Deep Supramolecular Language Processing for Co-Crystal Prediction.

Area of Science:

• Pharmaceutical Sciences
• Computational Chemistry
• Drug Discovery

Background:

• Suboptimal pharmacokinetic profiles affect approximately 40% of marketed drugs.
• Co-crystallization enhances drug physicochemical properties without impacting pharmacological activity.
• Identifying suitable co-crystal pairs is challenging due to the vast number of molecular combinations.

Purpose of the Study:

• To develop a novel deep learning approach, DeepCocrystal, for predicting co-crystal formation.
• To process chemical information from a supramolecular perspective using chemical language processing.
• To accelerate the discovery of new co-crystals for improved drug development.

Main Methods:

• Developed DeepCocrystal, a deep learning model utilizing molecular string representations.
• Trained and validated the model on predicting co-crystal formation.
• Employed explainable AI to understand the model's decision-making process.
• Integrated uncertainty estimation into the prediction framework.

Main Results:

• DeepCocrystal achieved a balanced accuracy of 78% in realistic prediction scenarios, outperforming existing models.
• Explainable AI confirmed the model learns chemically relevant supramolecular features.
• The model successfully identified two novel co-crystals of diflunisal in a prospective study.
• Uncertainty estimation guided the prospective discovery process.

Conclusions:

• DeepCocrystal effectively predicts co-crystal formation, accelerating the identification of promising drug candidates.
• Deep learning and chemical language processing offer powerful tools for pharmaceutical research.
• The developed model and its web application can benefit both academic and industrial drug development efforts.