DrugAgent: A multi-agent digital biosensor framework for interpretable virtual drug screening.

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery
• Biosensor technology

Background:

• Traditional virtual screening models are target-specific and require costly retraining for new assays.
• Large Language Model (LLM) approaches often lose molecular signal fidelity through textualization, limiting sensitivity and interpretability.
• Existing methods struggle with adaptability and interpretability in early-stage drug discovery.

Purpose of the Study:

• To introduce DrugAgent, a reconfigurable, multi-agent software biosensor framework for interpretable, gradient-free virtual screening.
• To provide an AI-powered sensing solution for Internet of Things (IoT)-enabled drug discovery.
• To overcome the limitations of target-specific, hard-wired models and ambiguous molecular textualization.

Main Methods:

• DrugAgent ingests raw physicochemical property signals for closed-loop sensing and decision-making.
• Utilizes a Signal Profiling Knowledge Base for sensor calibration.
• Employs a Diversity-Aware Retrieval module for robust decision thresholds.
• Features a hierarchical agent pipeline for noise filtering, feature selection, and interpretable output generation.

Main Results:

• Achieved 89.25% accuracy in gradient-free evaluation, outperforming standard LLM prompting baselines.
• Demonstrated robustness and practical utility through comprehensive evaluations, including early enrichment analysis (BEDROC = 0.835).
• Matched or outperformed supervised ML baselines in cold-start scenarios, highlighting advantages in data-limited settings.

Conclusions:

• DrugAgent offers a plug-and-play digital biosensor, rapidly repurposable across targets and adaptable to new data streams.
• Provides an interpretable and verifiable alternative to black-box virtual screening models.
• Represents a significant advancement for next-generation IoT-enabled digital drug discovery, especially in data-limited early stages.