Machine learning-guided identification and simulation-based validation of potent JAK3 inhibitors for cancer therapy

Abstract

Janus kinase 3 (JAK3) is a hematopoietic-specific kinase implicated in cytokine signaling and immune dysregulation and has recently been associated with cancer progression. However, selective and potent JAK3 inhibitors remain underdeveloped. In this study, we established a machine learning (ML)-based pipeline to identify novel JAK3 inhibitors with anti-cancer potential. A curated ChEMBL dataset of JAK3 inhibitors was used to train multiple ML classifiers, with the Random Forest model achieving the highest performance (AUC = 0.80, F1-score = 0.92). This model was applied to virtually screen 25,084 ChEMBL compounds, yielding 400 high-confidence candidates (prediction score > 0.9). Docking analysis identified ten top binders (binding affinity ≤ –8.5 kcal/mol), of which three CHEMBL49087, CHEMBL4117527, and CHEMBL50064 exhibited optimal ADMET profiles. These compounds underwent 200 ns molecular dynamics simulations, showing low RMSD (0.10–0.20 nm), stable binding conformations, and preserved protein compactness. MM/GBSA calculations revealed that CHEMBL4117527 displayed the strongest binding free energy (–29.5 kcal/mol), surpassing even the co-crystallized ligand (–17.7 kcal/mol). Our integrative approach combining machine learning, docking, pharmacokinetics, molecular dynamics, and free energy analysis presents a robust computational strategy for JAK3 inhibitor discovery. These findings support CHEMBL4117527 as promising candidates for further experimental evaluation in cancer therapeutics.