Abstract
Phosphoproteomics integrating data-independent acquisition (DIA) has enabled deep phosphoproteome profiling with improved quantification reproducibility and accuracy compared to data-dependent acquisition (DDA)-based phosphoproteomics. DIA data mining heavily relies on a spectral library that in most cases is built on DDA analysis of the same sample. Construction of this project-specific DDA library impairs the analytical throughput, limits the proteome coverage, and increases the sample size for DIA phosphoproteomics. Herein we introduce a novel deep neural network, DeepPhospho, which conceptually differs from previous deep learning models to achieve accurate predictions of LC-MS/MS data for phosphopeptides. By leveraging in silico libraries generated by DeepPhospho, we established a new DIA workflow for phosphoproteome profiling which involves DIA data acquisition and data mining with DeepPhospho predicted libraries, thus circumventing the need of DDA library construction. Our DeepPhospho-empowered workflow substantially expanded the phosphoproteome coverage while maintaining high quantification performance, which led to the discovery of more signaling pathways and regulated kinases in an EGF signaling study than the DDA library-based approach. DeepPhospho is provided as a web server to facilitate user access to predictions and library generation.
Introduction
In the field of proteomics, researchers rely on Mass Spectrometry (MS) to identify and quantify proteins. Data-Independent Acquisition (DIA) has emerged as a gold standard due to its superior reproducibility and accuracy compared to older methods. However, DIA faces a significant bottleneck: it requires a "spectral library"—a reference map—to identify peptides from raw data. Traditionally, researchers must build this map by performing labor-intensive Data-Dependent Acquisition (DDA) experiments on the same samples. This requirement consumes time and precious samples, often limiting the analysis to only the most abundant peptides. DeepPhospho addresses this by replacing physical library construction with high-accuracy deep learning predictions.
The DeepPhospho Architecture and Workflow
DeepPhospho architecture and the integrated DIA workflow. Source, p. 1
The core innovation of DeepPhospho is its hybrid architecture, which is the first to apply Transformers—the technology powering models like GPT—to peptide fragmentation prediction. The model is modular and designed to capture both local and global peptide characteristics:
- Bi-LSTM Network: A bidirectional Long Short-Term Memory network processes the peptide sequence to capture local context, such as how an amino acid relates to its immediate neighbors.
- Transformer Network: A Transformer module uses multi-head self-attention to capture long-range dependencies. This is crucial for phosphopeptides, where a modification at one end of a peptide can significantly influence how the entire molecule fragments in the mass spectrometer.
- Regressor: A linear regressor predicts fragment ion intensities (the peptide's "fingerprint") and indexed retention time (iRT), which indicates when the peptide will appear during the experiment.
To optimize for phosphoproteomics, the authors introduced specialized tokens for phosphorylated amino acids and a modified loss function that enforces chemical constraints, ensuring the model ignores "impossible" ions.
Key Results and Biological Insights
High-Fidelity Predictions
DeepPhospho achieved a median Pearson Correlation Coefficient (PCC) of 0.968 for fragment ion intensity and a median absolute error of only 1.74 units for retention time. Interestingly, the model's predicted spectra were often more accurate than experimental DIA libraries, as the model could pinpoint and correct false identifications in the experimental data by matching against synthetic standards.
Expanded Proteome Coverage
In a study of U2OS cell lines, the DeepPhospho-empowered workflow identified 32,511 phosphopeptides, a ~26% increase over the 25,814 identified using a standard DDA-based library. Despite this expanded coverage, the model maintained high quantification integrity, with a False Discovery Rate below 1% and high reproducibility (CV ~10%).
Discovery of New Signaling Pathways
The practical utility of this approach was demonstrated in an Epidermal Growth Factor (EGF) signaling study in RPE1 cells. DeepPhospho identified 7 additional enriched signaling pathways (including mTOR and MAPK) and several regulated kinases that were completely missed by the traditional DDA-based approach. This proves that the in silico library not only speeds up the workflow but also uncovers biological signals that are typically lost in physical library construction.
Conclusion
DeepPhospho provides a robust, DDA-free workflow for phosphoproteome profiling. By leveraging a hybrid bi-LSTM and Transformer architecture, it delivers the accuracy needed to bypass traditional bottlenecks, significantly expanding the depth and speed of biological discovery. The tool is available as a web server and on GitHub to facilitate widespread adoption in the proteomics community.
Summary updated