Mapping AAV capsid sequences to functions through function-guided in silico evolution

Abstract

Artificial intelligence (AI) offers significant potential to accelerate the functional engineering of adeno-associated virus (AAV) capsids in a time- and cost-efficient manner. However, existing approaches lack the ability to systematically map capsid sequences to multifunctional properties. To address this challenge, we developed ALICE, a knowledge-driven platform for in silico AAV capsid engineering that directly maps sequence to multiple functions. Our method incorporating a heuristic algorithm with contrastive learning principles, termed function-guided evolution (FE), to iteratively optimize high-performing capsid sequences generated by a naive language model toward desired functions. We elucidate FEs evolutionary mechanism, demonstrating its ability to navigate a function-guided landscape and generate capsids with tailored properties. After validating ALICEs ability to design murine-specific AAV capsids targeting Ly6a/Ly6c1 with limited training data, we developed ALICE-X by integrating a curriculum learning strategy. This upgrade facilitated exploration of sequence space beyond existing wet-lab datasets, enabling successful targeting of the human transferrin receptor 1 (hTfR1). The resulting engineered AAV variant, AAV.ALICE-H3, demonstrated improved viability, specific receptor targeting, and [~] 251-fold enhanced CNS tropism in human TFRC knock-in (hTFRC KI) mice compared to wildtype controls. This interpretable, knowledge-driven framework advances in silico AAV capsid design, offering broad applicability for translational gene therapy.