Dockerfile

# HydroPD — single-container Hugging Face Space (Docker SDK).
# One FastAPI process serves the built React UI (/) and the real inference API (/api),
# password-gated by $APP_PASSWORD. Built from a staging dir assembled by
# scripts/build_hf_space.py (mirrors the repo layout: hydropd/, website/, data/Used Papers/).
#
# Layout in the image (paths the code resolves relative to):
#   /app/hydropd                              the pipeline package (PYTHONPATH=/app)
#   /app/website/backend                      FastAPI app (registry -> ../src/data/generated)
#   /app/website/src/data/generated/models.json   the 30 usable models (registry source of truth)
#   /app/website/dist                         built frontend (app.py serves it)
#   /app/data/Used Papers/<H>/step4/models/*.joblib + <code>_oof.npy   (HYDROPD_ROOT=/app/data)

# ---- frontend build ----
FROM node:20-slim AS web
WORKDIR /web
COPY website/package.json website/package-lock.json ./
RUN npm ci
COPY website/ ./
RUN npm run build                       # -> /web/dist

# ---- python runtime ----
FROM python:3.10-slim
ENV PYTHONUNBUFFERED=1 \
    PYTHONPATH=/app \
    HYDROPD_ROOT=/app/data \
    HF_HUB_OFFLINE=1 \
    TRANSFORMERS_OFFLINE=1 \
    HF_HOME=/app/.cache/huggingface \
    TORCH_HOME=/app/.cache/torch \
    PORT=7860
WORKDIR /app
RUN apt-get update && apt-get install -y --no-install-recommends libgomp1 \
    && rm -rf /var/lib/apt/lists/*

# Pinned deps so the trained joblibs unpickle exactly (sklearn/xgboost minors are load-bearing).
COPY website/backend/requirements.txt /tmp/req.txt
# numpy<2 is load-bearing: torch 2.2.2 was compiled against NumPy 1.x and fails to
# initialise ("_ARRAY_API not found") under NumPy 2.x. pyahocorasick (module
# `ahocorasick`) + pyteomics are import-time deps of hydropd (index/proteome/biopep).
RUN pip install --no-cache-dir torch==2.2.2 --index-url https://download.pytorch.org/whl/cpu \
    && pip install --no-cache-dir "numpy<2" "scikit-learn==1.7.2" "xgboost==1.7.6" "fair-esm==2.0.0" \
       peptides pandas scipy pyarrow joblib huggingface_hub tokenizers pyahocorasick pyteomics \
       requests pyyaml \
    && pip install --no-cache-dir -r /tmp/req.txt

# Pfly training runtime (dlomix BiGRU on TensorFlow). CPU-only TF; pinned to the versions the
# `pfly` conda env validated (tf 2.15 / dlomix 0.2.7 / keras-2). TF 2.15 supports numpy<2, so it
# is consistent with the numpy pin above. Isolated layer: only imported on demand by the Pfly
# training branch (hydropd.train_adapter -> hydropd.pfly), so it never slows normal startup.
RUN pip install --no-cache-dir "tensorflow-cpu==2.15.0" "dlomix==0.2.7" datasets

COPY hydropd/ /app/hydropd/
COPY bioactivity_tools/ /app/bioactivity_tools/
COPY website/backend/ /app/website/backend/
COPY website/src/data/generated/ /app/website/src/data/generated/
COPY data/ /app/data/
COPY --from=web /web/dist /app/website/dist

# Pre-fetch the embedder weights at BUILD (so first request isn't a cold ~190 MB download).
# HF_HUB_OFFLINE is flipped on only for runtime; build fetches are allowed here.
RUN HF_HUB_OFFLINE=0 TRANSFORMERS_OFFLINE=0 python -c "import esm; esm.pretrained.esm2_t12_35M_UR50D()" || echo "esm2 prefetch skipped"
RUN HF_HUB_OFFLINE=0 python -c "from huggingface_hub import snapshot_download; snapshot_download('dzjxzyd/PepBERT-large-UniParc')" || echo "pepbert prefetch skipped"

EXPOSE 7860
WORKDIR /app/website/backend
CMD ["python", "-m", "uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]