ALICE

Artificial intelligence (AI) has been suggested to facilitate time- and cost-effective functional engineering of adeno-associated virus (AAV) capsid sequences. Nevertheless, an AI-empowered approach to identify AAV capsid sequence-to-multifunction relationships remains elusive. To overcome this challenge, we propose a machine-intelligent design method to map an AAV capsid sequence to multiple functions, thereby enabling direct in silico engineering of AAV capsids. To fuse multiple functions into a single capsid sequence, a heuristic algorithm coupled with contrastive learning and reinforcement learning, named function-guided evolution (FE), was introduced to steer further evolution of the high-performing capsid sequences generated by a naive language model toward functions. We then illustrated the evolutionary mechanism of the FE approach for function-guided generation of capsid sequences. Further optimization steers the evolution toward desired functions within a function-guided landscape. Despite the constraint of datasets of only 129 entries, we successfully constructed a model to map AAV capsid sequences to multiple functions of improved viability coupled with central nervous system (CNS) tropism. In vivo experiments confirmed that two of the top eight engineered variants exhibited enhanced viability and remarkable CNS tropism. This interpretable machine-intelligent design method represents a pioneering effort enabling direct in silico engineering of AAV capsids for effective gene delivery