CliPepPI: Scalable prediction of domain-peptide specificity using contrastive learning

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Domain-peptide interactions are crucial for cellular protein networks but challenging to predict due to short, fuzzy sequence motifs and limited experimental data.
• Existing structure-based methods are accurate but computationally intensive and difficult to scale.
• Data scarcity and bias from negative examples hinder the development of reliable prediction models.

Purpose of the Study:

• To develop a scalable and accurate computational model for predicting domain-peptide interaction specificity.
• To leverage sequence data and structural context for improved prediction performance.
• To create a versatile tool for large-scale proteomic analyses, including variant effect prediction.

Main Methods:

• Introduced CLIPepPI, a dual-encoder model using contrastive learning on sequence data.
• Initialized encoders from a protein language model (ESM-C) and fine-tuned with LoRA adapters for parameter efficiency.
• Augmented datasets with protein-protein interface data and incorporated structural information by marking interface residues.

Main Results:

• Achieved competitive performance across three independent benchmarks (PPI3D, ProP-PD, NES dataset).
• Demonstrated scalability through proteome-wide nuclear export signal (NES) scanning.
• Successfully applied the model to predict variant effects, distinguishing pathogenic from benign variants.

Conclusions:

• CLIPepPI provides a scalable, structure-informed approach for predicting domain-peptide specificity.
• The model generates meaningful embeddings suitable for large-scale proteomic analyses.
• CLIPepPI offers a valuable tool for understanding protein interactions and their role in disease.