File size: 2,061 Bytes
426041c
 
 
537e379
426041c
537e379
 
 
 
 
 
 
426041c
 
537e379
426041c
537e379
 
 
426041c
537e379
 
 
426041c
537e379
 
 
 
426041c
 
537e379
 
 
 
 
 
 
 
 
426041c
537e379
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
---
license: apache-2.0
library_name: pytorch
inference: false
tags:
  - nmr
  - nmr-spectroscopy
  - spectroscopy
  - chemistry
  - cheminformatics
  - deformable-detr
  - object-detection
---

# MolDeTr — chemistry-informed deep learning for ¹H NMR multiplet detection

MolDeTr is a 1D Deformable-DETR that reads a ¹H NMR spectrum window and returns the spin systems in it
directly: for each group of equivalent protons it gives the chemical shift (δ), the coupling (J), the
proton count, and the line width — in one forward pass, with no prior structure and no iterative fitting.

- **Paper:** [*Analytical Chemistry*, 2026](https://doi.org/10.1021/acs.analchem.5c03465)
- **Code:** <https://github.com/smidooo/MolDeTr>
- **Canonical release (weights + data):** Zenodo [10.5281/zenodo.21217101](https://doi.org/10.5281/zenodo.21217101). **This repo mirrors the checkpoint from that deposit for convenience — Zenodo is authoritative.**

## What's here

One file: `model_spin_system_ABCDEFG_exp2.pth` (~974 MB), the trained checkpoint. It is byte-identical to
the file in the Zenodo deposit (MD5 `faf842d1a1d8beae67e0544e28f226b5`).

## Usage

The model is custom (a 1D detection transformer), so it runs through the repo code rather than a standard
`transformers` pipeline:

```bash
git clone https://github.com/smidooo/MolDeTr && cd MolDeTr
pip install -e .
huggingface-cli download smidooo/moldetr model_spin_system_ABCDEFG_exp2.pth --local-dir moldetr/model
python scripts/predict.py --demo          # or: python app.py  (Gradio Detect + Simulate app)
```

See the [repository README](https://github.com/smidooo/MolDeTr) for the input contract (a 6144-point,
5.12 points/Hz, 1200 Hz window) and the interactive app.

## Benchmark

On the experimental test set (13 ROIs across 12 spectra, 80–600 MHz, vs. ground truth): median |Δδ|
**0.89 Hz**, median |ΔJ| **0.20 Hz**, and **93.5 %** proton-count accuracy.

## License & citation

Apache-2.0. If MolDeTr helps your work, please [cite the paper](https://doi.org/10.1021/acs.analchem.5c03465).