Classification of ULK1 inhibitors and SAR analysis by machine learning methods

Unc-51 like kinase 1 (ULK1), a key regulator of autophagy initiation, is a novel target for anticancer drug design. In this work, we collected 846 ULK1 inhibitors with IC₅₀ values from 30 references. Based on ECFP_4, MACCS fingerprints, and Mordred descriptors, we established a list of classification models by using Support Vector Machine (SVM), Random Forest (RF), extreme Gradient Boosting (XGBoost) and Deep Neural Networks (DNN). Additionally, several Fingerprint and Graph Neural Network (FP-GNN) models were also constructed using mixed molecular fingerprints and molecular graph. A total of 39 classification models were developed. Model_1D_1, an ECFP4-based DNN model, performed the best, achieving accuracies over 95% and Matthews correlation coefficient (MCC) over 0.9 on both validation and test sets. The applicability domain calculated by weighted Euclidean distance indicated that Model_1D_1 could reliably predict the activity for over 84% compounds in both training and test sets. We conducted structure-activity relationship (SAR) analysis through K-means and SHAP. The dataset’s molecular structures were classified into 7 subsets by K-means clustering. We identified three high-activity subsets sharing a common scaffold, 2-amino-4-(2-thienyl)-5-(trifluoromethyl)pyrimidine. SHAP analysis highlighted critical molecular fragments influencing activity, enhancing our understanding of model predictions and providing a theoretical basis for optimizing ULK1 inhibitors.