SpeCollate

SpeCollate learns cross-modal similarity between experimental mass spectra and peptide sequences using a deep cross-modal similarity network to improve peptide identification accuracy in MS-based proteomics.

Key Features:

• Deep learning approach: Employs a deep cross-modal similarity network trained on labeled MS data to learn the similarity function between experimental spectra and peptides.
• Shared Euclidean subspace transformation: Transforms experimental spectra and peptide sequences into fixed-size embeddings in a shared Euclidean subspace for direct similarity comparison.
• Custom SNAP-loss function: Optimizes network training with a custom SNAP-loss function tailored to discriminate subtle similarities between spectra and peptides.
• Online hardest negative mining: Incorporates online hardest negative mining to select challenging negative examples during training.
• Extensive training dataset: Trained on 4.8 million sextuplets derived from the NIST and MassIVE peptide libraries.

Scientific Applications:

• Peptide identification: Improves peptide-spectrum match (PSM) accuracy and peptide identification in MS-based proteomics.
• Enhanced recovery at low FDR: Recovers more PSMs and unique peptides than Crux and MSFragger, including under stringent false discovery rate (FDR) thresholds (<1%).
• Novel peptide discovery: Identifies peptides not reported by existing database search methods, enabling additional peptide identifications.

Methodology:

The method trains a deep cross-modal network on 4.8 million sextuplets from NIST and MassIVE, learning a similarity function from labeled MS data by transforming spectra and peptides into a shared Euclidean subspace and optimizing with a custom SNAP-loss and online hardest negative mining.