Predictionof Solute Descriptors for Linear SolvationEnergy Relationships Using K-Nearest Neighbors, Group Contributions,and Graph-Convolutional Neural Networks

Abstract Linear solvation energy relationship (LSER) models are nowadays often used to predict physicochemical properties of chemicals, such as partition coefficients, retention factors in chromatography, and solubilities. Due to their mechanistic foundation and transferability across phases, LSER models are particularly valuable for predicting partition coefficients in data-poor systems where experimental data sets are not available. However, their broader applicability is currently constrained by the limited availability of experimentally determined solute descriptors. We developed three different models, based on three different approaches, to predict the solute descriptors S, E, A, B, and L for LSER applications: (1) a group contribution model that includes an initial screening algorithm to identify functional groups as structural patterns, (2) a k-nearest neighbors model, and (3) a graph-convolutional neural network. All models were developed on the same curated data set for each solute descriptor. An independent test set was used to evaluate the overall model performance, with rmse values ranging from 0.08–0.13 for A, 0.10–0.15 for B, 0.17–0.23 for S, 0.09–0.19 for E, and 0.25–0.45 for L across the three approaches. By enabling the prediction of these descriptors directly from molecular structure, our modeling framework addresses a key bottleneck that has so far limited the scalability and broader application of the LSER models. In addition, we investigated whether a consensus approach would enhance overall prediction quality. Additionally, the predicted solute descriptors were directly used to derive environmentally and analytically relevant partition coefficients, demonstrating that reliable LSER-based property predictions are feasible even for chemicals lacking experimental descriptor data or large property-specific training sets. By enabling descriptor generation at scale, this work improves the practical applicability of LSER modeling and strengthens its role as a transferable and data-efficient tool for environmental fate and risk assessment.