roman-bushuiev / DreaMS

DreaMS (Deep Representations Empowering The Annotation of Mass Spectra)

DreaMS (Deep Representations Empowering the Annotation of Mass Spectra)

Source code for the paper "Emergence of molecular structures from repository-scale self-supervised learning on tandem mass spectra".

This GitHub repository is a work in progress. We are planning to transform it into a user-friendly Python package by the end of May 2024.

How to install

Run the following code from the command line.

git clone https://github.com/roman-bushuiev/DreaMS.git; cd DreaMS
. scripts/install.sh
. scripts/download_models.sh

git clone command will download this GitHub repository and install.sh will install it. download_models.sh script will download pre-trained DreaMS models from Zenodo. The installation script will create a conda environment named dreams. If you are not familiar with conda or do not have it installed, please refer to the official documentation.

How to use

To compute DreaMS representations for MS/MS spectra from .mgf file, run the following Python code.

from dreams.api import compute_dreams_embeddings
dreams_embeddings = compute_dreams_embeddings('data/examples/example_5_spectra.mgf')

The resulting dreams_embeddings object is a matrix with 5 rows and 1024 columns, representing 5 1024-dimensional DreaMS representations for 5 input spectra stored in the .mgf file.

Work in progress

• Wrap the repository into a pip package.
• Import utilities to matchms.
• DreaMS Atlas exploration demo.
• Upload weights of all models.
• Provide scripts to collect/download GeMS datasets.
• Extend dreams.api with more functionality (e.g. attention heads analysis).
• Add tutorial notebooks.
• Upload Murcko splits and detailed tutorial notebook.