Leveraging Biological Networks for Robust Gene Perturbation Prediction

Predicting the effects of gene perturbations on cellular gene expression is a fundamental challenge in functional genomics, crucial for understanding gene function and therapeutic development. While deep learning models, including large-scale "foundation models," hold promise for capturing complex non-linear relationships, they often struggle with extrapolation tasks, such as forecasting unseen double gene perturbations or the impact of unobserved genes, frequently underperforming simpler linear models. To address this, we propose the AdaptivePerturbPredictor (APP), a lightweight deep learning framework that integrates biological network information via Graph Neural Networks (GNNs), incorporates a novel perturbation-specific attention mechanism, and employs a parsimonious linear decoder. Our extensive experiments on public single-cell RNA-sequencing datasets demonstrate APP's superior performance in challenging extrapolation scenarios. Specifically, APP demonstrated superior performance, achieving lower \( L_2 \) distances for both double gene perturbation prediction and unseen single gene perturbations, significantly outperforming additive baselines, linear models, and complex existing deep learning models. Ablation studies confirm the critical contributions of GNNs and the attention mechanism, while analyses highlight APP's robustness to network quality and significantly enhanced computational efficiency compared to larger models. APP represents a significant advancement, offering a robust, interpretable, and scalable solution for gene perturbation prediction, particularly in out-of-distribution contexts.