Biosynfoni: a biosynthesis-informed and interpretable lightweight molecular fingerprint.

Area of Science:

• Computational chemistry
• Cheminformatics
• Natural product drug discovery

Background:

• Natural products are vital sources of bioactive molecules.
• Current molecular fingerprints lack specificity for natural product features, hindering accurate predictions.
• Improved interpretability is needed for natural product-specific machine learning models.

Purpose of the Study:

• To develop a natural product-specific molecular fingerprint that improves prediction interpretability.
• To assess the performance of this new fingerprint against established methods.
• To enable better examination of feature importance in machine learning models for natural products.

Main Methods:

• Development of a novel molecular fingerprint, Biosynfoni, based on biosynthetic building blocks.
• Evaluation of Biosynfoni's performance in estimating biosynthetic distance and classifying natural products.
• Comparison of Biosynfoni against MACCS, Morgan, and Daylight-like fingerprints.

Main Results:

• Biosynfoni demonstrates superior performance in biosynthetic distance estimation compared to standard fingerprints.
• The Biosynfoni fingerprint achieves performance on par with top methods for natural product classification.
• Biosynfoni's design facilitates visualization of substructures and their biosynthetic origins, enhancing interpretability.

Conclusions:

• A natural product-specific fingerprint using biosynthetic building blocks offers enhanced interpretability.
• Biosynfoni provides valuable insights into natural product structures and their classification.
• This approach improves the explainability of machine learning models in natural product research.