JENNIE

A free web tool that predicts whether a missense variant is loss-of-function (LOF), gain-of-function (GOF), or neutral, using a supervised classifier on ESM-2 embeddings trained on the GLOF benchmark.

Live: https://victormaricato-jennie.static.hf.space

Links

• Web tool: https://victormaricato-jennie.static.hf.space
• Classifier weights: https://huggingface.co/victormaricato/glof-lof-gof-classifier
• GLOF benchmark dataset: https://huggingface.co/datasets/victormaricato/glof
• Benchmark experiments: https://github.com/victormaricato/jennie-paper

How it works

Zero-shot foundation-model likelihoods predict pathogenicity but not the LOF-vs-GOF direction, so JENNIE serves a supervised classifier for the 3-class call: it mean-pools ESM-2 650M embeddings of the wild-type and mutant sequences and feeds the concatenation to a gradient-boosted model. Direction (GOF vs LOF) AUROC is about 0.76 on held-out genes under gene-disjoint cross-validation, so the direction call should be read as a weak prior. Variants already present in GLOF return their expert-curated label.

Layout

api/      FastAPI backend (classifier + Redis result/embedding cache + RQ queue
          + rate limiting), Dockerized
ui/       Next.js 15 + shadcn (new-york) static front end

Run locally

Backend (needs a Redis instance; point REDIS_URL at it):

cd webapp/api && pip install -r requirements.txt
uvicorn app.main:app --port 8000   # API
python worker.py                   # RQ worker (separate process)

Frontend:

cd webapp/ui && pnpm install
NEXT_PUBLIC_API_BASE=http://localhost:8000 pnpm dev

Deployment and security

The front end is a static site rebuilt and redeployed by GitHub Actions on push to main only (never on pull requests), so contributions cannot reach deploy secrets. See SECURITY.md. MIT-licensed.