MS2CNN

MS2CNN predicts tandem mass spectrometry (MS/MS) spectra from features derived from protein sequences to improve peptide identification in proteomics.

Key Features:

• Deep Learning Architecture: Non-linear regression model implemented with deep convolutional neural networks (CNNs) for MS/MS spectral prediction.
• Feature Integration: Incorporates amino acid composition, predicted secondary structure, and physicochemical properties including isoelectric point, aromaticity, helicity, hydrophobicity, and basicity.
• Training and Validation: Trained using five-fold cross-validation on a large-scale human Higher-energy Collisional Dissociation (HCD) MS/MS dataset acquired by Orbitrap LC-MS/MS and sourced from the National Institute of Standards and Technology (NIST).
• Evaluation Dataset: Evaluated on an independent test dataset derived from human HeLa cell lysate LC-MS experiments.
• Performance Metrics: Reported cosine similarity of 0.690 and Pearson correlation coefficient of 0.632, outperforming MS²PIP (cosine 0.647, Pearson 0.601) and comparable to pDeep.
• Complex Spectra Handling: Exhibits improved prediction for higher charge-state peptides (3+), expanding peptide-space and spectral-library coverage.
• Software Framework: Implemented in Python using Keras 2.0.4 and TensorFlow 1.1.0.

Scientific Applications:

• Peptide and protein identification: Supports peptide identification in LC-MS/MS workflows using Orbitrap HCD experiments by providing predicted MS/MS spectra.
• Spectral library searches: Improves sensitivity and coverage of spectral library searches through more accurate predicted spectra.
• Expanded peptide-space: Enhances detection and analysis of higher charge-state (3+) peptides to broaden proteomic coverage.
• Biomarker discovery and disease diagnostics: Enables more comprehensive spectral analysis in studies aimed at biomarker discovery and disease-related proteomics.

Methodology:

MS2CNN is a deep convolutional neural network trained as a non-linear regression model using five-fold cross-validation on a large-scale human HCD MS/MS Orbitrap LC-MS/MS dataset from NIST and evaluated on an independent human HeLa cell lysate LC-MS dataset; implementation uses Python with Keras 2.0.4 and TensorFlow 1.1.0.