Generic Context-Aware Group Contributions.

Area of Science:

• Computational Chemistry
• Quantitative Structure-Property Relationships (QSPR)

Background:

• Molecular properties are often predictable from structural features using regression models.
• Existing group contribution methods are limited to specific compound classes and require manual feature engineering.
• These limitations hinder the application of predictive models in exploring vast chemical spaces, especially with generative approaches.

Purpose of the Study:

• To develop a generic and automated group contribution method for molecular property prediction.
• To overcome the limitations of traditional, manually curated group contribution approaches.
• To enable efficient exploration of large chemical spaces using predictive modeling.

Main Methods:

• Proposed a generic group contribution method named Context AwaRe Group cOntribution (CARGO).
• Employed LASSO regression to iteratively identify significant structural regressors of increasing size.
• Anchored context-dependent contributions around a reference edge to mitigate ambiguities and prevent overcounting.

Main Results:

• The CARGO method demonstrated stable results on artificial and chemical thermodynamics datasets.
• Achieved prediction accuracies comparable to established regression techniques.
• Generated interpretable additive contributions for chemical bonds and context-dependent correction terms.

Conclusions:

• CARGO offers a versatile and automated approach to group contribution modeling.
• The method enhances the ability to predict molecular properties across diverse chemical structures.
• CARGO facilitates more effective exploration of chemical spaces for property prediction and molecular design.