mtaku3/openscience/nmr-compound-inference

NMR Compound Inference

Use this skill to analyze a local 1D NMR dataset, generate a figure and peak list, search standard reference libraries, and infer the most likely compound or a short ranked candidate list.

Best fit

• 1H and 13C 1D spectra
• Bruker processed pdata, Bruker raw fid, NMRPipe 1D files, or two-column ppm,intensity text exports
• Samples with one dominant component

Required inputs

• local spectrum path
• preferably nucleus, solvent, and field strength
• any sample context that narrows candidate chemistry

Dependencies

Install the dependencies into the active Python environment:

python3 -m pip install nmrglue numpy matplotlib

This skill depends on the upstream nmrglue package and does not bundle the nmrglue source code.

Workflow

1. Run the helper script to generate the evidence package:

python3 scripts/nmr_candidate_report.py \
  --input "<path-to-spectrum>" \
  --output-dir "./outputs/$(date +%F)" \
  --nucleus "1H"

Optional metadata for better matching:

python3 scripts/nmr_candidate_report.py \
  --input "<path-to-spectrum>" \
  --output-dir "./outputs/$(date +%F)" \
  --nucleus "1H" \
  --solvent "D2O" \
  --field-mhz 600

1. Review the generated outputs:

• spectrum.png — Plot of the processed spectrum
• peak_table.csv — Extracted peak positions and intensities
• analysis_summary.json — Structured summary for downstream processing
• analysis_report.md — Human-readable report with search recommendations

1. Read references/standard-libraries.md and search the relevant library sites:

• BMRB Metabolomics: https://bmrb.io/metabolomics/
• HMDB 1D NMR: https://hmdb.ca/spectra/nmr_one_d
• SDBS: https://sdbs.db.aist.go.jp/SearchResult.aspx
• nmrshiftdb2: https://nmrshiftdb.nmr.uni-koeln.de/nmrshiftdbhtml/t1.html

1. Return a report with:

• problem type
• what data was analyzed
• plot path
• major peaks
• library links searched
• top 1-3 candidate compounds
• confidence, mismatch notes, and what would disambiguate the call

Comparison rules

When comparing against reference libraries:

1. Match nucleus first
2. Match solvent second
3. Match field strength third
4. If those do not align, downgrade confidence explicitly
5. If the spectrum suggests multiple components, do not force a single-compound answer

Guardrails

• Do not overclaim exact identity from a weak or low-resolution spectrum.
• Treat this as ranked inference unless the library match is unusually strong.
• Penalize mismatches in nucleus, solvent, field strength, or obvious peak pattern.
• If the input is raw FID and the result is magnitude-mode rather than properly phased, say that explicitly.

Upstream references

• nmrglue GitHub: https://github.com/jjhelmus/nmrglue
• nmrglue documentation: https://nmrglue.readthedocs.io/