metadata

title: 🧪 HOMO‑LUMO Gap Predictor
emoji: 🧬
colorFrom: indigo
colorTo: purple
sdk: docker
app_file: app.py
pinned: false

HOMO‑LUMO Gap Predictor — Streamlit + Hybrid GNN

Live demo: huggingface.co/spaces/MooseML/homo-lumo-gap-predictor
Code: https://github.com/MooseML/homo-lumo-gap-predictor

This web app predicts HOMO–LUMO energy gaps from molecular SMILES using a trained hybrid Graph Neural Network (PyTorch Geometric and RDKit descriptors).
It runs on Hugging Face Spaces via Docker and on any local machine with Docker or a Python ≥ 3.10 environment.

Features

Predict gaps for single or batch inputs (comma / newline SMILES or CSV upload)
Shows up to 20 molecules per run with RDKit 2‑D depictions
Download full predictions as CSV
Logs all predictions to a lightweight SQLite DB (/data/predictions.db)
Containerised environment identical to the public Space

Quick start

1 Use the hosted Space

Open the live URL.
Paste SMILES or upload a CSV (1 column, no header).
Click Run Prediction → results and structures appear; a CSV is downloadable.

Heads‑up: on the free HF tier large files (> ~5 MB) can take 10–30 s to upload because of proxy buffering.
Local Docker runs are instant, see below:

2 Run locally with Docker (mirrors the Space)

git clone https://github.com/MooseML/homo-lumo-gap-predictor.git
cd homo-lumo-gap-predictor
docker build -t homolumo .
docker run -p 7860:7860 homolumo
# open http://localhost:7860

3 Run locally with Python (no Docker)

git clone https://github.com/MooseML/homo-lumo-gap-predictor.git
cd homo-lumo-gap-predictor
pip install -r requirements.txt
streamlit run app.py 
# app on http://localhost:8501

Input guidelines

Format	Example
Text area	`O=C(C)Oc1ccccc1C(=O)O`
CSV	One column, no header: `CCO` `Cc1ccccc1`

Invalid or exotic SMILES are skipped and listed in the terminal log (RDKit warnings)

Project files

├── app.py          – Streamlit front‑end
├── model.py        – Hybrid GNN loader (PyTorch Geometric)
├── utils.py        – RDKit helpers & SMILES→graph
├── Dockerfile      – identical to the Hugging Face Space
└── requirements.txt

The Docker image creates /data (writable, 775) for the persistent SQLite DB when a volume is attached.

Model in brief

Architecture: AtomEncoder and BondEncoder → GINEConv layers → global mean pooling → dense head
Descriptors: six RDKit physico‑chemical features per molecule
Training set: OGB PCQM4Mv2
Optimiser / search: Optuna hyperparameter sweep

Roadmap

Stream chunked CSV parsing to improve upload speed on the public Space
Toggle between 2‑D and 3‑D (3Dmol.js) molecule renderings
Serve the model weights from the HF Hub instead of bundling in the image

Author

Matthew Graham — @MooseML Feel free to open issues or contact me for collaborations.

HOMO‑LUMO Gap Predictor — Streamlit + Hybrid GNN