Hi, I'm Raphaël. I work on
AI for biology.

The dream is models good enough that a lot of experimental work could start in silico: faster wet-lab campaigns, AI in the loop when deciding which experiments to run, and maybe a bit more scientific discovery along the way. The virtual cell is one avenue toward that.

What bugs me most is how badly our models do out of distribution. Put them in a context or condition they weren't trained on and they often fail, in ways I'd argue are worse than how we humans cope. Faced with something new, we lean on what we already know, reason about how it's similar to or different from the situation in front of us, and we're often roughly right. We're far from perfect, sometimes we grab the wrong prior and do worse than chance, but the gap between that and how a model behaves in an unfamiliar biological context feels like something worth trying to close.

I also suspect AI-assisted experimental design is where progress is likeliest in the next few years, before the harder generalization questions are solved, given the amount of effort being put into it. So I'm trying to keep one foot in each.