Stable-Shift: Biologically Structured Prediction of Transcriptional Responses to Unseen Gene Perturbations

Stable-Shift introduces a structured machine learning method for predicting how genes respond to perturbations without requiring experimental data from those genes. The approach outperforms existing methods like GEARS on benchmark datasets, achieving 0.592 cosine similarity, and demonstrates the value of integrating biological context through graph neural networks for genomic prediction tasks.

Stable-Shift addresses a fundamental challenge in functional genomics: predicting transcriptional responses for genes that have never been experimentally perturbed. This problem mirrors common machine learning extrapolation challenges, but with biological complexity. The method's innovation lies in decomposing the prediction into two stages—learning a low-rank response basis from known perturbations, then contextualizing new genes within that basis using biological networks and ontologies. This structured approach outperforms GEARS, a competitive baseline, across multiple evaluation metrics and datasets, suggesting that embedding biological domain knowledge into latent representations improves generalization.

The research addresses a critical bottleneck in genomics research. CRISPR screening and perturbational studies generate massive datasets, but experimentally testing every gene-perturbation pair remains prohibitively expensive. Predictive methods could accelerate functional genomics by prioritizing which genes warrant experimental validation, reducing wet-lab costs and accelerating biological discovery timelines.

The performance gains are modest but consistent—improving from 0.569 to 0.592 cosine similarity—and robustness across different benchmarks (graph-aware, residualized, gene-space, Norman-dataset) strengthens confidence. However, acknowledged limitations matter: lower accuracy in gene-space predictions and sensitivity to sparse graph neighborhoods restrict applicability to genes with limited interaction data or in understudied pathways. These constraints mean the method performs best for well-characterized genes in well-mapped biological networks.

For computational biology and drug discovery, this work validates that structured prediction frameworks incorporating biological priors outperform purely statistical approaches. Future iterations addressing the graph sparsity problem could expand the method's scope to less-characterized genes, potentially accelerating therapeutic target discovery pipelines.

• →Stable-Shift achieves 0.592 cosine similarity on gene perturbation prediction, outperforming GEARS baseline across multiple benchmark datasets
• →The method integrates STRING interactions, network topology, expression statistics, and Gene Ontology annotations via graph convolution to contextualize unseen genes
• →Performance gains are modest but consistent, with ±0.008 variance across five evaluation splits, indicating stability
• →Limitations in sparse graph neighborhoods and gene-space predictions restrict applicability to well-characterized genes in established biological networks
• →Results support structured latent-response prediction as superior to unstructured statistical approaches for genomic extrapolation tasks