Quantitative structure-activity relationship analysis using deep learning based on a novel molecular image input technique

Quantitative structure-activity relationship (QSAR) analysis uses structural, quantum chemical, and physicochemical features calculated from molecular geometry as explanatory variables predicting physiological activity. Recently, deep learning based on advanced artificial neural networks has demonstrated excellent performance in the discipline of QSAR research. While it has properties of feature representation learning that directly calculate feature values from molecular structure, the use of this potential function is limited in QSAR modeling. The present study applied this function of feature representation learning to QSAR analysis by incorporating 360° images of molecular conformations into deep learning. Accordingly, I successfully constructed a highly versatile identification model for chemical compounds that induce mitochondrial membrane potential disruption with the external validation area under the receiver operating characteristic curve of ≥0.9.