Improving Measures of Chemical Structural Similarity Using Machine Learning on Chemical-Genetic Interactions

Area of Science:

• Computational chemistry
• Cheminformatics
• Systems biology

Background:

• Structural similarity is often used to predict biological activity.
• Benchmarking molecular fingerprints and similarity coefficients is crucial but limited by data.
• Chemical-genetic interaction data offers a proxy for biological similarity.

Purpose of the Study:

• To systematically benchmark molecular fingerprints and similarity coefficients.
• To identify optimal methods for predicting biological activity from chemical structures.
• To evaluate the utility of chemical-genetic data for this purpose.

Main Methods:

• Systematic benchmarking of 11 molecular fingerprint encodings and 13 similarity coefficients.
• Utilized large-scale chemical-genetic interaction data from yeast (Saccharomyces cerevisiae).
• Developed a machine learning pipeline using support vector machines.

Main Results:

• Performance varied significantly across different fingerprint and coefficient combinations.
• All-shortest path fingerprints paired with Braun-Blanquet similarity showed superior, robust performance.
• The machine learning pipeline achieved a fivefold improvement over unsupervised methods.

Conclusions:

• Chemical-genetic data is a powerful basis for refining molecular fingerprints.
• Optimized fingerprinting and similarity methods enhance prediction of biological functions.
• This approach improves the identification of bioactive molecules.