We have released a trained MIST-CF model (using the public NPLIB1/CANOPUS dataset). This can be downloaded (quickstart/download\_model.sh) and used to predict a set of 10 spectra from CASMI22, as included at data/demo_specs.mgf using the following commands:

Model output will be saved in quickstart/mist_cf_out/. This model may be less performant than the model trained on the commercial NIST20 Library (particularly for Orbitrap or higher resolution data). Download links to models trained on NIST20 models are available upon reasonable request to users with a NIST license.

MIST-CF scores the agreement between a precursor formula candidate and an unknown spectrum. Refer to notebooks/demo_mist_cf.ipynb for an interactive demo of how to use MIST-CF and also set user-defined formula candidates.

Four key datasets were used in the process of this paper:

1. biomols: A dataset of biologically relevant molecules that we used to learn a fast filter model
2. NPLIB1: A public natural products dataset extracted from the GNPS database. NPLIB1 is used for model training and evaluation.
3. NPLIB1 + NIST: A proprietary dataset of NPLIB1 and NIST20 spectra. This dataset is used for model training and evaluation.
4. CASMI 2022: A dataset of positive mode spectra from the CASMI 2022 challenge. This dataset is used for prospective analysis.

We provide instructions for downloading and extracting datasets biomols, NPLIB1, and CASMI 2022 as we preprocessed them.

For a more in depth view and understanding of our preprocessing pipeline, we refer the reader to preprocessing/run_all.sh and other relevant preprocessing scripts.

Fast filter model is trained using a large database of molecules and chemical formula extracted from varous sources prepared by Duhrkop et al..

NPLIB1 is a public natural products dataset extracted from the GNPS database. NPLIB1 is used for model training and evaluation. We refer to this in the directory structure as "canopus_train"

CASMI-2022 dataset is a well-accepted and recent benchmark.

Most of our data processing pipelines refer to a dataset nist_canopus, that includes both NIST and NPLIB1 jointly. The following instructions provide a simple demo of how to train a model on the NPLIB1 only (canopus_train) alone.

Training a fast filter model:

Training a mist cf model:

The exact source locations for each of the training and predict files is also listed for reference:

FastFilter

1. Train: src/mist_cf/fast_form_score/train.py
2. Predict: src/mist_cf/fast_form_score/predict.py

MIST-CF

1. Train: src/mist_cf/mist_cf_score/train.py
2. Predict: src/mist_cf/mist_cf_score/predict.py
3. Prospective analysis: src/mist_cf/mist_cf_score/predict_mgf.py

FFN

1. Train: src/mist_cf/ffn_score/train.py
2. Predict: src/mist_cf/ffn_score/predict.py

Xformer

1. Train: src/mist_cf/xformer_score/train.py
2. Predict: src/mist_cf/xformer_score/predict.py

The entries below provide a record of our experiments used in our initial MIST-CF paper. While not intended to be re-run exactly as shown below due to the requirement of the nist_canopus subfolder and data, these demonstrate the parameters and call signatures we utilized.

Evaluate fast filter performance of reducing formula candidates.

Experiment pipeline:

1. Hyperopt fast filter model: run_scripts/hyperparams/find_fast_params.sh
2. Edit fast filter config file: configs/fast_filter.yaml
3. Train fast filter model: run_scripts/fast_filter/launch_fast_train.sh
4. Evaluate fast filter prediction: run_scripts/fast_filter/launch_fast_pred.py

Benchmark MIST-CF performance with baseline models.

Experiment pipeline:

1. Hyperopt all models
   • run_scripts/hyperparams/find_mist_cf_params.sh
   • run_scripts/hyperparams/find_ffn_params.sh
   • run_scripts/hyperparams/find_xformer_params.sh
2. Edit model config files:
   • configs/mist_cf_canopus.yaml
   • configs/ffn_canopus.yaml
   • configs/ms1_canopus.yaml (hyperparameter same as ffn)
   • configs/xformer_canopus.yaml
3. Train models:
   • run_scripts/benchmarking/train_mist_cf.sh
   • run_scripts/benchmarking/train_ffn.sh
   • run_scripts/benchmarking/train_ms1.sh
   • run_scripts/benchmarking/train_xformer.sh
4. Evaluate models: run_scripts/benchmarking/eval_models.py

Show that few MS2 peaks are sufficient to learn candidate formula ranking.

Experiment pipeline:

1. Edit MS2 peak config file: configs/mist_cf_canopus_nist_max_subpeak.yaml
2. Train models: run_scripts/max_subpeak/train_mist_cf_subpeak.sh
3. Evaluate models: run_scripts/max_subpeak/eval_models.py

Compare MIST-CF and SIRIUS using a single split of NPLIB1 test data.

Experiment pipeline:

1. Run SIRIUS prediction: run_scripts/sirius_compare/sirius_1_run.py
2. Format SIRIUS output: run_scripts/sirius_compare/sirius_2_wrangle.py
3. Run MIST-CF prediction: run_scripts/sirius_compare/mist_cf_1_predict.sh
4. Evaluate performance: run_scripts/sirius_compare/eval_models.py

Compare MIST-CF and SIRIUS on CASMI-2022.

Experiment pipeline:

1. Run SIRIUS prediction: run_scripts/casmi22_eval/run_sirius.sh
2. Format SIRIUS output: run_scripts/casmi22_eval/wrangle_sirius.py
3. Run MIST-CF prediction: run_scripts/casmi22_eval/run_mist_cf.sh
4. Evaluate performance: run_scripts/casmi22_eval/eval_models.py

Analysis scripts can be found in analysis/ for evaluating model predictions analysis/evaluate_pred.py.

Additional analyses used for figure generation were conducted in notebooks/.

If you use this repository, please consider citing both our work and the original SIRIUS papers, as we still rely on their deterministic tool for formula enumerations:

1. Bocker, Sebastian, et al. "SIRIUS: decomposing isotope patterns for metabolite identification." Bioinformatics 25.2 (2009): 218-224.
2. Duhrkop, Kai, et al. "SIRIUS 4: a rapid tool for turning tandem mass spectra into metabolite structure information." Nature methods 16.4 (2019): 299-302.