ClinSift / README.md
Vasumathi Varadararajan
fix Space color metadata
fb04ba2
|
Raw History Blame Contribute Delete
2.95 kB
---
title: ClinSift
emoji: 🧪
colorFrom: blue
colorTo: gray
sdk: docker
app_port: 7860
pinned: false
---
# ClinSift — real-world drug-safety filter
Screens a drug or a novel compound against adverse-event signals and returns a
single **gate decision** (PASS / FLAG / HOLD) for a drug-development pipeline.
The UI and the API are one deployment: open the Space and it talks to its own backend.
- **Known drug** → observed cross-reactivity & class side-effects (from EHR-derived signals)
- **Novel compound (SMILES)** → SSI-DDI interaction prediction + **substructure attribution**
(which atoms drive the interaction) + nearest known drugs via MoLFormer-XL
- Every response is also a JSON contract a pipeline can gate on programmatically
The interaction model is a proof-of-concept (SSI-DDI, ~0.79 AUC on a public
BIOSNAP subset). Outputs are predictions, not clinical advice.
---
## ⚠️ DATA-USE COMPLIANCE — read before making this public
This public build ships **only** non-patient artifacts:
`molformer_embeddings.npz`, `ssiddi_best.pth`, and the aggregate
`crossreactivity_similarity.json` / `sideeffect_cohesion.json` (population-level,
no patient identifiers).
It **excludes** the patient-level MIMIC-derived files
(`patient_allergies.json`, `patient_crossreactivity_refined.json`, which contain
note IDs). MIMIC's PhysioNet Data Use Agreement **prohibits public redistribution**
of patient-level derivatives. `.gitignore` blocks these from being committed.
- **Public Space:** leave those files out. Allergy cross-reactivity shows
"requires patient data" — that is correct and expected.
- **Private Space (only if your DUA permits):** set the Space to *private*, then
add the two files to `data/`. Confirm with your PhysioNet DUA / data steward first.
---
## Run locally
```bash
pip install -r requirements.txt
uvicorn app:app --port 8000
# open http://localhost:8000
```
## Deploy to your HuggingFace account
You run these (they need your HF login; never share your token):
```bash
pip install huggingface_hub
huggingface-cli login # paste your WRITE token when prompted
# create a Docker Space under your account
huggingface-cli repo create clinsift --type space --space_sdk docker
# push this folder
git init && git lfs install
git lfs track "*.npz" "*.pth"
git add .gitignore .gitattributes
git add Dockerfile requirements.txt README.md app.py ssiddi_model.py ui.html data/
git commit -m "ClinSift: real-world drug-safety filter"
git remote add origin https://huggingface.co/spaces/<your-username>/clinsift
git push origin main
```
Large files (`.npz`, `.pth`) go through Git LFS — the `git lfs track` step above
handles that. First screen of a SMILES downloads MoLFormer-XL (~200 MB) once.
## API
```
GET /health
POST /screen {"drug":"penicillin"}
POST /screen {"drug_a":"warfarin","drug_b":"clopidogrel"}
POST /screen_smiles {"smiles":"CC1=C(...)","name":"CWR-001"}
```