shield_lock
JinDaGe
← Back to skills
extension
Category: Productivity & OfficeNo API key required

chem-nmr-analysis

Scripts for Wasserstein deconvolution of 1H NMR mixture spectra against reference spectra, reaction product prediction, time-series kinetics, and spectral plotting.

personAuthor: TashanworldhubOpenAPI
open_in_newRepositorydownloadDownload package

NMR Mixture Analysis

When to Use This Skill

The agent should use this skill's scripts when:

  • A workflow (e.g., reaction-to-nmr-quantification.md or nmr-reaction-kinetics.md) calls for deconvolution, product prediction, kinetics analysis, or spectral plotting.
  • The user already has reference spectra and a mixture spectrum and wants to quantify component proportions directly.
  • The user has multiple time-point spectra and wants to track reaction progress via NMR.

For end-to-end workflows that chain this skill with other skills, see: .agents/workflows/reaction-to-nmr-quantification.md and .agents/workflows/nmr-reaction-kinetics.md.

When NOT to Use This Skill

  • 13C NMR, 2D NMR (COSY, HSQC, etc.), or solid-state NMR -- this skill handles 1H solution-state NMR only.
  • Structure elucidation of unknown compounds -- this skill requires knowing (or predicting) what compounds are in the mixture. It does not identify unknowns from scratch.
  • Pure compound characterization -- if the user has a single pure compound and just wants to assign peaks, this skill is not appropriate. The agent should interpret the spectrum directly.
  • Mass spectrometry data -- despite the Wasserstein algorithm's origins in mass spec, this skill operates on NMR chemical shift axes only.
  • Digitizing spectrum images -- the agent should use the general-plot-digitizer skill for that step.
  • Predicting NMR spectra from SMILES -- the agent should use the chem-nmr-predict skill for that step.
  • Resolving compound names to SMILES -- the agent should use the drug-db-pubchem skill for that step.

Scripts Reference

| Script | Purpose | Key Inputs | Key Outputs | |---|---|---|---| | predict_products.py | Predict reaction products via ReactionT5 (HuggingFace API) | --reactant_smiles, --reagent_smiles | JSON with predicted product SMILES | | deconvolve.py | Wasserstein deconvolution of mixture against references | mixture file + reference files + --protons | proportions, Wasserstein distance, plot | | kinetics.py | Time-series deconvolution across multiple time points | --refs, --timepoints, --times | kinetics.csv + kinetics_plot.png | | plot.py | Overlay or stack NMR spectra for visual comparison | spectrum files + --labels | plot image | | spectra.py | I/O utilities (imported by other scripts, not called directly) | -- | -- |

predict_products.py

The agent should use this script to predict reaction products from reactant and reagent SMILES via the ReactionT5 model.

# Env: nmr-agent
export HF_TOKEN=<token>
python .agents/skills/chem-nmr-analysis/scripts/predict_products.py \
  --reactant_smiles "C1CCC(=O)C1" \
  --reagent_smiles "[BH3-]" \
  --output <research_dir>/predicted_products.json

deconvolve.py

The agent should use this script to determine mole fractions of known components in a mixture spectrum via Wasserstein-distance deconvolution.

# Env: nmr-agent
python .agents/skills/chem-nmr-analysis/scripts/deconvolve.py \
  mixture.csv ref_borneol.xy ref_isoborneol.xy \
  --protons 18 18 \
  --names "borneol" "isoborneol" \
  --baseline-correct \
  --plot <research_dir>/deconvolution_result.png \
  --json

kinetics.py

The agent should use this script when the user has crude NMR spectra recorded at multiple time points during a reaction.

# Env: nmr-agent
python .agents/skills/chem-nmr-analysis/scripts/kinetics.py \
  --refs ref1.xy ref2.xy \
  --timepoints t0.csv t10.csv t20.csv \
  --times 0 10 20 \
  --time_unit min \
  --protons 18 18 \
  --names "reactant" "product" \
  --baseline_correct \
  --output_dir <research_dir>/kinetics/

plot.py

The agent should use this script to overlay or stack spectra for visual inspection before or after deconvolution.

# Env: nmr-agent
python .agents/skills/chem-nmr-analysis/scripts/plot.py \
  mixture.csv ref_borneol.xy ref_isoborneol.xy \
  --labels "Mixture" "borneol" "isoborneol" \
  --title "Mixture vs References" \
  --output <research_dir>/spectra_overview.png

Key Arguments for deconvolve.py

| Argument | Required | Description | |---|---|---| | --protons | Yes | Number of 1H protons per molecule for each reference component. Critical for converting area fractions to mole fractions. The agent must look this up from the molecular formula or count from the SMILES. | | --names | No | Human-readable labels matching the order of reference files. The agent should always provide these for interpretable output. | | --baseline-correct | No | Shifts each spectrum so minimum intensity = 0. The agent should use this for digitized spectra or SPINUS-predicted spectra. | | --kappa | No | Denoising penalty (default 0.25). The agent should not change this unless instructed. | | --plot | No | Output plot path. The agent should always generate a plot. | | --json | No | Emit machine-readable JSON output. The agent should always use this. |

Interpreting Results

The deconvolution output contains proportions and a Wasserstein distance (WD) indicating fit quality.

If/Then rules for Wasserstein distance:

  • If WD < 0.05 -- good fit. The agent should report proportions with confidence.
  • If 0.05 < WD < 0.15 -- acceptable fit. The agent should report proportions but note the fit quality and suggest possible causes (minor missing components, baseline noise).
  • If WD > 0.15 -- poor fit. The agent should:
    1. Check if a component is missing (compare overlay plot for unmatched peaks).
    2. Check if there is a ppm calibration offset between mixture and references.
    3. Ask the user if there are additional species in the mixture not accounted for.
    4. Not report proportions as reliable.

If proportions do not sum to ~1.0 -- the agent should note that the "noise" fraction represents unmatched signal and explain what it might be.

Verification: After deconvolution, the agent must inspect the deconvolution plot, check the residual panel for large residuals, and verify that proportions are chemically reasonable. If results contradict known chemistry, the agent should flag this to the user rather than silently accepting.

Input Format Requirements

All spectrum files must be two-column numeric data (ppm, intensity):

  • .csv -- comma-delimited (auto-detected)
  • .xy -- tab-delimited (auto-detected)
  • .tsv -- tab-delimited
  • No header row required; delimiter is auto-detected from content.

If the user provides a Mnova export -- the agent should add --mnova flag to deconvolve.py.

Environment

All scripts in this skill use the nmr-agent conda environment:

mamba activate nmr-agent

Install: conda-envs/nmr-agent/install.sh

Required packages: numpy, scipy (>= 1.7), matplotlib, rdkit, requests, nmrsim, scikit-learn.

HF_TOKEN (for ReactionT5 product prediction): the agent should check if HF_TOKEN is set before attempting product prediction. If not set, the agent should ask the user to provide it or provide product SMILES directly.

Failure Modes

| Failure | Symptom | Agent Action | |---|---|---| | SPINUS returns no atoms | chem-nmr-predict prints FAILED for a compound | The SMILES may be invalid or the molecule too large. The agent should verify the SMILES and retry, or ask the user for a measured reference spectrum. | | ReactionT5 returns no products | predict_products.py returns empty products list | The agent should use its own chemistry knowledge to suggest products and ask the user to confirm. | | Wasserstein distance very high (> 0.15) | Deconvolution result unreliable | Missing component, ppm offset, or baseline issue. The agent should investigate and not report proportions as reliable. | | Proportions are all near zero except one | One component dominates | May be correct (e.g., >95% product), or may indicate missing starting material reference. The agent should check. | | nmrsim simulation fails | Warning in chem-nmr-predict output | Falls back to stick spectrum (shifts only, no multiplet structure). The agent should note reduced accuracy of that reference. | | kinetics curves are non-monotonic | Composition jumps up and down over time | Likely a mislabeled time point, phasing issue, or missing component. The agent should investigate individual spectra. |

References

  • Ciach, M. et al., "Masserstein: linear resampling of mass spectra by optimal transport", Rapid Commun. Mass Spectrom., 2020.
  • Domzal, B. et al., "Magnetstein: Wasserstein-distance NMR mixture analysis", Anal. Chem., 2024.
  • Sagawa, Y. et al., "ReactionT5: a large-scale pretrained model towards chemical reaction prediction", arXiv, 2023.
  • Dhawan, N. et al., "Synthesis of Isoborneol", World J. Chem. Educ., 2022.

Author: Jesus Diaz Sanchez Contact: GitHub @jdsanc