Multimodal Information-Driven Heterogeneous Graph Neural Networks for Protein-Ligand Binding Affinity Prediction

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Accurate prediction of protein-ligand binding affinity (PLA) is crucial for effective drug screening.
• Existing methods struggle to adequately model the multimodal interactions influencing binding affinity.

Purpose of the Study:

• Introduce AtomBind, a multimodal information-driven heterogeneous graph neural network framework.
• Improve the accuracy and consistency of protein-ligand binding affinity prediction.

Main Methods:

• Construct an atomic-level heterogeneous graph integrating sequence, 3D structure, and language model representations.
• Employ Intra Encoder (variational graph autoencoder, equivariant graph neural network) for intra-molecular analysis.
• Utilize Inter Encoder (graph diffusion convolution, graph Transformer) for inter-molecular interaction modeling.

Main Results:

• AtomBind demonstrated superior predictive consistency and reduced errors on test sets compared to existing models.
• Ablation studies confirmed the framework's efficiency and robustness in multimodal data integration.
• Analysis validated the effectiveness of pretrained language models within the AtomBind framework.

Conclusions:

• AtomBind offers a robust and efficient method for multimodal protein-ligand binding affinity prediction.
• The model shows strong generalization capabilities and potential for practical applications in drug discovery.
• AtomBind facilitates deeper analysis of protein pockets, intermolecular interactions, and model interpretability.