Gradient-aware modeling advances AI-driven prediction of genetic perturbation effects

Predicting the transcriptional effects of genetic perturbations across diverse contexts is a central challenge in functional genomics. While single-cell perturbational assays such as Perturb-seq have generated valuable datasets, exhaustively profiling all perturbations is infeasible, underscoring the need for predictive models. We present GARM (Gradient Aligned Regression with Multi-decoder), a machine learning (ML) framework that leverages gradient-aware supervision to capture both absolute and relative perturbational effects. Across multiple large-scale datasets, GARM consistently outperforms leading approaches--including GEARS, scGPT, and GenePert--in predicting responses to unseen perturbations within and across contexts. Complementing this, we show that widely used evaluation metrics substantially overestimate performance, allowing trivial models to appear predictive. To address this, we introduce perturbation-ranking criteria (PrtR) that better reflect model utility for experimental design. Finally, we provide insight into gene-specific predictability, revealing pathways and gene classes systematically easier or harder to predict, with implications for model development and biological interpretation. Together, these advances establish a unified methodological and conceptual framework that improves perturbation modeling, sets rigorous evaluation standards, and provides biological insight into gene-specific predictability in functional genomics.