F15 VCell

We worked on the Arc Institute Virtual Cell Challenge: given a gene knockdown, predict how the cell’s expression profile changes. Scoring is official Arc cell-eval on three metrics - DES (did you recover the right differential genes?), PDS (can you tell knockouts apart?), and MAE (overall error). We built our own Early-S baseline, then ran a fast loop of invent → implement → run on Azure GPUs → rescore with cell-eval. Across roughly twenty experiment waves we took official test DES from 0.1 baseline to about 0.5 (where 0.3 is considered an impressive result in the field that is very challenging to beat), with MAE as low as 0.017. That’s a big step on DES relative to common published ranges people cite for this problem. PDS results are less impressive, and are an area that we will focus on improving next, but remain above the baseline. The point of the project isn’t a new foundation model, but showing that with a hard public evaluator and a lot of measured iteration, you can push the biologically relevant DES axis a long way in days, and be clear about where the method still fails (perturbation identity). But the fact that we were able to implement and conduct hundreds of experiments is genuinely remarkable and we would not have been able to do that in a week were we not using Claude.