Topology-driven negative sampling enhances generalizability in protein-protein interaction prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Accurate protein-protein interaction (PPI) prediction is crucial for understanding diseases and identifying drug targets.
• Current machine learning (ML) models face limitations due to insufficient negative samples, shortcut learning, and poor generalizability.

Purpose of the Study:

• To introduce a novel approach for strategic sampling of protein-protein noninteractions (PPNIs).
• To develop a high-throughput ML pipeline, UPNA-PPI, for efficient screening of billions of interactions.
• To improve the generalizability and interpretability of PPI predictions.

Main Methods:

• Leveraging higher-order network characteristics for strategic PPNI sampling.
• Integrating unsupervised pre-training with Topological PPNI (TPPNI) samples in the UPNA-PPI pipeline.
• Utilizing network topology insights for a negative sampling methodology in graph ML.

Main Results:

• Improved generalizability and interpretability of PPI predictions using UPNA-PPI with TPPNI samples.
• Enhanced identification of potential protein binding sites on amino acid sequences.
• Facilitated transferability of ML predictions across protein families and homodimers.

Conclusions:

• UPNA-PPI establishes a foundation for negative sampling in graph ML.
• The method strengthens the prioritization of screening assays for drug discovery.
• This approach advances the field of protein interaction prediction and analysis.