Fig. 3: Comparison of in-DEL hold-out test performances of ML models.
From: Evaluation of DNA encoded library and machine learning model combinations for hit discovery
The models were trained using data from three DELs (80%) and tested using in-DEL hold-out set (20%). The feature representation for the molecules was 2048 bits Morgan fingerprints for MLP, SVM, RF, and XGB. The ChemProp model internally generated graphical neural network-based features to represent the molecules (Methods: Feature representation). The reported balanced accuracy, MCC, F1 score, and recall is reported for (a) multi-layer perceptron, (b) support vector machine, (c) random forest, (d) extra-gradient boosting and (e) graphical neural network (ChemProp) models. Values indicate the binary classification performance (Methods: ML performance evaluation metrics) of the five ML models in correctly predicting orthosteric DEL binders of CK1α and CK1δ.
