Simple Magnitude Signals Beat Deep Encoders in CRISPRi Prediction Study

A new machine learning study challenges the prevailing deep-learning approach to predicting CRISPRi perturbation effects. Researchers Mehrdad Shoeibi and Niloofar Yousefi, in their arXiv paper "Response Magnitude as a Dominant Signal for Held-Out CRISPRi Perturbation Effect Prediction," analyzed the Virtual Cell Challenge (VCC) benchmark under a strict held-out target-gene split. Their key finding: the target outcome—log Anderson-Darling distance from non-targeting controls—is strongly predictable from four deterministic scalar functions of the 2,000-dimensional input. While a deep MLP encoder with full input collapses toward the marginal training mean, a simple linear regression on just these four magnitude scalars exceeds the strongest x-only classical model. Adding Random Forest on input plus the four scalars pushes performance even further, substantially outpacing the deep proof-of-concept encoder. Two control experiments attribute the gain specifically to per-row alignment, not added dimensionality.

The paper also examines transfer to external CRISPRi screens. Using zero-shot transfer with target-gene endpoints rebuilt from single-cell data, magnitude-only predictors transfer positively, whereas expression-only predictors are negative or unresolved. Exposing magnitude to the deep encoder improves transfer over its expression-only counterpart, but the encoder still cannot beat a four-scalar linear regression on the same features. The authors also note that the Anderson-Darling column measures transcriptome-wide response breadth rather than target-gene effect strength, clarifying what the benchmark actually scores. For the ML community, this is a sobering reminder that simple, interpretable features often dominate deep models in biological perturbation prediction—and that benchmark design must account for response magnitude signals before claiming model superiority.