Hadrien Mary

A socratic dialogue over the utility of DNA language models (Part 2 of 2) link is here: owlposting.com/p/a-socratic-… part 2 is out! this time focused on genome generation. i think i tie myself a bit more to Evo 2 specifics to explain this, but i try to explain the context of the field too. maybe too much context, given that this post is 4.9k words part 1 was focused on variant pathogenicity prediction. which was interesting, but left me feeling a bit unsure about the point of these types of models. especially given that conservation scores weren't too far away in nearly matching accuracy i think the role of DNA language models in genome generation is a fair bit more interesting, since there's no reliable, wet-lab method to really do it at scale...other than to train a black box model on to do exactly that. its definitely early days and the results given here aren't especially actionable to anybody, but i think ive updated a fair bit more positively to the prospect of models like Evo 2. it feels very much like the blue sky research that places like the @arcinstitute should be pursuing