Distance-Aware Molecular Property Prediction in Nonlinear Structure-Property Space.
Journal:
Journal of chemical information and modeling
Published Date:
Jul 14, 2025
Abstract
Molecular property prediction with limited data in novel chemical domains remains challenging. We introduce an approach based on the hypothesis that prediction difficulty increases systematically with distance from well-characterized regions in an appropriately defined structure-property space. Our framework combines nonlinear structure-property space embedding with distance-aware domain classification and uncertainty quantification. We create a structure-property embedding connecting molecular similarity with prediction difficulty, implement distance-aware classification balancing precision and true positive rate, and provide distance-based uncertainty estimates scaled by molecular similarity. Across four ecotoxicity data sets, our local models reduced root mean squared error by 28-48% for truly in-domain molecules compared to global models, with strong correlations ( = 0.40-0.62) between distance and prediction error. In a biolubricant base oil property application, our approach reduced prediction error by 29% compared to a global model and outperformed transfer learning and standard machine learning approaches. This framework's focus on relevant domains and distance-calibrated uncertainty estimates for limited, heterogeneous chemical data makes it broadly applicable across applications, such as toxicity prediction, drug discovery, and materials design.
