Predicting PROTAC degradation activity and selectivity of effective E3 ligase through harnessing a combination of...

Proteolysis targeting chimeras (PROTACs) are drugs that recruit target proteins and E3 ligases to leverage the ubiquitin-proteasome system (UPS) for specific degradation of proteins. This approach has received significant attention because of their ability to target previously considered "undruggable" proteins. Predicting PROTAC degradation activity and the selection of E3 ligase often entail tests and trials to delve into the vast chemical space. To advance the understanding of protein degradation, this study encapsulates a comprehensive analysis of PROTAC molecules to evaluate a 30-machine learning model based on AtomPair fingerprinting to predict PROTAC degradation activity (15 models) and selectivity of effective E3 ligases (15 models) using the KNIME platform. Decision tree (DT), XGBoost tree ensemble (XGB), tree ensemble (TE), probabilistic neural network (PNN), and logistic regression (LR) algorithms were tested with three sampling strategies: (i) random; (ii) stratified; and (iii) synthetic minority oversampling technique (SMOTE), and the results of 30 models were then particularly compared to analyze model's distinct characteristics. In predicting PROTAC degradation activity, the XGB model trained with stratified sampling achieves 82.77% prediction accuracy, and Cohen's kappa measures of 0.584 on the test set. The XGB model trained with stratified sampling demonstrated the best predictive performance, achieving an area under the receiver operating characteristic (ROC) curve (AUC) measures of 0.811 in five-fold cross-validation, making it more balanced and significant at precisely finding the protein degradation traits of active PROTACs. On the other hand, XGB model trained with random sampling outperformed the other four machine learning algorithms in the selection of E3 ligase, with promising accuracy values of about 95.38%, and Cohen's kappa measures of 0.910 on the test set. XGB and TE models trained using SMOTE demonstrated strong performance with an AUC of 0.965 in five-fold cross-validation, making it an ideal choice for the selectivity of CRBN ligase. XGB models trained using SMOTE outperformed other machine learning models, achieving AUC scores of 0.960 in five-fold cross-validation and proven effective for the selectivity of VHL ligase. This study demonstrates that our approach provides a valuable perspective on improved prediction of PROTAC degradation activity and selectivity of effective E3 ligase in the pursuit of all possible interaction mechanisms, representing an important step towards understanding the behavior of PROTACs for wider clinical success. Our workflow is available at https://github.com/sanjeevipandiyan1990/PROTAC-activity-and-E3-ligase-selectivity.