Prediction of Ligand Binding to Transthyretin Using Machine Learning Algorithms and Low-Dimensional Molecular...

Area of Science:

• Computational toxicology and cheminformatics.
• Development of predictive models for chemical safety assessment.

Background:

• Transthyretin (TTR) is a critical serum transport protein; its disruption by environmental chemicals can cause endocrine system dysregulation.
• Predicting TTR binding computationally is challenging due to limited data and experimental variability.
• Robust predictive models are needed for broad chemical space analysis.

Purpose of the Study:

• To develop and validate a machine learning approach for predicting the percent displacement of ANSA from human TTR.
• To create a robust, computationally efficient model applicable to diverse chemical compounds.
• To facilitate high-throughput toxicity screening for various applications.

Main Methods:

• Utilized computationally efficient, low-dimensional (0D-2D) molecular descriptors and fingerprints, avoiding 3D conformational analysis.
• Benchmarked individual machine learning methods and optimized hyperparameters using Optuna for CatBoost and XGBoost.
• Developed a consensus model integrating complementary learning strategies from gradient boosting algorithms.

Main Results:

• The consensus model achieved an RMSE of 21.60 on the blind test set, ranking 15th out of 79 teams in the Tox24 Challenge.
• Feature importance analysis highlighted distinct learning strategies between CatBoost and XGBoost.
• Chemical space analysis confirmed test compounds were within the training data distribution, with minimal outliers.

Conclusions:

• Low-dimensional molecular descriptors combined with optimized consensus machine learning methods yield competitive predictive performance.
• The developed pipeline offers a practical approach for high-throughput toxicity screening.
• A freely accessible web server was created to enable rapid toxicity prediction for Tox21 and Tox24 endpoints.