MetFID

Code to train and run MetFID on our benchmark splits. Adapted from the paper Fan, Ziling, et al. "MetFID: artificial neural network-based compound fingerprint prediction for metabolite annotation: Z. Fan et al." Metabolomics 16.10 (2020): 104..

MetFID is a feed-forward network that predicts a 528-bit fingerprint (MACCS + OpenBabel FP3 + FP4) from an MS/MS spectrum. Candidates are ranked by how closely their fingerprint matches the predicted one.

Setup

conda create -n metfid python=3.11 -y
conda activate metfid
pip install torch pytorch-lightning numpy pandas tqdm matchms
conda install -c conda-forge rdkit openbabel

Running

Preprocess spectra. Positive and negative spectra are processed separately. The fold CSV needs a FOLD column with one row per spectrum, in the same order as the MGF.

python prepare_data.py --mgf spectra.mgf --fold_csv folds.csv.gz \
    --val_fold 9 --test_fold 10 --dataset_name spectraverse --out_dir data

Train. A separate model is trained for each ion mode. Test-set predictions from both are combined into <out_dir>/predictions/<dataset_name>_testfold10_valfold9_predictions.csv.

python train_model.py --data_dir data/split_10 --dataset_name spectraverse --out_dir output/split_10

Get fingerprints for the candidate sets. The input is a JSON of {query_smiles: [candidate_smiles, ...]}.

python prepare_candidate_fp.py -i candidates.json -o candidates_metfid.json --skip_failures

Score and rank candidates. Arguments are positional:

python retreival.py preds.csv candidates_metfid.json scores.txt ranks.txt

Each line of the output files is a spectrum title followed by the scores (or ranks) for its candidates. Ties are broken at random.