The proposed machine learning based approach could predict the Log (IC50) values of small molecule Protein-Protein Interaction Modulators (PPIMs).

Performance Measurement

Pearson Correlation Coefficient (PCC) and Spearman’s Rank Correlation (SRC) measurement have been used for evalution: 5-fold cross validation training/testing and blind sets to any types of PPIs and two specific PPIs:Mdm2/P53 and Bcl2/Bak. The SVM RBF, polynomial and linear kernel with different parameters including regularization parameter and error parameters were used for building up regression models.

Datasets

Target PPI Training/Testing dataset Blind validation dataset
Large PPIs dataset 1401 (Training/Testing & Blind)
Mdm2/P53 450 64
Bcl2/Bak 450 107

Results and brief description

We compiled 1401 of small chemical for 8 different target PPIs like Mdm2/P53, Bcl2/Bak, c-Myc/Max, BRD2/AcK, Cyclophilin, MLL, Rad51, and TNFa. We developed SVM RBF models using of 1200 small chemicals for any type of PPIs in 5-fold cross-validation technique. The SVM RBF model achieved better Pearson correlation coefficient (PCC) value of 0.73 than specific models for Mdm2/P53 and Bcl2/Bak. The best model was also evaluated on the blind dataset that contains random 201 small chemicals and we found better Pearson correlation coefficient values of 0.63. In summary, the model can easily predict the IC50 values of small chemicals targeting any PPIs. The model was not specific to any particular PPIs.

However, in case of specific Mdm2/P53 dataset, we found that the best SVM RBF model for Mdm2/P53 get 0.58 as Max. Pearson correlation coefficient (PCC)

Figure 1. The density plot of experimentally verified IC50 values and predicted IC50 values of Mdm2/data set using SVM RBF kernel

Figure 2. The regression plot of Mdm2/data set using SVM RBF kernel. The R-square value is 0.33.

Similarly, in case of specific Bcl2/Bak dataset, we found the best SVM RBF model for Bcl2/Bak get 0.54 as Max Pearson correlation coefficient (PCC).

Figure 3. The density plot of experimentally verified IC50 values and predicted IC50 values of Bcl2/Bak data-set using SVM RBF kernel

Figure 4. The regression plot of Bcl2/Bak data-set using SVM RBF kernel. The R-square value is 0.29.

The small chemicals PPIMs which are already pursued the clinically trials

Clinical verified Example Target PPI Experimental Log 10(IC50) value (nM) Predicted Log 10(IC50) value (nM) Prediction confidence
Nutlin-3A (CID: 11433190) Any PPI 0.6127838567 0.71202931 High
Nutlin-3A (CID: 11433190) Mdm2/P53 0.6127838567 1.713178 High
Navitoclax; ABT-263 (CID: 24978538) Bcl2/Bak 0.5185139399 1.2658233 High

Result interpretation of PPIM-IC50pred server

Datasets prediction of Log 10 (IC50) values (PPIM-IC50pred)
Any PPI, Mdm2/P53 and Bcl2/Bak
  • -2.0 < Predicted Log 10 (IC50) value < 0.0 Very high
  • 0.0 < Predicted Log 10 (IC50) value < 3.0 High
  • 3.0< Predicted Log 10 (IC50) value < 4.0 Medium
  • 4.0 < Predicted Log 10 (IC50) value < 6.0 Low

© 2018 Bose Institute, Kolkata. All rights reserved

For queries please contact Dr. Sudipto Saha (ssaha4@jcbose.ac.in, ssaha4@gmail.com)

Last updated on 14th June, 2019

Best viewed in
Min. Screen resolution: 1024 x 768