Fused framework for predicting binding affinity to ɑ-synuclein using three-dimensional molecular graphs and attention

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Aberrant alpha-synuclein (α-syn) aggregation is central to synucleinopathies.
• Effective small-molecule inhibitors are needed to block α-syn aggregation.
• Current deep learning models lack 3D stereochemical and geometric insights.

Purpose of the Study:

• To develop a high-precision deep learning predictor for α-syn binding affinity.
• To leverage 3D structural information for improved molecular representation.
• To provide a robust computational tool for screening α-syn inhibitors.

Main Methods:

• Compiled a dataset of 9628 α-syn-binding compounds from ChEMBL.
• Developed a deep learning framework integrating Graph Attention Networks (GAT) and Graph Convolutional Networks (GCN).
• Utilized 3D Euclidean interatomic distances as edge features and incorporated an atomic attention layer with a composite loss function (MSE + Pearson correlation).

Main Results:

• Achieved an average MSE of 0.1751 and Pearson correlation of 0.5703.
• Outperformed baseline Graph Neural Networks (GNNs) and a previous α-syn-specific predictor.
• Demonstrated robustness and generalization to novel chemical scaffolds via bootstrap calibration and cluster-exclusion split.

Conclusions:

• The GAT-GCN framework offers an accurate, robust, and interpretable tool for α-syn inhibitor screening.
• This work advances 3D structure-based molecular representation learning in drug discovery.
• The developed model aids in identifying potential therapeutics for synucleinopathies.