app.py

import gradio as gr
import joblib
import numpy as np
import pandas as pd
from propy import AAComposition, Autocorrelation, CTD, PseudoAAC
import torch
from transformers import BertTokenizer, BertModel
from lime.lime_tabular import LimeTabularExplainer
from math import expm1

# Load AMP Classifier and Scaler
model = joblib.load("RF.joblib")
scaler = joblib.load("norm (4).joblib")

# Load ProtBert
tokenizer = BertTokenizer.from_pretrained("Rostlab/prot_bert", do_lower_case=False)
protbert_model = BertModel.from_pretrained("Rostlab/prot_bert")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
protbert_model = protbert_model.to(device).eval()

# Define selected features (146 RFE-selected features)
selected_features = ["_SolventAccessibilityC3", "_SecondaryStrC1", "_SecondaryStrC3", "_ChargeC1", "_PolarityC1",
"_NormalizedVDWVC1", "_HydrophobicityC3", "_SecondaryStrT23", "_PolarizabilityD1001", "_PolarizabilityD2001",
"_PolarizabilityD3001", "_SolventAccessibilityD1001", "_SolventAccessibilityD2001", "_SolventAccessibilityD3001",
"_SecondaryStrD1001", "_SecondaryStrD1075", "_SecondaryStrD2001", "_SecondaryStrD3001", "_ChargeD1001",
"_ChargeD1025", "_ChargeD2001", "_ChargeD3075", "_ChargeD3100", "_PolarityD1001", "_PolarityD1050",
"_PolarityD2001", "_PolarityD3001", "_NormalizedVDWVD1001", "_NormalizedVDWVD2001", "_NormalizedVDWVD2025",
"_NormalizedVDWVD2050", "_NormalizedVDWVD3001", "_HydrophobicityD1001", "_HydrophobicityD2001",
"_HydrophobicityD3001", "_HydrophobicityD3025", "A", "R", "D", "C", "E", "Q", "H", "I", "M", "P", "Y", "V",
"AR", "AV", "RC", "RL", "RV", "CR", "CC", "CL", "CK", "EE", "EI", "EL", "HC", "IA", "IL", "IV", "LA", "LC", "LE",
"LI", "LT", "LV", "KC", "MA", "MS", "SC", "TC", "TV", "YC", "VC", "VE", "VL", "VK", "VV",
"MoreauBrotoAuto_FreeEnergy30", "MoranAuto_Hydrophobicity2", "MoranAuto_Hydrophobicity4",
"GearyAuto_Hydrophobicity20", "GearyAuto_Hydrophobicity24", "GearyAuto_Hydrophobicity26",
"GearyAuto_Hydrophobicity27", "GearyAuto_Hydrophobicity28", "GearyAuto_Hydrophobicity29",
"GearyAuto_Hydrophobicity30", "GearyAuto_AvFlexibility22", "GearyAuto_AvFlexibility26",
"GearyAuto_AvFlexibility27", "GearyAuto_AvFlexibility28", "GearyAuto_AvFlexibility29", "GearyAuto_AvFlexibility30",
"GearyAuto_Polarizability22", "GearyAuto_Polarizability24", "GearyAuto_Polarizability25",
"GearyAuto_Polarizability27", "GearyAuto_Polarizability28", "GearyAuto_Polarizability29",
"GearyAuto_Polarizability30", "GearyAuto_FreeEnergy24", "GearyAuto_FreeEnergy25", "GearyAuto_FreeEnergy30",
"GearyAuto_ResidueASA21", "GearyAuto_ResidueASA22", "GearyAuto_ResidueASA23", "GearyAuto_ResidueASA24",
"GearyAuto_ResidueASA30", "GearyAuto_ResidueVol21", "GearyAuto_ResidueVol24", "GearyAuto_ResidueVol25",
"GearyAuto_ResidueVol26", "GearyAuto_ResidueVol28", "GearyAuto_ResidueVol29", "GearyAuto_ResidueVol30",
"GearyAuto_Steric18", "GearyAuto_Steric21", "GearyAuto_Steric26", "GearyAuto_Steric27", "GearyAuto_Steric28",
"GearyAuto_Steric29", "GearyAuto_Steric30", "GearyAuto_Mutability23", "GearyAuto_Mutability25",
"GearyAuto_Mutability26", "GearyAuto_Mutability27", "GearyAuto_Mutability28", "GearyAuto_Mutability29",
"GearyAuto_Mutability30", "APAAC1", "APAAC4", "APAAC5", "APAAC6", "APAAC8", "APAAC9", "APAAC12", "APAAC13",
"APAAC15", "APAAC18", "APAAC19", "APAAC24"]

# --- FIX (LIME): seed the random background so explanations are reproducible
# across Space restarts. (Loading a real saved training sample here would
# produce more faithful weights; see build_lime_background.py for that path.)
np.random.seed(42)
sample_data = np.random.rand(500, len(selected_features))
explainer = LimeTabularExplainer(
    training_data=sample_data,
    feature_names=selected_features,
    class_names=["AMP", "Non-AMP"],
    mode="classification",
    random_state=42,
)

# Feature extraction function
def extract_features(sequence):
    sequence = ''.join([aa for aa in sequence.upper() if aa in "ACDEFGHIKLMNPQRSTVWY"])
    if len(sequence) < 10:
        return "Error: Sequence too short."

    try:
        dipeptide_features = AAComposition.CalculateAADipeptideComposition(sequence)
        filtered_dipeptide_features = {k: dipeptide_features[k] for k in list(dipeptide_features.keys())[:420]}
        ctd_features = CTD.CalculateCTD(sequence)
        auto_features = Autocorrelation.CalculateAutoTotal(sequence)
        pseudo_features = PseudoAAC.GetAPseudoAAC(sequence, lamda=9)

        all_features_dict = {}
        all_features_dict.update(ctd_features)
        all_features_dict.update(filtered_dipeptide_features)
        all_features_dict.update(auto_features)
        all_features_dict.update(pseudo_features)

        feature_df_all = pd.DataFrame([all_features_dict])
        normalized_array = scaler.transform(feature_df_all.values)
        normalized_df = pd.DataFrame(normalized_array, columns=feature_df_all.columns)

        if not set(selected_features).issubset(normalized_df.columns):
            return "Error: Some selected features are missing."

        selected_df = normalized_df[selected_features].fillna(0)
        return selected_df.values
    except Exception as e:
        return f"Error in feature extraction: {str(e)}"

# MIC prediction function
def predictmic(sequence):
    sequence = ''.join([aa for aa in sequence.upper() if aa in "ACDEFGHIKLMNPQRSTVWY"])
    if len(sequence) < 10:
        return {"Error": "Sequence too short or invalid."}

    seq_spaced = ' '.join(list(sequence))
    tokens = tokenizer(seq_spaced, return_tensors="pt", padding='max_length', truncation=True, max_length=512)
    tokens = {k: v.to(device) for k, v in tokens.items()}

    with torch.no_grad():
        outputs = protbert_model(**tokens)
        embedding = outputs.last_hidden_state.mean(dim=1).squeeze().cpu().numpy().reshape(1, -1)

    bacteria_config = {
        "E.coli": {"model": "coli_xgboost_model.pkl", "scaler": "coli_scaler.pkl", "pca": None},
        "S.aureus": {"model": "aur_xgboost_model.pkl", "scaler": "aur_scaler.pkl", "pca": None},
        "P.aeruginosa": {"model": "arg_xgboost_model.pkl", "scaler": "arg_scaler.pkl", "pca": None},
        "K.Pneumonia": {"model": "pne_mlp_model.pkl", "scaler": "pne_scaler.pkl", "pca": "pne_pca.pkl"}
    }

    mic_results = {}
    for bacterium, cfg in bacteria_config.items():
        try:
            # --- FIX (variable shadowing): renamed locals so the global `scaler`
            # and `model` (the AMP RF + its MinMax scaler) are NEVER overwritten.
            # The original code reused the names `scaler` and `model` here, which
            # silently broke the AMP classifier on every prediction after the
            # first MIC run.
            mic_scaler = joblib.load(cfg["scaler"])
            scaled = mic_scaler.transform(embedding)
            transformed = joblib.load(cfg["pca"]).transform(scaled) if cfg["pca"] else scaled
            mic_model = joblib.load(cfg["model"])
            mic_log = mic_model.predict(transformed)[0]
            mic = round(expm1(mic_log), 3)
            mic_results[bacterium] = mic
        except Exception as e:
            mic_results[bacterium] = f"Error: {str(e)}"

    return mic_results

# Main prediction function
def full_prediction(sequence):
    features = extract_features(sequence)
    if isinstance(features, str):
        return features

    prediction = model.predict(features)[0]
    probabilities = model.predict_proba(features)[0]

    try:
        class_index = list(model.classes_).index(prediction)
        confidence = round(probabilities[class_index] * 100, 2)
    except Exception:
        confidence = "Unknown"

    amp_result = "Antimicrobial Peptide (AMP)" if prediction == 0 else "Non-AMP"
    result = f"Prediction: {amp_result}\nConfidence: {confidence}%\n"

    # --- LIME first (per spec: LIME before SHAP in the HTML report).
    # explain_instance perturbs THIS single input sequence's feature row 2000
    # times and fits a local linear model; weights describe this specific input.
    try:
        explanation = explainer.explain_instance(
            data_row=features[0],
            predict_fn=model.predict_proba,
            num_features=10,
            num_samples=2000,
        )
        result += "\nTop Features Influencing Prediction (LIME):\n"
        for feat, weight in explanation.as_list():
            result += f"- {feat}: {round(weight, 4)}\n"
    except Exception as e:
        result += f"\nLIME explanation failed: {str(e)}\n"

    if prediction == 0:
        mic_values = predictmic(sequence)
        result += "\nPredicted MIC Values (μM):\n"
        for org, mic in mic_values.items():
            result += f"- {org}: {mic}\n"
    else:
        result += "\nMIC prediction skipped for Non-AMP sequences.\n"

    return result

# Gradio UI
iface = gr.Interface(
    fn=full_prediction,
    inputs=gr.Textbox(label="Enter Protein Sequence"),
    outputs=gr.Textbox(label="Results"),
    title="AMP & MIC Predictor + LIME Explanation",
    description="Paste an amino acid sequence (≥10 characters). Get AMP classification, MIC predictions, and LIME interpretability insights."
)

# --- FIX (launch): removed share=True. On Hugging Face Spaces the public URL
# is provided by the platform; share=True is for local dev only.
iface.launch()