title	BioAtlas API
emoji	🧬
colorFrom	red
colorTo	green
sdk	docker
app_port	7860
pinned	false

BioAtlas | Biomedical Research Intelligence

A living map of biomedical research: hybrid semantic search, cited evidence briefs, unsupervised topic discovery and real publication trends over 14,693 PubMed studies from 24 fields, 2014 to 2026.

• Live site: https://bio-med-research-helper-yzop.vercel.app
• API: https://sawabedarain-biomed-ai-backend.hf.space/docs

Note

Portfolio project, not a production service or medical advice. The API runs on Hugging Face's free CPU tier, so the first request after a quiet period can take about 30 seconds while it wakes up. Evidence briefs are extractive by default; set a GEMINI_API_KEY secret to add an optional grounded AI synthesis.

What it does

Hybrid search	Fine-tuned dense retrieval fused with BM25 (reciprocal rank fusion). Filter by field and year; facets show where results cluster.
Evidence briefs	Ask a question; get the most relevant sentences across the top studies (MMR selection, every sentence cited), quantitative findings, and an evidence profile (study designs, years, concepts).
Atlas	Interactive map of every study, positioned by meaning (UMAP). Light up any query on the map.
Topics	Research fronts discovered without labels (UMAP + HDBSCAN), named with class TF-IDF, ranked by growth.
Trends	True PubMed publication volumes per field, rising MeSH concepts, top journals.

The landing page is a live specimen under a fluorescence microscope, drawn on the GPU (WebGL): all 14,693 studies form a DNA double helix in green and red strands. Drag to spin it; hover any point to read the real study it stands for. Scrolling walks through how the system works: the helix uncoils into a rotating 3D space of meaning where each study takes its field's colour, a sample question lands in that space and its nearest studies light up, and the space flattens into the research map, annotated with the largest research fronts. The design follows the subject: a dark field, studies that glow like tagged cells, and GFP green and mCherry red as the accents.

Retrieval benchmark

Held-out evaluation: each of 1,514 unseen paper titles is a query against all 14,693 abstracts; the target is its own paper.

Model	Params	Recall@1	Recall@10	MRR@10
BM25 (keywords)	-	0.931	0.985	0.952
all-mpnet-base-v2 (v1 model)	110M	0.798	0.959	0.855
bge-small, off the shelf	33M	0.913	0.989	0.941
bge-small, fine-tuned here	33M	0.964	0.997	0.977
Hybrid: BM25 + fine-tuned (production)	33M	0.967	0.995	0.978

Fine-tuning: 26,257 title/MeSH-query to abstract pairs, MultipleNegativesRankingLoss, 3 epochs, 134 s on an RTX 5090. The deployed encoder is int8 per-channel ONNX (34 MB) and keeps Recall@1 at 0.961 vs 0.964 in fp32.

Architecture

pipeline/ (offline, GPU)                        backend/ (Hugging Face Space, CPU)        frontend/ (Vercel)
  ingest.py   year-stratified PubMed sample  ->   FastAPI                                  React + Vite + Tailwind
  counts.py   true PubMed volumes per field       int8 ONNX encoder + fast tokenizer       Lenis smooth scroll
  train_embed fine-tune bge-small (MNRL)          dense (numpy) + BM25 (scipy sparse)      WebGL specimen scene + Atlas
  build.py    vectors, BM25, topics, map,         RRF fusion, MMR briefs, topics, trends   static map.bin (no backend
              ONNX export + int8 quantisation                                             needed for the hero)

Why it is efficient: v1 loaded BART-large, BERTopic and a 110M-parameter encoder at startup on a free CPU. v2 serves a 33M-parameter encoder as int8 ONNX (no PyTorch in the image), precomputes topics and the map offline, and answers searches in milliseconds.

Run locally

conda env create -f environment.yml && conda activate biomed

# Rebuild the data and models (about 20 minutes with a GPU)...
bash pipeline/run_all.sh python
# ...or download the published artifacts from the Space
huggingface-cli download sawabedarain/biomed-ai-backend --repo-type space --include "artifacts/*" --local-dir backend

cd backend && uvicorn app.main:app --port 7860      # API on http://localhost:7860/docs
pytest tests -q

cd frontend && npm install && npm run dev           # site on http://localhost:5173

Deploy

• Backend: the root Dockerfile builds the Space; backend/artifacts/ is uploaded alongside the code.
• Frontend: Vercel project with root directory frontend/; VITE_API_URL points to the Space (defaults to it when unset).

Data: abstracts and metadata from NCBI PubMed via E-utilities. Abstracts remain the property of their publishers.