Zero-shot predictions are great, except for when it matters
I used to think that Evo2 likelihoods and the like are approximate conservation scores. And such scores are good for predicting pathogenicity or, in case of proteins, viability. You can't use them to predict eQTLs or anything else though
Yes, agreed, they're approximately reproducing amino acid frequencies in large scale multiple sequence alignments.
I’ve always had this intuition, but nice to see it studied and more carefully explained.
I used to think that Evo2 likelihoods and the like are approximate conservation scores. And such scores are good for predicting pathogenicity or, in case of proteins, viability. You can't use them to predict eQTLs or anything else though
Yes, agreed, they're approximately reproducing amino acid frequencies in large scale multiple sequence alignments.
I’ve always had this intuition, but nice to see it studied and more carefully explained.