Prediction of Inhibitory Activity against the MATE1 Transporter via Combined Fingerprint- and Physics-Based Machine...

Area of Science:

• Computational chemistry and drug discovery
• Pharmacology and toxicology

Background:

• Renal secretion is vital for drug excretion, with MATE1 and OCT2 transporters playing key roles.
• MATE1 transporter activity is closely linked to drug-drug interactions, necessitating accurate predictive models.
• Previous in silico models for MATE inhibitors had limited predictive power (ROC-AUC 0.78).

Purpose of the Study:

• To enhance the prediction accuracy of MATE1 inhibitors using a combined computational approach.
• To develop a robust model for identifying potential MATE1 inhibitors in early drug discovery.

Main Methods:

• Collected 225 compounds with known pIC50 values against MATE1.
• Employed a physics-based approach using Alpha-Fold protein structure and MM-GB/SA scoring.
• Developed Random Forest (RF) and message passing neural network models using ECFP4 fingerprints and graph representations, incorporating MM-GB/SA scores.

Main Results:

• The RF model, integrating ECFP4 and MM-GB/SA data, achieved the highest predictivity with an ROC-AUC of 0.833 ± 0.036 on the test set.
• The MM-GB/SA regression model showed a ROC-AUC of 0.742, with performance independent of the chemical space.
• Structural analysis revealed conserved interacting residues between inhibitors and substrates, suggesting similar binding modes.

Conclusions:

• Highly predictive classification models for MATE1 inhibitory activity were successfully developed.
• The combination of ECFP4 fingerprints and MM-GB/SA scores significantly improves prediction accuracy.
• Identified conserved binding interactions simplify experimental screening by reducing the need for diverse substrates.