Generative design of de novo AAV vectors with unique serological profiles

More than 5000 monogenic diseases are potentially curable by replacing a single defective gene with one that functions normally. Currently, the only clinically approved vector able to effectively deliver a therapeutic gene to a target tissue is Adeno-Associated Viruses (AAVs). However, despite decades of work, only a handful of gene replacement therapies have been approved, and many patients remain ineligible for even those. A major barrier to broad adoption of AAV therapies is pre-existing immunity to AAVs, mediated by neutralizing antibodies (nAbs). It is potentially possible to change exposed capsid surfaces in AAV capsids to avoid nAb recognition. However, traditional techniques such as rational design and directed evolution only allow modest changes. AAV function becomes compromised long before full nAb avoidance is achieved. Recent advances in protein design by generative modelling provide a new way to overcome this problem by learning AAV design constraints from natural sequences. We built a variational autoencoder that could learn these sequence constraints and generate functional AAV capsid proteins de novo. More than 255 designs were synthesized and tested in vitro in two design sets, with one set dedicated to exploring uncertain regions of the sequence space. Between the two campaigns, seven designs transduced human cells with >50% efficiency, of which four avoided nAb neutralization (NT50 < 1:50) in all human sera that were reactive to AAV. In one of these designs, sequence identity to the nearest natural AAV sequence was 18% divergent. Given the radical sequence diversity and its unique serological profile, this new capsid design would represent the first ever AAV serotype designed de novo.