Skip to content

INPUT_FORMAT.md

Latest commit

 

History

History
91 lines (68 loc) · 4.1 KB

INPUT_FORMAT.md

File metadata and controls

91 lines (68 loc) · 4.1 KB
Raw

Docs: README · Scope & limitations · Input format · Usage notes · Data schema

Running MolDeTr on your own spectra

MolDeTr works on one spectral window at a time. It does not read raw vendor files, pick peaks, or choose regions for you: you hand it a single preprocessed window and it returns the multiplets in it. To use your own data, resample each region to the format below.

Tip

The bundled examples/ files already follow this contract, and roi_S10_example.npz is a good template for your own exporter. Smoke-test with:

python scripts/predict.py --input examples/roi_S10_example.npz --plot

The input contract

Property Requirement
Length exactly 6144 points
Digital resolution 5.12 points/Hz (a 1200 Hz window)
Values real (absorption-mode). Complex input is reduced to its real part.
Intensity any global scale: each spectrum is min–max normalised, so overall receiver gain does not matter and no integral reference is needed. Relative intensities, SNR and line shape still matter.
Field strength not an input. MolDeTr works in Hz (largely field-agnostic), so 80–600 MHz spectra all map to the same 1200 Hz window.

moldetr/validation.py enforces the hard parts (length, finiteness); predict.py and the GUI call it and tell you exactly what to fix.

Why 1200 Hz

6144 points ÷ 5.12 points/Hz = 1200 Hz. In ppm that width depends on the spectrometer:

Field 1200 Hz in ppm
80 MHz ~15 ppm
300 MHz ~4 ppm
500 MHz ~2.4 ppm
600 MHz ~2 ppm

Choose a region no wider than 1200 Hz.

The rule that trips people up

Important

Keep coupling partners together. A window does not have to contain the whole molecule; it can hold several unrelated spin systems. But every proton that couples to a proton inside the window must also be inside the window. If a multiplet's coupling partner falls outside the region, the observed splitting points to a peak the model cannot see, and that multiplet will be predicted wrong. When you draw a region, make sure each spin system in it is complete.

Preparing a window

  1. Process the FID (the raw time-domain signal) to a phased, baseline-corrected real 1D spectrum. Any reader works: the open-source nmrglue library reads Bruker, Varian/Agilent, JCAMP, and more.
  2. Pick a region ≤ 1200 Hz wide that fully contains the spin systems you want.
  3. Resample that region to 5.12 points/Hz.
  4. Zero-pad (or crop) to 6144 points.
  5. Save it as a .npz (see keys below), or a plain .npy of the 6144-point array.
Bruker TopSpin recipe: phase and baseline before you export

MolDeTr is sensitive to phase and baseline distortion beyond its trained range, so get a clean absorption spectrum first:

  1. Fourier transform: ft (or efp with a window function).
  2. Phase: apk (automatic). Use apk0/apk1 for zero-/first-order only, apks for a robust variant, or apkf to phase on a region; fine-tune manually if the baseline still rolls.
  3. Baseline: absn (automatic, node-based; usually the best) or abs over the region; bcm and sab are alternatives. A flat baseline matters more than it looks.
  4. Export the real spectrum (e.g. read pdata/1 with nmrglue), then resample to 5.12 points/Hz and crop/pad to 6144 as above.

Heavy phase or baseline distortion is out of distribution and degrades the prediction; see SCOPE.md.

What the loaders read from an .npz

  • spectrum_padded (or spec): the 6144-point real spectrum. Required.
  • ppm_axis_padded: the per-point ppm axis; its first and last values calibrate the plot so shifts come out in ppm. Optional: without it, shifts are reported in Hz and points, or you can pass --ppm-left/--ppm-right to predict.py.
  • Anything else in the file is ignored.

Next: Usage notes explains how to read the output and how it fails.