Machine Learning Guided Drug Discovery: Comprehensive Application in Janus Kinase Inhibitor Design

Abstract

Conventional drug discovery is often plagued by resource-intensive processes, leading to significant attrition rates due to challenges such as efficacy shortcomings, safety concerns and suboptimal pharmacokinetics (PK). Meantime, the development of Machine Learning (ML) is heralding a transformation in this field. ML empowers precise predictions and de novo drug design, mitigating the necessity for costly and time-consuming experimental endeavors. In response to the challenges in drug discovery field, we introduced CoGT, an ensemble ML method adept at distinguishing JAK inhibitors from non-inhibitors. CoGT integrates conventional ML models, a graph-based model GraphVAE and a transformer-based model chemBERTa, resulting in state-of-the-art performance in predicting JAK inhibition. Subsequently, we designed a novel gastrointestinal (GI) Janus kinase (JAK) inhibitor for ulcerative colitis (UC) treatment with the application of ML and structure-tissue selectivity-activity- relationship (STAR). Employing the STAR system, we successfully designed a Class III candidate, MMT3-72, characterized by high selectivity for GI tissues and moderate potency to JAK isoforms. This innovative approach mitigated the stringent requirement for JAK isoform specificity, promising an effective and safe treatment for UC. Employing CoGT, we identified MMT3-72-M2, major metabolite of MMT3-72, as a potent JAK inhibitor. Subsequent experimental validation corroborated these predictions, emphasizing MMT3-72’s weak JAK inhibition and MMT3-72-M2’s effectiveness against JAK1/2 and TYK2. In vivo investigations further demonstrated MMT3-72’s superior efficacy, showcasing its targeted action and minimal systemic toxicity. Furthermore, we established a comprehensive ML framework to evaluate absorption, distribution, metabolism, excretion and toxicity (ADME-T) profiles cost effectively, addressing PK and toxicity concerns inherent in drug discovery. This ML approach facilitates concurrent prediction of multiple ADME-T properties, leveraging Graph neural networks (GNN)-based models to expedite drug candidate identification. Moreover, we harnessed ML to design JAK-specific inhibitors capable of targeting specific isoforms, aiming to mitigate toxicity concerns of JAK inhibitors. This ongoing work involves in silico drug design by using variational autoencoder and REINFORCE algorithm. Selection of potent drug candidates will be guided by the criteria set by the STAR, predicted ADME-T profiles, subsequently validation through experimental assays. In summary, our research underscores the profound potential of ML in accelerating drug discovery, aiming to make drug discovery practices more efficient and effective in delivering promising drug candidates at the forefront of pharmaceutical research.