Identifying Human Cytochrome P450 Inhibitors Using Multimodal Data and Transformer Encoder

In drug design, accurately predicting compound metabolic properties is vital for assessing drug candidates’ safety and efficacy. In this study, we introduce a novel framework, Multimodal Transformer Encoder Network (MTE-Net), which integrates and simultaneously learns information from molecular fingerprints and descriptors. The architecture of MTE-Net leverages the Transformer Encoder to model complex relationships between features, with the multi-head self-attention mechanism enabing the model to effectively capture long-range interactions and global dependencies. To evaluate the MTE-Net model, we conduct experiments on a large dataset containing more than 13,000 compounds extracted from the PubChem BioAssay Database. The balanced accuracy (BA), matthews correlation coefficient (MCC), precision (PRE), recall (REC) and F1-Score were decisive metrics used for model evaluation. The proposed model demonstrates competitive performance as indicated by MCC ranging from 0.5507 to 0.7030, and BA ranging from 0.7854 to 0.8515. Hence, MTE-Net holds the potential to become a dependable tool for identifying CYP450 inhibitors from non-inhibitors and expeditiously identifying compounds with potential side effects during the early stages of drug development.