MS-Spectral-Foundation

Biological Background (Spatial Omics Context)

This project is inspired by spatial omics studies of hepatocellular carcinoma (HCC), where MALDI-based spatial glycan imaging reveals strong molecular heterogeneity within tumor tissue (Wang et al., 2024).

Spatial omics provides the biological motivation by identifying disease-associated glycan patterns in tissue and serum.

This project focuses on the computational side: learning general spectral representations from serum LC-MS/MS data, biologically inspired by spatial omics findings.

Note that the current project does not analyze spatial imaging data directly; instead, spatial omics studies serve as biological motivation for studying molecular heterogeneity.

Biological Question

Do learned spectral embeddings reveal systematic molecular shifts between HCC and cirrhosis serum samples, as reflected by embedding-space distribution differences and exemplar spectra?

More specifically:

• Serum LC-MS/MS data contains thousands of MS/MS spectra reflecting underlying protein and glycoprotein states.
• Instead of relying on manually designed features, can we automatically learn representations from spectra themselves?
• Do these learned representations capture disease-related molecular patterns that reveal systematic embedding-space differences between HCC and cirrhosis?

This project explores a representation learning approach for proteomics-related mass spectrometry data inspired by spatial omics studies, aiming to understand whether self-supervised learning can reveal biologically meaningful structure from highly heterogeneous clinical samples.

Due to the limited number of patient samples (5 HCC samples and 5 cirrhosis samples), this project focuses on representation analysis rather than supervised disease classification.

---

Overview

Mass spectrometry experiments generate a large number of spectra, each containing pairs of m/z values and intensities. Traditionally, downstream analysis depends heavily on manual feature engineering and labeled datasets, which can be limiting for small or heterogeneous biomedical datasets.

This project takes a different approach:

Instead of directly modeling spatial omics images, this project focuses on learning transferable representations from MS/MS spectra derived from spatial-omics-related workflows.

1. Convert raw LC-MS/MS data into standardized MS/MS spectra (MGF format)
2. Treat each spectrum as a learning object
3. Use self-supervised masked learning to train a model that reconstructs missing spectral information
4. Extract fixed-dimensional embeddings that summarize each spectrum
5. Aggregate spectrum embeddings into sample-level representations
6. Perform downstream embedding-space analysis (distribution comparison and exemplar spectra discovery between HCC and cirrhosis) to evaluate whether the learned embeddings capture disease-related structure

In short:

Instead of manually defining features, we let the model learn how spectra are structured, and then examine whether the learned representations reveal systematic molecular differences between disease groups.

This work is positioned as a representation-learning foundation for future large-scale mass spectrometry analysis, where learned spectral embeddings can be reused across downstream molecular discovery tasks.

---

Input and Output

Input

The pipeline supports three levels of input data:

• .raw
  Raw LC-MS/MS files generated by the mass spectrometer.

• .mzML
  Open-format files converted from .raw, used for spectrum extraction.

• .mgf
  MS/MS spectra format used as the direct input for the self-supervised model.
  Each BEGIN IONS … END IONS block represents one spectrum.

Output

The pipeline produces:

• Processed spectra (MGF-based)
• Learned spectrum embeddings
• Sample-level embeddings
• Downstream embedding-space analysis results
  • Distribution differences between HCC and cirrhosis
  • Exemplar spectra highlighting maximal/minimal cross-group similarity
• Basic visualization (e.g., PCA / UMAP)

---

Project Goal (Simple Summary)

• Learn general-purpose spectral embeddings from MS/MS data
• Reduce dependence on manual feature engineering
• Test whether learned embeddings reveal systematic molecular shifts between HCC and cirrhosis
• Build a modular pipeline from raw MS data to downstream biological interpretation Note: for all the md documents except DDS.md, I used AI to format the documents such as organizing bullet points.

---

Example Data & Tutorial

Example Data

The notebook uses the bundled sample MGF file as example data:

• casanovo/sample_data/sample_preprocessed_spectra.mgf

This file contains a small set of MS/MS spectra in standard MGF format. To run the pipeline on a different dataset, replace the MGF_PATH variable in the tutorial notebook with any valid .mgf file.

Tutorial Notebook

A Jupyter notebook demonstrating the full happy-path pipeline is available here:

• tutorials/ms_spectral_foundation_tutorial.ipynb

The notebook walks through all pipeline stages:

1. Loading spectra (Module 1 — mgf_parse.py)
2. Preprocessing peaks (Module 2 — peak_filter.py)
3. m/z binning (Module 3 — bin_mz.py)
4. Initialising the SpectrumSSLv2 model
5. Running a masked forward pass
6. Computing MSE and Cross-Entropy losses
7. Extracting CLS spectrum embeddings
8. Visualising results (spectrum bar chart, predicted vs. true m/z scatter, PCA of embeddings)