Prediction of kinase inhibitors binding modes with machine learning and reduced descriptor sets

Abstract

Protein kinases are receiving wide research interest, from drug perspective, due to their important roles in human body. Available kinase-inhibitor data, including crystallized structures, revealed many details about the mechanism of inhibition and binding modes. The understanding and analysis of these binding modes are expected to support the discovery of kinase-targeting drugs. The huge amounts of data made it possible to utilize computational techniques, including machine learning, to help in the discovery of kinase-targeting drugs. Machine learning gave reasonable predictions when applied to differentiate between the binding modes of kinase inhibitors, promoting a wider application in that domain. In this study, we applied machine learning supported by feature selection techniques to classify kinase inhibitors according to their binding modes. We represented inhibitors as a large number of molecular descriptors, as features, and systematically reduced these features in a multi-step manner while trying to attain high classification accuracy. Our predictive models could satisfy both goals by achieving high accuracy while utilizing at most 5% of the modeling features. The models could differentiate between binding mode types with MCC values between 0.67 and 0.92, and balanced accuracy values between 0.78 and 0.97 for independent test sets.

Figure 1

Snapshots, obtained from KLIFS^–, for the binding site (with inhibitor) in each of the four binding modes (DFG: blue, $\alpha$C-helix: red (upper right), inhibitor: light green). (a) Type I: DJK bound to EGFR (PDB ID: 2J5F), (b) Type II: L09 bound to MAPK14 (PDB ID: 1WBN), (c) Type I$\frac{1}{2}$:EFQ bound to CDK2 (PDB ID: 3IGG), and (d) Type A: 0O7 (not shown) bound to PTK2 (PDB ID: 4EBV).

Figure 2

Testing and combining individual descriptor sets.

Figure 3

Average precision recall curves for the independent test sets in classification tasks. (a) ${\text{CT}}_{I-II}$, (b) ${\text{CT}}_{I-I\frac{1}{2}}$, (c) ${\text{CT}}_{II-I\frac{1}{2}}$, and (d) ${\text{CT}}_{A-(I+II+I\frac{1}{2})}$.

Figure 4

t-SNE plots for the features in classification task ${\text{CT}}_{II-I\frac{1}{2}}$. (a) All: 4135 features, (b) combined selected: 276 features, (c) final reduction: 50 features.

Figure 5

Data reduction and model building work flow in the proposed experimental framework.

Figure 6

Feature processing and reduction flow.