Predicting biomolecular interactions via a dual-stream graph neural network with motif constraint and diffusion-based...

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Biomolecular interactions, including RNA-Protein Interactions (RPIs) and Protein-Protein Interactions (PPIs), are essential for biological processes.
• Accurate prediction of these interactions is a significant challenge in computational biology.
• Existing deep learning methods face limitations in interpretability, generalization to new biomolecules, and handling sparse, noisy data.

Purpose of the Study:

• To develop a novel framework, DSG-BIP, for enhanced Biomolecular Interaction Prediction (BIP).
• To address the limitations of current methods regarding interpretability, generalization, and data robustness.

Main Methods:

• Employed a Dual-Stream Graph (DSG) neural network to model topological structure and node features separately.
• Integrated learnable motif constraints, dynamically optimized using sequence conservation and network context.
• Incorporated improved diffusion-based regularization and an adaptive masking mechanism for robustness against data sparsity and class imbalance.

Main Results:

• DSG-BIP achieved prediction performance comparable to state-of-the-art methods on RPI and PPI benchmark datasets.
• Demonstrated improved interpretability by dynamically optimizing motif constraints.
• Showcased enhanced generalization capabilities and robustness to sparse and noisy data.

Conclusions:

• DSG-BIP offers a significant advancement in predicting biomolecular interactions.
• The framework provides a more interpretable and robust approach compared to existing methods.
• DSG-BIP holds promise for advancing computational biology and drug discovery efforts.