Screening and Evaluation of Computationally Predicted JAK1-Binding Candidates from Multi-Source Chemical Space Using Deep Learning and Generative Models

Background/Objectives: Janus kinase 1 (JAK1) is a key mediator in inflammatory and autoimmune signaling pathways and represents a validated therapeutic target. Despite the availability of several JAK inhibitors, challenges in safety and efficacy persist, necessitating further computational prioritization of candidate molecules. Methods: We developed an integrated virtual screening workflow combining ensemble deep learning scoring (KPGT, KarmaDock, DrugCLIP), ADMET-AI filtering, and AutoDock Vina (v1.2.3) docking. The workflow was applied to the NPBS Atlas natural-product database and to molecules generated by 3DSMILES-GPT. Selected candidates were further evaluated by molecular dynamics simulations, DFT calculations, and synthetic accessibility assessment, with Filgotinib used as a reference compound. Results: Three natural-product-derived and three AI-generated molecules were prioritized. Among them, the MM-PBSA-estimated binding free energies were −26.251 kcal/mol for NPBS_111591, −19.858 kcal/mol for 3DSMILES_17365, and −19.593 kcal/mol for 3DSMILES_38767, compared with −29.253 kcal/mol for the reference compound Filgotinib. These MM-PBSA values were obtained under the same computational protocol and can provide relative computational comparisons, but pharmacological potency still requires validation by IC50 or biochemical binding assays. Conclusions: The study prioritizes computationally predicted JAK1-binding candidates and provides a multi-stage in silico workflow for further experimental evaluation.