Constant size descriptors for accurate machine learning models of molecular properties

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine learning models require effective molecular representations to predict thermodynamic and electronic properties.
• Existing methods often rely solely on molecular connectivity (graph-based) or 3D structure.

Purpose of the Study:

• To evaluate two classes of molecular representations for machine learning applications.
• To compare graph-based and 3D structure-based representations for predicting molecular properties.
• To introduce and assess the performance of Encoded Bonds features.

Main Methods:

• Utilized linear and kernel ridge regression models.
• Evaluated graph-based representations (e.g., bonding patterns) and 3D structure-based representations (e.g., Coulomb matrix, Bag of Bonds, Encoded Bonds).
• Constructed feature sets by combining graph and geometry-based features at different ranks.

Main Results:

• Graph-based features achieved a mean absolute error of 3.4 kcal/mol for atomization energies (QM7 dataset).
• 3D structure-based Encoded Bonds alone yielded 2.4 kcal/mol.
• Combining Encoded Bonds with graph features reduced the error to 1.19 kcal/mol, demonstrating superior performance.

Conclusions:

• Three-dimensional molecular structure information significantly enhances machine learning model performance for thermodynamic and electronic properties.
• Encoded Bonds offer a size-independent feature vector, enabling successful training on smaller molecules for prediction on larger ones.
• Hybrid approaches combining graph and 3D features provide the most accurate predictions.