TTT-MSMS models

Model checkpoints for the paper "Test-Time Tuned Language Models Enable End-to-end De Novo Molecular Structure Generation from MS/MS Spectra" (Mismetti et al.).

Code: https://github.com/rxn4chemistry/MultimodalAnalytical/tree/ttt-msms

Contents

msg/
├── preprocessor.pkl        # tokenizers/preprocessing, shared by both MSG models
├── pt/                     # pretrained on simulated spectra (MSG test molecules removed)
│   ├── pt.ckpt
│   └── config.yaml
└── ttt-msg/                # test-time trained on MSG, starting from msg/pt
    ├── ttt-msg.ckpt
    └── config.yaml
nplib1/
├── preprocessor.pkl        # tokenizers/preprocessing, shared by both NPLIB1 models
├── pt/                     # pretrained on simulated spectra (NPLIB1 test molecules removed)
│   ├── pt.ckpt
│   └── config.yaml
└── ttt-nplib1/             # test-time trained on NPLIB1, starting from nplib1/pt
    ├── ttt-nplib1.ckpt
    └── config.yaml

Model	Description
msg/pt	Pretrained on simulated MS/MS spectra, with all MSG test-set molecules excluded by 2D InChIKey matching
msg/ttt-msg	msg/pt after test-time training on MSG (seed 3247)
nplib1/pt	Pretrained on simulated MS/MS spectra, with all NPLIB1 test-set molecules excluded by 2D InChIKey matching
nplib1/ttt-nplib1	nplib1/pt after test-time training on NPLIB1 (seed 3247)

Inputs are the MS/MS spectrum and the molecular formula. Outputs are SMILES, generated with beam search.

Each config.yaml is the Hydra config used to train the model. Paths are relative to this folder. Before running, set data_path to your local copy of the dataset.

Usage

See the Github repository for instructions.

License

OpenMDW-1.0 (https://spdx.org/licenses/OpenMDW-1.0.html)