src/streamlit_app.py

import streamlit as st
import pandas as pd
from io import StringIO
from predictor import load_model, predict_from_df
from Bio import SeqIO
import torch
import os
from huggingface_hub import hf_hub_download

# ==============================
# 页面配置
# ==============================
st.set_page_config(page_title="🧬 Peptide–HLA Binding Predictor", layout="wide")

st.title("🧠 Peptide–HLA Binding Predictor")
st.markdown("""
Upload a **CSV** file with columns `Peptide` and `HLA`,  
or a **FASTA** file containing peptide sequences (headers optionally include HLA type).
""")

# ==============================
# 全局路径设置
# ==============================
CACHE_DIR = "/data/phla_cache"
MODEL_DIR = "/app/src"
UPLOAD_DIR = "/data/uploads"

for d in [CACHE_DIR, MODEL_DIR, UPLOAD_DIR]:
    os.makedirs(d, exist_ok=True)

# 环境变量（确保所有模型和 ESM 缓存写入 /data）
os.environ["HF_HOME"] = "/data/huggingface"
os.environ["TRANSFORMERS_CACHE"] = "/data/huggingface"
os.environ["TORCH_HOME"] = "/data/huggingface"
os.environ["ESM_CACHE_DIR"] = CACHE_DIR

# ==============================
# 模型加载函数（延迟加载 + 缓存）
# ==============================
@st.cache_resource
def load_model_cached():
    device = "cuda:0" if torch.cuda.is_available() else "cpu"
    local_path = os.path.join(MODEL_DIR, "model.pt")

    if not os.path.exists(local_path):
        st.warning("🔄 Model not found locally. Downloading from Hugging Face model repo...")
        # ⚠️ 使用 Model Repo，而不是 Space Repo
        local_path = hf_hub_download(
            repo_id="caokai1073/StriMap-model",  # 建议单独创建模型仓库
            filename="model.pt",
            cache_dir=MODEL_DIR
        )

    model, device = load_model(local_path, device=device)
    return model, device


# ==============================
# 上传文件（安全写入 /data/uploads）
# ==============================
uploaded_file = st.file_uploader("📤 Upload CSV or FASTA", type=["csv", "fasta"])

if uploaded_file:
    safe_name = uploaded_file.name.replace(" ", "_")
    temp_path = os.path.join(UPLOAD_DIR, safe_name)
    with open(temp_path, "wb") as f:
        f.write(uploaded_file.getbuffer())

    # ==============================
    # 文件解析
    # ==============================
    if safe_name.endswith(".csv"):
        df = pd.read_csv(temp_path)
    else:
        seqs = []
        for rec in SeqIO.parse(temp_path, "fasta"):
            header = rec.id
            seq = str(rec.seq)
            if "|" in header:
                hla, _ = header.split("|", 1)
            else:
                hla = "HLA-Unknown"
            seqs.append([seq, hla])
        df = pd.DataFrame(seqs, columns=["Peptide", "HLA"])

    st.write("✅ Uploaded data preview:")
    st.dataframe(df.head())

    # ==============================
    # 模型预测（延迟加载）
    # ==============================
    if st.button("🚀 Run Prediction"):
        with st.spinner("🔄 Loading model (this may take ~1 min first time)..."):
            model, device = load_model_cached()

        with st.spinner("Running inference..."):
            result_df = predict_from_df(df, model)

        st.success("✅ Prediction complete!")
        st.dataframe(result_df.head(10))

        # ==============================
        # 下载结果
        # ==============================
        csv = result_df.to_csv(index=False).encode("utf-8")
        st.download_button(
            "⬇️ Download results as CSV",
            data=csv,
            file_name="hla_binding_predictions.csv",
            mime="text/csv",
        )

# ==============================
# Debug / data check (optional)
# ==============================
if st.sidebar.button("📁 List /data files"):
    files = []
    for root, _, filenames in os.walk("/data"):
        for f in filenames:
            files.append(os.path.join(root, f))
    st.sidebar.write(files)