DiSPA: Differential Substructure-Pathway Attention for Drug Response Prediction

TL;DR

DiSPA tackles the problem of drug response prediction by coupling chemical substructures with pathway-level gene expression to model context-dependent mechanisms, addressing limitations of independent or late-fusion approaches. It introduces a bidirectional differential cross-attention framework that conditionally refines pathway-to-substructure and substructure-to-pathway interactions, suppressing spurious associations via a differential attention mechanism with a suppression parameter $\lambda$. Across the GDSC benchmark and CTRP data, DiSPA achieves state-of-the-art predictive accuracy for $IC_{50}$ values, with the largest gains under drug-blind and disjoint splits, and yields attention patterns that recapitulate pharmacophores and tissue organization. Moreover, a bulk-trained model transfers to spatial transcriptomics and single-cell data for zero-shot inference of region- and cell-type-specific drug sensitivity, enabling cross-scale pharmacogenomic analyses.

Abstract

Accurate prediction of drug response in precision medicine requires models that capture how specific chemical substructures interact with cellular pathway states. However, most existing deep learning approaches treat chemical and transcriptomic modalities independently or combine them only at late stages, limiting their ability to model fine-grained, context-dependent mechanisms of drug action. In addition, standard attention mechanisms are often sensitive to noise and sparsity in high-dimensional biological networks, hindering both generalization and interpretability. We present DiSPA, a representation learning framework that explicitly disentangles structure-driven and context-driven mechanisms of drug response through bidirectional conditioning between chemical substructures and pathway-level gene expression. DiSPA introduces a differential cross-attention module that suppresses spurious pathway-substructure associations while amplifying contextually relevant interactions. Across multiple evaluation settings on the GDSC benchmark, DiSPA achieves state-of-the-art performance, with particularly strong improvements in the disjoint-set setting, which assesses generalization to unseen drug-cell combinations. Beyond predictive accuracy, DiSPA yields mechanistically informative representations: learned attention patterns recover known pharmacophores, distinguish structure-driven from context-dependent compounds, and exhibit coherent organization across biological pathways. Furthermore, we demonstrate that DiSPA trained solely on bulk RNA-seq data enables zero-shot transfer to spatial transcriptomics, revealing region-specific drug sensitivity patterns without retraining. Together, these results establish DiSPA as a robust and interpretable framework for integrative pharmacogenomic modeling, enabling principled analysis of drug response mechanisms beyond post hoc interpretation.

Figures (4)

• Figure 1: Overview of the DiSPA framework. DiSPA consists of three major stages: feature encoding, dual--view differential cross--attention, and drug response prediction. Gene expression profiles are first mapped to KEGG pathways to construct pathway--level gene embeddings, while drug SMILES are decomposed into substructures and encoded alongside drug--level representations. In the dual--view cross--attention module, (View 1) pathway embeddings attend to drug substructures to capture pathway--specific chemical relevance, and (View 2) drug embeddings attend to pathway representations to model drug--conditioned biological context. The resulting pathway–substructure and drug–pathway representations are aggregated and passed through a multilayer perceptron (MLP) to predict drug response ($IC_{50}$) for each cell line–drug pair.
• Figure 2: Predictive performance and global organization of drug--cell line interactions learned by DiSPA. (a) Scatter plot comparing predicted and observed $\ln(IC_{50})$ values across all evaluated drug--cell line pairs under the Random split, demonstrating high regression fidelity across the full dynamic range. (b) Pairwise comparison of Pearson correlation coefficients (PCC) between DiSPA and DRPreter at the drug level (left) and cell--line level (right), showing the proportion of drugs and cell lines for which DiSPA achieves higher predictive correlation. (c) Kernel density estimates of substructure--level attention alignment scores across drugs, showing a bimodal distribution with subsets of compounds exhibiting strong positive correlation between attention weights and chemical substructure similarity (structure--driven) and others displaying weak correlation, indicative of more diffuse, context--dependent attention patterns. (d) Comparison of BertzCT structural complexity indices between drugs stratified by substructure–attention alignment, showing that compounds with low alignment exhibit higher molecular complexity than those with high alignment. (e) Density distributions of predicted $\ln(IC_{50})$ values for representative structure--driven and context--driven drugs across sensitive cell lines, highlighting divergent sensitivity patterns associated with different attention regimes. (f) Consensus clustering of cell--line attention patterns aggregated across all drugs, revealing a block--diagonal organization corresponding to major tissue lineages. (g) UMAP visualization of cell lines based on attention--derived similarity, showing clustering consistent with tissue annotations despite the absence of explicit tissue labels during model training.
• Figure 3: Mechanistic interpretation of substructure–pathway interactions learned by DiSPA. Representative case studies of chemically similar drug pairs with divergent predicted responses. Case 1 (KELLY, peripheral nervous system): (a) Chemical structures of UNC0638 and UNC0642 with highlighted regions associated with Path2Sub attention weights and potential activity cliffs. (b) Path2Sub attention weights showing differential emphasis on specific substructures. (c) Drug2Path attention comparison illustrating pathway--level consistency and divergence linked to activity differences. Case 2 (697, hematopoietic and lymphoid): (d) Chemical structures of Z--LLNle--CHO and MG--132 with highlighted activity--cliff regions and predicted $\ln(IC_{50})$ values. (e) Path2Sub attention distributions highlighting substructure--specific differences. (f) Drug2Path attention comparison demonstrating context--dependent pathway--level attention patterns.
• Figure 4: Transfer of bulk--trained drug response prediction to spatial and single--cell transcriptomics. (a) Counts of spatial domain--selective drugs in an invasive ductal carcinoma spatial transcriptomics dataset. (b) Overlap of domain--selective drugs ($p < 0.05$). (c) Top--ranked domain--selective compounds with ln(IC$_{50}$) differences. (d) Spatial maps of predicted sensitivity for representative tumor-- and invasive--selective drugs. (e) Cell type--resolved drug response heatmap in a colorectal cancer single--cell RNA--seq atlas. (f) UMAP of single cells based on predicted drug response profiles. (g) Distribution of cell type--selective drug counts. (h, i) Differential drug sensitivity across major cell types and their subtypes. Predictions are generated using a DiSPA model trained exclusively on bulk RNA--seq data.