Prediction of Inhibitory Activity against the MATE1 Transporter via Combined Fingerprint- and Physics-Based Machine Learning Models
The multidrug and toxin extrusion transporter (MATE1) is located in the renal epithelial cells and mediates efflux of organic cations, like metformin. It is understood that clinical inhibitors of organic cation transporters may also inhibit MATE1, which causes concern for potential drug-drug interactions. Therefore, being able to predict MATE1 inhibition at early stages in drug discovery could help guide medicinal chemists away from problem molecules (or at least flag those that should be tested in MATE1 before further development).
This article by Handa K. et. al. details a study that built various MATE1 predictive models using both extended-connectivity fingerprints with radius 4 (ECFP4) and MM-GB/SA scores as descriptors in Random Forest (RF) and Message Passing Neural Networks (MPNN), comparing the usefulness of the different models. Interestingly, the RF models performed better in terms of the ROC-AUC, accuracy, precision and specificity (but worse in sensitivity) than the MPNN models. The difference between the performance of the RF model built using only the ECFP4 descriptors vs that built using the ECFP4 descriptors + the MM-GB/SA descriptors was negligible; however, it was interesting that the MM-GB/SA descriptor in the combined model was identified as the most important feature. Also, although a MM-GB/SA score regression model suffered from low predictivity, it was the only model where the predictivity did not improve for test compounds of greater Tanimoto similarity to the training compounds, indicating that the model was independent of similarity to the training set. The authors were also able to take structural observations from the binding poses created from the MM-GB/SA simulations.
Overall, the paper shows that a combination of computational chemistry approaches are useful in MATE1 inhibitory prediction and that, although the predictivity of MM-GB/SA scores remains low, the inclusion of physics-based parameters is still desirable to expand the applicability domain and to guide our structural understanding of these systems.
Handa, K. et. al. Prediction of Inhibitory Activity against the MATE1 Transporter via Combined Fingerprint- and Physics-Based Machine Learning Models. JCIM. 2024. 64(18). 7068-7076.