Protein-protein interaction (PPI) inhibitors have a continued and increasing role in drug discovery. It is hypothesized that machine learning (ML) algorithms are able to classify or identify PPI inhibitors. In this work, we describe the performance of different algorithms broadly used in chemoinformatics to develop a classification model able to identify PPI inhibitors based on structural and physicochemical descriptors. We found that the classification algorithms have different performance according to different features employed in the training process: random forest (RF) models with the extended connectivity fingerprint radius 4 (ECFP4) had the best classification performance as compared to those models trained with ECFP6 o MACCS keys (166-bits). In general, logistic regression models had lower performance metrics than RF models, but ECFP4 was the representation most appropriate for linear regression. ECFP4 also yielded models with high performance metrics, in particular, with support vector machine (SVM). As part of this work, we constructed ensemble models based on the top-performing models. The pipeline code developed in this work and all results are freely available at https://github.com/BarbaraDiazE/PPI_ML.
Supplementary tables and figures
Table S1. PPI subfamilies and compounds. Table S2. RF setup information. Table S3. LRG setup information. Table S4. SVM setup information. Table S5. RF metrics values. Table S6. LRF metrics values. Table S7. SVM metrics values. Table S8. Statistical values of RF models. Table S9. Statistical values of LRG models. Table S10. Statistical values of SVM models. Table S11. RF models validation results. Table S12. LRG models validation results. Table S13. SVM models validation results. Figure S1. RF metrics heatmap. Figure S2. LRG metrics heatmap. Figure S3. SVM metrics heatmap.