README.md

---
license: cc-by-nc-4.0
library_name: pytorch
tags:
- chemistry
- cheminformatics
- optical-chemical-structure-recognition
- ocsr
- molecule-recognition
- smiles
- transformer
- swin-transformer
- minimum-risk-training
- molecular-graph
datasets:
- Keylab/COMO
metrics:
- exact_match
- tanimoto_similarity
---

# COMO: Closed-Loop Optical Molecule Recognition

COMO (Closed-loop Optical Molecule recOgnition) is a deep learning framework for
Optical Chemical Structure Recognition (OCSR). It recognizes chemical structure
diagrams from images and predicts SMILES strings with atom-level 2D coordinates
and bond matrices. COMO uses Minimum Risk Training (MRT) to directly optimize
molecular-level, non-differentiable objectives, closing the gap between
token-level training and molecular-level evaluation.

## Model Summary

- **Architecture:** Swin-B encoder → 6-layer Transformer decoder → bond MLP
- **Input:** 384×384 RGB image of a chemical structure diagram
- **Output:** SMILES string + atom coordinates + bond matrix
- **Vocabulary:** chartok_coords format (200 tokens: SMILES chars + 64 X/Y bins)
- **Parameters:** ~94M
- **Training data:** 1M PubChem + 652K USPTO (MLE) + 83K MolParser-SFT (MRT)

## Available Checkpoints

All checkpoints are from the **joint MLE+MRT** training pipeline (30 epochs,
interleaved MLE/MRT from scratch). Three reward variants are provided:

| Checkpoint | Reward Mode | Description |
|-----------|-------------|-------------|
| `models/tanimoto/final.pth` | Tanimoto | Morgan fingerprint Tanimoto similarity reward |
| `models/tanimoto/best.pth` | Tanimoto | Best validation epoch |
| `models/edit_distance/final.pth` | Edit Distance | Levenshtein string-similarity reward |
| `models/edit_distance/best.pth` | Edit Distance | Best validation epoch |
| `models/visual/final.pth` | Visual | Siamese visual-encoder cosine-similarity reward |
| `models/visual/best.pth` | Visual | Best validation epoch |

## Architecture

```
Image (384×384)
  → Swin-B backbone (ImageNet pretrained)
    → 2D sinusoidal positional encoding
      → 6-layer Transformer decoder (d=256, 8 heads)
        → chartok_coords tokens → SMILES + coordinates
        → Bond MLP (2-layer, GELU) → 7-class bond matrix
          → Graph reconstruction → canonical SMILES
```

The model outputs a molecular graph $G = (A, B)$ where:
- $A = \{(l_i, x_i, y_i)\}$ — atom SMILES labels with 2D image coordinates
- $B$ — pairwise bond types (none, single, double, triple, aromatic, wedge, dash)

## Training

### MLE Phase
- **Data:** 1M PubChem SMILES (synthetic) + 652K USPTO patent molecules
- **Augmentation:** Indigo-rendered images with random styles, functional group
  substitution, R-group insertion, wavy bonds, scan shadows, multilingual comments
- **Optimizer:** AdamW, lr=4×10⁻⁴ (encoder & decoder), weight decay=10⁻⁶
- **Schedule:** 2% linear warmup → cosine decay, batch size 64/GPU
- **Loss:** Label-smoothed cross-entropy (ε=0.1) + bond classification CE

### MRT Phase
- **Data:** 83K real-world molecular images (MolParser-SFT)
- **Candidates:** N=32 per image, multinomial sampling at τ=0.5
- **Reward weights:** validity=0.1, similarity=0.5, exact match=0.4
- **Sharpening:** α=1.0, loss weight λ=0.1
- **Schedule:** First 5 epochs MLE-only warmup, then interleaved MLE+MRT

## Evaluation Results

Exact match accuracy (%) on 10 benchmarks (COMO-Tanimoto variant):

| Benchmark | Images | Synthetic/Real | COMO-Tanimoto |
|-----------|--------|----------------|---------------|
| Indigo | 5,719 | Synthetic | 98.6 |
| ChemDraw | 5,719 | Synthetic | 96.5 |
| CLEF | 992 | Real (patents) | 94.8 |
| JPO | 450 | Real (patents) | 88.4 |
| UOB | 5,740 | Real (academic) | 98.0* |
| USPTO | 5,719 | Real (patents) | 93.4 |
| USPTO-10K | 10,000 | Real (patents) | 96.1 |
| Staker | 50,000 | Real | 87.4 |
| ACS | 331 | Real (publications) | 84.6 |
| WildMol-10K | 10,000 | Real (wild) | 77.1 |

*\*UOB results after tautomer standardization.*

See the [paper](#citation) for full comparison with MolScribe, MolParser,
SwinOCSR, and other baselines.

## Installation

```bash
pip install como-ocsr
```

## Usage

```python
import como

# Download checkpoint from HuggingFace:
# huggingface-cli download Keylab/COMO models/tanimoto/final.pth

model = como.load_model("models/tanimoto/final.pth", device="cuda")

# Single image prediction
smiles = como.predict(model, "molecule.png")
print(smiles)  # "CC(=O)O"

# Batch prediction
smiles_list = como.predict_batch(model, ["mol1.png", "mol2.png"])

# Benchmark evaluation
metrics = como.evaluate(model, "benchmark/USPTO/", "benchmark/USPTO.csv")
print(f"Exact Match: {metrics['postprocess/exact_match_acc']:.2%}")
```

Full documentation: [Github](https://github.com/netknowledge/COMO) [PyPI](https://pypi.org/project/como-ocsr/)

## Benchmarks

Benchmark datasets are available in the `benchmarks/` directory of this
repository. Each dataset contains `.png` images and a CSV file with columns
`image_id` and `SMILES`.

**Note:** These benchmarks are collected from existing public OCSR datasets.
Please refer to the original sources for attribution:

| Dataset | Source |
|---------|--------|
| USPTO, CLEF, JPO, UOB, Staker | [Rajan et al., 2020](https://github.com/Kohulan/OCSR_Review), [Xiong et al., 2023](https://github.com/jiachengxiong/alpha-Extractor) |
| Indigo, ChemDraw, ACS, Staker | [Qian et al., 2023](https://github.com/thomas0809/MolScribe) |
| USPTO-10K | [Morin et al., 2023](https://huggingface.co/datasets/docling-project/USPTO-30K) |
| WildMol-10K | [Fang et al., 2025](https://github.com/orgs/Chem-Struct-ML/repositories) |

## License

- **Model Weights:** CC BY-NC 4.0 (non-commercial use only)
- **Code:** MIT License
- **Benchmarks:** See original sources for applicable terms

## Citation

```bibtex
@article{lyu2026closed,
  title={COMO: Closed-Loop Optical Molecule Recognition with Minimum Risk Training},
  author={Lyu, Zhuoqi and Ke, Qing},
  journal={arXiv preprint arXiv:2604.23546},
  year={2026}
}
```