A Computational Workflow Integrating Explainable AI and Database Mining for the Prioritization of Antimicrobial Natural Products

Abstract

The escalating threat of antimicrobial resistance demands the discovery of antibacterial agents with unprecedented chemical scaffolds. Natural products have historically been the most prolific source of antibiotics, yet the vast majority of catalogued natural products remain untested for antimicrobial activity. Here we present Natural-Product Graph–language, Uncertainty-Aware Relational Discovery (NP-GUARD), a modular computational framework that reframes antimicrobial natural-product discovery as an uncertainty-aware, relational, multi-objective learning problem. NP-GUARD integrates six core components: provenance-tracked data curation, multi-view molecular representation, pathogen-conditioned multi-task prediction, explicit uncertainty quantification, stability-aware explainability, and conformal-Pareto prioritization, all consolidated within a queryable knowledge base. Of these components, the present study implements and benchmarks the data-curation, predictive, uncertainty, explainability and prioritization layers, including a first pathogen-conditioned masked multi-task graph model, while the full multi-view encoder and the knowledge base are specified for subsequent work. We report a proof-of-concept evaluation on public screening data forStaphylococcus aureus (38,995 compounds), Escherichia coli (2291) and Pseudomonas aeruginosa (29,986), together with human cytotoxicity data and a 406,959-compound screening library from COCONUT. Under repeated scaffold-aware splitting, baseline models achieve strong discrimination (AUROC 0.78–0.93) but reveal substantial split-to-split variance in minority-class performance; class-conditional conformal prediction attains near-nominal per-class coverage; and feature attributions show only moderate stability across independently trained models (top-20 Jaccard 0.54), underscoring a critical caveat for explanation-driven discovery. Screening the COCONUT library validates the uncertainty-aware activity ranking, achieving an 87-fold enrichment of known antibiotics in the top 100 candidates (128-fold within the fully characterized named-compound stratum), although analogue-exclusion analysis shows this signal is dominated by rediscovery of training-like chemistry. The novelty-seeking Pareto layer yields a structurally diverse 90-compound shortlist spanning 72 scaffolds, deliberately steered toward unexplored chemotypes; we emphasize that the shortlist constitutes prioritized hypotheses for experimental testing, not validated actives. NP-GUARD provides a transparent, reproducible and decision-oriented foundation for prioritizing natural products for experimental validation.