Genetic Perturbation Prediction Models Fail Out-of-Distribution: A Rigorous Benchmark and Systematic Evaluation

Perturbation prediction models - which forecast transcriptional responses to gene inactivation - are increasingly proposed as in silico surrogates for genome-scale drug discovery screens. Yet, rigorous evaluation of whether these models genuinely generalize beyond their training distribution has been lacking. Existing benchmarks rely predominantly on random train-test splits, obscuring systematic failure modes that only emerge under realistic out-of-distribution (OOD) conditions. Here we introduce EL-PERTURB (Extrapolative Learning for PERTURBation Prediction), a benchmarking framework that stress-tests 20 perturbation prediction models spanning six architecture classes on both essential gene and genome-wide Perturb-seq screens, using a novel stratified OOD splitting strategy that partitions train and test sets into functionally distinct regions of transcriptional space. We find that model performance degrades substantially under OOD conditions, and that architectural complexity - including single-cell foundation model embeddings - confers no reliable advantage over well-tuned simpler baselines. EL-PERTURB establishes a rigorous benchmark for perturbation effect prediction across a gradient of generalization difficulty, providing a guide for building models that can genuinely serve as surrogates for wet-lab screening.