IDSL_MINT

IDSL_MINT: Mass spectra INTerpretation by the Integrated Data Science Laboratory for Metabolomics and Exposomics (IDSL.ME) is a transformative mass spectrometry data processing framework. This innovative approach for mass spectrometry data processing has been constructed upon the transformer models delineated in the seminal paper, 'Attention is all you need'. IDSL_MINT has been meticulously engineered to predict molecular fingerprint descriptors and structures from MS/MS spectra in addition to forecasting MS/MS spectra from canonical SMILES. A key distinguishing feature of IDSL_MINT is its compatibility with any reference MS/MS data in .msp format to tailor IDSL_MINT models for various applications.

Table of Contents

• Features of IDSL_MINT
• Installation
• Workflow
• IDSL_MINT: Translating MS/MS Spectra into Molecular Fingerprints
• IDSL_MINT: Translating MS/MS Spectra into Canonical SMILES
• IDSL_MINT: Transforming Fingerprints into MS/MS Fragments
• Citation

Features of IDSL_MINT

1. Parameter selection for training and prediction through user-friendly and well-documented YAML files
2. Compatibility with .msp file formats.
3. Compatibility with various fingerprint descriptor methods.
4. Supports beam search inferencing.
5. Utilizes the power of the transformer model architecture.
6. Device-agnostic processing.

Installation

1. Installation of Prerequisites:

   a. Install PyTorch according to your system configurations. IDSL_MINT is device-agnostic and fully supports cuda GPU processing.

   b. Install RDKit.

2. Install the package:

   2.1. Option 1: pip

   • pip install git+https://github.com/idslme/IDSL_MINT
   • pip install IDSL_MINT

   2.2. Option 2: conda

   • git clone https://github.com/idslme/IDSL_MINT.git
   • cd IDSL_MINT
   • conda env create -f environment.yml
   • conda activate IDSL_MINT
   • pip install -e .

3. Update the Python PATH:

   export PATH="root/.local/bin:$PATH" --> root directory should be your system root directory.

Workflow

The IDSL_MINT framework encapsulates three transformative approaches to deeply interpret mass spectrometry data. Each of these methodologies can be effectively managed using designated model configuration yaml files. In the training step, weights of IDSL_MINT models are stored and updated in a designated directory on the decreasing trajectory of the training loss value to ensure optimal performance and accuracy. The yaml files are easy to update and model configuration is significantly simplified and commented. After configuring the model in the designated yaml file, run the below bash command to perform calculations. The IDSL_MINT package can automatically detect types of yaml file to run training or inference operations.

MINT_workflow --yaml /path/to/yaml/file

Important tips:

• IDSL_MINT can extract information from comment: and comments: entries in .msp files which enables this platform to process MoNA, GNPS, and other public library with any pre-treatment requirements.

• IDSL_MINT identifies chemical structures through SMILES: or InChI: labels without case sensitivity.

• In case multiple similar headers are present in a MSP block, the one with the longest content is selected for parsing.

• MSP blocks must include PrecursorMZ: row entries.

IDSL_MINT: Translating MS/MS Spectra into Molecular Fingerprints

IDSL_MINT includes a method to translate MS/MS spectra into molecular fingerprint descriptors. This method offers the option to calculate fingerprints using the Extended-connectivity fingerprints (ECFPs) or MACCS Keys RDKit methods from InChI and SMILES row entries. Another option to obtain molecular fingerprints is to parse the MSP files for the user-provided fingerprints. The following is an example of an Aspirin MSP block with custom fingerprint bits.

Name: Aspirin
Fingerprint: 15-53-85-157-246-322-329-343-444-464-553-708-763-785-799-821-847-1040-1139-1240-1250-1317-1348-1439-1450-1460-1475-1479-1502-1674-1693-1734-1841-1866-2046-2310-2329-2413-2627-2750-2755-2777-2782-2799-2901-2911-2915-3028-3049-3394-3412-3442-3514-3535-3557-3700-3737-3785-3972-3996
Synon: Acetyl salicilic acid
Synon: 2-acetyloxybenzoic acid
InChI: InChI=1S/C9H8O4/c1-6(10)13-8-5-3-2-4-7(8)9(11)12/h2-5H,1H3,(H,11,12)
Precursor_type: [M+H]+
Spectrum_type: MS2
PrecursorMZ: 181.0495
Instrument_type: LC-ESI-QFT
Instrument: Q Exactive Plus Orbitrap Thermo Scientific
Ion_mode: P
Collision_energy: 15 (nominal)
Formula: C9H8O4
MW: 180
ExactMass: 180.042258736
Num Peaks: 10
65.0385 0.217327
76.0304 0.107699
77.0383 0.124517
92.0255 0.129908
121.0283 0.125197
133.028 0.149192
149.0231 100.000000
163.0386 63.824575
167.0337 0.261816
181.0493 0.613766

Fingerprint row entries may be in any line in MSP blocks between Name and Num Peaks rows, and fingerprint bits must be dash-separated. This example represented Avalon fingerprint bits with nBits = 4096 for Aspirin MS/MS spectra.

To train an IDSL_MINT model with molecular fingerprint descriptors, download and fill a MINT_MS2FP_trainer.yaml file. Similarly, for model prediction, use MINT_MS2FP_predictor.yaml file.

A colab notebook was presented to demonstrate the performance of IDSL_MINT in training and predicting molecular fingerprint descriptors using MS/MS data.

IDSL_MINT: Translating MS/MS Spectra into Canonical SMILES

In this approach, InChI and SMILES row entries in the MSP blocks are converted into canonical SMILES using RDKit. Next, standard canonical SMILES are tokenized using a method similar to RXNFP. As long as InChI and SMILES row entries present in the MSP blocks are available, this approach may be used to train an IDSL_MINT model.

To train an IDSL_MINT model to predict molecular structures from MS/MS spectra, download and fill a MINT_MS2SMILES_trainer.yaml file. Likewise, for model prediction, use MINT_MS2SMILES_predictor.yaml file.

A colab notebook was presented to demonstrate the performance of IDSL_MINT in training and predicting canonical SMILES using MS/MS data.

IDSL_MINT: Transforming Fingerprints into MS/MS Fragments

This method is designed to translate fingerprints into MS/MS fragments using a transformer model. This approach contrasts with previous methods that predict fragment mass from fingerprints.

To train an IDSL_MINT model to predict MS/MS spectra from molecular structures, download and fill a MINT_FP2MS_trainer.yaml file. Likewise, for model prediction, use MINT_FP2MS_predictor.yaml file.

Citation

[1] Fakouri Baygi, S., Barupal, D.K. IDSL_MINT: a deep learning framework to predict molecular fingerprints from mass spectra. Journal of Cheminformatics, 2024, 16(8).