unknown-chemical-identification-in-complex-mixtures
Summary
Apply a deep learning model (MSGO) trained on pseudo SMILES-spectrum pairs to predict molecular structures from experimental mass spectra of unknown compounds in complex environmental samples. This skill enables structure elucidation of novel or unmeasured chemicals in real LC–QTOF wastewater datasets where reference libraries are incomplete.
When to use
You have experimental LC–QTOF mass spectra from complex environmental samples (e.g., wastewater) containing unknown or novel chemicals not present in reference databases, and you want to generate candidate molecular structures ranked by model confidence to guide laboratory verification or literature validation.
When NOT to use
- Input spectra are already matched to a curated reference library with high confidence — use database lookup instead.
- Spectra are from low-resolution instruments (e.g., ESI-quadrupole) lacking the mass accuracy and fragmentation detail needed for LC–QTOF-trained models.
- You require identified compounds only; novel structure predictions without experimental corroboration are unsuitable for your use case.
Inputs
- LC–QTOF mass spectra in CSV format (m/z, intensity pairs or tabular format with spectrum ID)
- Pre-trained MSGO model weights (for pfas or lipid variant)
- Polarity parameter (neg or pos) matching the ionization mode
Outputs
- CSV file containing top 10 predicted SMILES structures per spectrum with confidence scores
- Ranked molecular structure candidates for each unknown compound
How to apply
Load the pre-trained MSGO model (Python 3.7, Torch 1.7.1) and your experimental LC–QTOF spectra in CSV format. Preprocess the spectra to match the model's input format (normalize intensity, ensure m/z alignment). Run inference using the appropriate model variant (pfas for negative polarity, lipid for positive polarity) with a specified beam search size (typically 300–500) to generate top-k candidate SMILES structures ranked by model confidence score. Collect and rank predictions; validate top candidates against available reference compounds or literature annotations, or flag high-confidence novel predictions for orthogonal verification (e.g., NMR, synthesis).
Related tools
- MSGO (Pre-trained deep learning model for predicting SMILES from LC–QTOF mass spectra; loaded and invoked for inference on real wastewater spectra) — https://github.com/aaronma2020/MSGO
- cfmid (Fragmentation prediction tool used to generate 30k+ pseudo SMILES-spectrum training pairs)
- Python (Runtime environment (version 3.7) for spectrum preprocessing and model inference scripting)
- Torch (Deep learning framework (version 1.7.1) required to load and execute the MSGO model)
Examples
python tools/eval_standard.py --log_path ckpts/pfas --real_csv ./data/example/pfas.csv --out_csv ./pfas_results.csv --beam_size 500 --polar neg
Evaluation signals
- Output CSV contains top 10 predictions per spectrum with non-null confidence scores in descending order
- Predicted SMILES strings are chemically valid and parseable by standard cheminformatics tools (RDKit)
- For spectra with known reference compounds, predicted structures rank reference compound SMILES within top 5–10 predictions
- Confidence scores reflect model uncertainty; high-confidence predictions (e.g., top 1–3) show consistency with experimental m/z accuracy and fragmentation patterns characteristic of LC–QTOF
- Novel predictions flagged for manual review can be cross-validated against literature annotations or standard compound databases where available
Limitations
- Model is trained on pseudo SMILES-spectrum pairs generated by cfmid; performance on true unknowns depends on how well cfmid fragmentation patterns match real experimental spectra.
- Predictions are class-biased toward training data: the pfas variant is optimized for per- and polyfluoroalkyl substances, lipid variant for lipids; performance on structurally diverse unknowns is not guaranteed.
- Real-sample validation requires reference annotations or orthogonal methods (NMR, synthesis); model confidence alone does not confirm structure identity.
- Input spectra must match LC–QTOF preprocessing standards (intensity normalization, m/z calibration); raw instrument output may require format conversion.
Evidence
- [other] Can the MSGO model trained on pseudo SMILES-spectrum pairs successfully elucidate molecular structures in real wastewater samples from an LC–QTOF dataset?: "Can the MSGO model trained on pseudo SMILES-spectrum pairs successfully elucidate molecular structures in real wastewater samples from an LC–QTOF dataset?"
- [other] Perform inference using MSGO to generate predicted SMILES structures and corresponding confidence scores for each spectrum.: "Perform inference using MSGO to generate predicted SMILES structures and corresponding confidence scores for each spectrum."
- [readme] For Training, we use 30k+ pseudo smiles-specturm pairs generated by cfmid: "For Training, we use 30k+ pseudo smiles-specturm pairs generated by cfmid"
- [readme] For evaluation in real samples，we use one LC–QTOF dataset for wastewater samples to verify our model: "For evaluation in real samples，we use one LC–QTOF dataset for wastewater samples to verify our model"
- [readme] python tools/eval_standard.py --log_path ckpts/pfas --real_csv ./data/example/pfas.csv --out_csv ./pfas_results.csv --beam_size 500 --polar neg: "python tools/eval_standard.py --log_path ckpts/pfas --real_csv ./data/example/pfas.csv --out_csv ./pfas_results.csv --beam_size 500 --polar neg"
- [other] Validate predictions against reference compounds or literature annotations where available, or flag novel structure predictions for further verification.: "Validate predictions against reference compounds or literature annotations where available, or flag novel structure predictions for further verification."
- [readme] Then you can obatin a results csv file inluding top 10 predicts.: "Then you can obatin a results csv file inluding top 10 predicts."