# Machine Learning Prediction of Transthyretin Binding for Thyroid Hormone Transport Disruption for Chemical Risk Assessment

## TL;DR

This paper uses machine learning to predict how chemicals bind to a protein that transports thyroid hormones, helping identify potentially harmful chemicals.

## Contribution

The novel contribution is the development of a machine learning model to predict transthyretin binding affinity for endocrine-disrupting chemicals.

## Key findings

- The gradient boosting regressor model achieved an R2 of 0.89 on the training set for predicting TTR-binding affinity.
- Hydrophobicity, steric effects, and electronic properties are key factors in TTR disruption and stabilization.
- 97.5% of the test set and 96.0% of the validation set fell within the model's reliable descriptor space.

## Abstract

Endocrine-Disrupting Chemicals (EDCs) disrupt thyroid hormone (TH) synthesis, transport, metabolism, and action, thereby perturbing systemic endocrine homeostasis. Transthyretin (TTR) is a key TH transport protein that regulates circulating hormone distribution and tissue availability, particularly during critical developmental windows. Chemical interference with TTR-binding may alter TH bioavailability and represent a transport-mediated molecular initiating event within thyroid-axis perturbation. Despite widespread exposure, many thyroidal EDCs remain unidentified, and their health effects are difficult to assess due to multiple simultaneous exposures. To support endocrine hazard identification and chemical prioritization within risk assessment frameworks, we developed machine learning-based QSAR models during the Tox24 challenge, using a dataset of 1512 chemicals to predict TTR-binding affinity. Of these, 67% were used for training, 13% for testing, and 20% for validation. Molecular descriptors were selected by first removing highly correlated features and then ranking the remaining descriptors using mutual information regression. The leverage approach was applied to define the models’ applicability domain (AD). Five machine learning algorithms, including gradient boosting regressor (GBR), Random Forest, Lasso Regression, Support Vector Machine (SVM), and regularized SVM models, were developed. The GBR model demonstrated the best overall performance. This model achieved an R2 of 0.89 on the training set, 0.58 on the test set, and 0.55 on the validation set. The molecular descriptor analysis highlights hydrophobicity, steric effects, branching, connectivity, and ionization/electronic effects as the mechanistic basis for TTR disruption and stabilization, providing structural insight into features associated with thyroid hormone displacement. The AD analysis indicated that 97.5% of the test set and 96.0% of the validation set fell within the reliable descriptor space. Importantly, these predictions extend beyond model benchmarking by informing weight-of-evidence evaluations of thyroid-axis perturbation and supporting the prioritization of chemicals for targeted testing within non-animal new approach methodologies. Overall, this work highlights the application of in silico approaches for screening EDCs, supporting the prioritization and identification of potentially harmful chemicals.