usiGrabber: Automating the curation of proteomics spectra data at scale, making large datasets ready for use in machine learning systems

The paper introduces usiGrabber, a scalable and portable framework that automates the extraction and curation of large-scale proteomics spectra data from public repositories like PRIDE, demonstrating its effectiveness by rapidly constructing a massive phosphorylation dataset to retrain a machine learning classifier with performance comparable to manually curated models.

The Gist

Imagine you are a chef trying to create the world's best new recipe. You know that the secret ingredient is "phosphorylation" (a specific chemical change in proteins), but you need to taste millions of different dishes to learn exactly how it affects the flavor.

The problem? The world's biggest spice rack (a database called PRIDE) is open to the public, but it's a chaotic mess. It's like a library where books are thrown on the floor, labeled only by the city they came from, not by their contents. To find the specific pages about phosphorylation, a human chef would have to:

1. Walk into the library.
2. Pick up a random book (a research project).
3. Read every single page to see if it mentions the spice.
4. If it does, copy those pages by hand.
5. Repeat this for thousands of books.

This is slow, boring, and by the time you finish, the library has already added 10,000 new books. Most scientists just give up and use a tiny, old cookbook from 2017 because it's the only one they managed to copy.

Enter: usiGrabber (The Robot Librarian)

The authors of this paper built a robot called usiGrabber. Instead of a human reading every book, usiGrabber is a super-fast, automated system that does three things:

1. The Scanner (Extraction): It doesn't read the whole book. Instead, it scans the "Table of Contents" (the metadata files) of over 1,200 different research projects. It instantly finds the specific pages (spectra) that talk about phosphorylation and writes down their unique "barcode" (called a Universal Spectrum Identifier, or USI).

   • Analogy: Imagine a robot that can scan the spines of 800 million books in two days and pull out a list of exactly which pages contain the word "phosphorylation."

2. The Filter (Curation): It takes that massive list of barcodes and filters out the junk. It checks to make sure the pages are high quality and match what we are looking for.

   • Analogy: The robot sorts through the list and throws away any pages that are torn, written in the wrong language, or don't actually contain the spice we need.

3. The Collector (Download): Once it has the perfect list of barcodes, it goes back to the library, grabs only those specific pages, and assembles them into a brand-new, perfectly organized cookbook.

   • Analogy: Instead of downloading the entire 800-page library, it downloads only the 11 million specific pages it needs to build a new training dataset.

The Result: A New Cookbook in Two Days

The team used this robot to build a new "training dataset" for a machine learning model (a computer brain) designed to detect phosphorylation.

• Old Way: It would take a human team months or years to curate a dataset this size, and they'd be stuck using old data.
• usiGrabber Way: The robot did the whole job in less than two days.

They fed this fresh, massive dataset into the computer brain. The result? The new brain performed just as well as the old, famous models that were trained on tiny, outdated datasets. In fact, it was almost as good as a super-complex model that had been trained on double the data.

Why This Matters

Think of usiGrabber as a time machine for science.

• Before: Scientists were stuck using "vintage" data (like trying to learn to drive on a 1990s map).
• Now: They can instantly access the "live traffic" of today's research.

This means that as soon as a new experiment is published, usiGrabber can grab that data and help train better AI models immediately. It turns a chaotic, inaccessible mountain of data into a clean, usable stream of information, allowing artificial intelligence to finally learn from the full power of modern proteomics.

In short: They built a robot that can read the entire library of protein science in a weekend, pick out the exact pages you need, and hand you a perfect study guide, all while you sleep.

Technical Summary

1. Problem Statement

The field of mass spectrometry-based proteomics has generated an unprecedented volume of public data (e.g., 864 TB submitted to PRIDE in 2025 alone). However, this data remains largely inaccessible for modern machine learning (ML) applications due to several critical bottlenecks:

• Manual Curation Bottleneck: Constructing task-specific datasets requires deep domain expertise and tedious manual selection of spectra from specific projects. This process is time-intensive and does not scale.
• Legacy Data Reliance: Current state-of-the-art ML models (e.g., Casanovo, Modanovo, AHLF) rely on static datasets that are several years old (e.g., MassIVE-KB v1 from 2018, datasets from 2017). These models cannot easily incorporate newly published data.
• Repository Limitations: Public repositories like PRIDE are optimized for project-level deposition and discovery, not spectrum-level retrieval. They lack granular filtering capabilities for specific modifications (e.g., phosphorylation) or quality metrics without downloading massive raw files.
• Reanalysis Inefficiency: Alternative approaches like MassIVE-KB rely on full reanalysis pipelines, which are computationally expensive, storage-intensive, and often fail when attempting to download specific subsets (e.g., causing database crashes).

2. Methodology: The usiGrabber Framework

The authors propose usiGrabber, a scalable, portable, and extensible framework designed to automate the curation of proteomics datasets without requiring full reanalysis. The workflow consists of three primary stages:

A. Data Extraction and Indexing

• Source: The system targets the PRIDE repository, focusing on projects containing mzIdentML files (standardized result files) and Thermo Fisher Raw files.
• Parsing: A robust, adaptive parser extracts spectrum-level evidence from mzIdentML files. It retrieves:
  • File-level metadata (software used, QC thresholds).
  • Peptide Spectrum Match (PSM) details (sequence, post-translational modifications, charge state, mass-to-charge ratio).
  • Protein associations.
• Normalization: Data is enriched using ontologies (UNIMOD, NCBITaxon, MS) via the EMBL-EBI Ontology Lookup Service and linked to UniProt identifiers.
• Indexing: Extracted data is normalized into a common schema and indexed by Universal Spectrum Identifiers (USIs). This creates a local database instance containing over 800 million PSMs from ~1,200 projects.

B. Querying and Filtering

• Researchers can query the local database using simple filters to identify relevant USIs based on specific research questions (e.g., "phosphorylated peptides with Rank 1").
• This allows for the selection of specific subsets of data without downloading the entire repository.

C. Scalable Download and Validation

• Grouping: The download tool groups selected USIs by their source raw file.
• Retrieval: It downloads the specific raw files from PRIDE in parallel.
• Extraction: Spectra are extracted using ThermoRawFileParser (v2.0) and converted to ML-ready formats (Parquet or MGF).
• Quality Control (QC): A critical validation step compares the charge state in the extracted spectrum against the charge state in the USI metadata. Any raw file containing a mismatch is discarded to ensure data integrity.

3. Key Contributions

1. Scalable Automation: usiGrabber shifts the paradigm from manual curation of static datasets to automated, dynamic extraction. It enables the creation of task-specific datasets in days rather than months.
2. Lightweight Architecture: Unlike reanalysis pipelines, usiGrabber extracts metadata from existing result files, avoiding the computational cost of re-processing raw spectra.
3. Extensibility: The framework is designed to be modular, allowing researchers to build their own database instances and adapt parsers for new file formats or search tools.
4. Open Availability: The code, trained model weights, and the curated dataset are publicly available on GitHub and Zenodo.

4. Results

The authors validated usiGrabber by constructing a binary phosphorylation classifier dataset and retraining the AHLF model architecture.

• Extraction Performance:
  • Processed 1,270 projects (out of 4,116 suitable candidates) in approximately 49 hours.
  • Parsed over 800 million PSMs (throughput: ~18 million PSMs/hour).
  • Database diversity includes a wide range of species and instrument types (Figure 2).
• Dataset Construction:
  • Curated a phosphorylation-specific dataset of ~11 million spectra in under two days.
  • Applied strict QC filters (Rank 1, project-specific thresholds, charge state verification), reducing the initial 42.5 million potential PSMs to 10.9 million high-confidence PSMs.
• Model Performance:
  • Retrained the AHLF model on the new dataset.
  • Achieved a balanced accuracy of 0.78 and an F1 score of 0.71.
  • Comparison: The model performed comparably to the original AHLF ensemble (trained on 20 million manually curated spectra) and outperformed three of the four individual AHLF models (Alpha, Beta, Delta, Gamma). It was only ~3% lower than the combined ensemble, which used nearly double the training data.

5. Significance

• Bridging the Data Gap: usiGrabber solves the "data access bottleneck" in proteomics ML, enabling the use of the latest public data rather than relying on legacy datasets.
• Democratization: By providing a tool that runs on commodity infrastructure, it allows smaller research groups to construct high-quality, task-specific datasets without needing massive compute clusters for reanalysis.
• Future-Proofing: The framework supports the rapid integration of newly published data, ensuring that ML models in proteomics can evolve alongside the field's data generation rate.
• Validation of Automation: The study proves that automated extraction can yield datasets of sufficient quality to train competitive ML models, challenging the notion that manual curation is the only path to high-quality data.

In conclusion, usiGrabber represents a critical infrastructure advancement for the proteomics community, transforming raw repository data into machine-learning-ready assets efficiently and reliably.