Distance-Aware Molecular Property Prediction in Nonlinear Structure-Property Space

Area of Science:

• Computational chemistry
• Machine learning
• cheminformatics

Background:

• Predicting molecular properties with limited data in new chemical domains is a significant challenge.
• Existing methods often struggle with data scarcity and domain specificity.

Purpose of the Study:

• To develop a novel framework for molecular property prediction that addresses data limitations in novel chemical domains.
• To improve prediction accuracy by quantifying uncertainty based on domain relevance and molecular similarity.

Main Methods:

• Developed a nonlinear structure-property space embedding to link molecular similarity with prediction difficulty.
• Implemented distance-aware domain classification to balance precision and true positive rate.
• Incorporated distance-based uncertainty quantification scaled by molecular similarity.

Main Results:

• Local models reduced root mean squared error by 28-48% for in-domain molecules compared to global models across four ecotoxicity datasets.
• Demonstrated strong correlations (r = 0.40-0.62) between distance in the structure-property space and prediction error.
• Achieved a 29% reduction in prediction error for a biolubricant base oil property application, outperforming transfer learning and standard machine learning.

Conclusions:

• The proposed framework effectively handles limited and heterogeneous chemical data by focusing on relevant domains.
• Distance-calibrated uncertainty estimates provide reliable predictions for novel chemical entities.
• The approach is broadly applicable to various fields, including toxicity prediction, drug discovery, and materials design.