NetPolicy-RL: network-informed offline reinforcement learning for pharmacogenomic drug prioritization.

Area of Science:

• Computational Biology
• Pharmacogenomics
• Reinforcement Learning

Background:

• Large-scale pharmacogenomic screens generate vast drug response data.
• Existing computational methods often fail to align with practical drug testing limitations.

Purpose of the Study:

• To develop a biologically informed, decision-centric framework for pharmacogenomic drug prioritization.
• To integrate network diffusion modeling with offline reinforcement learning for improved drug selection.

Main Methods:

• Formulated drug selection as an offline contextual bandit problem.
• Incorporated mechanistic biological context via network diffusion (STRING, Reactome) and multi-omics data.
• Integrated network disruption scores with drug response data for state representation, optimized with an offline actor-critic architecture.

Main Results:

• NetPolicy-RL significantly outperformed global ranking heuristics and learning-to-rank baselines on held-out data.
• Achieved statistically significant improvements in per-cell Normalized Discounted Cumulative Gain (NDCG@10).
• Demonstrated substantial reductions in per-cell regret, improving NDCG@10 for 88.7% of cell lines compared to GlobalTopK.

Conclusions:

• Combining mechanistic network biology with offline policy learning offers an effective and interpretable approach for drug prioritization.
• The integration of network-derived features and empirical response signals is crucial for robust performance.
• NetPolicy-RL advances precision oncology by optimizing drug selection strategies.