HGLA: Biomolecular Interaction Prediction Based on Mixed High-Order Graph Convolution With Filter Network via LSTM...

Area of Science:

• Computational Biology
• Bioinformatics
• Network Science

Background:

• Biomolecular interaction prediction is crucial for understanding biological systems.
• Existing graph convolution methods face challenges in incorporating both immediate and high-order neighbors while mitigating noise.

Purpose of the Study:

• To propose a novel method, HGLA, for enhanced biomolecular interaction prediction.
• To address limitations in existing graph convolution approaches for handling high-order neighbors and noise.

Main Methods:

• Mixed high-order graph convolution with filter network via LSTM and channel attention (HGLA).
• Feature extraction using traditional Graph Convolutional Network (GCN) and Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing (MixHop).
• Integration of features through a filter network comprising LayerNorm, SENet, and LSTM for noise reduction and feature balancing.

Main Results:

• HGLA processes high-order features separately and effectively filters noise.
• The method achieves a better balance between basic and high-order features.
• HGLA demonstrates superior performance compared to state-of-the-art methods on four benchmark datasets.

Conclusions:

• HGLA offers a significant advancement in predicting biomolecular interactions.
• The proposed method effectively addresses the challenges of incorporating high-order neighbor information and noise reduction.
• HGLA provides a more robust and accurate approach for biomolecular network analysis.