A preprint finds protein language model embeddings retain biochemical signals linked to fitness prediction, while randomized projections do not.