Interpretable machine learning-driven identification of novel DENV NS2B-NS3 protease inhibitors through multi-stage virtual screening and experimental validation.

Dengue fever is a mosquito-borne viral infection caused by dengue virus (DENV). It has emerged as a worldwide health problem, afflicting millions of people each year throughout the tropical and subtropical regions. To date, there is no FDA-approved drug for the treatment of dengue fever, highlighting the urgent need to discover novel anti-dengue drugs. In this study, multiple machine learning models were constructed to predict the inhibitory activity of small molecules against DENV NS2B-NS3, a protease that is crucial for the replication of DENV. Among them, RF-ECFP and XGBoost-ECFP were identified as the optimal models. The SHapley Additive exPlanations method was introduced for the interpretation of predictive results. Following the initial machine learning predictions, a multi-step screening process including multi-level molecular docking, molecular dynamics simulations, and molecular orbital calculations was conducted, ultimately identifying six hit compounds from a library containing ten million small molecules. Molecular docking indicated that compound 5 could form stable interactions with the catalytic triad in the active site of DENV NS2B-NS3 protease. The surface plasmon resonance and enzymatic inhibition assays further revealed that compound 5 exhibited a Kd value of 31.2 µM and an IC₅₀ value of 31.44 µM against the DENV NS2B-NS3 protease. Our work suggests that compound 5 represents a potential lead compound for the further development of DENV NS2B-NS3 protease inhibitors.