CRISPRi Perturbation Prediction Driven by Response Magnitude

A new study investigates the prediction of transcriptomic effects from CRISPRi perturbations, particularly for target genes not seen during training. Previous work noted that simple baselines often match or exceed complex deep perturbation predictors. This research pinpoints a specific low-dimensional signal—response magnitude—as the key driver behind this phenomenon. The target metric, log Anderson-Darling distance, was found to be highly predictable from just four deterministic scalar functions derived from the 2,000-dimensional input. Deep learning models, even with full input access, tended to collapse towards the marginal training mean. A linear regression using only these four magnitude scalars outperformed the strongest classical models, and a Random Forest combining input with these scalars significantly surpassed a deep encoder. Crucially, magnitude-only predictors showed positive transferability to external CRISPRi screens and different cell types, whereas expression-only predictors performed poorly or inconsistently. The findings suggest that focusing on the magnitude of response, rather than just gene expression, is a more robust approach for predicting CRISPRi perturbation effects and improving transferability.