Similarity maps - a visualization strategy for molecular fingerprints and machine-learning methods.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Molecular fingerprint similarity is widely used but can lack transparency, especially with advanced methods.
• Scaffold hopping and machine learning models often obscure the basis of similarity calculations.
• Interpreting the 'why' behind molecular similarity is crucial for chemical intuition and model validation.

Purpose of the Study:

• To introduce similarity maps as a general strategy for visualizing atomic contributions to molecular similarity.
• To provide a transparent method for understanding similarity derived from both fingerprint-based and machine learning approaches.
• To enable better interpretation of structure-activity relationships and model predictions.

Main Methods:

• Development of a novel visualization technique called 'similarity maps'.
• Application of similarity maps to atom-pair and circular molecular fingerprints.
• Integration of similarity maps with machine learning models, including random forests and naive Bayes.
• Utilizing a dataset of dopamine D3 receptor ligands for demonstration.

Main Results:

• Similarity maps successfully visualize atomic contributions to similarity scores between molecules.
• The method provides interpretable insights into predictions made by machine learning models.
• Demonstrated effectiveness across different fingerprint types and machine learning algorithms.
• An open-source implementation facilitates broader adoption and application.

Conclusions:

• Similarity maps offer a powerful and generalizable tool for enhancing the interpretability of molecular similarity.
• This approach bridges the gap between intuitive similarity measures and complex machine learning predictions.
• Facilitates deeper understanding in drug discovery, particularly for lead optimization and scaffold hopping.