Machine Learning-Based Prediction of Transthyretin-Binding Activity and Its Application to Screening Food-Derived Compounds for ATTR Amyloidosis

Background/Objectives: Transthyretin amyloidosis (ATTR) is a progressive disease caused by the dissociation of the transthyretin (TTR) tetramer, leading to amyloid fibril formation. Although pharmacological stabilizers have been developed, preventive strategies for wild-type ATTR (ATTRwt) have not been established. This study developed a computational model to predict TTR-binding activity from chemical structures and applied it to the exploratory prioritization of food-derived compounds. Methods: A machine learning model was constructed using TTR–8-anilino-1-naphthalenesulfonic acid displacement assay data from the Tox24 Challenge. An integrated dataset compiled from multiple literature sources was used to compare predicted TTR-binding activity with amyloid fibril formation inhibition, and the trained model was applied to the PhytoHub database. Results: The model achieved a root mean square error of 21.34 and a mean absolute error of 15.9 percentage points on an external test dataset, a performance comparable to that of the top-scoring models in the Tox24 Challenge. The predicted TTR-binding activity showed positive correlations with inhibition of amyloid fibril formation, Pearson’s r = 0.673 and Spearman’s ρ = 0.602. The model prioritized 39 compounds with high predicted binding activity, predominantly polyphenols. Conclusions: These results suggest that the proposed in silico method may be useful for the exploratory prioritization of food-derived compounds with potential relevance to TTR stabilization.