Thread by @ideasbyjin

A "negative" result, but phenomenal thought piece from @naturalantibody / @antibodymap. TLDR predicting antibody-antigen interactions is pretty darn hard (1/5) #antibodies #machinelearning #alphafold

naturalantibody.com/use-case/deepm…

Predicting Ab-Ag interactions is a sub-problem of the protein-protein interaction problem. There are many facets to consider here, including but not limited to, identifying the correct antigen (let alone the correct epitope), the correct paratope, orientation, etc (2/5)

@antibodymap's team show first that true Ab-Ag pairs (i.e. those where we know the Ab binds antigen) and false Ab-Ag pairs (i.e. Ag was randomly given to an Ab), the pIDDT scores are incomparable, suggesting score-based discrimination is HARD. (3/5)

Next, the pIDDT score is used as a ranker; judging by the tweet above, one would expect this not to work well. Indeed, while predictions with high pIDDT scores are generally from correct Ab-Ag pairs, this is within the margin of error from random poses. (4/5)

I think -these- are the kind of manuscripts(? blog?) we sometimes need to celebrate! We shouldn't shy away from the fact it's a hard problem. While AF2 is a generationally important tool, it's not a one-size-fits-all black box. Back to the drawing board. (5/5)

Meant to say pIDDT scores are not discriminative*, oops