AHLF

AHLF learns fragmentation patterns and detects post-translationally modified peptides, including phosphorylation, directly from mass spectrometry fragmentation spectra using an end-to-end deep learning model.

Key Features:

• Deep Learning Architecture: Implements a Temporal Convolutional Network (TCN) with dilated convolution layers to model complex temporal dependencies in spectra.
• End-to-End Training: Trained end-to-end on a dataset of 19.2 million spectra compiled from multiple phosphoproteomic studies.
• Interpretability: Provides peak-level feature importances and identifies pairwise interactions between peaks that align with known peptide fragments.
• PTM Detection: Detects post-translational modifications such as protein phosphorylation directly from fragmentation spectra without requiring a database search.
• Performance: Outperforms existing state-of-the-art phosphoproteomics methods, increasing area under the ROC curve (AUC) by an average of 9.4%.
• Transfer Learning: Applies transfer learning to detect cross-linked peptides, achieving AUC values up to 94%.
• Generalization: Learns fragmentation patterns for modified and less-studied peptides to generalize across different types of mass spectrometry data.

Scientific Applications:

• Phosphoproteomics / Cell Signaling: Identification of phosphorylation events and analysis of phosphoproteomic datasets relevant to cell signaling studies.
• Cross-linking and Structural Biology: Detection of cross-linked peptides for protein structure analysis and structural proteomics.
• Discovery of Novel Modifications: Characterization of peptides with atypical or previously unknown fragmentation patterns without prior sequence or modification knowledge.

Methodology:

Uses a Temporal Convolutional Network with dilated convolutions trained end-to-end on 19.2 million phosphoproteomic spectra; produces peak-level importances and pairwise peak interactions and employs transfer learning for cross-linked peptide detection.