HAC-Net

HAC-Net predicts protein-ligand binding affinities using a hybrid attention-based neural network that combines 3-dimensional convolutional representations and graph convolutional feature aggregation to improve structure-based biomolecular property prediction.

Key Features:

• Novel Architecture: Integrates a 3-dimensional convolutional neural network (CNN) with channel-wise attention and two graph convolutional networks (GCNs) that use attention-based aggregation of node features.
• Benchmark Performance: Achieves state-of-the-art results reported on the PDBbind v.2016 core set for binding affinity prediction.
• Generalizability Assessment: Uses multiple train-test splits designed to maximize differences in protein structures, sequences, or ligand extended-connectivity fingerprints to evaluate generalization.
• Cross-Validation and Robustness Testing: Employs 10-fold cross-validation with a similarity cutoff applied to SMILES strings of ligands and includes evaluation on lower-quality data.
• Broad Applicability: Applicable to supervised learning problems for structure-based biomolecular property prediction beyond binding affinity.

Scientific Applications:

• Protein-Ligand Binding Affinity Prediction: Enables prediction of binding affinities for protein-ligand complexes to support rational drug design and protein engineering studies.
• Structure-Based Property Prediction: Supports other supervised structure-based biomolecular property prediction tasks that require combined spatial and graph representations.

Methodology:

HAC-Net combines a 3-dimensional CNN with channel-wise attention and two GCNs using attention-based aggregation of node features; it was trained and evaluated on the PDBbind v.2016 core set with multiple train-test splits that maximize differences in protein structures, sequences, or ligand extended-connectivity fingerprints, assessed via 10-fold cross-validation with a SMILES similarity cutoff, and tested on lower-quality data.