TL;DR
This paper introduces DeepPhospho, a novel deep learning model that generates in silico spectral libraries for phosphopeptide identification in mass spectrometry data. By bypassing the need for labor-intensive experimental libraries, DeepPhospho significantly expands the coverage of phosphoproteome profiling and enhances the discovery of biological insights, particularly in signaling pathway analysis.
What Problem Does This Paper Address?
Data-independent acquisition (DIA) mass spectrometry has become a powerful tool for comprehensive phosphoproteome profiling due to its high reproducibility and accuracy. However, a major bottleneck in DIA analysis is the reliance on high-quality spectral libraries, which are typically generated from time-consuming and resource-intensive data-dependent acquisition (DDA) experiments. Constructing these project-specific DDA libraries limits analytical throughput, reduces proteome coverage, and increases sample requirements. This paper addresses the challenge of creating accurate and comprehensive spectral libraries for DIA phosphoproteomics without the need for experimental DDA data.
Methods
The core of this work is the DeepPhospho deep neural network. Unlike previous models that often used simpler recurrent architectures with linear amino acid embeddings, DeepPhospho employs a hybrid architecture:
- Bi-LSTM Module: This module captures local peptide sequence information and generates an initial context-aware representation of each amino acid.
- Transformer Module: This module refines the peptide representation by using a multi-head self-attention mechanism, allowing it to capture long-range dependencies within the peptide sequence. This is a key innovation, as Transformers are more adept at handling variable-length sequences and complex relationships.
- Regressor Network: A linear regressor then predicts fragment ion intensities and indexed retention times (iRTs).
DeepPhospho also incorporates specific adaptations for phosphopeptides, including special tokens for phosphorylated amino acids and a modified loss function to account for the phosphate moiety. The authors also developed a workflow that leverages these in silico libraries for DIA data acquisition and mining, eliminating the need for DDA library construction.
Datasets
The authors utilized several datasets for training and evaluation:
- Training/Pre-training: Mouse brain DDA, Vero E6 DIA, yeast DIA, and human phosphopeptide RT datasets were used for pre-training DeepPhospho.
- Evaluation:
- RPE1 DDA and RPE1 DIA datasets (from a previous study)
- U2OS DIA dataset (from a previous study)
- Synthetic phosphopeptide datasets (human and yeast) for validation of predictions and quantification accuracy.
- A two-proteome model dataset (yeast/HeLa) for evaluating quantification accuracy and reproducibility.
They also leveraged public phosphoproteome and phosphosite databases (hPhosPepDB, hPhosSiteDB, yPhosPepDB) to generate predicted libraries.
Key Insights and Ideas
- Transformer for Peptide Representation: The integration of a Transformer module in DeepPhospho is a significant advancement, enabling a more nuanced understanding of peptide structure and long-range dependencies compared to purely recurrent models.
- In Silico Libraries Eliminate DDA Bottleneck: The core idea of generating spectral libraries in silico directly from DIA data and public databases bypasses the laborious and time-consuming experimental DDA library construction, dramatically accelerating phosphoproteomics workflows.
- DeepPhospho Can Identify False Positives: In some cases, DeepPhospho's predicted spectra showed much higher similarity to synthetic peptide spectra than to library spectra, suggesting that the library spectra might represent false identifications.
- Iterative Search Strategy: For large predicted libraries, an iterative search strategy was employed to reduce the search space and improve the sensitivity for identifying quantifiable phosphopeptides.
- Enhanced Biological Discovery: The use of DeepPhospho-generated libraries led to the identification of more regulated signaling pathways and kinases in an EGF signaling study, demonstrating the practical biological impact of improved phosphoproteome coverage.
Key Results
- Accurate Predictions: DeepPhospho achieved excellent agreement between predicted and experimental fragment ion intensities (median PCC = 0.968 for RPE1 DDA, 0.903 for RPE1 DIA) and accurate iRT predictions (median MAE = 1.74 units for RPE1 DDA, 1.86 units for RPE1 DIA).
- Outperforms Existing Models: DeepPhospho consistently outperformed other state-of-the-art models (pDeep2, DeepMS2, MS2PIP) in phosphopeptide fragment ion intensity prediction.
- Increased Proteome Coverage: Libraries generated by DeepPhospho (e.g., Lib 7) significantly increased phosphopeptide and phosphosite identifications compared to a project-specific DDA library (e.g., Lib 1), with gains of up to 25.9% in phosphopeptides and 23.1% in phosphosites.
- Maintained FDR Control: The use of DeepPhospho predicted libraries did not compromise false discovery rate (FDR) control, with FDRs comparable to or even lower than experimental libraries.
- Improved Quantification Accuracy and Reproducibility: In a two-proteome model, DeepPhospho predicted libraries demonstrated quantification accuracy and reproducibility equivalent to or better than the DDA library.
- Enhanced Biological Insights: In an EGF signaling study, DeepPhospho-based libraries identified significantly more regulated phosphosites, pathways, and kinases compared to the DDA library approach.
Limitations and Future Work
- Library Parameter Optimization: The authors note that the optimal parameter combination for generating predicted libraries from public databases (like hPhosPepDB) can vary and requires careful optimization for different datasets.
- Generalizability to Diverse Samples: While promising for human phosphoproteomics, further investigation is needed to assess the broad applicability of hPhosPepDB-derived libraries across various human sample types (tissues, body fluids, cell lines).
- Extension to Other Modifications: The authors envision that DeepPhospho's architecture could be adapted to predict non-phosphopeptides and peptides with diverse modifications, paving the way for efficient DIA workflows for global proteomics and other PTM studies.
- Targeted Applications: DeepPhospho is expected to be valuable for validating phosphopeptide identifications, developing targeted MS assays, and assessing FDR control in DIA analysis.
Summary updated