DiSPA: differential substructure-pathway attention for drug response prediction.

Area of Science:

• Computational biology
• Pharmacogenomics
• Machine learning

Background:

• Precision medicine requires accurate drug response prediction by integrating chemical and biological data.
• Existing deep learning models often fail to capture fine-grained interactions between drug substructures and cellular pathways.
• Standard attention mechanisms struggle with noisy and sparse biological data, limiting model generalization and interpretability.

Purpose of the Study:

• To develop a deep learning framework that models bidirectional interactions between chemical substructures and pathway-level gene expression for improved drug response prediction.
• To enhance the robustness and interpretability of models by suppressing spurious associations and focusing on context-relevant interactions.

Main Methods:

• Introduced Differential Substructure-Pathway Attention (DiSPA), a novel framework utilizing differential cross-attention.
• Modeled bidirectional interactions between chemical substructures and pathway-level gene expression.
• Evaluated performance on the GDSC benchmark and external datasets (CTRP), including disjoint and drug-blind splits.

Main Results:

• DiSPA achieved state-of-the-art performance on the GDSC benchmark, with significant improvements in disjoint and drug-blind settings, indicating enhanced robustness.
• Attention analysis revealed more selective and concentrated interactions compared to standard cross-attention.
• Demonstrated promising generalization on external datasets and potential for zero-shot application to spatial transcriptomics for region-specific drug sensitivity insights.

Conclusions:

• DiSPA offers a robust and interpretable approach for predicting drug response by effectively modeling chemical-biological interactions.
• The framework shows potential for advancing precision medicine and exploring drug sensitivity in complex biological contexts like spatial transcriptomics.