Classification of ULK1 inhibitors and SAR analysis by machine learning methods

Area of Science:

• Computational chemistry
• Drug discovery
• Oncology

Background:

• Unc-51 like kinase 1 (ULK1) is a crucial regulator of autophagy and a promising anticancer target.
• Developing effective ULK1 inhibitors requires robust predictive models for drug design.

Purpose of the Study:

• To build and evaluate computational models for predicting the activity of ULK1 inhibitors.
• To identify key structural features driving ULK1 inhibitor activity through structure-activity relationship (SAR) analysis.

Main Methods:

• Collected 846 ULK1 inhibitors with IC50 values.
• Developed 39 classification models using ECFP4, MACCS fingerprints, Mordred descriptors, SVM, RF, XGBoost, DNN, and FP-GNN.
• Performed SAR analysis using K-means clustering and SHAP values.

Main Results:

• The best performing model (Model_1D_1, an ECFP4-based DNN) achieved >95% accuracy and MCC >0.9.
• Model_1D_1 demonstrated reliable prediction for >84% of compounds within its applicability domain.
• Identified a common scaffold (2-amino-4-(2-thienyl)-5-(trifluoromethyl)pyrimidine) in high-activity subsets and highlighted critical active fragments.

Conclusions:

• A highly accurate DNN model was developed for predicting ULK1 inhibitor activity.
• SAR analysis provided insights into critical molecular fragments and scaffolds for optimizing ULK1 inhibitors.
• These findings offer a theoretical foundation for designing novel anticancer drugs targeting ULK1.