratio upon assembly. Thank you. Great. If there are no further questions, thank you, Nolan. Our next speaker is Kyle Ju. I'm waiting for his slides to come up. There you go. He'll be talking about ACE502, an AAV5 derived capsid engineered by in vivo directed evolution for robust CNS transduction and liver detargeting. Take it away. Thank you so much for your introduction. So before I start, I want to give thanks to the organized committee for having me this great opportunity to share our exciting data on ACE502, which is AAV5 derived capsid engineered for CNS, the robust CNS transduction and at the same time the liver detargeting. So my name is Kyle Ju. I'm the AAV capsid engineering lead at Aminoid Solution and also serve as a faculty member at Ewha Women's University in Korea. Let me briefly introduce our company, the Aminoid Solution. So you may guess what we are doing from the name. So Aminoid Solution is just the science driven biotech mainly focused on developing many different disease modifying kind of therapies, especially for Alzheimer's disease. So basically we have three different core interconnected programs. Mainly we are focusing on the brain tropic, AAV based gene therapy kind of approach and platform and also we are working for many different restoration of clear homeostasis for Alzheimer's disease treatment. Also we are developing some of the diagnostics for Alzheimer's as well too. So ACE502 is served as a foundational gene delivery vehicle for our therapeutic programs. We are mainly working on the capsid engineering. So this is how we discovered that this noble caps of the ACE502. Basically we are starting from the AAV5 and engineer two different variable regions at the capsid at the same time. Variable region 4 and variable region 8 and by substituting the amino acid sequences by using the random peptide display. After that we are using the sequential crucifixis in vivo screening in order to kind of reduce the risk of the single model the selection bias. And then to make sure that it's functional transduction in the brain. So we just once more just pursue the mRNA based screening and evaluation of the lead capsid variants in non-human primates by using the DNA bar coding. And after that the most important question is the protein level expression. So we are just selected the best single the capsid ACE502 by the individual protein level head to head evaluation and validation compared to many of the benchmark capsids. So based on the sequential the mouse to non-human primate in vivo screening. So we were able to progressively enrich many different CNS tropic capsid variants over the AAV5 wild time. So as you can see that the ACE502 and 501 located in a very high fold kind of increase in the average of the enrichment in a non-human primate. And also amino acid pre-perance the HIP map showing that they conserve the pattern of the sequences in two different the barial region of the capsid of AAV5. So across the species we have just enrichment for that many different variants. And if you're looking a little more closely into our in vivo kind of the screening strategy. So we're starting from the much bigger and larger kind of volume of the library which is 10 to the 7 variants. And after the two round of the mouse in vivo screening and then we just selected the 10,000 variants which is successfully the transduced mouse brain and synthesized and we inject into the non-human primate. It's a sign of mongus mechac. And then we further screen them out by using a dual selection criteria which is the CNS tissue enrichment score and also at the same time it's a liberty targeting score which is upper right kind of the corner of the box. The where's we identified our desired variants that we are looking for. And based on that this kind of the in vivo screening so we were able to identify top performing 28 capsids. And then we once again we just do an mRDNA based evaluation for the lead capsids for make sure that is a functional transduction. So as you can see the ACE 502 and 501 could achieve around a 70 fold increase in enrichment compared to the wild type of AAV5 and AAV9 as well in the non-human primates. And also we co-injected published kind of two of the published well known benchmark capsids and compared to our capsid. So interestingly we can have even higher transduction compared to those benchmark and the competitor capsids. And this HIP map showing that our ACE 502 and 501 could transduce successfully in a different region of the brain including the frontal cortex and motor cortex and deeper reach deep brain region which is hippocampus and sub-stentia nigra and so on. And then we see that in peripheral tissues including the liver and heart including the TRG as well. So ACE 501 and 502 is significantly reduced to signal compared to wild type of the AAV5 and AAV9 in mRDNA level of expression. So this is comprehensive mRDNA expression profile compares ACE 502 against all benchmark capsid included wild types. So dark blue is ACE 501 and light blue is ACE 502. So you can see that it's consistently maintained a higher transduction in all the region of the brain compared to the AAV9 wild type from the cerebellum and cortex and deep brain like hippocampus and sub-stentia nigra and thalamus and brain stem and including the spinal cord as well. And this baseline is the AAV5 and then you can see that green dot is AAV9 wild type. And then you can see that peripheral tissue especially in liver and heart. ACE 501 and 502 significantly reduced that mRDNA expression compared to the wild type AAV9. So we believe that it has a feature of the safety profile of the liver targeting. So we move on to the most important question is what about the protein level expression? So we have two different animals for each group and then we have a total seven groups of the non-human primate study including the non-injection. So ACE 501 and 502 and we include the two different kind of competitor capsid as well which is well known and ACE AAV5, AAV9 wild type. So we just delivered the GFP on their CMV promoter and IV injection at the dose of 1E13 vector genome per kilogram. So as you can see that ACE 502 at this dose can achieve around 100 fold increase in the whole brain as a protein level expression compared to the wild type of AAV5 and AAV9. And we just increase that higher the dose of the ACE 502 which is three times higher of the lower dose. So as you can see that around the 3.9 fold increase in a protein level expression in the whole brain which indicating that the ACE 502 could transduce the cell in a dose dependent manner. And more greater thing is in the liver. So we can deliver at the dose of the 1E13. So you can see that around the ACE 502 can reduce the liver accumulation by 99 percent. And greater thing is when we just deliver that higher dose of ACE 502 which is three times higher but still we can maintain the 99 percent of the target in compared to the wild type of AAV5 and also AAV9 as well too. So as you can see, so Western blood is confirmed that our ACE 502 can transduce many different region of the brain in those dependent manner. So kind of cortex region and deep brain region, the hippocampus and sub-central nigra. So we can see pretty much the signal compared to the many different benchmark capsules. And in the liver you can see that's very minimized and reduced signal compared to our CNS tissues. So this fluorescent imaging is kind of provide visually kind of compelling demonstration of ACE 502 in greater transduction in whole brain in non-human primate. So compared to AAV5 wild type, AAV9 has a slightly brighter signal of the whole brain. And then very similar level of expression of the published capsid of two of those, the capsid at this dose. And at the same dose in ACE 502 you can see the brighter signal in the whole brain in a non-human primate. And then even higher and brighter signal in a higher dose which is three times higher compared to the lower one. And we can see there is a very robust CNS and whole brain expression too. So we'll go a little more detail into the, look into the closely into the vision of each brain region. So you can see just so we just sustaining with that the neuron cells with a map to staining. And also we do in situ hybridization with the GFP with the red color. So in a cortex region you can see pretty much at the GFP expression of overlap with the neuron cells. And also in especially for the deep brain region which is a substantial nigra and thalamus we could observe that pretty much bright signal of the protein expression of ACE 502. So before they confirm with the immunohistochemistry to assay to make sure that this is not exaggerated any fluorescent imaging. So very very similar kind of result to be observed. And then it's broadly displaying the whole brain. And then next question is how much percentage of the neuron cells can be transduced by ACE 502. So it depends on the region of that of the brain. From that 20 to 50 percent of the neuron cells can be transduced by that the ACE 502. And interesting is if you compare to the transduction of the wild type AB5 it can be enhanced at around 180 fold increase especially for substantial nigra and thalamus which is that kind of deep brain region. So next one is we just had to have the kind of evaluation for the liver. So as you can see that AB5 is pretty much a chip expression over the liver. And it's very similar the pattern of the result in AB9 as well. And little bit reduced signal in two different capsid the competitor capsid as well. For ACE 502 at the same dose we can achieve around 99 percent of the targetting compared to the wild type which is a very kind of compelling data that we have. So currently we're working for identifying the receptor. Hopefully it's a find some new receptor. It could be another but transfer or anything else. So based on that we're going to get the result very soon. So based on that we can just pursue the rational design for next generation of ACE 502. So this is a wrap up. I mean the slide for the take a message. So basically whole brain of the non-human primate we could achieve around 100 fold increase at the specific dose and CNS transduction. At the same time we can lower that 99 percent of that liver targeting and liver accumulation. So we believe that ACE 502 could be promising kind of capsid for robust CNS transduction at the same time that liver targeting. So that's what we are prepared today. So thank you so much for your attention. I'm happy to answer any question you may have. Thank you. Thanks Guy. The floor is open. Did you look at the vector genome distribution? Did you look at the vector genome distribution? That's a good question. We are just working on that. So basically we are just see that whole that the western blood and different fluorescent imaging. But we want to quantify that how much kind of the vector genome can deliver to the each cells of the each different region of the brain. Yeah, especially in the liver because AV5 notoriously doesn't transduce the liver well. Yeah, that's a good. Yeah. So we are going to work on that. Yeah. Any other questions from the floor? Maybe I'll take this opportunity to ask the question myself. Interested to know the thinking behind the choice of the base capsid. Why AV5 and sort of a follow up question to that. These published capsid that you benchmark with, are they also AV5 based? No. The reason why we chose AV5 is yeah, it's well known as a preg. It's immunity is much lower compared to the AV9. Even though it's AV9 known to the going to the better way to the brain. That's why we're starting from that. And then relatively lower accumulation to the liver compared to the AV9. And all benchmark capsid is AV9 derived capsid. Perfect. Thank you, Kai. Thank you. Our next speaker is Zhenhua Wu. Yeah, perfect. He'll be talking to us. He's from exogenesis. He'll be talking to us about transferrin receptor targeting AV, achieved superior and broad CNS transduction in multiple species after intravenous administration. Thank you very much. Thanks for the invitation for this great opportunity to present here at ASGCT. So this is our forward looking stem in. Our exogenesis bio is AV based gene therapy and we're going to be talking about the random library screening. There are many different ways to screen novel AV capsid for CNS penetrating capsid. The random library is really great. At one time you can screen up to one million virus at one time. However, there is also caveats. Whenever there is hits you really don't know the mechanism. You really don't know which receptor binds to. So with that we really think about what we can screen seeing as penetrating capsid differently. And this is the new system that we generated. We called it receptor targeting AV engineering system. In this particular system what we do is that we come up with a transfer receptor binding library. It's a peptide library. And with that transferring preferring library we insert into AV and then we screen directly into monkey and see whether with a known mechanism whether you have a better chance to find a CNS penetrating capsid. And we have a lot of, we use heavily on AI in terms of generating the peptide library. I'm going to go through the details of the AI process. At the end we found a very good CNS penetrating capsid we called RTAV001. And when we got the capsid we really tested the affinity against the human, TFR monkey, TFR and also mouse TFR to see whether there is any difference. And the end result is the affinity against all three species are very, very similar. So we're very encouraged by that because we know if that particular capsid works in mouse it should work in monkey and also should work in human. And also we later on did a TEM study to see whether we can figure out the binding pocket of the particular receptor and as shown here actually we did figure out the binding pocket is actually first is allosteric. It does not hinder the interaction of transferring binding to transfer receptor. And also we figured out the monoclonal magnets why this particular association really render the specificity towards the TFR receptor. So we know exactly where it binds to, we know exactly how it binds to transfer receptor. And we put this particular capsid into mice. As you can see here from the top left is the GDNA results. As we can see we see 30 to 60 percent fold increase in terms of, I'm sorry, 20 to 30 percent increase on the GDNA. And on the top right is the MRI expression and we see up to 60 fold increase in terms of the RNA expression. At the same time we also took a look of the GDNA distribution and MRI expression in the liver and in the muscle because we have some concerns that TFR targeting AV might increase the distribution to the muscle and the liver. As a matter of fact we see about a three to four fold decrease in the liver and muscle wise we see exactly the same distribution expression compared to AV9. Of course we put into mice and we did the immunohistochemical staining and as we can see on the top panels are the AV9 staining. There is not much of green transgene expression. On the bottom is the RTAV001. It really transduces very efficiently across the brain regions and we quantify it about 50 to 80 percent of the neurons are actually transduced. At the same time we also take a look of the transduction of microglia and astrocytes and indeed RTAV001 can transduce some astrocytes and microglia but in terms of the percentage it's less than five percent, much less compared to neuronal transduction. Maybe because the RTAV001 targeting TFR and neuronal has TFR receptor on the surface of their neurons. Also we really take a look of the RTAV expression whether that induced any histological change to the brain and based on this pathological studies we really don't see much change in the neurons. So prove that this is safe in mouse. Next is we really put that RTAV into monkey studies and because there are a lot of reports in terms of if you switch monkey species the AV transduction will change. So we use actually two species, monkey, son of monkey and also rhesus monkey. In these studies we co-dose RTAV001 with AV9 at the same time. Each of them is a full E13 VG per kg dose and when we combine them both capsules contains the same transgene but with different tags. So at the end of the study we can really use IHC to study which capsule expressing at what level and of course we take all the organs and all the tissues and do all the DNA, mRNA and the protein expression study. This is the results from rhesus monkey. Again on the top left is GDNA. In the middle is mRNA and on the right is the GDNA and mRNA results in all the other tissues. Very, very striking difference between AV9 versus RTAV001 and for mRNA expression wise we reach up to 250 fold increase across all the brain regions very consistently including the spinal cord and when we look at the DNA distribution across other brain, other tissues only liver again had a three to four decrease in distribution and mRNA expression and across all other regions including spleen, including muscle, including other organs are all almost the same. So has slightly liver targeting property for RTAV001 and again this is immunohistochemistry training. As you can see this is really across all the brain regions RTAV001. It was very effectively transducing neuronal cells compared to AV9. Almost no positive staining was observed and even including the spinal cord up to 70% of the spinal neuronal, motor neurons are really transduced and when we quantify the neuronal transduction about 30% to 70% of neurons are transduced depending on different brain regions. Highly efficient. This is the spinal monkey study and the same results so I'm not going to go through the details and this is staining exactly look like similar to RISCUS monkey. We also took a look of the potential liability for anemia, liability for this particular capsule because as we know transferring targeting can cause anemia and of course reticulocytes reduction and red blood cells reduction. As you can see from here really don't see any concerns of anemia and at the same time because of this AV9 plus RTAV total dose is AE13 BG per kg so as expected we see some AST and ALT transient increase but this is really expected. So this is really very well tolerated capsid in monkey. We also took a look of the histopathology of this particular capsule in monkey brains and also liver as well and we really don't see any concerns in terms of morphological changes after RTAV001 dosing and the same thing for RISCUS monkey as well. Again we now put into the capsule into a transfer receptor, human transfer and knocking animals as you can see from here. This is highly efficient in human TFR knocking animals as well compared to white type. The transduction rate is almost look like almost the same so we know that that also works on human TFR even in the TFR knocking animal. Also we put that into human cells. We use the HGK cells which express TFR highly and on the top is the AV9 transduction on HGK and really nothing really transduces or very low transduction rate on HGK at that MOI but at same LMI we can see really tremendous intonation of the capsid in the HGK cells. So again this is a second sort of indirect proof that it might work in human, at least it binds to human receptors. And we took a look of the manufacturability of this particular capsule and compared to AV9 the manufacturability is really very similar including the upstream yield and also the upstream full and empty capsid very similar to AV9 so we think this is highly manufacturable as well. So next thing is we really want to put this RTAV01 into human testing and the first indication we choose is SMA. So we did this particular study in SMA transgenic animals. As you know SMA transgenic animals if you don't rescue majority of the mice will die before three weeks on the right as you can see the survivor curve. The black line is the animals that without any treatment or with vehicle treatment the animal dies very quickly and two red lines are AV9 version of the vectors expressing SMN1 protein and they can rescue the survivor really dramatically especially at 2E14 VG per kg dose. So this is a highly efficient can push the survivor curve all the way to 120 days but even at a four times lower dose with RTAV001 the survivor curve rescue effect is really really significant compared to the AV9 version of the vector. So we believe it would really work in the SMA model and on the left is the distribution data even at a ten times lower dose RTAV001 can have much higher tissue distribution in all brain regions. So this is again indicating the RTAV001 is really good capsid for the SMA treatment and currently the capsid is under the development R&D development and we plan to put this particular product into human testing hopefully by the end of the year or beginning of next year. This is the conclusion basically RTAV001 is seeing as penetrating AV with a really different mechanism and very known mechanism toward BBB and we believe this particular capsid is highly manufacturable and we are on the go in the end of the studies and put that into human testing. And thank you very much. I would address any questions you might have. Floor's open. Hi that was a great talk. I was wondering if you looked at the immunogenicity profile of the capsid in non-human primates if it was comparable to AV9 or not. Very good question we did and we believe this is very similar to AV9. Great talk. So you only can. Can we take a question from the other side we'll keep switching. Okay. Hi great talk. How did you land on the I think nanobody that was inserted into AV9 for RTAV01. I think you just glossed over that very briefly. The screening method I guess. Yeah. The library the TFR binding library we generated consists of really different binding peptides. So we have a very long, short, including like nanobody you know candidates. So there's a really variety of candidates in that particular library. And among this library we found this particular AV capsid RTAV001. And how was this library designed? We as I mentioned we use a lot of AI tools. Again there's many different steps. We generate the scaffold first then filling the sequences. And really we're trying to predict the binding affinity between the binding peptides versus our transfer receptor. Thank you. Yeah. Thank you. Great talk. So you inserted the affinity peptide in the well-tapped AV9 on one loop right? Correct. Yeah. So you'll keep the other side unchanged. You don't do the mutation like reducing the liver targeting? No. We didn't do liver targeting mutation. There's only one insertion of the TFR binding peptides. Okay. How that you know you only insert TFR affinity peptide and it can you know enhance BBB crossing. When the AV cross into the brain how can they mainly targeting the neural cells? Yeah. As I mentioned the neuronal because neurons have a high expression of transfer receptor. And then maybe that's the reason neuron has better uptake of the RTAV001 compared to astrocytes and microglial cells. Yeah. This is one explanation and microglial and astrocytes are highly turnover cells and within a month maybe some of the transduction are gone already. Okay. Another small question. How long is the peptide? How long is the peptide? Yeah. We have various of different peptides. In this case the peptide is really big. It's not seven or ten or twelve amino acids. Really long big peptides. And the fundamental reason is because smaller peptide will have smaller interface with the transfer receptor and any change on the receptor for example species change. We may disrupt this binding. That's why we're looking for a bigger binding pocket. And in our TEM study we finally map out the binding sequence or binding epto of that interaction. And the binding sequence, the sequence is 100% conserved in mouse, in monkey and in human. That's probably why we saw similar affinity in mouse, monkey and human receptors. Okay. Thank you. Thanks. I apologize. We are a little bit out of time. I encourage the questioners to ask the question after the session to the speakers. And let's maybe thank Zenduwa for his presentation. Thank you very much. Our next speaker is Angela Enrica Arajot. She's going to be talking about the future of CAR T cells towards the in vivo generation of stable CAR T T cells with T cell tropic adenos associated viral vectors. Thank you for your kind introduction and thank you to the organization to give me the opportunity today to present my work about in vivo generation of CAR T cells using IV vectors. So T cell based gene and cell therapies are mainly focused on the use of CAR T in the ecological field. But as we know the field is currently expanding for the treatment of autoimmune disease, infection disease. And of course we can also modify the cells for the treatment of immunodeficiencies. And currently the genetic modification of T cells is based on the ex vivo modification usually using sorry, integrative vector like lentiviral vector or nucleofection. But this type of process time consuming, very costly, depend on the presence of specialized center. And this, of course, can decrease the accessibility of the type of therapies and modify in vivo. The T cells directly in the bloodstream of the patient can help reduce some of this problematic. And one of the vector that's been proposed for this type of terraces are AVs. And as is the focus of today's session AV can modified to retargeting. And in my project we start with a AV2 peptide display library with a seven amino acid peptide insertion in the second of I speak of AV2. And this library was then selected in vivo in non-human primates. And we isolate different T cells populations in particular in focus on naive and memory cells of the CD4 and CD8 T cells. And from the NGS analysis of the DNA isolated from T cells we identified 21 candidates that were particularly enriched on the T cells. And we can see here the enrichment for our 21 candidates in the different population. And we can see as our very strict condition for the first selection round allow us to directly select our candidate after just one selection round in vivo. And generally the variants were enriched in all the different populations. But some show particular preference for one of more T cells population like for example variant 6 just in the CD4 cells memory cells. Or variant 9 in the CD4 different CD4 populations. And to give you overview how we test and confirm the trophism of our 21 candidates we first test them in vitro in human T cells. And this allows us already to reduce the number of candidates from 21 to 12 that were equipped with GFP with a unique barcode for each of them. And we then test them in human blood for both efficiency and specificity. And we perform also a bio distribution in vivo in 6 mice. And for this experiment we were able to identify two final candidates that we renamed H3 and H6. One particularly efficient and one particularly specific for the transduction of human T cells. So now let's go back and see why we choose these two candidates. And we can see here our test on human primary T cells in vitro where a different GOI so vector genome per cells. Our variants outperform AV2 wild type in both CD4 and CD8 population that we can see in blue and in green. And the results were even more impressive when we compare these two variants with AV2 and AV6 that is considered the gold standard in human blood. So after transduction in human blood we isolate also the T cells and the T cells population to look for both efficiency and selectivity. And of course we can see here in the T cells CDNA that AV6 is doing the job quite well as expected. But if we compare the expression at the CDNA level with the GDNA level amount we can see that our variants are much more effective in T cells, in expressing on T cells than AV6. And if we go even forward and look at the ratio between the expression on the T cells to the non-T cells population we can see here that AV6 is not specific at all. While our H3 and in particular H6 show a very clearly selectivity for the T cells population. And this was confirmed when also with flow cytometry for GXP expression when we test the variant separately in human blood. And we can see here that AV6 in black clearly has a better expression efficiency in non-T cells when it's instead over performed by our variants in the T cells population. In the black 6 mi in our bio-distribution we can see here in light green and dark green that two of our vectors show a particular enrichment in the T cells and CD4 and CD8 population. And this is confirmed when we look at the GDNA level. And we can see that the GDNA population is isolated for both the spleen and the blood of the mice. Angiotri is one of these two candidates and this confirms that our variants are able to target in T cells in different species. And the other variants, variant 2, share a very close similar peptide insertion that also confirms the strongness of our selection. And the AV present a limitation. They lack an integrator so they remain usually an epizomal form inside the cells. But this will mean that we lose color expression for every cell division in the activated T cells. And as been shown of course that for CAR T cells therapies the persistence in vivo of the CAR T is linked to the success of this type of therapies. So what is our solution? And our solution is using our selected vectors in a dual IVA platform where one vector will encode for the transposase and the other will encode for the CAR in a transposable element. And the idea is that when both vectors enter the T cells the CAR will be integrated and we will have a stable CAR T cells. And the transposase, transposone approach is already in clinical trials using nucleofection so ex vivo production of CAR T cells. And our idea is to combine this approach with our AV vectors to a stable in vivo CAR T cells generation. And we already show and test this approach in vitro, in primary T cells. And we can see in the central figure that we test two different CAR constructs and we were able to obtain more than 20% of CD19 positive CAR T cells with a dual vector approach or more than 10% of our ROR1 CAR T cells. And on the figure on the left we can see that the CAR expression was stable maintained for more than four weeks in the presence of the transposase vector when in its absence the expression of the CAR T cells decline over time. Also the CAR T cells produced with our angiotri vector that we can see red were at the same level or outperform the nucleofection produced CAR T, so the current standard method using this technology. And we can see in the last picture on the right also the AV produced CAR T cells show a better viability to the one produced using the standard nucleofection method. And maybe you noticed that we also test AV6 as comparison and it was still outperforming our variance for the percentage of the CAR T cells produced. However when we test the CAR T for cytotoxicity assay for specific cell killing in both CD19 or ROR1 positive T cells our angiotri CAR T show similar or even better killing activity than the AV6 produced one. And in both cases they were outperforming the nucleofection produced ones. So I'm already to my summary so I hope that I show you today that we have identified two AV candidates that have increased production, efficiency and trophies for T cells both in vitro, ex vivo and in vivo. And that a dual AV approach can be used with these candidates to have in vivo stable CAR T generation. And now after our testing we are now moving forwards with the in vivo evaluation of our approach in humanized mice and we are really hopeful that our system can become easily adaptable product form for different CAR constructs and this can help reduce the cost and making CAR T scalable off the shelf product that can increase their accessibility in the world. And I want to finish my presentation thanking especially the CAR T team that is responsible between MHH in ANOVA and the University of Woosburg of this dual AV approach and of course all my group in ANOVA and all our collaborators and my PI in the GAP UNI and I want to thank all of you for listening to me today and now I'm open to answer any question. Floor is open. Thank you for the presentation. We're already struggling and making everything we can to decrease the cost of making one AV. Here you're doubling the cost of the dose. So I mean it's not the first time I see it. It works for sure. But now we are more seeing trying approaches where if we're using two vectors trying to use other type of vectors that maybe are decreasing the cost. So there's already examples of that in clinical such as VLP, VEDV, LNPs as well combined with the AV or there's also I've seen people being able to reduce the size of the transposase or miniaturize so it fits into everything fits into one AV. So I would suggest if you want really to go for affordability and accessibility to the patient that you try to think about that strategy because this is not going to solve it. It was great talk. Yes, thank you so much. It's an interesting point. But just to comment on this for now we are going at least what we saw for the in vitro results and what we plan to test in vivo if it works. We are going to the same dose that people are testing in animals for just one AV approach but like not integrative and for now it's working so we didn't have to increase the dose. But yeah of course it's a great point. Thank you. Yeah, very nice work. I'm Wen from VCU. I have two questions. The first one I guess you have some justification why you use AV2 for the library not AV6 because AV2 not strong enough to transduce T cell. Yeah, it's true but AV6 it's very efficient when tested in vitro already in vivo it's clearly struggling to do the job even more than AV2 because it's diluted and targeting the liver. And for what we saw at least it's when we go already in the blood or in vivo in the mice our vector it's able to do the job much much better. And last question in the mouse model maybe you have the point I missed that. What's the maximum efficiency for mouse T cell in vivo you can get? In the mouse primary T cell so what's the maximum transducing efficiency? Oh yeah for that experiment it was barcode so at the end the results was obtained through the NGS reads and we compared we're in a stage and we compared the different variants between them so we don't have a single capsid efficiency for the mouse model. But we are currently both in this project and other project that were born from my capsid testing the efficiency in mouse of the single variant so we'll be in my next talk. Thank you. Excellent if there are no further questions thank you. Thank you. Thank you. Our next speaker is Samantha Howard from Alexion. She is going to be presenting an AV capsid presenting a miniaturized anti transfer receptor antibody enables systemic CNS wide distribution with minimal liver tropism and cross species conservation. So take it away. Thank you. Hi everyone my name is Samantha Howard. I'm a senior scientist in biology in the genomic medicine unit at Alexion AstraZeneca and we're based in Cambridge. I'm excited to be here today to share with you some of our bio distribution data from evaluating novel capsid that enables broad CNS distribution and Liberty targeting. Here's my disclosures. So just AV is a novel. You can actually capture that provides broad distribution and middle liver expression. It's been developed by JCR pharmaceuticals and license to Alexion AstraZeneca just AV crosses the blood brain barrier via transfer mediated transcytosis which is a clinically validated pathway with other modalities including those that has been already in the clinic by JCR pharmaceuticals. So it's been designed to engage both semologous and human TFR which does enable animal models for multiple preclinical studies and how it actually engages transfer receptor is it contains a miniaturized antibody that recognizes HTFR. And just to orient you with the rest of the data in this deck payload that we use to evaluate by distribution has a neuronal specific promoter. So first we want to assess how the just AV performs in vitro and to do that we compare just AV versus AV nine and both contain the same gene. We're able to see greater transaction with just AV versus AV nine and a human neuroblastoma cell line and that was sustained across all the MI tested and also with the ability to see these high transaction and are in expression rates at these low levels of the brain. So we're able to see a little bit of a better understanding of the gene and the gene. So next we want to look at how this capsule performs in vivo. And so for these two cases we're going to look at just AV versus AV nine and just AV versus AV nine and a human neuroblastoma cell line and that was sustained across all the MI tested and also with the ability to see these high transaction and are in expression rates at these lower MI's. So next we want to look at how this capsule performs in vivo. And so for these capsules selection studies what we did was evaluated the just AV with the same trans genes versus AV nine in HTFR knock in mice in a line that was developed by JCR and Cinemalgus macaque. Both animals receive a single IV injection of the AV and were taken out for a set endpoint with mice in vitro. For evaluating the bio distribution we paired both biochemistry and histological analysis in particular having one hemisphere sent for the biochemistry workflow either micro dissected in the mouse or tissue punches in the NHP and the other side for the And so for these two cases we're going to look at just AV versus AV nine and a human neuroblastoma cell line and just AV versus AV nine in HTFR knock in mice in vitro. And so for these cases we paired both biochemistry and histological analysis in particular having one hemisphere sent for the biochemistry workflow either micro dissected in the mouse or tissue punches in the NHP and the other side for histology and linking molecular and spatial bio distribution readouts such as vector genome copies, RNA expression, protein expression or histology looking at transgene RNA scope with tissue hybridization as well as new end immunohistochemistry to understand distribution in their own since we are using a neuronal promoter. So now into the data. So on the left hand side we have data from the HFR knock in mice and what we did here was compare either vehicle, AV nine, PHB a known mouse, Bb penicillin capsid as well as just AV at two different doses. What we were able to see was significant increase in brain vector genome copies and transgene RNA expression significantly above AV nine and in addition we also saw significant Liberty targeting in the mouse. And then on the right hand side we have the data in the NHP and so again this was a single IV injection with endpoints of four or 13 weeks and comparing AV nine versus just AV and the filled circles are the four week endpoints, open circles are the 12, 13 week endpoints. And so we're able to also see increased vector genome copies and transgene RNA across various brain regions at various levels but still significantly above AV nine as well as significant liver detargety about 213 fold less than AV nine. And this also sustained between the two different doses. So while biochemistry analysis does allow us to understand the bio distribution and RNA DNA protein relationships we are still limited to the bulk expression so we did want to pivot to including the spatial analysis to have a more thorough characterization of bio distribution with the cell type of regional specificity. So here are some example images from the different bio distribution studies in the mouse and the NHP so on the left hand side again we have the mouse RNA scopes staining and then the right hand side is the NHP and so you can see that there is broad distribution definitely different patterns and intensities so we went after going ahead and quantifying that. So using a series of AI image analysis we built different algorithms for each of the different brain regions that have the different patterns of distribution and we're able to quantify the percentage of transgene RNA expression and then also paired that in subsequent slides looking at the new end positive cells to quantify an estimated level of neuronal transduction. So you can see in the mouse there is significant expression in some regions getting close to 100 percent at that dose too and then in the NHP you can also see a dose dependent increase across the various doses with some regions also at the top doses reaching closer to that 100 percent. But it's important to note too that we did see that just that the just AAV transcecment in the mouse is about fivefold greater than the NHP when matching at the same dose. However what was nice to see is that the regional patterns are pretty conserved with certain regions such as the pons and thalamus trending to be the highest expression in both the species. So considering that the pathology is as we all are very familiar with here about 95 percent between the synote TFR and the human TFR we know that small sequence differences can make a difference in impacting efficiency but still it's nice that just AAV remains one of the few transfer in BbP capsids with translatability across both species which does enable various preclinical workflows. So in conclusion I share with you that just AAV is a novel BbP-henticin capsid modified and developed by JCR pharmaceuticals that contains miniaturized anti-TFR antibody that provides broad brain distribution and liberty targeting. And just AAV recognizes both the synomologous macaque and the human TFR receptors which does enable use of the same type of antibody. And in the capsid selection study we did observe this put in neuronal transduction as well as spatial analysis does support this as well. Also realize I didn't quite mention yet that we also did some of the pathology reviews and we didn't see any TA-associated findings so that was a little bit of a surprise. So our ongoing work includes these datasets to characterize bio-distribution in specific neuronal populations which will then support indication selection and enabling pre-IND packages. So that's a really nice example of what AAV does. Thank you. Sorry, but I also acknowledge everyone. Yeah, really appreciate all the work. Let's take any questions. Floor is open. Hi there, nice talk. So can you comment on the manufacturability of these constructs? Do you see any aggregation in your final product? Can you comment on the homogeneity of your final product? Definitely there is some different nuances with novel capsids always independent of the type of antibody that we have. Thank you. Very nice talk. Good job. So I've got a question for you. When compared to AAV and AAV9, how does targeting the liver really affect the So it's definitely lower with just AAV. Is that your question? Yes. So we saw liver detargeting. Do you have the numbers? The ratio relatively? Oh yeah, it was in the deck. I am curious if you've measured the seroreactivity of this capsid compared to AAV9? It's comparable. So you kind of hinted at this at the end, but I'm curious if the antibody targeting for the transfer and receptor also changes the nature of cell specificity within the tissues where you didn't have 100% transduction. Do you get any kind of change in cell specificity that you would expect out of some of the serotypes or are you still looking into that? Yeah, we're still looking into that, but as commented earlier, we know that transfer receptor highly expressed on neurons and that could also be, there could be some differences there as well. Next. Hi, great talk. I'm going to ask you a question I couldn't ask before to the transfer receptor. The previous talk that I said, it seems to be a disconnect between the engagement for transcytosis, you know, crossing the barrier and then transduction in other tissues that are known to express the transfer receptor, liver, muscle. Is there a mechanistic understanding that you can comment on to understand this? Yeah, I think you're getting at a really big question in the field, but I think there's some excellent work out there showing that affinity makes a big difference. So, you know, each even small sequence changes can make a big impact on affinity. So I'm sure that is a huge part of it. It's like a sweet spot that is better for transcytosis, let's say, but it keeps you away from other tissues that are bigger and may suck up. Exactly. Yeah. And there could also be, you know, different paths of expression and perhaps even different tissues have different affinities that are needed to get into different regions. So I think it's a big field of research. So thank you. I have a question for you. Since it's an antibody based sort of binding to TFR, as I understand that, can you comment maybe a little bit on the structural aspects of this binding? Do you bind the same spot that the transplant binds these receptors or do you bind away from it? The reason I'm asking is because obviously we know there are capsules that bind, you know, and don't bind human or vice versa. And I'm trying to understand if you sort of nail down that sweet spot. Yeah, I can't comment on like the exact area, but it's conserved between the species. So it was taken into account. OK, there are no further questions in the room. Thank you, Samantha. Thank you. We have the last speaker for the session, Rebecca Winkler. Rebecca is going to talk to us about the variable region one of AB9 capsid is a permissive site for peptide insertion. Whenever you're ready, Rebecca, take it away. OK, thank you all for coming today and thank you for the opportunity to present our research. So as he mentioned, this is a more of a proof of concept study looking at variable region one. No disclosures. For a quick review, the AB capsid is a three dimensional sphere made up of 60 subunits. And those the subunits are one of three isoforms, VP1, VP2, VP3, that would result from different start sites in the cap gene. And then also coded for in the cap gene is the AAP protein, which is not part of the final capsid, but it's involved in capsid assembly. And that protein is interesting because it's in a one plus frame shift from the viral proteins. So I think like the jargon is a little confusing because VP1, VP2, VP3 are viral proteins. The VR one through nine are variable regions, which are different surface regions that would vary between serotypes. And when people use like a peptide library or something for capsid engineering, typically they will insert the peptide either into variable region eight or variable region four. We instead decided to look at variable region one to see if that was going to be a region that could be used for capsid engineering. So we're starting with an AV9. And then the so in the box on the left, the sequences in blue are the variable region. And then the red is the peptide that we inserted. And so, like I said, this is a proof of concept study. We used a flag tag peptide because it has a antibody that recognizes it very well. And it also has an enzyme that will digest it. So we have two different binding partners, two different methods that we can use. --- ## AAV Trafficking {#aav-trafficking} **File:** `AAV_Trafficking.mp4` **Type:** Video transcript (Whisper medium) ### Pre-analysis AAV Trafficking session. Expected: receptor interactions, intracellular trafficking, ML models for tropism prediction. ### Full text / transcript Welcome everybody to the session new insights into AAV trafficking we have in the session and Just as a reminder Do not use flash photography in the session Probably distracting and I don't think I need to remind everybody anybody about the Wi-Fi anymore at this point So let's just go straight ahead to the first speaker, which is Sloan one From UMass Chan Medical School My name is Sloan one Reporting our study on behalf of dr. Tweet Monty and sway Unfortunately, she can't make to the meeting today. So I Presenting on behalf of her on our study Advancing capacity engineering and the transcriptional regulation towards decoding the role of a VB1 you So the awkward gene therapy is at the forefront of the gene therapy Especially since the first AV gene therapy product laxative and I get approval from FDA And then the majority of the awkward gene therapy is conducted through the subrational injection but with a number of limitations such as Limited to therapeutic efficacy and then the complexity of the rational surgery as a alternative intravitory injection Theoretically can deliver more of our vectors --- ## Lir AAV LLM {#lir-aav-llm} **File:** `Lir_AAV_LLM.mp4` **Type:** Video transcript (Whisper medium) ### Pre-analysis Lir group — AAV + LLM. Expected: large language models applied directly to AAV capsid sequence design or functional prediction. ### Full text / transcript Our next speaker is Thomas Peacock. He's going to talk to us about navigating the AAV capsid sequence space using virus specific protein language models. Thank you. Good morning everyone. My name is Thomas Peacock. I am an AI research scientist at Lear Therapeutics and yes today I will introduce some of the work we've been doing developing virus specific protein language models for AAV capsid design. At Lear we are aiming to reimagine viral delivery starting with AAV. So we have developed our multimodal lab in the loop platform navigator to generate AI optimized libraries with high impact mutations not just in a single region but across the entire viral surface. These libraries are validated through a combination of our in vivo tissue screens for tropism and potency and in vitro assays for immune evasion with each experiment feeding new data back into the navigator platform to retrain and refine our models. Navigator itself is a holistic AI platform for viral vectors. Some of its core components include our virus specific DNA models that we use for library generation, our structural analysis suite that we use for rapid capsid assembly and large scale docking screens and our protein language models that I will talk about throughout the rest of this talk. We've already made some exciting progress using our pipeline and even in our early results we have observed clear improvements in both enrichment and tissue detargeting across our libraries. As a brief introduction to protein language models these are neural networks that map amino acid sequences into high dimensional vector representations to capture the patterns that link protein sequence to protein structure and protein function. To learn these patterns the model is trained on an objective like masked language modelling where it is required to predict amino acids that have been intentionally hidden in the input sequence by observing the surrounding sequence context. And by iteratively updating its own weights the aim is to reduce the difference between its predictions and the true sequence. And the real beauty of this strategy is that the model is learning relationships between proteins using sequence alone. And this allows us to leverage the millions and millions of protein sequences that we have collected as a scientific community to learn across the full diversity of protein space. However when it comes to very specific tasks such as AAV engineering this enormous diversity becomes something of a double edged sword. Popular protein language models like ESM2 are generalists. They are trained on representative sequences across the whole protein universe but they lack the dense coverage of AAV sequence variation that is important if we want to model for example tissue specificity or immune evasion. The plot on the left here shows a sample of the ESM2 training data with the viral sequences highlighted in pink. And then on the right hand side we have subtracted the background protein families that were shown in purple and supplemented the data with additional publicly available viral sequences including AAV sequences. And it's clear I think that not only are the AAV sequences not present in the ESM2 training data but they also largely fall out of distribution of the training data. So to address this problem we have retrained the ESM2 architecture using our viral data set shown here to create a specialist model that is tailored to AAV engineering tasks. We have benchmarked our specialist model that we are calling Corsair against the public protein GM viral data set and we have found that Corsair shown here in green outperforms ESM2 in pink across all the metrics that we've computed and for all the model sizes that we have tested. While it is important for us to validate our models on public benchmarks our main motivation is to apply them to our own in vivo data sets. We have refined Corsair on results from tissue experiments that we have done in mice and this allows us to use Corsair directly to design our new variant libraries. To apply a protein language model to our own endpoint tasks like tissue tropism or liver avoidance we need to transfer the learned representation into our AAV predictive setting and we do that by adapting the pre-trained network that we've previously been talking about with our new label data sets. The simplest way to do this is through fine-tuning. This is where we extend the network shown in pink with a task specific head shown here in green and this time we continue the training but for a regression or classification objective rather than the masked language modeling objective that we discussed earlier and we do that using our labeled examples. When fine-tuning there is a key trade-off between speed and performance so allowing all of the weights of the model to update offers us maximum flexibility to adapt to the new task but is computationally expensive whereas we can freeze some of these model weights to make the model faster and lighter but it will limit the ability of the model to adapt. We've compared performance when freezing all of the weights, some of the weights and none of the weights as part of our benchmarking. We've also compared these models to the LoRa fine-tuning method shown on the right hand side here and under LoRa we insert small trainable matrices into the model while keeping the base network frozen and it's these matrices that update through the training process rather than the full network. The aim here is essentially to act as a shortcut achieving high performance by keeping the training fast and memory efficient. The results of this benchmarking showed that while reducing the number of parameter updates through freezing does improve training efficiency it generally comes at the cost of performance. The plot here shows the effect in one of our heart tissue datasets but we see a similar pattern across tissues, across both classification and regression tasks and across different model sizes. In contrast the LoRa fine-tuning method tends to preserve the performance close to full fine-tuning while significantly reducing training time so as a result this is generally the approach we have taken for our future model iterations. When comparing Corsair and ESM2 on our in-house tissue datasets we observed that Corsair improves performance across tissues. The radial plot on the right here shows varying spearmor and correlations when predicting relative transduction across one of our recent tissue libraries and this reflects our targeted library design where in this case we intended to probe the fitness landscape of a localized region on the capsid surface. Across the panel we see a meaningful effect size improvement using Corsair which gives us a more stable foundation for modeling transduction and identifying both high and low performing candidates. While the improvement on the previous slide may appear modest I think it's worth highlighting the effect that this can have in large libraries. So here is an example, a toy simulation that we ran where we start with an initial pool of 5 million variants which is a fairly typical of our current pipeline. Then if we assume that 1% of those are hits and that we have two models we are comparing, models A and B, that achieve spearmor and correlations of 0.50 and 0.52 respectively then when we screen the variants with those two models and we select the top 40,000 to proceed to experimental testing we find that model B identifies over 400 additional hits compared to model A so not an insignificant increase I think for just that small 2% difference in spearmor and correlation. Alongside refining our models with more complete AAV datasets we are continuing to iterate on our model design to further improve performance. One recent example of this is our multi-tissue model. So this builds upon the core set architecture but here we are training a single model jointly across all tissues. So this not only improves our training efficiency because now we are just training a single model rather than the separate models for every single tissue but it also allows the model to use cross tissue patterns to hopefully improve its predictive performance. The plot on the right here again shows the spearmor and correlation when predicting relative transduction but this time for our multi-task model compared to our single task models in green and pink respectively and we found that the current version of our multi-task Corsair model outperforms our single task models on all tissues except for liver and spleen. To sum up our protein language module is just one component of the broader navigator platform for comprehensive modelling of the AAV virus capsid. It sits alongside our other models and our library generation tools which together enable the design of variant libraries that we validate in vivo and produce the data that drive the next generation of models which power our navigator engine. To conclude I'd like to thank the rest of the team at LEA. A special thank you to Stephen for his work on the modelling. I'll be here at ASGCT for the rest of the week along with our co-founders Killian and Hugh so please do come and find us if you would like to know more about the work we're doing or if you are interested in potential partnerships or collaborations. Enjoy the rest of the conference and thank you very much for listening. Thank you very much. If there are any questions from the audience I have a very naive question. I think you did mention the Corsairs can also be used for immunization. Have you tried any of that and have you put them to test? Yes so we have generated immune data sets and we do train or we refine the model on those data sets. I didn't have time to show any of those results today. We have shown them at a previous conference. So yeah we see kind of a similar performance I think so we are not getting like sort of perfect performance but we are identifying candidates that we can take forward to future screens. In our most recent immune study we did identify a pool of candidates that had a specific motif that was quite interesting where they like the variants that performed well in immunization sort of shared this motif. Thank you. I think we have a lot of questions from the audience now. Hi I'm Jason from Sanofi. Great talk. Could you reiterate some more on the training data used for training Corsair like specifically for the AAV variants? Were they like all different AAV stereotypes? Were they all like insertions or just like slightly mutated versions of a specific serotype? Yeah so for the pre-training step which uses the huge number of unannotated sequences ESM2 uses the UniREF50 data set which essentially clusters protein sequences so you miss out on a lot of virus sequences that share similar features. So part of our data set has been just general viruses both double-stranded and single-stranded viruses to kind of keep our data set somewhat diverse rather than just AAV sequences. We have added all of the AAV like wild type serotypes into the data set and then we've really looked for as much AAV data as we can find so some of the data is well a lot of the data is AAV2 and AAV9 just because that's the sort of data that's available. We are working in AAV9 at the moment so that's that's sort of fine for us. How this would translate to other serotypes in the future is kind of an open question for us. Hi thanks for the talk. I was interested in your plot where you showed that AAV lives in a different part of the latent space than the other viral sequences. There was still a lot of structure and I was wondering if that recapitulated what we know about AAV serotypes, biogenetic relationships or even like species tropism. Yes so I think that the way that you plot this is kind of not misleading but like I think we so we this was not really the latent space what we did was we we did a KMA we split all the sequences by KMAs and then sort of plotted based on frequency of KMAs here so I think that does mean that you see this kind of distributed pattern for the AAV where you have a lot of single mutations that's spreading out the data. The majority of the ESM2 training data that while the viral population is double stranded DNA viruses so I think this is why it clusters quite heavily in the center and the AAV sort of spreads out around the sides but yes I what we didn't want to do was to run all of this data through ESM2 to plot the AAV data compared to the ESM2 data because then you've I mean it's very difficult to tell whether that it falls out of distribution or not when doing that. Hi Dan Cox thank you for the talk. Well one question when you said you fine-tuned the model after the mass language modeling on specific tasks you were meaning transduction tasks a series of tasks? Right yeah so we have calculated various different endpoints the example I showed here is we we take the number of variants that we find in just the virus before it goes into tissue and then we we check the log fold change after we sequence the tissue and then compare it relative to wild type. Would you try it on multiple tissues at once? I have so we have done that for the multitask model we use all the data at once there we give the model information about which tissue that data has come from with the idea being that if it has that small amount of information it might be able to compare sequences across tissues at the same time. I haven't tried blindly using all of the data from all the tissues sort of sort of unlabeled as one big data set I think it'd be quite interesting to try though. I see thank you. Thank you. Thank you Andrew. Thank you very much. So this is the last and final. --- ## ShapeTX AAV5engineering {#shapetx-aav5engineering} **File:** `ShapeTX_AAV5engineering.mp4` **Type:** Video transcript (Whisper medium) ### Pre-analysis ShapeTX — AAV5 engineering. Expected: proprietary ML-based platform for AAV5 capsid diversification and functional screening. ### Full text / transcript which recognizes the transplant receptor of both humans and NHPs to cross the blood-brain barrier in people. Thank you Andrew. Thank you. All right. My name is Andrew Dunn. I'm from Shape Therapeutics. Thank you for attending this talk. Let's get started. All right. There we go. Maybe. Okay. Forward looking statements and all that. Okay. So at Shape Therapeutics one of the main pillars of the company is on capsid development. So I've been working over the last few years to develop CNS targeting capsids following systemic administration. Our platform uses AAV5. And so we employ direct mutagenesis rather than loop insertion. So we directly mutate capsids of loop 8 and these are around the residues that were previously identified by the McKinna group to interact with sialic acid. So under the theory that you could probably, if you mutate these regions, retarget from sialic acid to another receptor that hopefully is of interest and benefit to your program. So using this method we previously reported on a capsule that we call DB1 or deep brain 1. And then with a relatively moderate dose of 1E13 viral genomes per kilogram, the cinnamalgus macaque, we were able to use and deliver our RNA fix payload and achieve fairly significant on-target editing of our target transcript for our Parkinson's disease program, specifically in deep brain structures of interest to us like the substantia nigra. So now that you have a fairly interesting biologic that you want to take forward before deploying copious amounts of capital on clinical trials, it's probably helpful to ask the question whether or not it might work in humans. And so to that end it's important to know what the primary receptor is that is mediating your capsid. So over the last year I've been onboarding and optimizing an LCMS technique to try to identify the target of these novel capsids. So again, these capsids, they were mutated randomly so we did not start a priori with a particular target in mind. So this is just a simple technique. You grind up your tissue, extract your protein, add your AAV as bait, pull down your AAV, purify the protein, send it off for LCMS and then try to identify the target from the soup that comes out. And LCMS is a very powerful technique. It's very sensitive but it can also be very dirty. And so you achieve a lot of hits. Now this is an interaction assay, it's not a functional assay so we don't know if any of these hits are actually allowing for an enhanced functional transduction when expressed. So we identified hundreds of candidates and then filtered those down rationally based upon what might act as a receptor. Does it have a transmembrane domain? Is it GPI anchored? And over all of these hits where we then delivered a luciferase transgene in vitro, we cloned all of these hits individually and then expressed them across hex and chose and then delivered DB1, transduced to DB1, transduced with another shaped capsid, another CNS tropic shaped capsid, as well as wild type V. And there was one receptor at the very last plate that we ran that provided significant enhancement of functional transduction in vitro that was specific and that was unique to DB1 and that was transferrin receptor. And so with that in mind or with that in hand, you can start to ask the question, does it have really any chance to work in humans if we think that transferrin receptor is the primary receptor that is mediating our method of action crossing the blood-brain barrier? Does it interact with the human orthologs? So from previous studies that we know that DB1 works fairly well in old world monkeys and African green and Cinnamalgus macaques, but we also know that it doesn't work in mice. And so then we can ask the question pretty easily, will it work in humans? Do we think? Does it have a chance of working in humans? And so we ran an ortholog screen where you can just take all of the different orthologs of transferrin receptor and then express them individually in cell lines and then see what the functional transduction is when you over-express those receptors. And pretty obviously, unfortunately for DB1, that it did not appear to functionally interact with the human ortholog. We had excellent maintenance across the old world monkeys, but we did not have a functional transduction boost when over-expressing the human ortholog. So then we wanted to ask the question, where on what location of transferrin receptor might the DB1 be interacting with? And so we made this assumption or this hypothesis that it might be interacting with the apical domain of transferrin receptor. So if you run an alignment of Cinnamalgus transferrin receptor with human transferrin receptor, you can start to see in the apical domain a couple of locations of divergent amino acids. And so when you start to look at a structure of transferrin receptor, these divergent amino acids actually exist around the tops of loops, the loop that might be interacting, with a capsid of interest. So we broke these up into blocks, one, two, three, and four. block either contains one or a few divergent amino acids. Then you can start to create a chimera where you take the Cinnamalgus ortholog, that block from Sinnoh, graph that onto the human ortholog, and start to see if that block might rescue functional transduction. So that's what we did. So we have the full human or the human apical domain, and then grafting the full Sinnoh onto the human ortholog, as well as each of the blocks individually where we take the block from the Cinnamalgus ortholog onto the human ortholog. And there was one block from all of these four blocks that rescued functional transduction of DB1 when over-expressing that human Sinnoh chimera, and that was block A. What's nice about block A is there is a single divergent amino acid when comparing Sinnohs to humans, and that is at position 208. It is a glycine in the Cinnamalgus macaques, and it is an arginine in humans. That is one of the most divergent pairs of amino acids that you can get. So the fun continues for us. So we wanted to try to jump the speciation barrier. To do that, we ran single and double-site saturated mutagenesis of DB1 to see if we could cross over and hopefully bridge from Sinnohs into humans, to see if we can get something that works on both. And panning in vitro, where we express individually either the human transfer intercept or over-expressing human transfer intercept or over-expressing the Cinnamalgus ortholog relative to the parental cell line, the unedited cell line. So we were able to identify a population that appeared to gain this function against. --- ## TuningReceptorInteractions Caltech {#tuningreceptorinteractions-caltech} **File:** `TuningReceptorInteractions_Caltech.mp4` **Type:** Video transcript (Whisper medium) ### Pre-analysis Caltech — Tuning receptor interactions. Expected: computational/experimental approaches to engineering AAV-receptor binding specificity. ### Full text / transcript transduction so basically how to get into the blood brain barrier thank you Alex thank you so much for that warm welcome and good morning everyone my name is Jin Hyeong Alex Chung and I'm from the Grotto Narrow Lab and I'm excited to share with you my talk titled tuning AAV interactions with AVR and LRP6 toggles vectors between vascular and neuron selective tropism. A lot of the recent efforts in the field have focused on identifying receptors that are being mediated by AAVs for effective transportation across the BBB. There are receptors such as Ly6A for PHPB LRP6 for AB9X1 and 9P31 for CAR4 for 9P31 amongst others and even more recently a lot of focus have been garnered towards targeting these receptor for more effective engineering so that we have increased predictability when we're thinking about future therapeutics. As we're going into this landscape of thinking about future therapeutics I believe that there is something that we should also consider which is how binding affinity to these receptors might play an important factor when we're thinking about predicting tropism. Work that's done in the antibody and nanobody engineering field has shown that maybe there could be a Goldilocks zone for binding affinity to a receptor for effective transectosis across the BBB. In work done by Bianlei et al. they show that some engineered antibodies that target transferrin receptors with too high of a binding affinity ends up getting stuck in these endothelial cells and don't actually make it across effectively through the BBB. Conversely those binders that are too weak to bind to transferrin receptor actually never made it through the BBB because perhaps they were being out competed by endogenous ligands against the transferrin receptor and thus they found that this medium zone of moderate binding to the transferrin receptor was the optimal binding affinity to help for effective transectosis across the BBB and we wondered if the same principle could apply to AAVs as it's trying to cross the blood-brain barrier as well. And for this I'm focusing on predominantly two receptors AAVR and LRP6. And we're also looking at our engineered vector AV9X1 and how we can start engineering future variants off this vector in hopes that it can become more neurotropic. AAV9X1 was a previously engineered vector from our lab which targets AAVR and LRP6 for effective transduction of the endothelial cells specifically in the CNS of mice when they're delivered systemically. And we wondered if this targeting of endothelial cells is because of the fact that AAV9X1 is binding too tightly to AAVR and LRP6 and end up getting stuck in these endothelial cells. And therefore by weakening its interaction to AAVR and LRP6 could we see more effective transduction cytosis across the BBB. So first I wanted to go and independently weaken its interaction with AAVR in hopes to create a more neurotropic vector. And this stemmed from recent work that was published in our lab which showcased that CAPB10 which is a further evolved vector from HPEB which modified the VR4 loop to have further detargeting from PKD2, the binding domain of AAVR. This change actually was the reason why there was increased specificity in targeting neurons in the CNS of mice when this vector CAPB10 was delivered systemically. And we wondered if by introducing this CAPB10 loop into AAV9X1 we could also have this dramatic decrease in its binding affinity to AAVR. Unfortunately by just simply substituting the same VR4 loop onto AAV9X1 did not dramatically decrease its binding affinity to AAVR and therefore we need to come up with novel ways to detarget from AAVR. And luckily we were able to recently solve the cryo-EM structure of its interaction AAV9X1's interaction with AAVR and we were able to identify key residues that might be responsible for its binding to AAVR. And by targeting these locuses for effective rational engineering we were able to create this beta mutation and by introducing this beta mutation to AAV9X1 we were able to see a dramatic decrease in its binding affinity to AAVR. And when we delivered this vector AAV9X1 beta that we're calling into mice systemically we were able to see that this vector became a pretty specific vector in targeting neurons. And I think many of you guys might be thinking that AAV9X1 beta is a pretty bad vector when it's targeting neurons altogether because it's not targeting that many cells right? And I thought the same thing when I saw these vectors. And so I wanted to play around with the second receptor LRP6 and to see if by weakening LRP6's interaction with AAV9X1 we could have a more increased potency when it comes to targeting neurons. However unlike AAVR and AAV9X1 we don't have a solved cryo-EM structure to identify the specific loci. And for that we actually ended up using a more traditional method by having a scanning tumor library across a VR8 loop of AAV9X1 to create mutants that could have various binding affinities to LRP6. We subjected this library into magnetic bead conjugated E1 E2 subdomains of LRP6 and pulled them down and we were able to analyze the enrichment values of the capsids that bound most tightly to LRP6 and those that did not. And as expected we were able to see that AAV9X1 is near the top of the ranking profile because we know that AAV9X1 binds with LRP6 and conversely AAV9 tends to be at the bottom of this ranking profile. And so you guys can think of this ranking profile as a proxy for binding affinity towards LRP6. And when we delivered some of the vectors that were more similar in ranking to AAV9X1 such as UX5 into mice brain, into mice systemically we were able to see that its transduction profile is predominantly still endothelial. Conversely, variants such as UX1 which are similar in ranking to AAV9 and very little transduction altogether. And when we looked at vectors with higher ranking than UX5 or AAV9X1 we saw that it is still pretty endothelial in its targeting. However, when we started to look at variants that were in between this AAV9X1 binding affinity and AAV9 we were able to see vectors that are starting to transcytose across the BBB and start targeting some of these astrocytes. And I also thought that this vector as many of you guys are thinking right now is a pretty messy vector. It is permeating through the BBB but it's not at all specific in terms of the tropism that we would want to see. And so we wondered by combining these insights that we gained from understanding the interaction with AAVR and LRP6 we could see a highly potent neurotropic vector. And just to summarize, starting from AAV9X1 when we weaken independently just AAVR by introducing this beta mutation we saw a sparsely neuronal vector. And then when we created this vector with weakened interaction with LRP6 but not per se with AAVR we were able to get this more permeable vector that goes across the BBB. And so we wondered by introducing the same B10 mutation that we introduced to AAV9X1 could we see a strongly neuronal vector. And when we did introduce this beta mutation to UX13 we were able to see this increased potency when targeting neurons and it is predominantly transusing only neurons when quantified. Overall just to summarize my talk I was able to start playing around independently these two dials AAVR binding affinity and LRP6 binding affinity to get to looking at how these binding affinis are determining its tropism altogether. And I hope we can take this into consideration as we're going into future therapeutics research and trying to have more predictable outcomes with our engineering efforts. Thank you so much and I open the floor up for questions. Questions from the audience? Hi, did you look at vector genomes in peripheral organs especially at early time points? We analyzed these animals after three to four weeks post delivery so we didn't really look at really early time points after delivery. But an important point that you bring up is the peripheral organs. Something unique that we saw with introducing this beta mutation into these AAV variants was that we're starting to see increased transduction of the peripheral nervous system as well. You guys can see pretty much in this brain cell as one representation of how it would perform like the other peripheral vectors that were produced from our lab. And when we looked at organs such as the distal colon we saw increased transduction profile there as well. Okay that's true both on the transduction level as well as on the vector genome per deployed genome. We didn't do an analysis about looking at quantifying the number of genomes that were found. We currently are at the stage of just looking at the protein level of expression. Okay thanks. Thank you very interesting talk. I would like to know what your thoughts are on the actual role of AVR in like why do you see this difference in cell targeting given that AVR should be expressed in every cell and there's even some indication that it might be more assorting signal than the actual entry receptor. I think that's a really important question to consider. Frankly I don't quite understand what's happening here either but I know there are a lot of really good talks that are happening around the role of AVR in transduction. I think AVR in general is very critical for transduction altogether. As many of you guys know knockouts of AVR completely diminishes systemic delivered AVV transduction and I think these vectors are still definitely binding to AVR but having this goldie log zone of binding maybe perhaps allows AVR to play more of a trafficking rule rather than more of a sticky web for targeting off target regions of other organs that we don't care about. Right so you think you still transduce what you enter astrocytes but they don't express in them with your new vectors? We haven't looked at the genomic level yet and I think we're hoping to go with methods that were developed such as specter or zombie AV zombie to look into the genomic level as well to see where these vectors are actually ending up but at the current moment we don't really know. Okay thank you. Could I ask you a quick question? Yes of course. So if you're talking about modulating and thinking about the Goldilocks effect, analytically how would you come to this Goldilocks composition? So I think that's also another thing. Are you looking just into one species and one strain of mice or have you looked into Vabc as well because we have a lot of precedents regarding C57 BL6 you know having a very different tropism versus Vabc being very different. Definitely so regarding your first point about how are we defining the Goldilocks zone we're actually hoping to take more of a molecular approach by determining the KD values of their binding interactions using assay such as BLI and by doing so we can start to have this defined region of the prefer optimal rate of binding for transcytosis. As for the second question I think it's really important because I work in non-traditional organisms as well. Full disclosure these animals are C57s that we tested them in but interestingly LRP6 if you guys don't know is a highly conserved receptor with over 98% sequence similarity between mice and humans and so hopefully the same patterns do hold true in other strains of mice as well but we definitely do have to test them in vivo to see if this is true. Yeah and perhaps you're going up to NHBs as well the next. That'll be really exciting if we can. Thank you very much. --- ## ASGCT recap Georg {#asgct-recap-georg} **File:** `ASGCT recap Georg.pptx` **Type:** PowerPoint slide text ### Pre-analysis Georg Feichtinger's personal ASGCT 2026 recap. Key findings, highlighted talks, relevance for Parvotec ML project. ### Full text / transcript ### Slide 1 1423 ASGCT 2026 Parvotec relevant talks and posters ### Slide 2 I. Chronic pain gene therapy ASGCT 2026 - Abstracts, talk slides and posters ### Slide 3 3 Relevant oral presentations Chronic pain gene therapy ### Slide 4 4 SereNeuro Oral abstracts ### Slide 5 5 SereNeuro (key slides attached) Oral abstracts ### Slide 6 6 Exgenesis Bio (key slides attached) Oral abstracts ### Slide 7 7 Sanofi (no poster pdfs available) Poster abstracts ### Slide 8 8 Encoded (poster attached) Poster abstracts ### Slide 9 9 Sangamo (poster attached) Poster abstracts ### Slide 10 10 Sangamo (poster attached) Poster abstracts ### Slide 11 11 Borea Therapeutics (poster attached) Poster abstracts ### Slide 12 Abstract 487 AAV-Mediated RNA Interference Targeting Nav1.7 Provides Durable Relief of Hyperalgesia in Preclinical Models of Chronic Pain Exgenesis Bio Inc, China (CJ Song) ### Slide 13 13 Abstract 487 – Exgenesis Bio Inc ### Slide 14 14 Abstract 487 – Exgenesis Bio Inc ### Slide 15 15 Abstract 487 – Exgenesis Bio Inc ### Slide 16 16 Abstract 487 – Exgenesis Bio Inc ### Slide 17 17 Abstract 487 – Exgenesis Bio Inc ### Slide 18 18 Abstract 487 – Exgenesis Bio Inc ### Slide 19 19 Abstract 487 – Exgenesis Bio Inc ### Slide 20 20 Abstract 487 – Exgenesis Bio Inc ### Slide 21 21 Abstract 487 – Exgenesis Bio Inc ### Slide 22 22 Abstract 487 – Exgenesis Bio Inc ### Slide 23 23 Abstract 487 – Exgenesis Bio Inc ### Slide 24 Abstract 537 Human Sensory Neuron Profiling Enables KCNQ2 Gene Therapy for Neuropathic Pain SereNeuro, USA (Tea Soon Park) ### Slide 25 25 ### Slide 26 26 ### Slide 27 27 ### Slide 28 28 ### Slide 29 29 ### Slide 30 30 ### Slide 31 31 ### Slide 32 Abstract 1312 NociPro: A Novel Modality-Agnostic Promoter Platform for Precise Cellular Targeting of Nociceptor Sensory Neurons in Gene Therapy for Chronic Pain Encoded Therapeutics, USA (poster) ### Slide 33 33 ### Slide 34 Abstract 1309 ST-503 Nonclinical Safety Studies Evaluating Zinc Finger Repressors Regulating the Expression of the Nav1.7 Gene for Treatment of Small Fiber Neuropathy Sangamo Therapeutics, USA ### Slide 35 35 ### Slide 36 Abstract 1033 A Modular Platform for Chemical Capsid Functionalization of AAV Vectors, Enabling Receptor-Guided Targeting and Programmable Payload Release for Pain Gene Therapy Borea Therapeutics, Italy ### Slide 37 37 ### Slide 38 II. Capsid engineering ASGCT 2026 - Abstracts, talk slides and posters ### Slide 39 39 Elements of our engine present at ASGCT Parvotec Capsid Discovery Engine Everyone R-Scan for Motoneurons (poster) ? ### Slide 40 40 Bonus: Imo most relevant talk: miRNA design for AAVs 50% ### Slide 41 41 Relevant oral presentations Capsid engineering ### Slide 42 42 Relevant oral presentations Capsid engineering ### Slide 44 44 ### Slide 45 45 ### Slide 47 47 ### Slide 48 48 ### Slide 50 50 ### Slide 51 51 ### Slide 53 53 ### Slide 54 54 ### Slide 55 55 ### Slide 57 57 ### Slide 58 58 ### Slide 59 59 ### Slide 60 60 ### Slide 61 61 ### Slide 62 62 ### Slide 63 63 ### Slide 64 64 ### Slide 65 65 ### Slide 66 66 ### Slide 67 67 ### Slide 68 68 ---