Can deep learning algorithms enhance the prediction of solute descriptors for linear solvation energy relationship approaches?

Abstract

Experimental solute descriptors for about 8,000 chemicals are currently available to apply physicochemical property predictions based on linear solvation energy relationship (LSER) models. The solute descriptors can be predicted by fragmental-based quantitative structure-property relationship (QSPR) models. However, the predictions are problematic for larger chemical structures, including multiple functional groups. We developed deep neural networks (DNNs) as alternative prediction models based on graph representations of the chemicals. The root mean square errors rmses range between 0.11 and 0.46 for the different solute descriptors. The predictions of the solute descriptors were compared to predictions from the QSPR of LSERD (an online database) and ACD/Absolv (a commercial software). We further investigated the predictive power of all tools based on three different datasets of experimentally determined partition coefficients, namely the octanol-water partition coefficient (Kow), the octanol-air partition coefficient (Koa), and the water-air partition coefficient (Kwa). Additionally, we used two different sets of retention data for GC and LC to evaluate the results of all prediction tools. All prediction tools perform comparably well with rmses of ∼ 1.0 log unit for the Kow dataset (12,010 chemicals) and ∼ 1.3 log units for the Kwa dataset (696 chemicals), for example. Nevertheless, larger chemical structures are predicted poorly by each approach. We recommend to use the novel DNN model as a complementary prediction tool.