Deep Learning-Guided Discovery of Selective JAK2-JH2 Allosteric Inhibitors: Integration of MLP Predictive Modeling, BREED-Based Library Design, and Computational Validation

Mebarka Ouassaf^1*afaf zekri¹Shafi Ullah Khan²Kannan RR Rengasamy³Bader Y Alhatlani^4*

• ¹University of Biskra, Biskra, Algeria
• ²Universite de Caen Normandie, Caen, France
• ³North-West University, Potchefstroom, South Africa
• ⁴Qassim University, Buraydah, Saudi Arabia

The Janus kinase 2 pseudokinase domain (JH2) is a critical therapeutic target in hematologic and oncologic disorders, offering a unique avenue for allosteric modulation. In this study, we developed a robust multilayer perceptron-based deep learning model trained on 1,200 active and inactive JAK2-targeting compounds, validated through rigorous internal and external evaluations. Leveraging a BREED-inspired fragment hybridization strategy, we generated an enriched library of 6,210 molecules derived from known JAK2 inhibitors. Sequential virtual screening guided by the MLP model, pharmacokinetic profiling (QED, SAS), and molecular docking yielded three high-potential candidates: BRD1, BRD2, and BRD3. Among these, BRD1 exhibited superior binding affinity, conformational stability, and selectivity for key JH2 residues, outperforming the reference ligand 36H. Molecular dynamics simulations and ADMET predictions further validated BRD1's stability and low off-target risks. These computational insights nominate BRD1 as a promising allosteric JAK2-JH2 inhibitor candidate, though experimental validation of its therapeutic efficacy remains essential. (All code and models are openly accessible at GitHub/profnabila/mlp). Keywords: JAK2 pseudokinase domain (JH2); Allosteric inhibition; Deep learning; Virtual screening; Molecular dynamics simulations.