Compressing Chemistry Reveals Functional Groups

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Traditional chemical functional groups are widely used but their explanatory utility lacks formal assessment.
• Computational learning theory suggests data compression can reveal good explanations.

Purpose of the Study:

• To formally assess the utility of traditional chemical functional groups.
• To develop an unsupervised learning algorithm for discovering explanatory molecular substructures.
• To improve bioactivity prediction models.

Main Methods:

• An unsupervised learning algorithm based on the Minimum Message Length (MML) principle was developed.
• The algorithm searched for substructures that compress approximately three million biologically relevant molecules.
• The algorithm was applied to 24 bioactivity prediction datasets to discover dataset-specific functional groups.

Main Results:

• Discovered substructures include known functional groups and novel, larger patterns with specific functions.
• Dataset-specific functional groups were identified.
• Fingerprints from these groups significantly outperformed standard representations (e.g., MACCS, Morgan) in bioactivity regression tasks.

Conclusions:

• The MML-based approach effectively identifies meaningful chemical substructures.
• Data-driven functional groups offer superior performance in predictive modeling for bioactivity.
• This method provides a powerful tool for chemical explanation and drug discovery.