Great post! I missed most of the discussion of this competition, so I wanted to ask: are there any interesting and useful for science solutions on top or hacking won?
Hey Alex! The Virtual Cell Challenge just released their results! It seems like the winning methods all actually relied on some biology to create their models, although the top methods didn't bother trying to get the magnitude of perturbations correct (since it didn't matter for PDS or DES, and it's near impossible to improve on the MAE). The winners will likely release their fully methodology soon! The Altos lab won the "generalist" prize, which takes into account more metrics, and they used a U-net architecture trained on >7 million cells and fine tuned on H1 stem cells. But, since they're a biotech company, I'm not sure they'll ever release their full model.
Great post! I missed most of the discussion of this competition, so I wanted to ask: are there any interesting and useful for science solutions on top or hacking won?
Hey Alex! The Virtual Cell Challenge just released their results! It seems like the winning methods all actually relied on some biology to create their models, although the top methods didn't bother trying to get the magnitude of perturbations correct (since it didn't matter for PDS or DES, and it's near impossible to improve on the MAE). The winners will likely release their fully methodology soon! The Altos lab won the "generalist" prize, which takes into account more metrics, and they used a U-net architecture trained on >7 million cells and fine tuned on H1 stem cells. But, since they're a biotech company, I'm not sure they'll ever release their full model.