Deep learning prediction boosts phosphoproteomics-based discoveries through improved phosphopeptide identification.

Area of Science:

• Proteomics
• Computational Biology
• Cancer Research

Background:

• Shotgun phosphoproteomics offers high-throughput analysis of phosphopeptides but is limited by low identification rates in data analysis.
• Accurate phosphopeptide identification and phosphosite localization are crucial for understanding cellular signaling pathways and disease mechanisms.

Approach:

• Developed DeepRescore2, a computational workflow integrating deep learning-based retention time and fragment ion intensity predictions.
• Leveraged state-of-the-art computational methods as a benchmark for performance evaluation.

Key Points:

• DeepRescore2 significantly increases the number of correctly identified peptide-spectrum matches by 17% in synthetic datasets.
• Identified 19%-46% more phosphopeptides in biological datasets compared to benchmark methods.
• In a liver cancer dataset, DeepRescore2 identified 30% of significantly altered phosphosites and 60% of prognosis-associated phosphosites missed by the benchmark workflow.
• Uniquely identified EGFR hyperactivation as a novel therapeutic target in poor-prognosis liver cancer, experimentally validated.

Conclusions:

• Deep learning integration in DeepRescore2 substantially improves phosphopeptide identification and phosphosite localization accuracy.
• The workflow facilitates novel biological discoveries and enhances the potential of shotgun phosphoproteomics for clinical applications.