Predicting the equivalent alkane carbon number of oils using graph neural networks and quantum mechanical descriptors

Oils are key components in personal care products, typically used as emollients to moisturize the skin and prevent dryness. The hydrophobicity of oils is crucial for designing tailor-made surfactant/oil/water (SOW) systems and can be quantified by the equivalent alkane carbon number (EACN). While various quantitative structure–property relationship (QSPR) models have been proposed, they require manual selection of informative descriptors. Recently, the largest publicly available data set with 121 EACN values was curated [Delforce et al. (2022), ACS Omega]. Herein, we apply graph neural networks (GNNs) for EACN prediction and propose a descriptor-augmented graph neural network (DA-GNN), that utilizes semi-empirical quantum mechanical (QM) descriptors. We first observe that the simple GNN model achieves highly accurate predictions. Afterwards, we find that incorporating QM descriptors enhances the accuracy of standard GNNs, resulting in improved EACN predictions by the DA-GNN model. However, the accuracy remains slightly lower than that of manually derived QSPR models, highlighting the need for sufficient data when training GNNs. We also experimentally measure the EACN of two cosmetic-type oils and analyze the predictive performance of both GNN models. The DA-GNN model accurately predicts their EACN values, making it practical for future applications in designing SOW systems. • Graph neural networks for predicting the Equivalent Alkane Carbon Number of Oils. • Augmenting graph neural networks with semi-empirical quantum mechanical descriptors. • Comparative predictive performance to state-of-the-art quantitative structure-property relationship models.