README.md

---
title: MarkushGlyph
emoji: 🧪
colorFrom: gray
colorTo: purple
sdk: gradio
sdk_version: 6.23.1
app_file: app.py
short_description: Extract Markush CXSMILES from chemical structure images
python_version: "3.12"
startup_duration_timeout: 30m
---

# MarkushGlyph — Markush Structure Recognition

This Space demonstrates **MarkushGlyph**, a vision-language model (LoRA fine-tune of
[Qwen3.5-2B-Base](https://huggingface.co/Qwen/Qwen3.5-2B-Base)) that converts images
of patent Markush chemical structures into [CXSMILES](https://docs.chemaxon.com/display/docs/chemaxon-extended-smiles-and-smarts.md) strings.

## How it works

1. Upload an image of a chemical structure (Markush or ordinary molecule).
2. The model extracts the structure and returns it in the XML format
   `<markush><cxsmi>...</cxsmi><stable>...</stable></markush>`.
3. The CXSMILES core and stable-group definitions are displayed separately for
   easy copying.

## Model

- **Base model:** [Qwen/Qwen3.5-2B-Base](https://huggingface.co/Qwen/Qwen3.5-2B-Base)
- **LoRA adapter:** [EdisonScientific/MarkushGlyph](https://huggingface.co/EdisonScientific/MarkushGlyph) (r=128, alpha=128)
- **Code:** [github.com/EdisonScientific/glyph](https://github.com/EdisonScientific/glyph) (Apache-2.0)

## Example images

The example images are from the [glyph repository](https://github.com/EdisonScientific/glyph)
(Apache-2.0 license):

- `markush_m2s_13.png` — a Markush structure crop from the M2S benchmark
- `imatinib.png` — the drug imatinib (C₂₉H₃₁N₇O)
- `cholesterol_ester.png` — a cholesterol ester structure