PepFormer

PepFormer predicts peptide detectability in shotgun proteomics experiments using an end-to-end Siamese network that integrates Transformer attention mechanisms and gated recurrent units (GRUs) to analyze peptide sequences directly without relying on external physicochemical properties or sequential composition features.

Key Features:

• Hybrid Architecture: Integrates Transformer attention mechanisms with gated recurrent units (GRUs) to capture complex sequence relationships.
• End-to-End Siamese Network: Uses an end-to-end Siamese network to learn from paired peptide sequence inputs and improve discrimination.
• Contrastive Learning: Employs contrastive learning with a novel loss function to optimize representation separation and enhance generalization.
• Cross-Species Transferability: Demonstrates transferability for predicting peptide detectability across species including Homo sapiens and Mus musculus.
• Interpretable Model Representations: Learns embedded peptide sequence representations that visualization analyses show capture discriminative latent information.

Scientific Applications:

• Shotgun proteomics: Predicts peptide detectability to inform peptide identification and quantification workflows in shotgun proteomics experiments.
• Comparative and multi-species proteomics: Enables comparative proteomic studies by providing transferable detectability predictions across species such as Homo sapiens and Mus musculus.
• Experimental design and data interpretation: Supports experimental design and interpretation by prioritizing peptides with higher predicted detectability.

Methodology:

Uses an end-to-end Siamese network combining Transformer attention mechanisms and GRUs, trained with contrastive learning optimizing a novel loss function to learn embedded peptide representations that were analyzed via visualization.