Efficiently predicting high resolution mass spectra using graph neural networks

This repository is supporting material for our ICML 2023 paper. We provide code to reproduce our results here, public-domain datasets we used + processing notebooks, and trained model weights for GrAFF-MS pluis our NEIMS implementation. Dependencies are listed in requirements.txt.

To just run GrAFF-MS using our trained weights:

python run-graff-ms.py lightning_logs/graff-ms/version_0/checkpoints/epoch=96-step=27257.ckpt example_input.tsv example_output.msp --min_probability 0.001

Each line of example_input.tsv yields an individual spectrum in example_output.msp, and should be four tab-delimited fields: a unique identifier for the spectrum, a SMILES string, a precursor type ([M+H]+ or [M-H]-) and a normalized collision energy (see Thermo Scientific NCE technical note). The --min_probability flag is an intensity cutoff to keep the output files small.

We do not include NIST-20 or files derived from it (apart from our train-test splits and model weights) for licensing reasons. NIST-20 can be purchased from GC Image ("MS/MS only" option) and extracted into a file hr_msms_nist.MSP using the Library Conversion Tool for MS Windows, making sure "2008 MS Search compatible" is turned off.

Once hr_msms_nist.MSP has been generated, and placed in ./data/nist-20/, it must be parsed into a dataframe:

python preprocess-nist.py data/nist-20/hr_msms_nist.MSP data/nist-20/inchikeys.pkl

To then train GrAFF-MS:

python train-graff-ms.py data/nist-20/hr_msms_nist.pkl

And to train the NEIMS baseline:

python train-neims.py data/nist-20/hr_msms_nist.pkl

To reproduce our evaluations, first run the preprocessing notebooks in the ./data/ subfolders:

casmi-16/casmi-16.ipynb
chembl/chembl.ipynb
gnps/gnps.ipynb
cfm-id/cfm-id.ipynb

Then run evaluate-models.ipynb.

Note: this codebase is provided for the purpose of reproducing the results shown in our ICML manuscript. If you are a user of mass spectrometry interested in a spectrum prediction model with a more polished interface, consider the excellent work of our colleagues in the Coley group: https://github.com/samgoldman97/ms-pred

murphy17/graff-ms

Efficiently predicting high resolution mass spectra using graph neural networks