app.py

import os
# --- Prevent SIGSEGV (exit 139) from TensorFlow + PyTorch native lib clashes ---
# TF and torch each bundle their own OpenMP/MKL; loaded together they can collide
# and crash at the C level. These settings make them coexist and reduce memory.
os.environ.setdefault("KMP_DUPLICATE_LIB_OK", "TRUE")
os.environ.setdefault("OMP_NUM_THREADS", "1")
os.environ.setdefault("MKL_NUM_THREADS", "1")
os.environ.setdefault("OPENBLAS_NUM_THREADS", "1")
# Quiet TensorFlow logs (must be set before importing tensorflow)
os.environ.setdefault("TF_CPP_MIN_LOG_LEVEL", "3")
os.environ.setdefault("TF_ENABLE_ONEDNN_OPTS", "0")

import gradio as gr
import joblib
import numpy as np
import pandas as pd
from propy import AAComposition, CTD
import sys
import json
import subprocess

# ---------------------------------------------------------------------------
# LAZY LOADING — keeps the free 16GB Space from OOM-ing at startup.
# Only the TensorFlow AMP model is loaded in THIS process. ProtBert/PyTorch
# run in a SEPARATE process (mic_worker.py) to avoid a native-library clash
# between TensorFlow and PyTorch that caused SIGSEGV (exit 139).
# ---------------------------------------------------------------------------
_amp_model = None
_amp_scaler = None


def get_amp_model():
    global _amp_model, _amp_scaler
    if _amp_model is None:
        from tensorflow.keras.models import load_model
        _amp_model = load_model("Comb1_aac_ctd_RFE_selected_features_model.keras")
        _amp_scaler = joblib.load("Comb1_aac_ctd_RFE_selected_features_scaler.joblib")
    return _amp_model, _amp_scaler


# ---------------------------------------------------------------------------
# The EXACT 343 features the scaler was fit on, IN THE EXACT TRAINING ORDER.
# The scaler was fit on a numpy array (no stored names), so order is critical:
# we must select these columns in this order BEFORE calling scaler.transform().
# ---------------------------------------------------------------------------
selected_features = [
    "_PolarizabilityC1", "_PolarizabilityC2", "_PolarizabilityC3",
    "_SolventAccessibilityC1", "_SolventAccessibilityC2", "_SolventAccessibilityC3",
    "_SecondaryStrC1", "_SecondaryStrC2", "_SecondaryStrC3",
    "_ChargeC1", "_ChargeC2", "_ChargeC3",
    "_PolarityC1", "_PolarityC2", "_PolarityC3",
    "_NormalizedVDWVC1", "_NormalizedVDWVC2", "_NormalizedVDWVC3",
    "_HydrophobicityC1", "_HydrophobicityC2", "_HydrophobicityC3",
    "_PolarizabilityT12", "_PolarizabilityT13", "_PolarizabilityT23",
    "_SolventAccessibilityT12", "_SolventAccessibilityT13", "_SolventAccessibilityT23",
    "_SecondaryStrT12", "_SecondaryStrT13", "_SecondaryStrT23",
    "_ChargeT12", "_ChargeT13", "_ChargeT23",
    "_PolarityT12", "_PolarityT13", "_PolarityT23",
    "_NormalizedVDWVT12", "_NormalizedVDWVT13", "_NormalizedVDWVT23",
    "_HydrophobicityT12", "_HydrophobicityT13", "_HydrophobicityT23",
    "_PolarizabilityD1001", "_PolarizabilityD1025", "_PolarizabilityD1050",
    "_PolarizabilityD1075", "_PolarizabilityD1100",
    "_PolarizabilityD2001", "_PolarizabilityD2025", "_PolarizabilityD2050",
    "_PolarizabilityD2075", "_PolarizabilityD2100",
    "_PolarizabilityD3001", "_PolarizabilityD3025", "_PolarizabilityD3050",
    "_PolarizabilityD3075", "_PolarizabilityD3100",
    "_SolventAccessibilityD1001", "_SolventAccessibilityD1025",
    "_SolventAccessibilityD1050", "_SolventAccessibilityD1075",
    "_SolventAccessibilityD1100",
    "_SolventAccessibilityD2001", "_SolventAccessibilityD2025",
    "_SolventAccessibilityD2050", "_SolventAccessibilityD2075",
    "_SolventAccessibilityD2100",
    "_SolventAccessibilityD3001", "_SolventAccessibilityD3025",
    "_SolventAccessibilityD3050", "_SolventAccessibilityD3075",
    "_SolventAccessibilityD3100",
    "_SecondaryStrD1001", "_SecondaryStrD1025", "_SecondaryStrD1050",
    "_SecondaryStrD1075", "_SecondaryStrD1100",
    "_SecondaryStrD2001", "_SecondaryStrD2025", "_SecondaryStrD2050",
    "_SecondaryStrD2075", "_SecondaryStrD2100",
    "_SecondaryStrD3001", "_SecondaryStrD3025", "_SecondaryStrD3050",
    "_SecondaryStrD3075", "_SecondaryStrD3100",
    "_ChargeD1001", "_ChargeD1025", "_ChargeD1050",
    "_ChargeD1075", "_ChargeD1100",
    "_ChargeD2001", "_ChargeD2025", "_ChargeD2050",
    "_ChargeD2075",
    "_ChargeD3001", "_ChargeD3025", "_ChargeD3050",
    "_ChargeD3075", "_ChargeD3100",
    "_PolarityD1001", "_PolarityD1025", "_PolarityD1050",
    "_PolarityD1075", "_PolarityD1100",
    "_PolarityD2001", "_PolarityD2025", "_PolarityD2050",
    "_PolarityD2075", "_PolarityD2100",
    "_PolarityD3001", "_PolarityD3025", "_PolarityD3050",
    "_PolarityD3075", "_PolarityD3100",
    "_NormalizedVDWVD1001", "_NormalizedVDWVD1025",
    "_NormalizedVDWVD1050", "_NormalizedVDWVD1075",
    "_NormalizedVDWVD1100",
    "_NormalizedVDWVD2001", "_NormalizedVDWVD2025",
    "_NormalizedVDWVD2050", "_NormalizedVDWVD2075",
    "_NormalizedVDWVD2100",
    "_NormalizedVDWVD3001", "_NormalizedVDWVD3025",
    "_NormalizedVDWVD3050", "_NormalizedVDWVD3075",
    "_NormalizedVDWVD3100",
    "_HydrophobicityD1001", "_HydrophobicityD1025",
    "_HydrophobicityD1050", "_HydrophobicityD1075",
    "_HydrophobicityD1100",
    "_HydrophobicityD2001", "_HydrophobicityD2025",
    "_HydrophobicityD2050", "_HydrophobicityD2075",
    "_HydrophobicityD2100",
    "_HydrophobicityD3001", "_HydrophobicityD3025",
    "_HydrophobicityD3050", "_HydrophobicityD3075",
    "_HydrophobicityD3100",
    "A", "R", "N", "D", "C", "E", "Q", "G", "H", "I",
    "L", "K", "M", "F", "P", "S", "T", "W", "Y", "V",
    "AR", "AD", "AQ", "AG", "AL", "AK", "AF", "AP", "AT", "AV",
    "RA", "RC", "RE", "RG", "RI", "RL", "RS", "RT", "RV",
    "NR", "NC", "NG", "NI", "NP", "NS", "NY", "NV",
    "DR", "DN", "DC", "DE", "DG", "DF", "DS", "DT", "DY",
    "CR", "CN", "CD", "CC", "CI", "CL", "CK", "CT", "CY", "CV",
    "EA", "ER", "ED", "EC", "EE", "EG", "EI", "EL", "EK",
    "EF", "EP", "ET", "EV",
    "QN", "QF", "QV",
    "GA", "GR", "GC", "GE", "GG", "GI", "GL", "GK", "GF", "GP", "GY",
    "HA", "HP", "HT",
    "IA", "IR", "ID", "II", "IL", "IF", "IP", "IS", "IV",
    "LA", "LR", "LD", "LC", "LG", "LI", "LK", "LM", "LF",
    "LS", "LT", "LY", "LV",
    "KA", "KN", "KC", "KG", "KI", "KL", "KK", "KP", "KY",
    "MA", "MD", "ME", "MI", "MK", "MF", "MP", "MS", "MV",
    "FR", "FE", "FQ", "FG", "FL", "FF", "FS", "FT", "FY", "FV",
    "PA", "PR", "PC", "PE", "PL", "PK", "PP", "PS", "PV",
    "SA", "SR", "SD", "SC", "SG", "SH", "SI", "SL", "SP", "ST", "SY",
    "TA", "TR", "TC", "TE", "TQ", "TG", "TI", "TL", "TP", "TS", "TV",
    "WA",
    "YN", "YD", "YC", "YQ", "YG", "YP",
    "VA", "VR", "VD", "VC", "VE", "VG", "VI", "VL", "VK",
    "VS", "VT", "VY", "VV"
]
assert len(selected_features) == 343, f"Expected 343 features, got {len(selected_features)}"


def keras_predict_proba(X):
    """Return probabilities as [P(Non-AMP), P(AMP)] for LIME (X already scaled)."""
    amp_model, _ = get_amp_model()
    preds = amp_model.predict(X, verbose=0)
    if preds.ndim == 1 or preds.shape[1] == 1:
        preds = preds.reshape(-1, 1)
        return np.hstack([1 - preds, preds])  # sigmoid output assumed = P(AMP)
    return preds


def extract_features(sequence):
    """Compute CTD + AAC, select the 343 training columns IN ORDER, then scale."""
    sequence = ''.join([aa for aa in sequence.upper() if aa in "ACDEFGHIKLMNPQRSTVWY"])
    if len(sequence) < 10:
        return "Error: Sequence too short."

    try:
        _, amp_scaler = get_amp_model()

        # Compute full feature pool
        ctd_features = CTD.CalculateCTD(sequence)
        aac = AAComposition.CalculateAADipeptideComposition(sequence)

        # Merge everything into one lookup dict
        pool = {}
        pool.update(ctd_features)
        pool.update(aac)

        # Verify all needed features are present
        missing = [f for f in selected_features if f not in pool]
        if missing:
            return f"Error: Missing features from propy: {missing[:5]}..."

        # Build the 343-wide row IN THE EXACT TRAINING ORDER, THEN scale.
        ordered_values = [pool[f] for f in selected_features]
        feature_row = np.array(ordered_values, dtype=np.float64).reshape(1, -1)

        scaled = amp_scaler.transform(feature_row)  # scaler expects exactly 343 cols
        return scaled.astype(np.float32)
    except Exception as e:
        return f"Error in feature extraction: {str(e)}"


def predictmic(sequence):
    """Run MIC prediction in a SEPARATE process (mic_worker.py).

    This isolates PyTorch/ProtBert from TensorFlow, preventing the native-library
    crash (exit 139) and keeping peak memory low. The worker prints a JSON dict on
    its last stdout line; we parse and return it.
    """
    sequence = ''.join([aa for aa in sequence.upper() if aa in "ACDEFGHIKLMNPQRSTVWY"])
    if len(sequence) < 10:
        return {"Error": "Sequence too short or invalid."}

    try:
        # First run downloads ProtBert (~1.6GB), so allow a generous timeout.
        proc = subprocess.run(
            [sys.executable, "mic_worker.py", sequence],
            capture_output=True,
            text=True,
            timeout=900  # 15 minutes; mostly for the one-time model download
        )
    except subprocess.TimeoutExpired:
        return {"Error": "MIC prediction timed out (model download may still be in progress; try again shortly)."}
    except Exception as e:
        return {"Error": f"Failed to start MIC worker: {str(e)}"}

    if proc.returncode != 0:
        # Worker crashed; surface stderr tail for debugging
        tail = (proc.stderr or "").strip().splitlines()[-3:]
        return {"Error": f"MIC worker exited with code {proc.returncode}. {' '.join(tail)}"}

    # Parse the last non-empty stdout line as JSON
    out_lines = [ln for ln in (proc.stdout or "").splitlines() if ln.strip()]
    if not out_lines:
        return {"Error": "MIC worker produced no output."}

    try:
        return json.loads(out_lines[-1])
    except Exception:
        return {"Error": f"Could not parse MIC worker output: {out_lines[-1][:200]}"}


def full_prediction(sequence):
    print("[CHECKPOINT] full_prediction called", flush=True)
    features = extract_features(sequence)
    if isinstance(features, str):
        print("[CHECKPOINT] extract_features returned error:", features, flush=True)
        return features
    print("[CHECKPOINT] features extracted OK, shape:", features.shape, flush=True)

    amp_model, _ = get_amp_model()
    print("[CHECKPOINT] AMP model loaded, running predict...", flush=True)
    raw_pred = amp_model.predict(features, verbose=0)
    print("[CHECKPOINT] AMP predict done:", raw_pred, flush=True)

    if raw_pred.ndim == 1 or raw_pred.shape[1] == 1:
        prob_amp = float(raw_pred.flatten()[0])  # sigmoid output assumed = P(AMP)
        if prob_amp >= 0.5:
            prediction = 1
            confidence = round(prob_amp * 100, 2)
        else:
            prediction = 0
            confidence = round((1 - prob_amp) * 100, 2)
    else:
        class_idx = int(np.argmax(raw_pred[0]))
        prediction = class_idx
        confidence = round(float(raw_pred[0][class_idx]) * 100, 2)

    # Label convention: 1 = AMP, 0 = Non-AMP (swap if your model is reversed)
    amp_result = "Antimicrobial Peptide (AMP)" if prediction == 1 else "Non-AMP"
    result = f"Prediction: {amp_result}\nConfidence: {confidence}%\n"

    if prediction == 1:
        print("[CHECKPOINT] AMP detected, starting MIC (ProtBert)...", flush=True)
        mic_values = predictmic(sequence)
        print("[CHECKPOINT] MIC done:", mic_values, flush=True)
        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"

    try:
        from lime.lime_tabular import LimeTabularExplainer
        sample_data = np.random.rand(100, len(selected_features))
        explainer = LimeTabularExplainer(
            training_data=sample_data,
            feature_names=selected_features,
            class_names=["Non-AMP", "AMP"],
            mode="classification"
        )
        explanation = explainer.explain_instance(
            data_row=features[0],
            predict_fn=keras_predict_proba,
            num_features=10
        )
        result += "\nTop Features Influencing Prediction:\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"

    return result


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."
)

iface.launch()