app.py

import gradio as gr
import joblib
from huggingface_hub import hf_hub_download
import pandas as pd
import numpy as np
from collections import Counter

# --- Download model from HF Hub ---
repo_id = "Ym420/Peptide-Function"
model_filename = "xgb_multilabel_model_full.pkl"

model_path = hf_hub_download(repo_id=repo_id, filename=model_filename)
model_package = joblib.load(model_path)

# --- Unwrap model dict ---
model_dict = model_package['model']           # dict: {'Gram+': XGBClassifier, ...}
feature_columns = model_package['feature_columns']

# --- Metadata (all restored) ---
aa_list = model_package.get('aa_list', [])
dipeptides = model_package.get('dipeptides', [])
hydrophobicity_scale = model_package.get('hydrophobicity_scale', {})
eisenberg_scale = model_package.get('eisenberg_scale', {})
aa_mass = model_package.get('aa_mass', {})
aa_charge = model_package.get('aa_charge', {})
aa_boman = model_package.get('aa_boman', {})
aa_flexibility = model_package.get('aa_flexibility', {})
aa_polarizability = model_package.get('aa_polarizability', {})
aa_aliphatic = model_package.get('aa_aliphatic', {})
aa_deltaG = model_package.get('aa_deltaG', {})

# --- Dynamic TARGET_CELLS ---
TARGET_CELLS = list(model_dict.keys())  # automatically detects all targets

# --- Feature extraction (future-proof) ---
def extract_features_app(seq: str) -> pd.DataFrame:
    seq = seq.upper()

    # --- 1. Dipeptide composition ---
    count = Counter([seq[i:i+2] for i in range(len(seq)-1)])
    total = max(len(seq)-1, 1)
    dipep_features = {dp: count.get(dp, 0) / total for dp in dipeptides}

    # --- 2. Physicochemical features ---
    def g(aa, table): return table.get(aa, 0)
    def h(dp, table): return (g(dp[0], table) + g(dp[1], table)) / 2.0

    dipeptides_seq = [seq[i:i+2] for i in range(len(seq)-1)]

    if len(seq) < 2:
        physchem_features = {
            'mw': 0, 'charge': 0, 'hydro': 0, 'aromatic': 0, 'pI': 0,
            'instability': 0, 'hydro_moment': 0, 'aliphatic': 0,
            'boman': 0, 'flexibility': 0, 'polarizability': 0, 'deltag': 0
        }
    else:
        mw = np.mean([h(dp, aa_mass) for dp in dipeptides_seq])
        charge = np.mean([h(dp, aa_charge) for dp in dipeptides_seq])
        hydro = np.mean([h(dp, hydrophobicity_scale) for dp in dipeptides_seq])
        aromatic = np.mean([(dp[0] in 'FWY') + (dp[1] in 'FWY') for dp in dipeptides_seq]) / 2.0
        pI = np.mean([h(dp, {aa: 7 + (int(aa in 'KRH') - int(aa in 'DE')) for aa in aa_list}) for dp in dipeptides_seq])
        instability = np.mean([((dp[0] in 'DEKR') + (dp[1] in 'DEKR')) / 2.0 for dp in dipeptides_seq])
        hydro_moment = np.sqrt(np.mean([(h(dp, eisenberg_scale))**2 for dp in dipeptides_seq]))
        aliphatic = np.mean([h(dp, aa_aliphatic) for dp in dipeptides_seq])
        boman = np.mean([h(dp, aa_boman) for dp in dipeptides_seq])
        flexibility = np.mean([h(dp, aa_flexibility) for dp in dipeptides_seq])
        polarizability = np.mean([h(dp, aa_polarizability) for dp in dipeptides_seq])
        deltag = np.mean([h(dp, aa_deltaG) for dp in dipeptides_seq])

        physchem_features = {
            'mw': mw, 'charge': charge, 'hydro': hydro, 'aromatic': aromatic, 'pI': pI,
            'instability': instability, 'hydro_moment': hydro_moment, 'aliphatic': aliphatic,
            'boman': boman, 'flexibility': flexibility, 'polarizability': polarizability, 'deltag': deltag
        }

    # --- Combine features ---
    all_features = {**dipep_features, **physchem_features}

    # --- Align with feature_columns ---
    df = pd.DataFrame([[all_features.get(col, 0) for col in feature_columns]], columns=feature_columns)
    df = df.astype('float32')
    return df

# --- Prediction function ---
def predict_peptide(sequence: str):
    seq = "".join(sequence.split()).upper()
    if not seq:
        return []

    X = extract_features_app(seq)

    table = []
    for target in TARGET_CELLS:
        clf = model_dict.get(target)
        if clf is not None:
            prob = clf.predict_proba(X)[0][1]  # positive-class probability
            table.append([target, round(float(prob), 4)])
        else:
            table.append([target, None])
    return table

# --- Gradio Interface ---
custom_css = """
footer, .footer {display:none !important;}
"""

with gr.Blocks(css=custom_css, theme="default") as demo:
    gr.Markdown("## Peptide Antimicrobial Predictor\nEnter a peptide sequence to predict efficacy/toxicity.")

    seq_input = gr.Textbox(label="Enter Peptide Sequence")

    with gr.Row():
        predict_btn = gr.Button("Predict", variant="primary")
        clear_btn = gr.Button("Clear")

    table_output = gr.Dataframe(
        headers=["Target Cell", "Probability of Efficacy/Toxicity"],
        datatype=["str","number"],
        interactive=False
    )

    predict_btn.click(fn=predict_peptide, inputs=seq_input, outputs=table_output)
    clear_btn.click(fn=lambda: ("", []), outputs=[seq_input, table_output])

    # API endpoint for iOS app
    gr.api(predict_peptide, api_name="predict_peptide")

if __name__ == "__main__":
    demo.launch(show_error=True)