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

Virtual screening can be viewed as an in silico biosensor for early-stage drug discovery, where physicochemical descriptors act as multi-modal "signals" and the screening model serves as the sensing and decision unit. However, most existing computational biosensors are target-specific and "hard-wired" (trained) to a fixed assay setting, requiring costly data collection and retraining when the biological context changes. Meanwhile, recent Large Language Model (LLM) attempts often "textualize" molecules (e.g., SMILES prompting), which compresses high-fidelity physicochemical signals into ambiguous token sequences, leading to low sensitivity and limited verifiable interpretability. We introduce DrugAgent, a reconfigurable, multi-agent software biosensor framework that supports interpretable, gradient-free virtual screening, aligning with the vision of AI-empowered sensing in the Internet of Things. Analogous to a reconfigurable sensor array, DrugAgent directly ingests raw physicochemical property signals (e.g., predicted binding affinity, lipophilicity) and performs closed-loop sensing and decision making through three coordinated components: (i) an automated Signal Profiling Knowledge Base built from statistical priors to provide sensor-level calibration, (ii) a Diversity-Aware Retrieval module that constructs balanced evidence for robust decision thresholds under class imbalance, and (iii) a hierarchical agent pipeline that filters noise, selects discriminative features, and outputs activity readouts with human-verifiable rationales. Without any model fine-tuning or gradient updates, DrugAgent functions as a plug-and-play digital biosensor that can be rapidly repurposed across targets and updated as new assay data streams in. Experiments on public benchmarks demonstrate that DrugAgent achieves 89.25% accuracy under gradient-free evaluation, significantly outperforming standard LLM prompting baselines while providing an interpretable and verifiable alternative to black-box virtual screening models for next-generation IoT-enabled digital drug discovery. Comprehensive evaluations including early enrichment analysis (BEDROC = 0.835), binding affinity noise sensitivity testing, cross-dataset validation on MoleculeNet benchmarks, and quantitative interpretability assessment further demonstrate the robustness and practical utility of the framework. Under a cold-start setting where predicted binding affinity is unavailable to supervised baselines, DrugAgent - which retains access to binding affinity through its reasoning pipeline - matches or outperforms all supervised ML baselines, highlighting its advantage in data-limited early-stage discovery.