A bond-aware graph neural network integrated with docking, pharmacophore modeling and molecular dynamics for JAK2 inhibitors affinity prediction.

The Janus kinase (JAK)/STAT signaling pathway plays a pivotal role in cancer biology as well as in inflammatory and autoimmune disorders such as psoriasis. Recent advances in biomedical research and targeted therapies have highlighted the importance of computational approaches for accelerating the discovery of selective kinase inhibitors. This study aimed to develop a robust computational framework for predicting the inhibitory potency of JAK2 ligands and for analyzing their binding interactions using structure-based methods. A curated dataset of 1869 chemically valid JAK2 ligands with experimentally reported Ki values was compiled, standardized, and converted to pKi. Using this dataset, a bond-aware graph neural network (GNN) was trained and evaluated for pKi prediction. Top-ranked predicted ligands were further examined via molecular docking, pharmacophore modeling, and molecular dynamics simulations to assess their interactions within the JAK2 ATP-binding site. The proposed model achieved strong predictive performance, yielding an average test-set R2 of 0.91 ± 0.01, MAE of 0.14 ± 0.01, and RMSE of 0.26 ± 0.02 across repeated data splits. Structure-based analyses supported the predicted binding poses and identified key stabilizing interactions within the JAK2 ATP-binding site. Overall, this integrative computational framework provides a reliable approach for predicting JAK2 inhibitory potency and offers mechanistic insights that may support the computational prioritization of candidate molecules for future experimental evaluation.