همایش ملی بیوانفورماتیک ایران

صفحه اصلی / 4th international edition and 13th Iranian Conference on Bioinformatics

Machine Learning-Driven Discovery of JAK2 Inhibitors from ChEMBL Databank

نویسندگان :

Negar Abdolmaleki¹ Hamid Mahdiuni²

1- دانشگاه رازی 2- دانشگاه رازی

کلمات کلیدی :

Janus Kinase 2 Inhibitors،Virtual Screening،Machine Learning،Random Forest،ECFP4 Fingerprints

چکیده :

This study aims to identify potent JAnus Kinase 2 (JAK2) inhibitors using machine learning models for virtual screening of the small molecule subset of the ChEMBL databank. JAK2 is a key player in the JAK-STAT signaling pathway, which regulates immune responses and inflammation. Hence, JAK2 inhibitors have significant potential for treating autoimmune disorders, inflammatory diseases, and certain cancers (Lv and Qi, 2024). The machine learning models utilized in this study included Support Vector Machine (SVM), XGBoost, and Random Forest (RF). For training the models, datasets of 6,847 active JAK2 inhibitors were sourced from ChEMBL (Gaulton and Hersey, 2017), BindingDB (Gilson and Liu, 2016), and PubChem (Kim and Chen, 2019), along with 6,500 inactive compounds from the DUD-E database (Mysinger and Carchia, 2012). Each compound was labeled with an activity status (1 for active and 0 for inactive). Molecular characteristics were represented using extended connectivity fingerprints (ECFP4) (Baptista and Correia, 2022), alongside molecular descriptors such as molecular weight, polar surface area (PSA), and logP. The datasets were processed using RDKit to extract ECFP4 fingerprints and additional descriptors. The performance of the models was evaluated using several metrics, including accuracy and area under the curve (AUC). The Random Forest model achieved the highest performance, with a testing set accuracy of 0.9970 and an AUC of 0.9996. The SVM model achieved an accuracy of 99.63% and an AUC of 99.93%, while the XGBoost model had an accuracy of 99.55% and an AUC of 99.85%. Therefore, according to the performance data, the Random Forest model was used for virtual screening on a large-scale compound database containing 1,930,555 molecules. The model identified 99,653 compounds as potential JAK2 inhibitors (active) and classified the remaining 1,830,902 as inactive. These findings demonstrate that combining ECFP4 fingerprints and molecular descriptors with the Random Forest model highlights the effectiveness of machine learning-driven virtual screening in accelerating drug discovery for JAK2 inhibitors.