Augmenting Chemical Databases for Atomistic Machine Learning by Sampling Conformational Space.

Area of Science:

• Computational chemistry
• Chemical informatics
• Machine learning applications

Background:

• Machine learning (ML) is integral to chemical space exploration.
• Model performance is highly dependent on the training database.
• Existing databases may not be optimal for specific chemical prediction tasks.

Purpose of the Study:

• To investigate the impact of database augmentation on ML model performance for chemical tasks.
• To identify optimal strategies for selecting molecules and conformations for database enhancement.
• To establish a baseline for rational chemical database creation.

Main Methods:

• Systematic augmentation of an initial restricted database (iRD) with new molecules and conformations.
• Generation of molecular conformations at varying temperatures.
• Evaluation of prediction performance using metrics like Kullback-Leibler (D_KL) and Jensen-Shannon (D_JS) divergence, and Wasserstein distance (W_1).

Main Results:

• Adding a small percentage (1%) of conformations generated at 300 K consistently improved model performance across various chemical tasks.
• Redundant molecules and highly deformed structures in the augmentation set negatively impacted prediction quality.
• Energy and bond distributions were effectively analyzed using divergence and distance metrics.

Conclusions:

• Rational augmentation of chemical databases is crucial for enhancing ML model performance.
• Careful selection of molecules and conformations, particularly including low-percentage, diverse conformations, is key to improving predictive accuracy.
• The study provides a framework for creating or augmenting synthetic chemical databases for ML applications.