app.py

#!/usr/bin/env python3
"""
Flexynesis Tissue VAE — Web Application (compressed, real-model inference)
=========================================================================
Upload a bulk RNA-seq gene-expression matrix → UBERON tissue-of-origin
predictions + 121-dim latent embeddings.

Runs the trained supervised VAE via a 53 MB int8 TorchScript bundle
(vae_tissue_int8.torchscript.pt) — exact inference, no flexynesis dependency,
fits the Hugging Face free tier. Trained on 118,263 tissue-curated samples
(TCGA, GTEx, ARCHS4), 42 UBERON tissues, 94.9% balanced accuracy.

Author: Amit Pande, MDC Berlin/BIMSB
"""
import json
from pathlib import Path
from collections import Counter

import numpy as np
import pandas as pd
import streamlit as st
import torch
import joblib

MODEL_DIR = Path(".")

st.set_page_config(page_title="Flexynesis Tissue VAE", page_icon="🧬", layout="wide")
st.title("🧬 Flexynesis Tissue VAE")
st.markdown(
    "Upload a bulk RNA-seq gene expression matrix to classify tissue-of-origin "
    "using a supervised VAE trained on **118,263 tissue-curated samples** from TCGA, GTEx, and ARCHS4 "
    "across **42 UBERON tissue categories** (94.9% balanced accuracy, 121-dim latent space)."
)


@st.cache_resource
def load_model():
    model = torch.jit.load(str(MODEL_DIR / "vae_tissue_int8.torchscript.pt"))
    model.eval()
    art = joblib.load(MODEL_DIR / "vae_tissue.artifacts.joblib")
    gene_list = list(art["feature_lists"]["gex"])
    scaler = art["transforms"]["gex"]
    label_mapping = {int(k): v for k, v in
                     json.loads((MODEL_DIR / "label_mapping.json").read_text()).items()}
    return model, gene_list, scaler, label_mapping


try:
    model, gene_list, scaler, label_mapping = load_model()
    st.success(
        f"✅ Model loaded · {len(gene_list):,} genes · "
        f"{len(label_mapping)} tissue classes · exact VAE inference (int8 TorchScript)"
    )
except Exception as e:
    st.error(f"Model load error: {e}")
    st.info("Ensure vae_tissue_int8.torchscript.pt, vae_tissue.artifacts.joblib, "
            "and label_mapping.json are in the repository root.")
    st.stop()


def orient_matrix(df):
    in_cols = len(set(df.columns) & set(gene_list))
    in_rows = len(set(df.index) & set(gene_list))
    if in_rows > in_cols:
        df = df.T
    return df


st.markdown("---")
col1, col2 = st.columns([2, 1])
with col1:
    uploaded = st.file_uploader(
        "Upload CSV/TSV (genes × samples or samples × genes)",
        type=["csv", "tsv", "txt"],
        help="HGNC gene symbols. Log2-transformed expression values.")
with col2:
    st.markdown(f"""
    **Expected input:**
    - HGNC gene symbols
    - {len(gene_list):,} genes used by model
    - Log2-transformed expression
    """)

if uploaded:
    sep = "\t" if uploaded.name.endswith((".tsv", ".txt")) else ","
    df = pd.read_csv(uploaded, index_col=0, sep=sep)
    st.write(f"**Uploaded:** {df.shape[0]:,} × {df.shape[1]:,}")
    st.dataframe(df.iloc[:5, :5], use_container_width=True)

    if st.button("🚀 Classify Tissues", type="primary"):
        with st.spinner("Running the VAE..."):
            df = orient_matrix(df)
            overlap = len(set(df.columns) & set(gene_list))
            st.write(f"Gene overlap: **{overlap:,}/{len(gene_list):,}** "
                     f"({100*overlap/len(gene_list):.1f}%)")
            if overlap < 1000:
                st.warning("Low gene overlap — results may be unreliable.")

            # --- Input scale check: model expects log2(TPM) ---
            common_for_scale = [g for g in gene_list if g in df.columns]
            if common_for_scale:
                vals = df[common_for_scale].to_numpy(dtype=float)
                vmax = float(np.nanmax(vals))
                p99 = float(np.nanpercentile(vals, 99))
                st.caption(f"Input value range: 99th pct = {p99:.1f}, max = {vmax:.1f} "
                           "(log2(TPM) is typically 0–20).")
                if vmax > 30 or p99 > 25:
                    st.warning(
                        "⚠️ Input values look larger than expected for **log2 scale**. "
                        "This model expects **log2-transformed expression** (e.g. log2(TPM+1)). "
                        "Raw TPM/counts will give unreliable predictions — please log2-transform first."
                    )

            aligned = pd.DataFrame(0.0, index=df.index, columns=gene_list)
            common = [g for g in gene_list if g in df.columns]
            aligned[common] = df[common].values
            aligned = aligned.fillna(0)
            X = torch.tensor(scaler.transform(aligned.values), dtype=torch.float32)

            with torch.no_grad():
                logits = model(X)
            nan_idx = {v: k for k, v in label_mapping.items()}.get("nan")
            if nan_idx is not None:
                logits[:, nan_idx] = -1e9
            probs = torch.softmax(logits, dim=1)
            pred_idx = logits.argmax(dim=1)
            pred_labels = [label_mapping[int(i)] for i in pred_idx]
            conf = probs.max(dim=1).values.numpy()

        st.markdown("---")
        st.subheader("📊 Results")
        results = pd.DataFrame({
            "Sample": df.index,
            "Tissue": pred_labels,
            "Confidence": [f"{c:.1%}" for c in conf],
        })
        st.dataframe(results, use_container_width=True, height=400)

        st.subheader("Tissue Distribution")
        st.bar_chart(pd.Series(pred_labels).value_counts())

        st.download_button("📥 Predictions (CSV)", results.to_csv(index=False),
                           "flexynesis_predictions.csv", "text/csv")

st.markdown("---")
st.caption(
    "Flexynesis Tissue VAE (v3) · Akalin Lab, MDC Berlin/BIMSB · "
    "118,263 samples · 42 UBERON tissues · 94.9% balanced accuracy · "
    "github.com/BIMSBbioinfo/flexynesis_tissue_vae_manuscript"
)