Learning drug synergy through environment-conditioned feature modulation.

Area of Science:

• Computational biology
• Pharmacology
• Artificial intelligence

Background:

• Drug combinations are vital for overcoming cancer resistance.
• Current deep learning models for drug synergy lack dynamic cell-drug environment modeling.
• This limitation hinders the prediction of environment-specific synergistic effects.

Purpose of the Study:

• To develop a deep learning framework, Env-Syn, that dynamically models cell-drug interactions for improved synergy prediction.
• To capture how the cellular environment modulates drug representations for environment-specific synergy.

Main Methods:

• Env-Syn utilizes Environment-Conditioned Feature Modulation.
• A Residual Feature-wise Linear Modulation (R-FiLM) module performs affine transformations on drug representations.
• These transformations are conditioned on paired drugs and cellular environments.

Main Results:

• Env-Syn consistently outperforms state-of-the-art methods in benchmark evaluations.
• The model demonstrates exceptional generalization, with AUROC and AUPRC exceeding 0.81 in leave-drug-out settings for unseen drugs.
• Env-Syn shows strong cross-dataset reliability and identified 8 literature-supported novel drug combinations out of 15 predictions.

Conclusions:

• Env-Syn is an effective computational tool for discovering drug synergy.
• The framework accurately models drug-drug-cell interactions, improving prediction of synergistic effects.
• Env-Syn shows promise for identifying novel, effective cancer drug combinations.