DiffDR: A Diffusion-based Deep Learning Framework for Accurate Drug Response Imputation and Feature Selection.

Area of Science:

• Computational biology
• Genomics
• Pharmacology

Background:

• Molecular features significantly impact cellular responses to therapies, influencing drug sensitivity and resistance.
• Understanding these molecular influences is key to explaining varied treatment outcomes.
• Integrating multi-omics data aids in biomarker discovery and drug response prediction but faces modeling challenges due to data complexity.

Purpose of the Study:

• To develop a robust and interpretable computational framework for predicting drug response.
• To model complex dependencies within multi-omics data for enhanced prediction accuracy.
• To provide mechanistic insights into how molecular features affect drug response.

Main Methods:

• Introduced DiffDR, a diffusion-based framework utilizing an energy-constrained diffusion module.
• Modeled multi-omics features and drug representations, efficiently propagating information without explicit graph structures.
• Incorporated a gradient-based interpretability module for attributing predictions to specific omics features.

Main Results:

• DiffDR achieved superior predictive performance over existing state-of-the-art methods.
• Ablation studies confirmed the significant contribution of the energy-constrained diffusion mechanism to predictive accuracy.
• The framework effectively handles high-dimensional multi-omics data.

Conclusions:

• DiffDR successfully captures cross-modal molecular dependencies.
• The framework offers interpretable insights into drug response mechanisms.
• DiffDR provides a robust approach for understanding molecular drivers of drug response.