Evaluating Molecular Similarity Measures: Do Similarity Measures Reflect Electronic Structure Properties?

Area of Science:

• Computational chemistry
• Cheminformatics
• Artificial intelligence in chemistry

Background:

• Big data, machine learning (ML), and generative artificial intelligence (AI) are increasingly used in chemical discovery.
• Molecular similarity, often measured by molecular fingerprints, is vital for database curation, diversity analysis, and property prediction.
• Current evaluations of similarity measures primarily use biological data, limiting their applicability to nonbiological domains like electronic structure properties.

Purpose of the Study:

• To develop and present a framework for evaluating the correlation between molecular similarity measures and molecular properties.
• To address the limitations of existing evaluation methods by focusing on nonbiological properties.
• To provide a publicly available dataset and evaluation framework for the chemical community.

Main Methods:

• Utilized a dataset of over 350 million molecule pairs with associated electronic structure, redox, and optical properties.
• Developed a framework based on neighborhood behavior and kernel density estimation (KDE) analysis.
• Systematically evaluated the correlation between various molecular fingerprint generators, distance functions, and molecular properties.

Main Results:

• Demonstrated the effectiveness of the proposed framework in quantifying the relationship between similarity measures and properties.
• Identified variations in how well different similarity measures capture correlations with specific molecular properties.
• Established a comprehensive evaluation of fingerprint-based similarity metrics using a large-scale, nonbiological property dataset.

Conclusions:

• The proposed framework offers a robust method for assessing molecular similarity measures beyond biological activity.
• Accurate quantification of molecular similarity is essential for reliable AI-driven chemical discovery.
• The publicly available dataset and framework will facilitate further research and development in cheminformatics and AI.