When scientists predict how a cell will change after a genetic perturbation, they usually measure accuracy using a correlation score over all genes. But most genes do not change at all, so a model can get a high score just by correctly predicting the unchanging genes, making it look far more accurate than it really is.
The researchers compared three methods: a popular model called GEARS, a simple baseline that always predicts the average change, and a corrected metric that looks only at changed genes. On a standard test, GEARS scored 0.977 on all genes but only 0.076 on changed genes alone, while the simple baseline scored 0.267, over three times better.
This means current models are not learning meaningful cause-and-effect relationships between perturbations and gene changes; they are just good at guessing the default state. The finding matters because without proper evaluation, progress on building predictive cell models will be misleading, and simple baselines could outperform complex AI systems in real-world use.