Li, S.; Ruan, Y.; Yang, X.; Wen, Z.; Saigo, H.

In transcriptomics-driven drug discovery, upstream predictors of chemical-induced transcriptional profiles (CTPs) are typically trained with gene-wise objectives, whereas downstream interpretation relies on pathway-level, rank-based statistics such as Gene Set Enrichment Analysis (GSEA). This objective mismatch destabilizes pathway conclusions under prediction errors: small ranking perturbations can flip enrichment direction or distort pathway ordering. To bridge this gap, we present differentiable GSEA (dGSEA), a training-compatible surrogate that maps predicted gene-level scores to pathway enrichment with well-behaved gradients. Technically, dGSEA replaces discrete ranking operations with temperature-controlled soft sorting, smooth prefix accumulation, and differentiable extremum aggregation. Critically, to preserve the statistical semantics of classical GSEA, we introduce sign-specific robust permutation normalization (dNES) with optional k-calibration. For computational efficiency, a scalable Nystrom-window approximation (nyswin) reduces the quadratic bottleneck to near-linear complexity, enabling genome-scale evaluation. Empirically, across synthetic benchmarks and LINCS L1000 signatures, dGSEA matches classical GSEA accuracy with improved numerical stability. When incorporated as an auxiliary objective for SMILES-to-transcriptome prediction, dGSEA improves pathway-level agreement (macro correlation 0.257 -> 0.306; sign accuracy 0.620 -> 0.641) without compromising gene-level performance, providing a practical mechanism for pathway-aware optimization in transcriptomic prediction pipelines.