Comparing Multiple Machine Learning Algorithms and Metrics for Estrogen Receptor Binding Prediction.

Area of Science:

• Environmental Chemistry
• Computational Toxicology
• Biochemistry

Background:

• Endocrine-disrupting chemicals (EDCs) pose health risks, but in vitro/in vivo screening is expensive and slow.
• Quantitative structure-activity relationship (QSAR) models and machine learning (ML) offer faster, cost-effective alternatives.
• ML models are increasingly reliable for predicting chemical properties like estrogen receptor (ER) binding.

Purpose of the Study:

• To exhaustively compare various ML algorithms, chemical feature sets, and evaluation metrics for predicting ER binding.
• To assess the performance of classic ML algorithms versus Deep Neural Networks (DNNs) for ER binding prediction.
• To identify the most effective computational approaches for prioritizing chemicals for EDC testing.

Main Methods:

• Utilized public datasets curated via cheminformatics software (Assay Central).
• Employed diverse chemical features: binary fingerprints (ECFP6, FCFP6, ToxPrint, MACCS) and RDKit descriptors.
• Applied classic ML algorithms (Naive Bayes, AdaBoost, Random Forest, SVM) and DNNs, evaluating with multiple metrics (accuracy, AUC, F1-score, etc.).

Main Results:

• DNNs showed higher accuracy on the training set.
• DNNs and classic ML models performed similarly in cross-validation and external testing.
• Random Forest demonstrated superior performance on the validation set when ranked by metric and dataset.

Conclusions:

• Classic ML algorithms are sufficient for developing high-quality predictive models of ER activity.
• The choice of chemical features and specific ML algorithm may have less impact than anticipated for ER binding prediction.
• Computational methods provide a viable strategy for screening and prioritizing chemicals for endocrine disruption potential.