Designing new hit series of JAK3 inhibitors using generative AI, reinforcement learning, and molecular dynamics

Janus kinase 3 (JAK3) is a crucial non-receptor tyrosine kinase involved in the JAK/STAT signaling pathway. Its dysregulation is associated with cancers, autoimmune diseases, and immunodeficiencies, making JAK3 inhibitors a promising therapeutic option. This study introduced a framework for designing new JAK3 inhibitors using generative AI, reinforcement learning, and molecular dynamics. We constructed a chemical space of 3000 compounds and navigated it with reinforcement learning. This process generated 13 Murcko scaffolds, of which five were novel and synthesizable. Each scaffold was used to generate 9000 compounds, which were filtered to select 70 potential candidates targeting JAK3. Molecular dynamics simulations were performed to calculate MM-GBSA scores for these compounds. In a benchmark for ritlecitinib, we found that scaffolds NS_1813 and NS_2063 showed better binding affinities, with 20 compounds meeting our selection criteria. ADMET predictions were performed on compounds from NS_1813 and NS_2063 to evaluate their pharmacokinetic and safety profiles. From these, six lead compounds were chosen for detailed molecular dynamics and docking pose analyses to assess their conformational stability and binding interactions within the JAK3 active site. These compounds are promising candidates for further development, including chemical synthesis, followed by in vitro and in vivo testing to evaluate their potential as JAK3-targeting therapeutic agents.