Comprehensive top-down mass spectral repository enables pan-dataset analysis and top-down spectral prediction

This paper introduces TopRepo, the first comprehensive repository of over 18 million top-down mass spectra from diverse species and instruments, which enables large-scale proteoform analysis and significantly enhances both proteoform identification and deep learning-based spectral prediction.

The Gist

Imagine you are trying to identify every single person in a massive, bustling city. In the past, scientists had to take a photo of each person, cut them into tiny puzzle pieces (like chopping a person into their fingers, nose, and toes), and then try to guess who they were based on those scattered bits. This is called "bottom-up" analysis. It's useful, but you lose the full picture of who that person really is—did they have a specific haircut? A scar? A unique tattoo?

Top-Down Mass Spectrometry (TD-MS) is like taking a photo of the whole person intact. It lets scientists see the entire "proteoform" (a protein with all its specific modifications, like a unique outfit or a scar) in one go. This is incredibly powerful for understanding how our bodies work, but there was a huge problem: nobody had a big enough "Wanted Poster" book to help identify these people.

Without a massive library of reference photos, scientists were flying blind, trying to match the whole person to a tiny, blurry sketch.

Enter TopRepo: The Ultimate "Wanted Poster" Library

This paper introduces TopRepo, a massive new digital library created by researchers Kun Li, Kaiyuan Liu, and their team. Think of TopRepo as the world's largest, most comprehensive "Wanted Poster" collection for proteins.

• The Scale: They didn't just collect a few photos; they gathered 18 million spectral "photos" from 12 different species (including humans, mice, and bacteria) using 8 different types of high-tech cameras (mass spectrometers).
• The Curation: From that mountain of data, they cleaned and organized 5.4 million high-quality, annotated entries. Now, instead of guessing, scientists can look up a protein and find its exact "fingerprint" in the book.

What Did They Do With This Library?

The team didn't just build the library; they used it to solve three major problems:

1. The "Pan-Dataset" Detective Work
Because they had so much data from so many different sources, they could spot patterns that were invisible before.

• Analogy: Imagine looking at 100 photos of people from one city and seeing a trend. Now imagine looking at 100 photos from 12 different countries. You can suddenly see global trends, like "Most people in this region cut their hair on the left side" or "People in this city tend to wear red hats."
• The Result: They discovered new details about how proteins are processed in our bodies, such as how the "N-terminal" (the start of the protein) gets chopped off or modified, which is crucial for understanding diseases.

2. The "Super-Search" Engine
Before TopRepo, if a scientist wanted to identify a protein, they could only search against a small library (maybe 3,000 photos). If the protein wasn't in that small book, they couldn't find it.

• The Upgrade: They built a new search engine using the massive TopRepo library.
• The Result: When they tested it, the new library helped them identify 41.5% more proteins than the old, small library could. It's like upgrading from a local phone book to the entire internet; suddenly, you can find people you never knew existed.

3. The "Crystal Ball" (AI Prediction)
This is perhaps the coolest part. They used the library to train a Deep Learning AI called TD-Pred.

• The Analogy: Imagine you show a child 5 million photos of cars and ask them to guess what a new car will look like before it's even built. The AI learned the "rules" of how proteins break apart and what their "fingerprints" look like.
• The Result: The AI can now predict what a protein's spectrum should look like before the experiment is even run. This helps scientists design better experiments and identify proteins with much higher confidence. It's like having a crystal ball that tells you exactly what you'll see in the microscope.

Why Does This Matter?

For a long time, Top-Down Mass Spectrometry was like a Ferrari with no map. It was a powerful tool, but scientists didn't have enough data to drive it effectively.

TopRepo is the map.

• It fills the gap in our knowledge of human biology.
• It helps us find the "bad guys" (disease-causing protein variations) faster.
• It teaches AI how to predict biological outcomes, speeding up drug discovery and medical research.

In short, this paper gives the scientific community the biggest, most detailed reference book ever created for intact proteins, turning a difficult, guess-heavy process into a precise, data-driven science.

Technical Summary

1. Problem Statement

Top-down mass spectrometry (TD-MS) offers unique advantages over bottom-up MS (BU-MS) for characterizing intact proteoforms and combinatorial post-translational modifications (PTMs) without enzymatic digestion. However, the field faces a critical bottleneck: the lack of large-scale, comprehensive spectral libraries.

• Unlike BU-MS, which has established resources like ProteomeTools and NIST, TD-MS lacks sufficient data to train deep learning (DL) models for spectral prediction or to support robust spectral library searching.
• Existing databases are too small to enable "pan-dataset" analyses or to teach DL models the complex spectral similarity functions required to distinguish true from false identifications.
• Current TD-MS workflows suffer from limited confidence in PTM localization and site assignment due to insufficient annotated data.

2. Methodology

The authors developed TopRepo, a comprehensive repository and analysis pipeline, and TD-Pred, a deep learning model for spectral prediction.

A. TopRepo Construction

• Data Aggregation: The repository aggregates 3,671 raw MS files from 33 publications, covering 12 species (including human, mouse, E. coli, yeast, etc.) and 8 types of mass spectrometers (primarily Orbitrap and FT-ICR).
• Processing Pipeline:
  1. Conversion: Raw files converted to centroided mzML using msconvert.
  2. Deconvolution: Spectral deconvolution and feature detection performed using TopFD. For FAIMS data, this generated multiple msalign files per raw file.
  3. Identification: Deconvoluted spectra searched against UniProt proteomes using TopPIC to identify Proteoform-Spectrum Matches (PrSMs).
  4. Quality Control: Filtering applied at 1% False Discovery Rate (FDR) at both spectrum and proteoform levels.
  5. Annotation: Fragment ions in msalign and MGF files were annotated based on identified proteoform sequences, allowing for mass shift and PTM analysis.
• Scale: The final repository contains 18.2 million MS/MS spectra, with a curated library of 5.4 million annotated spectra.

B. TD-Pred Model (Deep Learning)

• Architecture: A hybrid model integrating Convolutional Neural Networks (CNNs) and Transformers.
  • Input Encoding: Proteoform sequences (up to 200 residues) are one-hot encoded and augmented with residue mass, normalized position, and sequence length.
  • CNN Subnetwork: Eight parallel modules with kernel sizes 2–9 capture local sequence dependencies.
  • Meta-Information: Experimental metadata (instrument type, fragmentation method, precursor charge state, collision energy) is embedded and added to every sequence position.
  • Transformer: Six encoder and six non-autoregressive decoder layers predict the spectrum.
• Output Representation:
  • Backbone Representation: An $(L-1) \times 60$ matrix representing relative abundances of N- and C-terminal fragment ions across charge states 1–30.
  • Simplified Representation: An $(L-1) \times 2$ matrix summing total N- and C-terminal abundances (used to improve accuracy for high charge states/long sequences).

3. Key Results

A. Repository Statistics and Pan-Dataset Analysis

• Identification Rates: The average spectral identification rate across projects was 30.0%, varying significantly by sample complexity (e.g., 86.1% for E. coli vs. 57.2% for complex human tissues).
• Proteoform Characteristics:
  • Identified 311,248 unique proteoforms from 19,318 proteins.
  • Truncation Analysis: Only 16.3% of identified proteoforms were "complete" (full-length). The majority showed N-terminal, C-terminal, or internal truncations, often matching trypsin digestion patterns, suggesting endogenous enzymatic activity during sample prep is a major factor.
  • N-terminal Processing: Detailed analysis of N-terminal Methionine Excision (NME) and N-terminal Acetylation (NTA) confirmed known substrate specificities of Methionine Aminopeptidases (MAP1/MAP2) and N-terminal Acetyltransferases (NATs) across human and E. coli.
  • Signal Peptides: Successfully identified signal peptide cleavage sites in 209 human proteins.
• Reproducibility:
  • Protein-level reproducibility between datasets was moderate (39–72%).
  • Proteoform-level reproducibility was low (often ≤17% between different labs/projects), primarily due to dataset-specific truncated proteoforms. This highlights the need for standardized sample preparation.

B. Impact on Spectral Library Searching

• The authors constructed a large-scale human HCD library (HUMAN-HCD, ~259k spectra) and compared it to a small, single-dataset library (SW480-2D, ~5k spectra).
• Result: Using the large-scale library increased proteoform identifications by 41.5% (4,454 vs. 3,148) compared to the small library, demonstrating that scale is critical for sensitivity in TD-MS.

C. TD-Pred Performance

• Training: The model was trained on ~600k spectra (CID and HCD).
• Accuracy:
  • Achieved a validation cosine similarity of 0.821 using the backbone representation.
  • Accuracy improved to 0.867 (CID) and 0.840 (HCD) when using the simplified representation (removing charge state prediction).
• Limitations: Prediction accuracy decreased for high charge states (>20) and long proteoforms (>180 residues) due to data scarcity and spectral complexity.

4. Significance and Contributions

1. First Large-Scale TD-MS Repository: TopRepo is the largest public repository of top-down spectra to date, providing a foundational resource for the community.
2. Enabling Deep Learning: The dataset enables the training of DL models (like TD-Pred) for in silico spectral prediction, a capability previously limited in TD-MS.
3. Enhanced Identification: The study proves that large-scale spectral libraries significantly outperform single-dataset libraries, increasing proteoform identification sensitivity by over 40%.
4. Biological Insights: The pan-dataset analysis reveals systematic issues in current TD-MS workflows, specifically the prevalence of sample-induced truncations and the challenges in cross-study reproducibility.
5. Future Roadmap: The authors identify key areas for improvement, including better deconvolution algorithms, inclusion of Time-of-Flight (TOF) data, and extending DL models to handle diverse PTMs and retention time prediction.

5. Conclusion

TopRepo and the associated TD-Pred model represent a major leap forward for top-down proteomics. By aggregating 18 million spectra, the authors have created the necessary infrastructure to move TD-MS from a niche technique to a high-throughput, data-driven field capable of comprehensive proteoform characterization and accurate PTM localization.