The State of AI Antibody Design

Summary

I often say that I think that antibody design is one of the better use cases for AI drug discovery at present, because you start with a lot of inherent advantages: antibodies as a class are evolutionarily optimized for long plasma half-lives, and the long list of approved antibody drugs shows you that success is out there to be had. But this is a relative scale! I still believe that antibodies are one of the more realizable AI design stories, but that doesn’t mean you can just walk over and start scooping up the profits. There are still a lot of things that can go wrong, starting with picking the right target (the recent anti-IL-6 failure for cardiovascular indications is an example of that one). Then you have to worry about specificity for your target, because “It’s an antibody” is not a plan to get around that step. And there’s immunogenicity: having an antibody be recognized as a foreign protein and then setting off an immune response against it just seems unfair, considering how antibodies in general are supposed to be sort of a privileged class in that regard. But it most certainly does happen. And given the variability of the immune system across a large patient population, it surely will happen with some people. But how many? And how bad will the consequences be? That’s up to you to find out, and computational methods are going to be of limited help there. So it’s still hard and still expensive, and this recent paper is a good illustration of that. It’s a report on the AIntibody challenge , where 511 AI-designed antibodies (from 29 different organizations) were put through a series of rigorous tests. All of them target the SARS-CoV-2 receptor binding domain (RBD), about whose structure we know far, far more than anyone ever would have wanted to before the pandemic. There were three different sorts of challenge here. The first was affinity maturation of a known antibody sequence recognizing the RBD, using a set of improved sequences from yeast display assays. Only the parent antibody’s affinity was provided - it itself had been experimentally improved, but the matured sequence from that effort was not provided to the participants. The idea was to improve affinity without compromising other known developability factors. The results here was that the overall distribution of the AI submissions were worse than the experimentally-derived antibodies. Most (…