One class classification as a practical approach for accelerating π-π co-crystal discovery

Area of Science:

• Computational Materials Science
• Machine Learning in Chemistry
• Crystallography

Background:

• Machine learning models are transforming materials design, but data-driven approaches struggle with imbalanced datasets due to a lack of negative synthetic data.
• Unsuccessful synthesis attempts often go unreported, creating biases in datasets used for predictive modeling in materials discovery.

Purpose of the Study:

• To apply one-class classification methodology to overcome limitations posed by imbalanced datasets in materials design.
• To identify an effective workflow for discovering emerging materials, specifically weakly bound polyaromatic hydrocarbon co-crystals, which represent a small, well-defined class.
• To accelerate materials discovery by reducing the search space and guiding synthetic efforts.

Main Methods:

• An extensive study of various one-class classification algorithms was conducted to identify the most suitable workflow.
• A two-step approach was developed: training on known co-crystal data from the Cambridge Structural Database and scoring potential new pairs from the ZINC15 database.
• Interpretability techniques were employed to understand molecular properties driving co-crystallization.

Main Results:

• The one-class classification approach effectively handles imbalanced datasets in materials design.
• The workflow successfully identified high-ranking potential co-crystal pairs from a vast chemical space.
• Two novel co-crystals, pyrene-6H-benzo[c]chromen-6-one (1) and pyrene-9,10-dicyanoanthracene (2), were discovered and validated.

Conclusions:

• One-class classification is a powerful tool for discovering materials within small, specific classes, even with limited negative data.
• The developed methodology significantly accelerates materials discovery by efficiently screening potential candidates.
• Understanding the molecular drivers of co-crystallization enhances the predictive power and applicability of machine learning in chemistry.