import gradio as gr import joblib import numpy as np import pandas as pd from propy import AAComposition, CTD import tensorflow as tf from tensorflow.keras.models import load_model import torch from transformers import BertTokenizer, BertModel from lime.lime_tabular import LimeTabularExplainer from math import expm1 # Load AMP Classifier (Keras) and Scaler model = load_model("Comb1_aac_ctd_RFE_selected_features_model.keras") scaler = joblib.load("Comb1_aac_ctd_RFE_selected_features_scaler.joblib") # Load ProtBert (for MIC prediction) 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 (AAC + CTD, RFE-selected) # Note: 'Activity' is the target label and is excluded from input features 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', 'A', 'R', 'N', 'D', 'C', 'E', 'Q', 'G', 'H', 'I', 'L', 'K', 'M', 'F', 'P', 'S', 'T', 'W', 'Y', 'V', 'AA', 'AR', 'AN', 'AD', 'AC', 'AE', 'AQ', 'AG', 'AH', 'AI', 'AL', 'AK', 'AM', 'AF', 'AP', 'AS', 'AT', 'AW', 'AY', 'AV', 'RA', 'RR', 'RN', 'RD', 'RC', 'RE', 'RQ', 'RG', 'RH', 'RI', 'RL', 'RK', 'RM', 'RF', 'RP', 'RS', 'RT', 'RW', 'RY', 'RV', 'NA', 'NR', 'NN', 'ND', 'NC', 'NE', 'NQ', 'NG', 'NH', 'NI', 'NL', 'NK', 'NM', 'NF', 'NP', 'NS', 'NT', 'NW', 'NY', 'NV', 'DA', 'DR', 'DN', 'DD', 'DC', 'DE', 'DQ', 'DG', 'DH', 'DI', 'DL', 'DK', 'DM', 'DF', 'DP', 'DS', 'DT', 'DW', 'DY', 'DV', 'CA', 'CR', 'CN', 'CD', 'CC', 'CE', 'CQ', 'CG', 'CH', 'CI', 'CL', 'CK', 'CM', 'CF', 'CP', 'CS', 'CT', 'CW', 'CY', 'CV', 'EA', 'ER', 'EN', 'ED', 'EC', 'EE', 'EQ', 'EG', 'EH', 'EI', 'EL', 'EK', 'EM', 'EF', 'EP', 'ES', 'ET', 'EW', 'EY', 'EV', 'QA', 'QR', 'QN', 'QD', 'QC', 'QE', 'QQ', 'QG', 'QH', 'QI', 'QL', 'QK', 'QM', 'QF', 'QP', 'QS', 'QT', 'QW', 'QY', 'QV', 'GA', 'GR', 'GN', 'GD', 'GC', 'GE', 'GQ', 'GG', 'GH', 'GI', 'GL', 'GK', 'GM', 'GF', 'GP', 'GS', 'GT', 'GW', 'GY', 'GV', 'HA', 'HR', 'HN', 'HD', 'HC', 'HE', 'HQ', 'HG', 'HH', 'HI', 'HL', 'HK', 'HM', 'HF', 'HP', 'HS', 'HT', 'HW', 'HY', 'HV', 'IA', 'IR', 'IN', 'ID', 'IC', 'IE', 'IQ', 'IG', 'IH', 'II', 'IL', 'IK', 'IM', 'IF', 'IP', 'IS', 'IT', 'IW', 'IY', 'IV', 'LA', 'LR', 'LN', 'LD', 'LC', 'LE', 'LQ', 'LG', 'LH', 'LI', 'LL', 'LK', 'LM', 'LF', 'LP', 'LS', 'LT', 'LW', 'LY', 'LV', 'KA', 'KR', 'KN', 'KD', 'KC', 'KE', 'KQ', 'KG', 'KH', 'KI', 'KL', 'KK', 'KM', 'KF', 'KP', 'KS', 'KT', 'KW', 'KY', 'KV', 'MA', 'MR', 'MN', 'MD', 'MC', 'ME', 'MQ', 'MG', 'MH', 'MI', 'ML', 'MK', 'MM', 'MF', 'MP', 'MS', 'MT', 'MW', 'MY', 'MV', 'FA', 'FR', 'FN', 'FD', 'FC', 'FE', 'FQ', 'FG', 'FH', 'FI', 'FL', 'FK', 'FM', 'FF', 'FP', 'FS', 'FT', 'FW', 'FY', 'FV', 'PA', 'PR', 'PN', 'PD', 'PC', 'PE', 'PQ', 'PG', 'PH', 'PI', 'PL', 'PK', 'PM', 'PF', 'PP', 'PS', 'PT', 'PW', 'PY', 'PV', 'SA', 'SR', 'SN', 'SD', 'SC', 'SE', 'SQ', 'SG', 'SH', 'SI', 'SL', 'SK', 'SM', 'SF', 'SP', 'SS', 'ST', 'SW', 'SY', 'SV', 'TA', 'TR', 'TN', 'TD', 'TC', 'TE', 'TQ', 'TG', 'TH', 'TI', 'TL', 'TK', 'TM', 'TF', 'TP', 'TS', 'TT', 'TW', 'TY', 'TV', 'WA', 'WR', 'WN', 'WD', 'WC', 'WE', 'WQ', 'WG', 'WH', 'WI', 'WL', 'WK', 'WM', 'WF', 'WP', 'WS', 'WT', 'WW', 'WY', 'WV', 'YA', 'YR', 'YN', 'YD', 'YC', 'YE', 'YQ', 'YG', 'YH', 'YI', 'YL', 'YK', 'YM', 'YF', 'YP', 'YS', 'YT', 'YW', 'YY', 'YV', 'VA', 'VR', 'VN', 'VD', 'VC', 'VE', 'VQ', 'VG', 'VH', 'VI', 'VL', 'VK', 'VM', 'VF', 'VP', 'VS', 'VT', 'VW', 'VY', 'VV' ] # Wrapper to make Keras model behave like a sklearn classifier for LIME def keras_predict_proba(X): """Return probabilities for both classes as [P(Non-AMP), P(AMP)].""" preds = model.predict(X, verbose=0) if preds.ndim == 1 or preds.shape[1] == 1: preds = preds.reshape(-1, 1) # Assuming sigmoid output = P(AMP); adjust if your model is reversed. return np.hstack([1 - preds, preds]) return preds # Dummy data for LIME 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" ) # Feature extraction function (AAC + CTD only) 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: # AAC: 20 single AAs + 400 dipeptides = 420 features dipeptide_features = AAComposition.CalculateAADipeptideComposition(sequence) filtered_aac = {k: dipeptide_features[k] for k in list(dipeptide_features.keys())[:420]} # CTD: Composition, Transition, Distribution ctd_features = CTD.CalculateCTD(sequence) all_features_dict = {} all_features_dict.update(ctd_features) all_features_dict.update(filtered_aac) 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): missing = set(selected_features) - set(normalized_df.columns) return f"Error: Missing features: {list(missing)[:5]}..." selected_df = normalized_df[selected_features].fillna(0) return selected_df.values.astype(np.float32) except Exception as e: return f"Error in feature extraction: {str(e)}" # MIC prediction function (unchanged) 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: 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 # Keras prediction raw_pred = model.predict(features, verbose=0) # Handle sigmoid (1 output) vs softmax (>=2 outputs) if raw_pred.ndim == 1 or raw_pred.shape[1] == 1: prob_amp = float(raw_pred.flatten()[0]) # assume output = P(AMP) if prob_amp >= 0.5: prediction = 1 # AMP confidence = round(prob_amp * 100, 2) else: prediction = 0 # Non-AMP 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 uses the opposite) amp_result = "Antimicrobial Peptide (AMP)" if prediction == 1 else "Non-AMP" result = f"Prediction: {amp_result}\nConfidence: {confidence}%\n" if prediction == 1: 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" # LIME explanation try: 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 # 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." ) iface.launch(share=True)