Deep learning-guided discovery of selective JAK2-JH2 allosteric inhibitors: integration of MLP predictive modeling, BREED-based library design, and computational validation

The JAK2 pseudokinase domain (JH2) is an important therapeutic target in hematologic and oncologic diseases, motivating the search for selective allosteric inhibitors. In this study, a multilayer perceptron (MLP) deep learning model was trained on 1,200 JAK2-targeting compounds and validated internally and externally, while a BREED-based fragment hybridization strategy generated 6,210 new molecules that were screened using MLP scoring, pharmacokinetic filters, and molecular docking. Three compounds–BRD1, BRD2, and BRD3–emerged as promising inhibitors, with BRD1 showing the strongest binding affinity, highest conformational stability, and best selectivity for key JH2 residues, surpassing the reference ligand 36H; MD and ADMET analyses further supported its stability and favorable safety profile. Overall, BRD1 is identified as a strong computational candidate for selective allosteric inhibition of JAK2-JH2, warranting future experimental validation, and all models and code are openly available.