Kernel approach to molecular similarity based on iterative graph similarity

Area of Science:

• Computational chemistry
• Cheminformatics
• Bioinformatics

Background:

• Molecular similarity is crucial for drug discovery and toxicology.
• Existing methods may not fully capture complex molecular relationships.
• Graph-based representations offer a powerful way to model molecules.

Purpose of the Study:

• To introduce a novel molecular similarity measure based on annotated molecular graphs.
• To develop an efficient iterative algorithm for computing this similarity.
• To validate the measure's performance in machine learning tasks.

Main Methods:

• Iterative graph similarity algorithm for molecular comparison.
• Optimal assignment strategies for graph matching.
• Support Vector Machine (SVM) classification and regression for evaluation.
• Application to pharmaceutical and toxicological datasets.

Main Results:

• The proposed iterative algorithm converges and yields a unique solution.
• An upper bound for iterations ensures desired precision.
• Empirical evidence supports the positive semidefiniteness of the similarity function.
• Promising results when used as a kernel in SVM for classification and regression.

Conclusions:

• The novel molecular similarity measure is computationally sound and effective.
• It demonstrates potential for improving predictive accuracy in pharmaceutical and toxicological studies.
• This graph-based approach offers a valuable tool for cheminformatics and drug design.