DrugReasoner: Interpretable drug approval prediction with a reasoning-augmented language model

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Machine learning for pharmaceutical development

Background:

• Drug discovery is resource-intensive, necessitating accurate early prediction of approval outcomes.
• Classical and deep learning models offer predictive power but lack interpretability.
• Interpretability is crucial for optimizing research investments and enhancing trust in AI-driven decisions.

Purpose of the Study:

• To develop and evaluate DrugReasoner, a reasoning-based large language model (LLM) for predicting small-molecule drug approval.
• To enhance the interpretability of AI models in drug discovery by providing step-by-step rationales.
• To improve the accuracy and robustness of AI-driven predictions in pharmaceutical decision-making.

Main Methods:

• Developed DrugReasoner using the LLaMA architecture, fine-tuned with group relative policy optimization (GRPO).
• Integrated molecular descriptors with comparative reasoning against similar approved/unapproved compounds.
• Evaluated performance using AUC and F1 scores on validation, test, and external independent datasets.

Main Results:

• DrugReasoner achieved strong performance with AUCs of 0.732 (validation) and 0.725 (test), outperforming conventional baselines.
• On an independent dataset, DrugReasoner achieved an AUC of 0.728 and F1-score of 0.774, surpassing the ChemAP model.
• The model provided interpretable, step-by-step rationales alongside predictions and confidence scores.

Conclusions:

• Reasoning-augmented LLMs like DrugReasoner offer a promising approach to interpretable and accurate AI-assisted drug discovery.
• DrugReasoner addresses a key bottleneck in AI-driven pharmaceutical decision-making by combining predictive accuracy with transparency.
• The model's robust performance and interpretability demonstrate its potential for real-world application in optimizing drug development pipelines.