How to Engineer Therapeutic AAVs
exploring public viral packaging + biodistribution data from Dyno Tx // how to use scientific plots to answer important biological questions about library scale data
In May 2023, thousands of scientists convene at the American Society of Gene and Cell Therapy (ASGCT) Annual Meeting to hear the latest progress in therapeutics.
Among the presentations is Eric Kelsic from Dyno Therapeutics. On the main stage, he gives an overview of Dyno’s platform. To the audience’s surprise, he unveils data for Dyno bCap1, the first version of Dyno’s AAV product for brain delivery, a milestone for both his company and the gene therapy space.
With striking, fluorescent images, he compares delivery using wildtype AAV9 serotype and Dyno bCap1. The contrast is stark. With machine learning guided engineering loops, Dyno built a capsid with 100x greater brain transduction, 10x greater liver detargeting and comparable production to existing AAV9 capsids.
Just this year, the NVIDIA CEO Jensen Huang, alluded to the power of such compounding progress in biology:
For the very first time in human history, biology has the opportunity to be engineering, not science. When something becomes engineering, and not science, it becomes less sporadic, and exponentially improving. It can compound on the benefits of previous years.
DynoTx’s success with bCap1 is an inspiring example of what this progress looks like in practice. The step function improvement in the therapeutic properties of bCap1 came from the iterative engineering of large libraries of AAV molecules.
Inspired by Dyno’s groundbreaking work, the team at Latch replicated one of Dyno's key papers by Ogden et al, which took a look at all single-codon mutants of the AAV2 cap gene across 735 positions. We took the public data, reproduced some of the key scientific results and had fun thinking about the same biological questions the Dyno team was confronted with years ago:
How do mutations affect viral production?
How do mutations affect biodistribution?
How can we use machine learning to generate new, optimal designs?
Along the way we dig into the guts of interesting visualizations and show how to use Latch Plots — a Python-based scientific plotting framework - to answer these concrete questions.
A primer on AAV
Adeno-associated viruses (AAVs) are small, naturally occurring viruses that are becoming powerful tools in gene therapy. Unlike other viruses, AAVs are non-pathogenic, meaning they don’t cause disease in humans, making them safer for therapeutic use. Additionally, AAVs can infect a wide range of cell types and don’t integrate into the host’s genome, which is important for clinical application.
These unique properties make AAVs ideal candidates for delivering therapeutic genes to target tissues, effectively acting as delivery vehicles that can reach and “infect” cells with carefully engineered DNA to address genetic diseases at the root level.
However, natural AAV capsids often have limited transduction efficiency and don’t deliver therapeutic genes with the efficiency needed for clinical use.
In a paper by Ogden et al., the authors tackled the transduction efficiency challenge by mapping out the AAV2 capsid's “fitness landscape.” This effort included:
Understanding viral production. The team created a library of mutant capsids, transfected the library into HEK293T cells and measured the quantity of purified virus.
Understanding biodistribution. They then injected mutant libraries into mice and analyzed tissue-specific distributions, they identified key mutations that enhanced targeting to select tissues, such as the liver or heart.
Generating optimal designs using machine learning: They used the previous datasets to create a new capsid library that outperformed random mutagenesis, yielding a diverse and viable set of designs.
Disclaimer: This paper contains many findings, but due to data availability from Gene Expression Omnibus (GSE), we were only able to reproduce key Figures 1 (b&d) and Figure 3 (d&c). We highly recommend reading the full paper yourself to explore all the results in detail!
Assessing Viral Production Efficiency
To evaluate how mutations impact capsid assembly and genome packaging, Ogden et al. transfected a plasmid library of variants into HEK 293T cells to produce recombinant AAV. The fitness of each variant was then calculated as normalized enrichment relative to the wild-type (WT) AAV2.
The figure below shows this fitness value for each variant. Positive selection indicates beneficial mutations enriched in the virus, likely enhancing capsid assembly or genome packaging. Negative selection suggests mutations that may impair virus production.
Right off the bat, we can observe all the red points on the upper quadrant that are more likely to yield higher transduction efficiency, instantly narrowing down the list of candidates to test.
Fitness heatmap for all single–amino-acid insertions, deletions, stop codons, and substitutions
To uncover the sequence motifs responsible for desirable packaging, we replicated a heatmap showing how fitness changes across the sequence of the gene. Here we combine counts for all synonymous codons of the same amino acid.
Most mutations decreased fitness, as indicated in blue. However, amidst this sea of blue, a smaller set of variants actually enhanced production. This map allows us to easily visualize which mutations are beneficial.
The author colored the 3D structure of the capsid with these fitness scores to visualize the geometry of these perturbations. Mutations in deeply embedded areas of the capsid showed strong negative selections. Buried residues are essential for maintaining the capsid’s structural integrity.
Mutations at more exposed surface areas of the capsid were better tolerated. Since these regions are more flexible, they can handle changes without destabilizing the entire structure.
Assessing in vivo biodistribution
To assess how AAV capsid mutations impact delivery in a living system, the authors injected a library of variants into mice and analyzed tissue-specific distribution across organs. They collected blood 1 hour post-injection and then waited 18 days before taking samples from spleen, liver, kidney, heart, and lung. They then sequenced the barcodes from purified virus genomes.
We performed dimensionality reduction with PCA, and projected the individual mutants onto PC1 and PC2, colored by tissue enrichment. Higher enrichment indicated that certain AAV capsids were more effective for specific tissue types.
The authors also three k-means clusters of mutants that showed increased distribution to specific tissues:
Cluster 1: Included well-known R585 and R588 mutants, which were low in the liver but high in the blood, heart, and kidney, as expected.
Cluster 2: Similar to Cluster 1 but specifically low in the heart.
Cluster 3: Low in blood and spleen but enriched across other tissues.
These clusters also differed in their structural locations on the capsid: mutations in Clusters 1 and 2 were concentrated near the 3-fold axis on the capsid surface, while Cluster 3 mutations were spread throughout buried positions in the capsid structure.
Using machine learning to design novel AAV capsids
Introducing multiple rational mutations without compromising capsid function was a great challenge several years ago. Even single amino acid changes were difficult to predict and could be harmful.
The team hypothesized that machine learning could predict the behavior of multiple perturbations using data from single mutations.
Using a machine learning guided design strategy (which was not discussed directly in the paper), the authors created a pooled library of 1271 ML-generated variants and 10,047 randomly generated mutants and measured liver biodistribution in mice.
The results were promising:
The graph above shows the fraction of mutants that target the liver more effectively than the wild type (WT). For randomly generated mutants (gray), this fraction drops to near zero by the time they accumulate three mutations. In contrast, the designed mutants (orange) maintain a positive fraction, even with up to nine mutations.
The bottom graph shows the spread of biodistribution values for random and designed mutants separated by number of mutations. Even at distances farther than four mutations, the designed approach outperforms random mutagenesis (pink box).
To Conclude
With the complexity of biological systems, turning drug discovery into an engineering discipline seems like a far-fetched dream. However Dyno’s computer-guided loops and use of engineering principles can produce mutant AAVs with desirable therapeutic properties. We hope you were able to appreciate to interesting scientific questions their team navigated years ago to produce this landmark result in AAV engineering.
If you want to use Plots for your team, request a free trial here.