A Lightweight Curriculum and Contrastive Learning Framework for Protein-Protein Interaction Prediction

Area of Science:

• Computational Biology
• Bioinformatics
• Network Science

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions, drug development, and disease research.
• Existing PPI prediction methods struggle with multi-scale network heterogeneity, topological biases, and high computational costs due to reliance on external data.
• Insufficient exploitation of subgraph-level semantic complexity hinders the accuracy of current models.

Purpose of the Study:

• To develop a lightweight and efficient framework for protein-protein interaction (PPI) prediction.
• To mitigate representational bias and improve the exploitation of subgraph-level semantic complexity in PPI networks.
• To enhance model robustness and generalization without relying on external data or handcrafted features.

Main Methods:

• JCCLPPI: A joint curriculum- and contrastive-learning framework for lightweight PPI prediction.
• PPI network-structure encoding module to mitigate topological bias and learn interpretable representations.
• Motif-based curriculum learning to incrementally introduce training samples by complexity, enhancing robustness.
• Graph neural network modules for local structural modeling and global context encoding during inference.

Main Results:

• JCCLPPI demonstrates improved model generalization on human PPI benchmark datasets (SHS27k, SHS148k).
• Achieved an approximate 4% increase in micro-F1 score compared to state-of-the-art methods.
• Significantly improved computational efficiency: 76% reduction in memory consumption and 35% reduction in inference time.

Conclusions:

• JCCLPPI offers a computationally efficient and accurate approach for protein-protein interaction prediction.
• The framework effectively addresses topological biases and structural heterogeneity in PPI networks.
• Provides a scalable foundation for therapeutic target prioritization and early-stage drug discovery.