MEDUSA

Machine learning Enabled Deisotoping and Untargeted Spectra Analysis

Mass spectrometry (MS) is a convenient, highly sensitive, and reliable method for the analysis of complex mixtures, which is vital for materials science, life sciences fields such as metabolomics and proteomics, and mechanistic research in chemistry. Although it is one of the most powerful methods for individual compound detection, complete signal assignment in complex mixtures is still a great challenge. The unconstrained formula-generating algorithm, covering the entire spectra and revealing components, is a “dream tool” for researchers. We present the framework for efficient MS data interpretation, describing a novel approach for detailed analysis based on deisotoping performed by gradient-boosted decision trees and a neural network that generates molecular formulas from the fine isotopic structure, approaching the long-standing inverse spectral problem. The workflow was successfully tested on three examples: fragment ion analysis in protein sequencing for proteomics, analysis of the natural samples for life sciences, and study of the cross-coupling catalytic system for chemistry.

Follow us

How to use it?

To start with MEDUSA, first, install the required packages running. It is recommended to create new virtual environment for this purpose.

pip install -r requirements.txt

If you want to build docs, you will also have to install Sphinx and furo theme, running

pip install sphinx
pip install furo

Then you will be able to

cd docs
make html

Built docs will be in the _build directory

Simple operations with mass-spectra

To open a spectrum run

from mass_automation.experiment import Experiment

exp = Experiment('spectrum.mzXML', n_scans=128, n_points=6)

Individual spectra can be accessed in list-like fashion

spectrum = exp[0]

Masses and intensities can be accessed manually

masses = spectrum.masses
ints = spectrum.ints

See more details on working with MEDUSA in documentation.

Train element regression/classification models

To prepare training data do the following actions:

• Subsample PubChem using script research/formula_generation/scripts/subsample_pubchem.py and generate list of formulas (RDKit is required)
• Use research/formula_generation/scripts/honest_subsampling.py to subsample data from the formula list
• Generate isotopic distributions using research/formula_generation/scripts/generate_fake_representation_multiprocess.py of its single-process version by running research/formula_generation/scripts/generate_fake_representation.py. These scripts contain some parameters for the spectra generation, which were used in the current study. This may be altered to match your needs ( e.g. lower resolution instrument).

The representations can then be used to train the models. Data, required to reproduce this work, is precalculated and is provided in the training_data.tsv.gz file.

The neural network training code is presented in the research/formula_generation/dl directory. The pretrained model can be loaded:

from mass_automation.formula.model import LSTM

model = LSTM.load_from_checkpoint('pretrained_model.ckpt')

See examples on using the models in research/formula_analysis_examples.

Train deisotoping models

To train your own model you have to create your own artificial spectra with research/deisotoping/generate_mixtures.py. After that, you can create your own dataset for learning with research/deisotoping/generate_dataset.py. See examples on using the models in research/deisotoping/MLDeisotoper_vs_LinearDeisotoper.ipynb.

Sample-oriented analysis

This analysis is useful to differentiate spectra and find similarities in them by applying unsupervised ML techniques. See examples on using the algorithms in research/clustering_examples.

Compound presence verification

This algorithm is useful to find isotopic distribution of ion in spectrum with knowing molecular formula. See examples on using the algorithm in research/compound_presence_verification.

Data requirements

Deisotoping, compound presence verification, sample-oriented analysis

These algorithms are not rely on the fine isotopic structure thus can be performed on usual HRMS spectra. However, the deisotoping models in data folder were trained with measurement mass errors typical for FT-ICR MS. To make model suitable for usual HRMS spectra, you have to train the model on synthetic data with bigger mass measurement errors.

Element classification and element regression

These algorithms heavily rely on the fine isotopic structure, observable by using FT-ICR/MS and Orbitrap instruments. The models can be retrained to match specific requirements of any type of the instruments. If fine isotopic structure is not observable, some elements still can be recognised, but model prediction will be significantly complicated.

Current development plan

• Publish MEDUSA on PyPI
• Add support for lower-resolution instruments
• Add support for MS/MS spectra

Where is the data?

All used data (including files required to run tests) is available on MEGA. The tea dataset is currently used in the ongoing research project and can be provided upon request.

How to cite it?

Boiko D.A., Kozlov K.S., Burykina J.V., Ilyushenkova V.V., Ananikov V.P., "Fully Automated Unconstrained Analysis of High-Resolution Mass Spectrometry Data with Machine Learning", J. Am. Chem. Soc., 2022, ASAP https://doi.org/10.1021/jacs.2c03631