Back

Learning from and improving upon high-throughput screens for protein fitness with Generative AI - Application to BBB-crossing AAV design

Kashyap, A.; Soniya, K.; Anubhooti, ; Kandavel, S.; Avadhani, K.; Kanishka, S.; Sharma, M.; Chodavarapu, P.

2025-02-05 bioinformatics
10.1101/2024.12.06.627153 bioRxiv
Show abstract

Deep Science is enabling high-throughput experimentation (HTE) to design novel biological entities with desired properties. E.g., blood-brain-barrier (BBB) crossing adeno-associated virus (AAV) vectors, needed for systemic delivery of gene therapies to brain cells, have been identified through innovative directed evolution assays such as M-CRE-ATE and TRACER. But, even these high-throughput experiments are only able to explore a miniscule portion of the large design space of biological entities. In this paper, we introduce autograd based maximization of protein fitness (AutoMaxProFit) to learn from and improve upon protein designs generated with high-throughput screens. Using a transformer based generative AI network and protein language models, we improve upon the design of a variant previously discovered through HTE, to yield 2x better enrichment in brain endothelial cells, as estimated by molecular dynamics (MD) simulations. This shows that Deep Tech models can learn from the observations generated by Deep Science experiments and go on to find more optimal design candidates for application in Biopharma.

Matching journals

The top 10 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.