Enhancing Transthyretin Binding Affinity Prediction with a Consensus Model: Insights from the Tox24 Challenge

Area of Science:

• Computational chemistry
• Toxicology
• Molecular modeling

Background:

• Transthyretin (TTR) is crucial for thyroid hormone transport and homeostasis.
• Exogenous compounds interacting with TTR can disrupt endocrine function and cause toxicity.

Purpose of the Study:

• To develop a deep learning-based consensus model for predicting TTR binding affinity.
• To evaluate the model's performance in the Tox24 challenge and assess its utility for identifying potential TTR binders.

Main Methods:

• Integrated three deep learning models (sPhysNet, KANO, GGAP-CPI) utilizing 2D, 3D, and protein-ligand interaction data.
• Developed a consensus model for enhanced predictive accuracy of TTR binding affinity.
• Utilized the standard deviation of ensemble outputs as an uncertainty estimate for predictions.

Main Results:

• The consensus model achieved an RMSE of 20.8 on the blind test set, ranking fifth in the Tox24 challenge.
• Incorporating additional data reduced the RMSE to 20.6 in a retrospective study.
• Prediction error and RMSE increased with model uncertainty, validating uncertainty as a confidence measure.

Conclusions:

• Combining regression models across different modalities significantly improves predictive accuracy for TTR binding affinity.
• The developed consensus model is a valuable tool for in silico prediction of TTR binders and their affinities.
• Uncertainty estimation provides a reliable measure of prediction confidence, aiding in risk assessment.