Bridging cheminformatics and water emergency response: An interpretable machine learning approach accelerates natural...

Harmful algal blooms (HABs) are increasingly threatening aquatic ecosystems and water resource sustainability, demanding urgent development of eco-friendly control strategies. While natural algaecides offer biodegradable alternatives to synthetic chemicals, conventional discovery methods remain labor-intensive and inefficient. Here, we present an innovative explainable machine learning (ML) framework integrating virtual screening with algal bloom emergency management needs. We curated a dataset of 9751 algal inhibition records encompassing diverse algal taxa from public database and literature, resulting in 1946 unique compounds for modeling. We developed a ensembled model using three optimized ML algorithms, along with advanced SHapley Additive exPlanations (SHAP) interpretability. The model achieved high predictive accuracy (AUC 0.889) and identified key molecular features governing algal inhibition, notably revealing the synergistic effects of molar refractivity (MolMR) and hydrophobicity (MolLogP) on antialgal activity. Experimental validation confirmed the significant inhibitory effects (46.7% to 58.5% at 10 mg/L) of 5 screened natural algaecides against Microcystis aeruginosa, including taxifolin from fruits and vegetables, and caryophyllene epoxide from various plants. Our open workflow bridges cheminformatics with aquatic ecosystem preservation, offering a scalable and sustainable solution for mitigating HABs and advancing water resource management strategies.