
This is why @amplesa argues we need more/better instruments, not more of the same data we can already collect, in order to get more from AIxBio: amplesa.substack.com/p/ai-fo…
This is an important paper to read for tamping down some of the "AI will solve biology" enthusiasm. The researchers pretrained 400 models on a corpus of 22.2 million cells and performed 6,400 evaluations. They found that current single-cell foundation-model approaches frequently plateaued at substantially smaller training sizes and did not exhibit clear data scaling laws analogous to language models. Importantly, this suggests that simply collecting more of the same kind of single-cell data is not necessarily the solution. Better coverage of biological states, conditions, perturbations and relevant endpoints may matter more. One of the key things to realize about biology is how emergent and heterogeneous it is, which means that simply gathering more data at one level is insufficient for developing foundation models, designing better drugs etc. We need *relevant* data at scale for many different levels and interactions, including single gene, multi-gene, gene-protein, protein-protein, cell-cell etc. A lot of this data - for instance spatial cell-cell data -is not easy to acquire at large scale, and we have just started to do it. And even when we have that, the reasoning models will have to figure out which one of these data types is most suitable for a particular problem. I see a lot of progress using AI for data analysis and gathering in biology in the future, but it's critical to recognize how the data problem for AI in biology is fundamentally different from the data problem in areas like image and text analysis (which is hard to begin with). nature.com/articles/s41592-0…




















