TCGA-BRCA - PAM50 Cancer Subtype Classification

Cohort Summary:

Stage	Methylation	mRNA	miRNA	Clinical	PAM50
Raw (features x samples)	20,107 x 885	18,321 x 1,212	503 x 1,189	1,098 x 18	1,087 x 1
Final aligned (samples x features)	769 x 20,106	769 x 16,758	769 x 354	769 x 17	769 x 1
After feature selection	769 x 400	769 x 200	769 x 100	769 x 17	769 x 1

Target Definition: PAM50 Subtype (5-class): LumA (n=419), LumB (n=140), Basal (n=130), Her2 (n=46), and Normal (n=34).

Dataset Source

• Omics Data: FireBrowse BRCA

• PAM50 Data: TCGAbiolinks

import pandas as pd
from pathlib import Path
root = Path("/home/vicente/Github/BioNeuralNet/TCGA_BRCA_DATA")

mirna_raw = pd.read_csv(root/"BRCA.miRseq_RPKM_log2.txt", sep="\t",index_col=0,low_memory=False)                            
rna_raw = pd.read_csv(root / "BRCA.uncv2.mRNAseq_RSEM_normalized_log2.txt", sep="\t",index_col=0,low_memory=False)
meth_raw = pd.read_csv(root/"BRCA.meth.by_mean.data.txt", sep='\t',index_col=0,low_memory=False)
clinical_raw = pd.read_csv(root / "BRCA.clin.merged.picked.txt",sep="\t", index_col=0, low_memory=False)

# display all shapes and first few rows of each dataset
display(mirna_raw.iloc[:3,:5])
display(mirna_raw.shape)

display(rna_raw.iloc[:3,:5])
display(rna_raw.shape)

display(meth_raw.iloc[:3,:5])
display(meth_raw.shape)

display(clinical_raw.iloc[:3,:5])
display(clinical_raw.shape)

TCGA-BioLink: Pam50

This section demonstrates how to use the TCGAbiolinks R package to access and download molecular subtype data. It begins by ensuring TCGAbiolinks is installed, then loads the package. It retrieves PAM50 molecular subtype labels using TCGAquery_subtype() and writes them to a CSV file.

  # Install TCGAbiolinks
  if (!requireNamespace("TCGAbiolinks", quietly = TRUE)) {
    if (!requireNamespace("BiocManager", quietly = TRUE))
      install.packages("BiocManager")
    BiocManager::install("TCGAbiolinks")
  }

  # Load the library
  library(TCGAbiolinks)

  # Download PAM50 subtype labels
  pam50_df <- TCGAquery_subtype(tumor = "BRCA")[ , c("patient", "BRCA_Subtype_PAM50")]
  write.csv(pam50_df, file = "BRCA_PAM50_labels.csv", row.names = FALSE, quote = FALSE)

pam50 = pd.read_csv(root /"BRCA_PAM50_labels.csv",index_col=0)

display(pam50.iloc[:5,:5])
display(pam50.shape)

Data Processing Summary

1. Transpose Data: All raw data (miRNA, RNA, etc.) is flipped so rows represent patients and columns represent features.

2. Standardize Patient IDs: Patient IDs in all tables are cleaned to the 12-character TCGA format (e.g., TCGA-AB-1234) for matching.

3. Handle Duplicates: Duplicate patient rows are averaged in the omics data. The first entry is kept for duplicate patients in the clinical data.

4. Find Common Patients: The script identifies the list of patients that exist in all datasets.

5. Sparse Feature Filter: Remove features where missingness exceeds the imputation limit (20%).

6. Final Data Aligment: All data tables are filtered down to only this common list of patients.

7. Impute Missing Values: Remaining missing data (NaNs) are estimated and filled using mean imputation.

8. Extract Target: The BRCA_Subtype_PAM50 column is pulled from the pam50 dataframe from the last cell. We assigned each label a numerical value.

from bioneuralnet.utils import m_transform, impute_knn
from bioneuralnet.utils import data_stats, sparse_filter

mirna = mirna_raw.T
rna = rna_raw.T
meth = meth_raw.T
clinical = clinical_raw.T

print(f"miRNA (samples, features): {mirna.shape}")
print(f"RNA (samples, features): {rna.shape}")
print(f"Methylation (samples, features): {meth.shape}")
print(f"Clinical (samples, features): {clinical.shape}")

def trim_barcode(idx):
    return idx.to_series().str.slice(0, 12)

# Standardized patient IDs across all files
meth.index = trim_barcode(meth.index)
rna.index = trim_barcode(rna.index)
mirna.index = trim_barcode(mirna.index)
clinical.index = clinical.index.str.upper()
clinical.index.name = "Patient_ID"

# Convert all data to numeric, coercing errors to NaN
meth = meth.apply(pd.to_numeric, errors='coerce')
rna = rna.apply(pd.to_numeric, errors='coerce')
mirna = mirna.apply(pd.to_numeric, errors='coerce')

# For any duplicate columns in the omics data -> average their values
meth = meth.groupby(meth.index).mean()
rna = rna.groupby(rna.index).mean()
mirna = mirna.groupby(mirna.index).mean()

# Drop unwanted columns
clinical.drop(columns=["Composite Element REF"], errors="ignore", inplace=True)
meth.drop(columns=["Composite Element REF"], errors="ignore", inplace=True)

# For any duplicate rows (patients) in the clinical data -> keep the first occurrence
clinical = clinical[~clinical.index.duplicated(keep='first')]

# Convert beta values to M-values using bioneuralnet utility with small epsilon to avoid log(0)
meth_m = m_transform(meth, eps=1e-7) 

print(f"\nMethylation shape: {meth_m.shape}")
print(f"RNA shape: {rna.shape}")
print(f"miRNA shape: {mirna.shape}")
print(f"Clinical shape: {clinical.shape}")

# Standardize column names
for df in [meth_m, rna, mirna]:
    df.columns = df.columns.str.replace(r"\?", "unknown_", regex=True)
    df.columns = df.columns.str.replace(r"\|", "_", regex=True)
    df.columns = df.columns.str.replace("-", "_", regex=False)
    df.columns = df.columns.str.replace(r"_+", "_", regex=True)
    df.columns = df.columns.str.strip("_")

# Patients common across all data files
common_patients = sorted(list(set(meth_m.index) & set(rna.index) & set(mirna.index) & set(clinical.index)))

# Subset to only common patients
X_meth = meth_m.loc[common_patients].copy()
X_rna = rna.loc[common_patients].copy()
X_mirna = mirna.loc[common_patients].copy()
clinical = clinical.loc[common_patients].copy()

# Calling data_stats on subset to prevent issues downstream
data_stats(X_mirna, "miRNA")
data_stats(X_rna, "RNA")
data_stats(X_meth, "Methylation")

# apply the sparse_filter and missing_fraction parameter
X_mirna = sparse_filter(X_mirna, missing_fraction=0.20)
X_rna = sparse_filter(X_rna, missing_fraction=0.20)
X_meth = sparse_filter(X_meth, missing_fraction=0.20)

targets = pam50['BRCA_Subtype_PAM50'] 
mapping_brca = {
    'LumA': 0, 
    'Her2': 1, 
    'LumB': 2, 
    'Basal': 3, 
    'Normal': 4
}
Y_label = targets.map(mapping_brca).to_frame(name="target")

# Secondary aligment after sparse filter
final_patients = sorted(
    set(X_meth.index) & 
    set(X_rna.index) & 
    set(X_mirna.index) &
    set(clinical.index) &
    set(Y_label.index)
)

# subsetting based on secondary alignment
X_meth = X_meth.loc[final_patients]
X_rna = X_rna.loc[final_patients]
X_mirna = X_mirna.loc[final_patients]
clinical = clinical.loc[final_patients]
Y_label = Y_label.loc[final_patients]

# impute the remaining missing values with knn
X_meth = impute_knn(X_meth, n_neighbors=5)
X_rna = impute_knn(X_rna, n_neighbors=5)
X_mirna = impute_knn(X_mirna, n_neighbors=5)

# Inspect the first 3 rows and 5 colums.

display(X_meth.iloc[:3,:5])
display(X_meth.shape)

display(X_rna.iloc[:3,:5])
display(X_rna.shape)

display(X_mirna.iloc[:3,:5])
display(X_mirna.shape)

display(clinical.iloc[:3,:5])
display(clinical.shape)

display(Y_label.value_counts())

Feature Selection

Unsupervised feature selection was performed across all three omic modalities using Laplacian Score filtering.

from bioneuralnet.utils import laplacian_score

meth_lap = laplacian_score(X_meth, n_keep=400)
rna_lap = laplacian_score(X_rna, n_keep=200)
mirna_lap = laplacian_score(X_mirna, n_keep=100)

# We inspected each of the following categorical featues and built a ordinal mapping

print(clinical["pathologic_stage"].value_counts())
print(clinical["pathology_T_stage"].value_counts())
print(clinical["pathology_N_stage"].value_counts())
print(clinical["pathology_M_stage"].value_counts())

import numpy as np
from bioneuralnet.utils import preprocess_clinical

columns_to_drop = [
    'Composite Element REF',
    'tumor_tissue_site',
    'gender',
    'date_of_initial_pathologic_diagnosis',
    'vital_status',
    'days_to_death',
    'days_to_last_followup',
    'days_to_last_known_alive',
    'radiation_therapy'
]

mappings = {
    'pathologic_stage': {
        'stage i': 1.0, 'stage ia': 1.1, 'stage ib': 1.2,
        'stage ii': 2.0, 'stage iia': 2.1, 'stage iib': 2.2,
        'stage iii': 3.0, 'stage iiia': 3.1, 'stage iiib': 3.2, 'stage iiic': 3.3,
        'stage iv': 4.0, 
        'stage x': np.nan, 'nan': np.nan
    },
    'pathology_T_stage': {
        't1': 1.0, 't1a': 1.1, 't1b': 1.2, 't1c': 1.3,
        't2': 2.0, 't2a': 2.1, 't2b': 2.2,
        't3': 3.0, 't3a': 3.1,
        't4': 4.0, 't4a': 4.1, 't4b': 4.2, 't4c': 4.3, 't4d': 4.4,
        'tx': np.nan, 'nan': np.nan
    },
    'pathology_N_stage': {
        'n0': 0.0, 'n0 (i-)': 0.0, 
        'n0 (mol+)': 0.1, 'n0 (i+)': 0.2,
        'n1mi': 0.5,
        'n1': 1.0, 'n1a': 1.1, 'n1b': 1.2, 'n1c': 1.3,
        'n2': 2.0, 'n2a': 2.1, 'n2b': 2.2,
        'n3': 3.0, 'n3a': 3.1, 'n3b': 3.2, 'n3c': 3.3,
        'nx': np.nan, 'nan': np.nan
    },
    'pathology_M_stage': {
        'm0': 0.0, 'cm0 (i+)': 0.5, 'm1': 1.0,
        'mx': np.nan, 'nan': np.nan
    }
}
continuous_cols = ['years_to_birth', 'number_of_lymph_nodes']

clinical_numeric = preprocess_clinical(
    X=clinical,
    drop_columns=columns_to_drop,
    continuous_columns=continuous_cols,
    ordinal_mappings=mappings)


print(f"\nNaN count: {clinical_numeric.isna().sum().sum()}")
print(f"\nInf count: {np.isinf(clinical_numeric.values).sum()}")
print(f"\nColumn dtypes:\n{clinical_numeric.dtypes}")

Data Availability

To facilitate rapid experimentation and reproduction of our results, the fully processed and feature-selected dataset used in this example has been made available directly within the package.

Users can load this dataset, bypassing all preceding data acquisition, preprocessing, and feature selection steps.

import pandas as pd
from bioneuralnet.datasets import DatasetLoader

tgca_brca = DatasetLoader("brca")
print(tgca_brca.shape)

X_methylation = tgca_brca["methylation"]
X_rna = tgca_brca["rna"]
X_mirna = tgca_brca["mirna"]
clinical_numeric = tgca_brca["clinical"]
Y_labels = tgca_brca["target"]

display(X_methylation.iloc[:3,:5])
display(X_rna.iloc[:3,:5])
display(X_mirna.iloc[:3,:5])
display(clinical_numeric.iloc[:3,:5])
display(Y_labels.iloc[:3,:5])

Reproducibility and Seeding

This utility function propagates the seed to all sources of randomness, including random, numpy, and torch.

from bioneuralnet.utils import set_seed

SEED = 8183
set_seed(SEED)

from bioneuralnet.network import threshold_network

omics_brca = pd.concat([X_methylation, X_rna, X_mirna], axis=1)

threshold_63_22 = threshold_network(omics_brca,b=6.3, k=22)
threshold_67_15 = threshold_network(omics_brca,b=6.7, k=15)

from bioneuralnet.network import NetworkAnalyzer

thre_22_analysis = NetworkAnalyzer(threshold_63_22, source_omics=[X_methylation, X_rna, X_mirna])
thre_22_analysis.basic_statistics(0.0001)
thre_22_analysis.hub_analysis(0.0001)
thre_22_analysis.find_strongest_edges(5)

thre_15_analysis = NetworkAnalyzer(threshold_67_15,  source_omics=[X_methylation, X_rna, X_mirna])
thre_15_analysis.basic_statistics(0.00001)
thre_15_analysis.hub_analysis(0.00001)
thre_15_analysis.find_strongest_edges(5)

# A targeted subset of clinical variables is selected. This can be modified as needed for testing/experiments

clinical_numeric = clinical_numeric.copy()[[
    'histological_type_infiltrating lobular carcinoma',
    'histological_type_infiltrating ductal carcinoma',
    'histological_type_metaplastic carcinoma',
    'race_black or african american',
    'race_white',
    'years_to_birth',
    'pathologic_stage',
    'number_of_lymph_nodes',
    'pathology_N_stage'
]]

Classification using DPMON: Training and Evaluation

from bioneuralnet.downstream_task import DPMON

output_dir_base = Path("/home/vicente/Github/BioNeuralNet/dpmon_gat_results/brca")

current_output_dir = output_dir_base / "threshold_63_22"
current_output_dir.mkdir(parents=True, exist_ok=True)

dpmon_params_base = {
    "adjacency_matrix": threshold_63_22,
    "omics_list": omics_brca,
    "phenotype_data": Y_labels,
    "phenotype_col": "target",
    "clinical_data": clinical_numeric,
    "tune_trials" : 20,
    "model": 'GAT',
    "tune": True, 
    "cv": True,   
    "n_folds": 5,
    "repeat_num": 5,
    "gpu": True,
    "seed": SEED,
    "output_dir": current_output_dir
}

gat_dpmon = DPMON(**dpmon_params_base)
gat_predictions, gat_metrics, gat_embeddings = gat_dpmon.run()

output_dir_base = Path("/home/vicente/Github/BioNeuralNet/dpmon_gcn_results/brca")

current_output_dir = output_dir_base / "threshold_63_22"
current_output_dir.mkdir(parents=True, exist_ok=True)

dpmon_params_base = {
    "adjacency_matrix": threshold_63_22,
    "omics_list": omics_brca,
    "phenotype_data": Y_labels,
    "phenotype_col": "target",
    "clinical_data": clinical_numeric,
    "tune_trials" : 20,
    "model": 'GCN',
    "tune": True, 
    "cv": True,   
    "n_folds": 5,
    "repeat_num": 5,
    "gpu": True,
    "seed": SEED,
    "output_dir": current_output_dir
}

gcn_dpmon = DPMON(**dpmon_params_base)
gcn_predictions, gcn_metrics, gcn_embeddings = gcn_dpmon.run()

from bioneuralnet.downstream_task import DPMON

output_dir_base = Path("/home/vicente/Github/BioNeuralNet/dpmon_gat_multiple_results/brca")
all_results_gat = []

# To test several graphs sequentially. We can loop over the graphs dynamically like this:

comparison_runs = [
    {"name": "threshold_63_22", "graph": threshold_63_22, "omics": omics_brca},
    {"name": "threshold_67_15", "graph": threshold_67_15, "omics": omics_brca},
]

for run_config in comparison_runs:
    graph_name = run_config["name"]
    A_full = run_config["graph"]
    omics = run_config["omics"]
    
    current_output_dir = output_dir_base / graph_name
    current_output_dir.mkdir(parents=True, exist_ok=True)

    dpmon_params_base = {
        "adjacency_matrix": A_full,
        "omics_list": omics,
        "phenotype_data": Y_labels,
        "phenotype_col": "target",
        "clinical_data": clinical_numeric,
        "tune_trials" : 20,
        "model": 'GAT',
        "tune": True, 
        "cv": True,   
        "n_folds": 5,
        "repeat_num":5,
        "gpu": True,
        "cuda": 0,
        "seed": SEED,
        "output_dir": current_output_dir
    }
    
    dpmon_tunned = DPMON(**dpmon_params_base)
    predictions_df, metrics, embeddings = dpmon_tunned.run()

    all_results_gat.append({
        "graph": graph_name,
        "predictions": predictions_df,
        "metrics": metrics,
        "embeddings": embeddings
    })

# Once all the experiments are done, we can review the results for each graph configuration:
for res in all_results_gat:
    graph_name = res["graph"]
    graph_metrics = res["metrics"]
    
    acc_row = graph_metrics.loc[graph_metrics['Metric'] == 'Accuracy'].iloc[0]
    f1_macro_row = graph_metrics.loc[graph_metrics['Metric'] == 'F1 Macro'].iloc[0]
    f1_weighted_row = graph_metrics.loc[graph_metrics['Metric'] == 'F1 Weighted'].iloc[0]
    recall_row = graph_metrics.loc[graph_metrics['Metric'] == 'Recall'].iloc[0]
    precission_row = graph_metrics.loc[graph_metrics['Metric'] == 'Precision'].iloc[0]
    auc_row = graph_metrics.loc[graph_metrics['Metric'] == 'AUC'].iloc[0]
    aupr_row = graph_metrics.loc[graph_metrics['Metric'] == 'AUPR'].iloc[0]
    
    acc_avg, acc_std = acc_row['Average'], acc_row['StdDev']
    f1_macro_avg, f1_macro_std = f1_macro_row['Average'], f1_macro_row['StdDev']
    f1_weighted_avg, f1_weighted_std = f1_weighted_row['Average'], f1_weighted_row['StdDev']
    recall_avg, recall_std = recall_row['Average'], recall_row['StdDev']
    precision_avg, precision_std = precission_row['Average'], precission_row['StdDev']
    auc_avg, auc_std = auc_row['Average'], auc_row['StdDev']
    aupr_avg, aupr_std = aupr_row['Average'], aupr_row['StdDev']

    print(f"\nResults for: {graph_name}")
    print(f"Accuracy (Avg +/- Std): {acc_avg:.4f} +/- {acc_std:.4f}")
    print(f"F1 Macro (Avg +/- Std): {f1_macro_avg:.4f}  +/- {f1_macro_std:.4f}")
    print(f"F1 Weighted (Avg +/- Std): {f1_weighted_avg:.4f} +/- {f1_weighted_std:.4f}")
    print(f"Recall: {recall_avg:.4f} +/- {recall_std:.4f}")
    print(f"Precision: {precision_avg:.4f} +/- {precision_std:.4f}")
    print(f"AUC: {auc_avg:.4f} +/- {auc_std:.4f}")
    print(f"AUPR: {aupr_avg:.4f} +/- {aupr_std:.4f}")

from sklearn.model_selection import StratifiedKFold, ParameterSampler, RepeatedStratifiedKFold
from sklearn.metrics import accuracy_score, f1_score, recall_score, roc_auc_score, average_precision_score, precision_score
from sklearn.preprocessing import StandardScaler, label_binarize
from sklearn.linear_model import LogisticRegression
from sklearn.neural_network import MLPClassifier
from sklearn.pipeline import Pipeline
from sklearn.base import clone
from scipy.stats import loguniform, randint
from sklearn.ensemble import RandomForestClassifier
from sklearn.tree import DecisionTreeClassifier
from xgboost import XGBClassifier
from sklearn.svm import SVC

SEED = 8183
y = Y_labels["target"]
X = pd.concat([omics_brca,clinical_numeric],axis=1)

print(f"Successfully created X vector with shape: {X.shape}")
print(f"Successfully created y vector with shape: {y.shape}")

pipe_lr = Pipeline([
    ('scaler', StandardScaler()),
    ('model', LogisticRegression(solver='lbfgs', max_iter=1000, penalty=None, random_state=SEED))
])

pipe_mlp = Pipeline([
    ('scaler', StandardScaler()),
    ('model', MLPClassifier(max_iter=500, early_stopping=True, n_iter_no_change=10, random_state=SEED))
])

pipe_xgb = Pipeline([
    ('scaler', StandardScaler()),
    ('model', XGBClassifier(eval_metric='logloss', tree_method='hist', max_bin=128, random_state=SEED))
])
pipe_rf = Pipeline([
    ('scaler', StandardScaler()),
    ('model', RandomForestClassifier(random_state=SEED))
])

pipe_svm = Pipeline([
    ('scaler', StandardScaler()),
    ('model', SVC(probability=True, random_state=SEED))
])

pipe_dt = Pipeline([
    ('scaler', StandardScaler()),
    ('model', DecisionTreeClassifier(random_state=SEED))
])

params_lr = {'model__penalty': ['l2'], 'model__C': loguniform(1e-4, 1e2)}

params_mlp = {
    'model__hidden_layer_sizes': [(100,), (100, 50), (50, 50)],
    'model__activation': ['relu', 'tanh'],
    'model__alpha': loguniform(1e-5, 1e-1),
    'model__learning_rate_init': loguniform(1e-4, 1e-2)
}
params_xgb = {
    'model__n_estimators': randint(50, 200),
    'model__learning_rate': loguniform(0.01, 0.3),
    'model__max_depth': randint(3, 7),
    'model__subsample': [0.8, 1.0], 
    'model__colsample_bytree': [0.8, 1.0]
}
params_rf = {
    'model__n_estimators': randint(100, 300),
    'model__max_depth': [10, 20, 30, None],
    'model__min_samples_split': randint(2, 10),
    'model__min_samples_leaf': randint(1, 5),
    'model__max_features': ['sqrt', 'log2']
}
params_svm = {
    'model__C': loguniform(1e-2, 1e3),
    'model__kernel': ['rbf', 'linear'],
    'model__gamma': ['scale', 'auto']
}

params_dt = {
    'model__max_depth': randint(3, 15),
    'model__min_samples_split': randint(2, 20),
    'model__criterion': ['gini', 'entropy']
}

models_to_tune = {
    "LogisticRegression": (pipe_lr, params_lr),
    "SVM": (pipe_svm, params_svm),
    "MLP": (pipe_mlp, params_mlp),
    "XGBoost": (pipe_xgb, params_xgb),
    "RandomForest": (pipe_rf, params_rf),
    "DecisionTree": (pipe_dt, params_dt),
}

all_results = {
    "LogisticRegression": {"acc": [], "f1_w": [], "f1_m": [], "recall": [], "auc": [], "precision": [], "aupr": []},
    "MLP": {"acc": [], "f1_w": [], "f1_m": [], "recall": [], "auc": [], "precision": [], "aupr": []},
    "XGBoost": {"acc": [], "f1_w": [], "f1_m": [], "recall": [], "auc": [], "precision": [], "aupr": []},
    "RandomForest": {"acc": [], "f1_w": [], "f1_m": [], "recall": [], "auc": [], "precision": [], "aupr": []},
    "SVM": {"acc": [], "f1_w": [], "f1_m": [], "recall": [], "auc": [], "precision": [], "aupr": []},
    "DecisionTree": {"acc": [], "f1_w": [], "f1_m": [], "recall": [], "auc": [], "precision": [], "aupr": []},
}



for model_name, (pipeline, param_dist) in models_to_tune.items():
    print(f"Evaluating model with nested CV: {model_name}")
    
    outer_cv = RepeatedStratifiedKFold(n_splits=5, n_repeats=5, random_state=SEED)

    # Inner fold loop is for finding the best hyperparameters per fold.
    for fold_idx, (train_idx, test_idx) in enumerate(outer_cv.split(X, y), start=1):
        X_train_outer, X_test_outer = X.iloc[train_idx], X.iloc[test_idx]
        y_train_outer, y_test_outer = y.iloc[train_idx], y.iloc[test_idx]
        inner_cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=SEED)
        
        best_score_fold = -np.inf
        best_params_fold = None

        # A fixed seed = same hyperparamters each fold.
        # No seed = different hyperparameters each fold. 
        # This adds more randomness, and may yield better generalization.
        param_sampler = list(ParameterSampler(param_dist, n_iter=20))
        
        for params in param_sampler:
            inner_scores = []
            
            for inner_train_idx, inner_val_idx in inner_cv.split(X_train_outer, y_train_outer):
                X_train_inner = X_train_outer.iloc[inner_train_idx]
                X_val_inner = X_train_outer.iloc[inner_val_idx]
                y_train_inner = y_train_outer.iloc[inner_train_idx]
                y_val_inner = y_train_outer.iloc[inner_val_idx]
                
                inner_pipeline = clone(pipeline)
                inner_pipeline.set_params(**params)
                inner_pipeline.fit(X_train_inner, y_train_inner)
                
                y_val_pred = inner_pipeline.predict(X_val_inner)
                score = f1_score(y_val_inner, y_val_pred, average='macro', zero_division=0)
                inner_scores.append(score)
            
            mean_score = np.mean(inner_scores)
            if mean_score > best_score_fold:
                best_score_fold = mean_score
                best_params_fold = params
        
        print(f"Outer fold {fold_idx}: best params (inner CV F1-M={best_score_fold:.4f})")
        print(f"{best_params_fold}")

        final_pipeline = clone(pipeline)
        final_pipeline.set_params(**best_params_fold)
        final_pipeline.fit(X_train_outer, y_train_outer)
        
        preds = final_pipeline.predict(X_test_outer)
        
        if hasattr(final_pipeline, "predict_proba"):
            proba = final_pipeline.predict_proba(X_test_outer)
        else:
            proba = None
        
        acc = accuracy_score(y_test_outer, preds)
        f1_w = f1_score(y_test_outer, preds, average='weighted', zero_division=0)
        f1_m = f1_score(y_test_outer, preds, average='macro', zero_division=0)
        recall = recall_score(y_test_outer, preds, average='macro', zero_division=0)
        precision = precision_score(y_test_outer, preds, average='macro', zero_division=0)
        
        auc = np.nan
        aupr = np.nan
        
        if proba is not None:
            try:
                if len(np.unique(y)) == 2:
                    auc = roc_auc_score(y_test_outer, proba[:, 1])
                    aupr = average_precision_score(y_test_outer, proba[:, 1])
                else:
                    auc = roc_auc_score(y_test_outer, proba, multi_class='ovr', average='macro')
                    y_test_bin = label_binarize(y_test_outer, classes=np.unique(y))
                    aupr = average_precision_score(y_test_bin, proba, average='macro')
            except Exception:
                auc = np.nan
                aupr = np.nan

        print(f"Fold {fold_idx} results: Acc={acc:.4f}, F1-W={f1_w:.4f}, "
            f"F1-M={f1_m:.4f}, Recall={recall:.4f}, Precision={precision:.4f}, AUC={auc:.4f}, AUPR={aupr:.4f}")
        
        all_results[model_name]["acc"].append(acc)
        all_results[model_name]["f1_w"].append(f1_w)
        all_results[model_name]["f1_m"].append(f1_m)
        all_results[model_name]["recall"].append(recall)
        all_results[model_name]["precision"].append(precision)
        all_results[model_name]["auc"].append(auc)
        all_results[model_name]["aupr"].append(aupr)

print("\nFINAL BASELINE RESULTS\n")
for model_name, metrics in all_results.items():
    avg_acc = np.mean(metrics["acc"])
    std_acc = np.std(metrics["acc"])
    avg_f1_w = np.mean(metrics["f1_w"])
    std_f1_w = np.std(metrics["f1_w"])
    avg_f1_m = np.mean(metrics["f1_m"])
    std_f1_m = np.std(metrics["f1_m"])
    avg_recall = np.mean(metrics["recall"])
    std_recall = np.std(metrics["recall"])
    avg_prec = np.mean(metrics["precision"])
    std_prec = np.std(metrics["precision"])
    avg_auc = np.nanmean(metrics["auc"])
    std_auc = np.nanstd(metrics["auc"])
    avg_aupr = np.nanmean(metrics["aupr"])
    std_aupr = np.nanstd(metrics["aupr"])
    
    print(f"\n{model_name}:")
    print(f"Accuracy: {avg_acc:.4f} +/- {std_acc:.4f}")
    print(f"Precision: {avg_prec:.4f} +/- {std_prec:.4f}")
    print(f"F1 Weighted: {avg_f1_w:.4f} +/- {std_f1_w:.4f}")
    print(f"F1 Macro: {avg_f1_m:.4f} +/- {std_f1_m:.4f}")
    print(f"Recall: {avg_recall:.4f} +/- {std_recall:.4f}")
    print(f"AUC: {avg_auc:.4f} +/- {std_auc:.4f}")
    print(f"AUPR: {avg_aupr:.4f} +/- {std_aupr:.4f}")