Chemical Topic Modeling: Exploring Molecular Data Sets Using a Common Text-Mining Approach.

Area of Science:

• Medicinal Chemistry
• Cheminformatics
• Data Science

Background:

• Big data presents significant challenges in organizing and searching vast molecular datasets generated by modern technologies.
• Existing methods for handling large molecule sets often compromise result interpretability.
• Medicinal chemistry faces increasing data volumes from DNA encoded libraries, peptide libraries, text mining, and in silico methods.

Purpose of the Study:

• To develop an intuitive and interpretable method for organizing large molecular datasets.
• To implement and evaluate a probabilistic topic modeling framework for chemical data.
• To enable the identification and retrieval of chemical series and relationships within large molecule sets.

Main Methods:

• Adoption and adaptation of the probabilistic topic modeling framework from text-mining for chemical data.
• Development of the CheTo (Chemical Topic) open-source implementation.
• Evaluation of the method's performance on various experiments and the ChEMBL22 dataset (1.6 million molecules).

Main Results:

• Successful assignment of large molecule sets to meaningful 'chemical topics' (e.g., 'proteins', 'DNA', 'steroids').
• Demonstrated ability to reproduce human-assigned concepts and retrieve specific chemical series.
• Creation of an intuitive visualization for chemical topics, outperforming traditional clustering methods.
• Efficient modeling of 1.6 million molecules into 100 topics within approximately one hour.

Conclusions:

• Topic modeling offers a powerful and interpretable approach for managing large chemical datasets.
• CheTo provides a robust and efficient solution for organizing and exploring molecular data in medicinal chemistry.
• The open-source implementation (CheTo) and datasets facilitate further research and application in cheminformatics.