Count-Based Morgan Fingerprint: A More Efficient and Interpretable Molecular Representation in Developing Machine...

Area of Science:

• * Cheminformatics and Computational Toxicology
• * Machine Learning in Chemistry

Background:

• * Traditional binary Morgan fingerprints (B-MF) represent the presence/absence of chemical substructures.
• * Limitations exist in B-MF's ability to capture quantitative information about substructure occurrences.
• * Accurate prediction of contaminant activity and properties is crucial for environmental and health risk assessment.

Purpose of the Study:

• * To introduce and evaluate the count-based Morgan fingerprint (C-MF) for representing chemical structures.
• * To develop and compare machine learning (ML) models using C-MF versus B-MF for predicting contaminant activities and properties.
• * To assess the impact of C-MF on model performance, interpretability, and applicability domain (AD).

Main Methods:

• * Development of C-MF to quantify atom group counts within chemical structures.
• * Application of six ML algorithms (ridge regression, SVM, KNN, RF, XGBoost, CatBoost) using both C-MF and B-MF.
• * Evaluation of models across 10 contaminant-related datasets, comparing predictive performance, interpretation (SHAP values), and AD.
• * Creation of the ContaminaNET platform for model deployment.

Main Results:

• * C-MF significantly outperformed B-MF in predictive performance on nine out of 10 datasets.
• * Performance gains were dependent on the ML algorithm and chemical diversity differences between fingerprints.
• * C-MF-based models provided enhanced interpretability by elucidating the impact of atom group counts.
• * C-MF-based models demonstrated an applicability domain comparable to B-MF-based models.

Conclusions:

• * C-MF offers a superior representation for chemical structures in ML tasks compared to B-MF.
• * The enhanced interpretability of C-MF aids in understanding structure-activity relationships.
• * C-MF-based models are robust and maintain a reliable applicability domain.
• * The ContaminaNET platform provides accessible C-MF-based predictive tools for contaminants.