README.md

---
title: TwinBooster
emoji: 👁️
colorFrom: gray
colorTo: indigo
sdk: gradio
sdk_version: 4.44.1
app_file: app.py
pinned: false
python_version: 3.10
---

## Local development

Install dependencies (Python 3.9+ recommended):

```
py -3 -m pip install -r requirements.txt
```

Run the Gradio demo:

```
py -3 app.py
```

Usage:
- Paste SMILES (one per line) and a bioassay description, then click **Run prediction**.
- Use **Download / refresh models** if the pretrained weights are not present yet.
- Download predictions as CSV or Excel from the output panel.

Weights:
- All weights are pulled from public Hugging Face repos on first run and cached under `~/.cache/twinbooster`:
  - Text encoder: `mschuh/PubChemDeBERTa-augmented`
  - LightGBM model: `HF_LGBM_REPO` env var (defaults to `mschuh/TwinBooster-lgbm`)
  - Barlow Twins model: `HF_BT_REPO` env var (defaults to `mschuh/TwinBooster-bt`)
- If the repos are public on HF LFS, no additional setup is required on Spaces; downloads are automatic and reused on subsequent runs.

## ZeroGPU setup (Hugging Face Spaces)
- In the Space settings, choose the **ZeroGPU** hardware tier and Python **3.10** or **3.12** (both supported on ZeroGPU).
- The GPU workload is isolated in `_gpu_predict`, decorated with `@spaces.GPU(duration=120)`, so a GPU is provisioned only for the inference step.
- Queueing is enabled with a single worker (`demo.queue(concurrency_count=1)`) to match the one-at-a-time ZeroGPU execution model.
- To avoid spending GPU time on downloads, click **Download / refresh models** once after each deployment to prefetch weights on CPU.
- If inference ever needs more time, adjust the `duration` parameter in `app.py`, but keep it as low as practical to respect ZeroGPU queue fairness.
- A lightweight, bundled `twinbooster` shim (version `0.3.1`) is shipped in `./twinbooster/` to avoid PyPI's Python 3.8 restriction on the official wheels. No external install is required; the public API used by the app is preserved.
- `huggingface_hub` is pinned to `<1.0.0` because Gradio OAuth still imports `HfFolder`, which was removed in 1.x; keep this constraint to avoid startup import errors on ZeroGPU.