README.md

metadata

title: MVP
emoji: 🏆
colorFrom: blue
colorTo: pink
sdk: streamlit
app_file: app.py
pinned: false
short_description: msms annotation tool
python_version: 3.11.7

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

🏆 MultiView Projection (MVP) for Spectra Annotation

Authors

Yan Zhou Chen, Soha Hassoun
Department of Computer Science, Tufts University

---

MVP is a framework for ranking molecular candidates given a spectrum. This repository provides the official implementation, pretrained models, and utilities for data preparation and training.

---

📑 Table of Contents

0. Quick Test
1. Install & Setup
2. Data Preparation
3. MassSpecGym Data Download
4. Using the Pretrained Model
5. Training from Scratch
6. References

---

🚀 Quick Test

Run MVP instantly with our interactive app for small-scale experiments.

---

⚙️ Install & setup

1. Clone the repository: git clone https://huggingface.co/spaces/HassounLab/MVP/
2. Install evironment or only key packages:

conda create -n mvp python=3.11
conda activate mvp
pip install -r requirements.txt

Key packages

• python
• dgl
• pytorch
• rdkit
• pytorch-geometric
• numpy
• scikit-learn
• scipy
• massspecgym
• lightning

---

📂 Data prep

We provide sample spectra data and candidates in data/sample. For preprocessing:

1. If using formSpec, compute subformula labels
2. Run our preprocess code to obatain fingerprints and consensus spectra files

# If using formSpec
python subformula_assign/assign_subformulae.py --spec-files ../data/sample/data.tsv --output-dir ../data/sample/subformulae_default --max-formulae 60 --labels-file ../data/sample/data.tsv
python data_preprocess.py --spec_type formSpec --dataset_pth ../data/sample/data.tsv --candidates_pth  ../data/sample/candidates_mass.json --subformula_dir_pth ../data/sample/subformulae_default/ --output_dir ../data/sample/

# If using binnedSpec
python data_preprocess.py --spec_type binnedSpec --dataset_pth ../data/sample/data.tsv --candidates_pth  ../data/sample/candidates_mass.json --output_dir ../data/sample/

We include sample subformula, fingerprint, and consensus spectra data in ../data/sample/.

Use our pretrained model

You can use our pretrained model (on MassSpecGym) to rank molecular candidates by providing the spectra data and a list of candidates.

After prepping your data, modify the params_binnedSpec.yaml or params_formSpec.yaml with your dataset paths:

# If using formSpec
python test.py --param_pth params_formSpec.yaml

# If using binnedSpec
python test.py --param_pth params_binnedSpec.yaml

We provide a notebook showing sample result files in notebooks/demo.ipynb

MassSpecGym data download

Our model is trained on MassSpecGym dataset. Follow their instruction to download the spectra and candidate dataset.

You can preprocess the MassSpecGym dataset as descirbed in the above section or download the preprocessed files as follows:

mkdir data/msgym/
cd data/msgym
wget https://zenodo.org/records/15223987/files/msgym_preprocessed.zip?download=1

Training from scratch

To train a model from scratch:

1. Prepare data as described in the data prep section
2. Modify the configuration in params file as necessary
3. Train using the following

# If using formSpec
python train.py --param_pth params_formSpec.yaml

# If using binnedSpec
python train.py --param_pth params_binnedSpec.yaml

---

📚 References

Preprint:Learning from All Views: A Multiview Contrastive Framework for Metabolite Annotation

---

📧 Contact

For questions, reach out to: [email protected]

=======