SPPIDER-seq: sequence-based partner-aware predictor of protein-protein interaction sites.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Protein Interaction Networks

Background:

• Traditional protein-protein interaction (PPI) site predictors often ignore crucial partner-specific context, especially for transient and disordered interactions.
• Interface specificity is vital for understanding biological recognition, particularly in dynamic and intrinsically disordered protein systems.

Purpose of the Study:

• To develop a partner-aware PPI site prediction framework that incorporates interacting partner information.
• To improve the accuracy of predicting interaction sites, especially for challenging disordered and transient protein complexes.

Main Methods:

• Introduced SPPIDER-seq, a novel framework utilizing pretrained ESM-2 embeddings and a cross-attention architecture.
• Developed two complementary models: a receptor-centric model for structured interfaces and a peptide-centric model for disordered, motif-driven binding.
• Trained and benchmarked models on curated, non-redundant protein-peptide interaction datasets from BioLiP.

Main Results:

• SPPIDER-seq achieved high performance on blind benchmarks, with AUROC up to 0.797 and MCC up to 0.269.
• Outperformed AlphaFold3 on peptide-mediated and disordered interfaces, demonstrating complementary performance on globular complexes.
• Analysis of 341 TP53 interaction partners revealed distinct, partner-specific interface patterns in both structured and disordered regions.

Conclusions:

• SPPIDER-seq effectively captures partner-specific context for accurate PPI site prediction, particularly for disordered interactions.
• The framework provides insights into the molecular basis of protein recognition across different interface types.
• SPPIDER-seq models, datasets, and code are publicly available, facilitating further research in protein interaction prediction.