app.py

import gradio as gr
import os
import pandas as pd
import tempfile
from pathlib import Path
from inference_wrapper import RNAFoldingPredictor
from huggingface_hub import hf_hub_download

# 🔽 Download the model weights from Hugging Face Hub
ckpt_file = hf_hub_download(
    repo_id="ptope/rna-model-weights",
    filename="best_val_model.pt"
)

# ✅ Set up model and static directory
MODEL_PATH = ckpt_file
STATIC_DIR = Path("static")
os.makedirs(STATIC_DIR, exist_ok=True)
predictor = RNAFoldingPredictor(MODEL_PATH)

# ✅ Load test sequences and map sequences to target IDs
test_sequences = pd.read_csv("test_sequences.csv")
test_seq_list = test_sequences["sequence"].dropna().unique().tolist()
sequence_to_target_id = {
    row["sequence"].strip().upper(): row["target_id"]
    for _, row in test_sequences.iterrows()
}

# ✅ Format predicted structure into standard PDB format
def create_pdb_from_prediction(prediction_df):
    pdb_lines = ["HEADER    RNA STRUCTURE PREDICTION"]
    atom_index = 1
    prev_c1_index = None
    for _, row in prediction_df.iterrows():
        resname = row['resname']
        resid = row['resid']
        x, y, z = row['x_1'], row['y_1'], row['z_1']
        c1_index = atom_index
        pdb_lines.append(f"ATOM  {atom_index:5d}  C1'  {resname} A{resid:4d}    {x:8.3f}{y:8.3f}{z:8.3f}  1.00  0.00           C")
        atom_index += 1
        base_offset = {'A': (1.5, 0.0, 0.5), 'C': (1.2, 0.3, 0.3), 'G': (1.7, -0.2, 0.7), 'U': (1.0, 0.5, 0.0)}.get(resname, (1.5, 0.0, 0.0))
        base_x, base_y, base_z = x + base_offset[0], y + base_offset[1], z + base_offset[2]
        base_index = atom_index
        pdb_lines.append(f"ATOM  {atom_index:5d}  N9   {resname} A{resid:4d}    {base_x:8.3f}{base_y:8.3f}{base_z:8.3f}  1.00  0.00           N")
        atom_index += 1
        pdb_lines.append(f"CONECT{c1_index:5d}{base_index:5d}")
        if prev_c1_index is not None:
            pdb_lines.append(f"CONECT{prev_c1_index:5d}{c1_index:5d}")
        prev_c1_index = c1_index
    pdb_lines.append("END")
    return "\n".join(pdb_lines)

# ✅ This function handles GPT model prediction and downloadable .pdb
def generate_and_return_file(sequence, description=""):
    sequence = sequence.strip().upper()
    if not sequence or not all(b in "ACGU" for b in sequence):
        return "Invalid input", None

    df = predictor.predict(sequence, description)
    pdb_text = create_pdb_from_prediction(df)
    filename = f"{next(tempfile._get_candidate_names())}.pdb"
    path = STATIC_DIR / filename

    with open(path, "w") as f:
        f.write(pdb_text)

    return "Download ready", str(path)

# ✅ Load precomputed .pdb from /pdbs folder for Mol* viewer
def serve_precomputed_pdb(sequence, description=""):
    sequence = sequence.strip().upper()
    if sequence not in sequence_to_target_id:
        return ""

    target_id = sequence_to_target_id[sequence]
    filename = f"{target_id}.pdb"

    # ✅ Here's the FIX: point viewer to /pdbs folder

    pdb_url = f"https://ptope--rna-folding-demo.hf.space/pdbs/{filename}"
    iframe_url = f"https://molstar.org/viewer/?loadFromUrl={pdb_url}"

    # Check
    print(f"[DEBUG] Final URL to check: {pdb_url}")
    print(f"[DEBUG] Final URL to check: {iframe_url}")
    
    iframe_html = f'<iframe src="{iframe_url}" width="100%" height="600px" style="border:1px solid #ccc;"></iframe>'
    return iframe_html

# ✅ Gradio app layout
with gr.Blocks() as demo:
    gr.Markdown("## 🧬 RNA 3D Viewer: Precomputed Visualization + Predictive Download")

    with gr.Row():
        with gr.Column():
            dropdown = gr.Dropdown(choices=test_seq_list, label="Select a Test Sequence")
            seq_input = gr.Textbox(label="RNA Sequence", lines=4)
            desc_input = gr.Textbox(label="Description (optional)", lines=1)
            dropdown.change(fn=lambda s: s, inputs=dropdown, outputs=seq_input)
            submit_btn = gr.Button("Submit")

        with gr.Column():
            status_output = gr.Textbox(label="Status")
            file_output = gr.File(label="Download .pdb")

    viewer_html = gr.HTML(label="Mol* Viewer")

    # ✅ Submit triggers BOTH file generation and structure visualization
    submit_btn.click(
        fn=generate_and_return_file,
        inputs=[seq_input, desc_input],
        outputs=[status_output, file_output]
    )

    submit_btn.click(
        fn=serve_precomputed_pdb,
        inputs=[seq_input, desc_input],
        outputs=[viewer_html]
    )

demo.launch()