A Machine Learning-Enhanced Ligand-Based Workflow Identifies Novel Cryptosporidium parvum Inhibitors

Abstract Cryptosporidium parasites are a leading cause of waterborne disease worldwide, causing life-threatening diarrhea and malnutrition in young children and people with HIV. Currently, nitazoxanide is the only FDA-approved treatment, but it is largely ineffective in severely immunocompromised patients, highlighting the need for new therapies. Here, we use a ligand-based virtual screening workflow that integrates Thompson Sampling, a pharmacophore model, and a machine learning classifier to screen a subset of Enamine’s REAL database comprising 51 million virtual compounds. Across two screening campaigns, 20 novel micromolar hits were identified among 69 tested compounds (29%), eight of which exhibited confirmed activity at or below 21 μM. These findings expand the chemical diversity of the original series by introducing new tail and head chemotypes and demonstrate the effectiveness of active learning strategies for improving ligand-based virtual screening performance.