{ "cells": [ { "cell_type": "markdown", "id": "d408841e", "metadata": {}, "source": [ "# TCGA-KIPAN - Binary Cancer Staging\n", "\n", "**Cohort Summary:**\n", "\n", "| Stage | Methylation | mRNA | miRNA | Clinical |\n", "| --- | --- | --- | --- | --- |\n", "| **Raw** (features x samples) | 20,117 x 867 | 18,272 x 1,020 | 472 x 1,005 | 20 x 941 |\n", "| **Final Aligned** (samples x features) | 658 x 20,116 | 658 x 18,272 | 658 x 472 | 658 x 19 |\n", "| **After feature selection** | 633 x 400 | 633 x 200 | 633 x 100 | 633 x 19 |\n", "\n", "**Target Definition:** Binary cancer stage: Early (Stages I/II, n=417) vs. Late (Stages III/IV, n=216).\n", "\n", "\n", "**Data Source:** [FireBrowse KIPAN](http://firebrowse.org/?cohort=KIPAN)" ] }, { "cell_type": "code", "execution_count": null, "id": "1c56af27", "metadata": {}, "outputs": [], "source": [ "import pandas as pd\n", "from pathlib import Path\n", "root = Path(\"/home/vicente/Github/BioNeuralNet/KIPAN\")\n", "\n", "mirna_raw = pd.read_csv(root/\"KIPAN.miRseq_RPKM_log2.txt\", sep=\"\\t\",index_col=0,low_memory=False) \n", "rna_raw = pd.read_csv(root / \"KIPAN.uncv2.mRNAseq_RSEM_normalized_log2.txt\", sep=\"\\t\",index_col=0,low_memory=False)\n", "meth_raw = pd.read_csv(root/\"KIPAN.meth.by_mean.data.txt\", sep='\\t',index_col=0,low_memory=False)\n", "clinical_raw = pd.read_csv(root / \"KIPAN.clin.merged.picked.txt\",sep=\"\\t\", index_col=0, low_memory=False)\n", "\n", "# display all shapes and first few rows of each dataset\n", "display(mirna_raw.iloc[:3,:5])\n", "display(mirna_raw.shape)\n", "\n", "display(rna_raw.iloc[:3,:5])\n", "display(rna_raw.shape)\n", "\n", "display(meth_raw.iloc[:3,:5])\n", "display(meth_raw.shape)\n", "\n", "display(clinical_raw.iloc[:3,:5])\n", "display(clinical_raw.shape)" ] }, { "cell_type": "markdown", "id": "e13c6054", "metadata": {}, "source": [ "## Data Processing Summary\n", "\n", "1. **Transpose Data:** All raw data (miRNA, RNA, etc.) is flipped so rows represent patients and columns represent features.\n", "2. **Standardize Patient IDs:** Patient IDs in all tables are cleaned to the 12-character TCGA format (e.g., `TCGA-AB-1234`) for matching.\n", "3. **Handle Duplicates:** Duplicate patient rows are averaged in the omics data. The first entry is kept for duplicate patients in the clinical data.\n", "4. **Find Common Patients:** The script identifies the list of patients that exist in all datasets.\n", "5. **Extract & Map Target:** The `pathologic_stage` column is pulled from the clinical data and mapped to a binary variable (0 for Early Stage, 1 for Late Stage). Patients without valid stage data are dropped.\n", "6. **Final Data Alignment:** All data tables are filtered down to only the common list of patients who also have valid staging data." ] }, { "cell_type": "code", "execution_count": null, "id": "33acffac", "metadata": {}, "outputs": [], "source": [ "from bioneuralnet.utils import m_transform\n", "\n", "mirna = mirna_raw.T\n", "rna = rna_raw.T\n", "meth = meth_raw.T\n", "clinical = clinical_raw.T\n", "\n", "print(f\"miRNA (samples, features): {mirna.shape}\")\n", "print(f\"RNA (samples, features): {rna.shape}\")\n", "print(f\"Methylation (samples, features): {meth.shape}\")\n", "print(f\"Clinical (samples, features): {clinical.shape}\")\n", "\n", "def trim_barcode(idx):\n", " return idx.to_series().str.slice(0, 12)\n", "\n", "# standarized patient IDs across all files\n", "meth.index = trim_barcode(meth.index)\n", "rna.index = trim_barcode(rna.index)\n", "mirna.index = trim_barcode(mirna.index)\n", "clinical.index = clinical.index.str.upper()\n", "clinical.index.name = \"Patient_ID\"\n", "\n", "# convert all data to numeric, coercing errors to NaN\n", "meth = meth.apply(pd.to_numeric, errors='coerce')\n", "rna = rna.apply(pd.to_numeric, errors='coerce')\n", "mirna = mirna.apply(pd.to_numeric, errors='coerce')\n", "\n", "# for any duplicate columns in the omics data, we average their values\n", "meth = meth.groupby(meth.index).mean()\n", "rna = rna.groupby(rna.index).mean()\n", "mirna = mirna.groupby(mirna.index).mean()\n", "\n", "#drop unwanted columns\n", "clinical.drop(columns=[\"Composite Element REF\"], errors=\"ignore\", inplace=True)\n", "meth.drop(columns=[\"Composite Element REF\"], errors=\"ignore\", inplace=True)\n", "\n", "# for any duplicate rows in the clinical data, we keep the first occurrence\n", "clinical = clinical[~clinical.index.duplicated(keep='first')]\n", "\n", "# convert beta values to M-values using bioneuralnet utility with small epsilon to avoid log(0)\n", "meth_m = m_transform(meth, eps=1e-7) \n", "\n", "print(f\"\\nMethylation shape: {meth_m.shape}\")\n", "print(f\"RNA shape: {rna.shape}\")\n", "print(f\"miRNA shape: {mirna.shape}\")\n", "print(f\"Clinical shape: {clinical.shape}\")\n", "\n", "for df in [meth_m, rna, mirna]:\n", " df.columns = df.columns.str.replace(r\"\\?\", \"unknown_\", regex=True)\n", " df.columns = df.columns.str.replace(r\"\\|\", \"_\", regex=True)\n", " df.columns = df.columns.str.replace(\"-\", \"_\", regex=False)\n", " df.columns = df.columns.str.replace(r\"_+\", \"_\", regex=True)\n", " df.columns = df.columns.str.strip(\"_\")\n", " \n", " df.fillna(df.mean(), inplace=True)\n", "\n", "# to see which pateints are common across all data files\n", "common_patients = sorted(list(set(meth_m.index)&set(rna.index)&set(mirna.index)&set(clinical.index)))\n", "\n", "print(f\"\\nFound: {len(common_patients)} patients across all data types.\")\n", "\n", "# subset to only common patients\n", "X_meth = meth_m.loc[common_patients]\n", "X_rna = rna.loc[common_patients]\n", "X_mirna = mirna.loc[common_patients]\n", "clinical = clinical.loc[common_patients]\n", "\n", "# extract target labels from clinical data\n", "raw_stages = clinical['pathologic_stage'].astype(str).str.lower().str.strip()\n", "\n", "# map directly to binary: early (0) vs late (1)\n", "stage_mapping = {\n", " 'stage i': 0,\n", " 'stage ii': 0,\n", " 'stage iii': 1,\n", " 'stage iv': 1\n", "}\n", "\n", "# apply mapping and drop any patients without a valid stage \n", "target_labels = raw_stages.map(stage_mapping).dropna().to_frame(name=\"target\")\n", "target_labels['target'] = target_labels['target'].astype(int)\n", "final_patients = sorted(list(set(common_patients) & set(target_labels.index)))\n", "\n", "print(f\"\\nFiltered down to {len(final_patients)} patients with valid staging data.\")\n", "\n", "# final alignment\n", "X_meth = meth_m.loc[final_patients]\n", "X_rna = rna.loc[final_patients]\n", "X_mirna = mirna.loc[final_patients]\n", "Y_labels = target_labels.loc[final_patients]\n", "\n", "clinical_processed = clinical.loc[final_patients].copy()\n", "clinical_processed.drop(columns=['pathologic_stage'], inplace=True, errors='ignore')\n", "\n", "print(\"\\nTarget Distribution (0: Early Stage, 1: Late Stage):\")\n", "print(Y_labels['target'].value_counts())" ] }, { "cell_type": "code", "execution_count": null, "id": "4ca526c4", "metadata": {}, "outputs": [], "source": [ "display(X_meth.iloc[:3,:5])\n", "display(X_meth.shape)\n", "\n", "display(X_rna.iloc[:3,:5])\n", "display(X_rna.shape)\n", "\n", "display(X_mirna.iloc[:3,:5])\n", "display(X_mirna.shape)\n", "\n", "display(clinical_processed.iloc[:3,:5])\n", "display(clinical_processed.shape)\n", "\n", "display(Y_labels.value_counts())" ] }, { "cell_type": "markdown", "id": "2a1f4b0b", "metadata": {}, "source": [ "## Feature Selection\n", "\n", "Unsupervised feature selection was performed across all three omic modalities using Laplacian Score filtering." ] }, { "cell_type": "code", "execution_count": null, "id": "20cc77cc", "metadata": {}, "outputs": [], "source": [ "from bioneuralnet.utils import laplacian_score\n", "\n", "meth_lap = laplacian_score(X_meth, n_keep=400)\n", "rna_lap = laplacian_score(X_rna, n_keep=200)\n", "mirna_lap = laplacian_score(X_mirna, n_keep=100)\n" ] }, { "cell_type": "code", "execution_count": null, "id": "0b8fa26d", "metadata": {}, "outputs": [], "source": [ "import numpy as np\n", "from bioneuralnet.utils import preprocess_clinical\n", "\n", "columns_to_drop = [\n", " 'Hybridization REF', 'Composite Element REF', 'value', \n", " 'tumor_tissue_site', 'date_of_initial_pathologic_diagnosis', \n", " 'vital_status', 'days_to_death', 'days_to_last_followup', \n", " 'days_to_last_known_alive', \n", " 'pathologic_stage', 'pathology_T_stage', 'pathology_N_stage', 'pathology_M_stage',\n", " 'karnofsky_performance_score', \n", " 'number_pack_years_smoked',\n", " 'year_of_tobacco_smoking_onset'\n", "]\n", "\n", "continuous_cols = ['years_to_birth']\n", "\n", "clinical_numeric = preprocess_clinical(\n", " X=clinical_processed,\n", " scale=True,\n", " impute=True,\n", " drop_columns=columns_to_drop,\n", " continuous_columns=continuous_cols)\n", "\n", "\n", "print(f\"\\nNaN count: {clinical_numeric.isna().sum().sum()}\")\n", "print(f\"\\nInf count: {np.isinf(clinical_numeric.values).sum()}\")\n", "print(f\"\\nColumn dtypes:\\n{clinical_numeric.dtypes}\")\n", "print(clinical_numeric.describe())" ] }, { "cell_type": "markdown", "id": "f0d52489", "metadata": {}, "source": [ "## Data Availability\n", "\n", "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.\n", "\n", "Users can load this dataset, bypassing all preceding data acquisition, preprocessing, and feature selection steps." ] }, { "cell_type": "code", "execution_count": null, "id": "78e99139", "metadata": {}, "outputs": [], "source": [ "import pandas as pd\n", "from bioneuralnet.datasets import DatasetLoader\n", "\n", "tgca_kipan = DatasetLoader(\"kipan\")\n", "display(tgca_kipan.shape)\n", "\n", "# The dataset is returned as a dictionary. We extract each file independently based on the name( Key).\n", "X_methylation = tgca_kipan[\"methylation\"]\n", "X_rna = tgca_kipan[\"rna\"]\n", "X_mirna = tgca_kipan[\"mirna\"]\n", "clinical_numeric = tgca_kipan[\"clinical\"]\n", "Y_labels = tgca_kipan[\"target\"]\n", "\n", "display(X_methylation.iloc[:3,:5])\n", "display(X_rna.iloc[:3,:5])\n", "display(X_mirna.iloc[:3,:5])\n", "display(clinical_numeric.iloc[:3,:5])\n", "display(Y_labels.iloc[:3,:5])" ] }, { "cell_type": "markdown", "id": "4c68ddb7", "metadata": {}, "source": [ "## Reproducibility and Seeding\n", "\n", "This utility function propagates the seed to all sources of randomness, including `random`, `numpy`, and `torch`." ] }, { "cell_type": "code", "execution_count": null, "id": "dafbebb2", "metadata": {}, "outputs": [], "source": [ "from bioneuralnet.utils import set_seed\n", "\n", "SEED = 8183\n", "set_seed(SEED)" ] }, { "cell_type": "code", "execution_count": null, "id": "dc91d2b6", "metadata": {}, "outputs": [], "source": [ "from bioneuralnet.network import correlation_network, similarity_network\n", "\n", "omics_kipan = pd.concat([X_methylation, X_rna, X_mirna], axis=1)\n", "\n", "similarity_22 = similarity_network(omics_kipan, k=22)\n", "spearman_12 = correlation_network(omics_kipan, method=\"spearman\", k=12)" ] }, { "cell_type": "code", "execution_count": null, "id": "413e038d", "metadata": {}, "outputs": [], "source": [ "from bioneuralnet.network import NetworkAnalyzer\n", "\n", "sim_22_analysis = NetworkAnalyzer(similarity_22, source_omics=[X_methylation, X_rna, X_mirna])\n", "sim_22_analysis.basic_statistics(0.0001)\n", "sim_22_analysis.hub_analysis(0.0001)\n", "sim_22_analysis.find_strongest_edges(5)\n", "\n", "spear_12_analysis = NetworkAnalyzer(spearman_12, source_omics=[X_methylation, X_rna, X_mirna])\n", "spear_12_analysis.basic_statistics(0.0001)\n", "spear_12_analysis.hub_analysis(0.0001)\n", "spear_12_analysis.find_strongest_edges(5)" ] }, { "cell_type": "code", "execution_count": null, "id": "2fc4aea0", "metadata": {}, "outputs": [], "source": [ "# A targeted subset of clinical variables is selected. This can be modified as needed for testing/experiments\n", "\n", "clinical_numeric = clinical_numeric.copy()[[\n", " 'radiation_therapy_yes',\n", " 'histological_type_kidney clear cell renal carcinoma',\n", " 'histological_type_kidney papillary renal cell carcinoma',\n", " 'gender_male',\n", " 'ethnicity_not hispanic or latino',\n", " 'years_to_birth',\n", " 'race_white'\n", "]]\n", "\n", "print(clinical_numeric.head())" ] }, { "cell_type": "code", "execution_count": null, "id": "4906eda5", "metadata": {}, "outputs": [], "source": [ "from bioneuralnet.downstream_task import DPMON\n", "\n", "output_dir_base = Path(\"/home/vicente/Github/BioNeuralNet/dpmon_sage_results/kipan\")\n", "\n", "current_output_dir = output_dir_base / \"similarity_22\"\n", "current_output_dir.mkdir(parents=True, exist_ok=True)\n", "\n", "dpmon_params_base = {\n", " \"adjacency_matrix\": similarity_22,\n", " \"omics_list\": omics_kipan,\n", " \"phenotype_data\": Y_labels,\n", " \"phenotype_col\": \"target\",\n", " \"clinical_data\": clinical_numeric,\n", " \"tune_trials\" : 20,\n", " \"model\": 'SAGE',\n", " \"tune\": True, \n", " \"cv\": True, \n", " \"n_folds\": 5,\n", " \"repeat_num\": 5,\n", " \"gpu\": True,\n", " \"seed\": SEED,\n", " \"output_dir\": current_output_dir\n", "}\n", "\n", "gcn_dpmon = DPMON(**dpmon_params_base)\n", "gcn_predictions, gcn_metrics, gcn_embeddings = gcn_dpmon.run()\n" ] }, { "cell_type": "code", "execution_count": null, "id": "a9173702", "metadata": {}, "outputs": [], "source": [ "from bioneuralnet.downstream_task import DPMON\n", "\n", "output_dir_base = Path(\"/home/vicente/Github/BioNeuralNet/dpmon_sage_results/kipan\")\n", "\n", "current_output_dir = output_dir_base / \"spearman_12\"\n", "current_output_dir.mkdir(parents=True, exist_ok=True)\n", "\n", "dpmon_params_base = {\n", " \"adjacency_matrix\": spearman_12,\n", " \"omics_list\": omics_kipan,\n", " \"phenotype_data\": Y_labels,\n", " \"phenotype_col\": \"target\",\n", " \"clinical_data\": clinical_numeric,\n", " \"tune_trials\" : 20,\n", " \"model\": 'SAGE',\n", " \"tune\": True, \n", " \"cv\": True, \n", " \"n_folds\": 5,\n", " \"repeat_num\": 5,\n", " \"gpu\": True,\n", " \"seed\": SEED,\n", " \"output_dir\": current_output_dir\n", "}\n", "\n", "gcn_dpmon = DPMON(**dpmon_params_base)\n", "gcn_predictions, gcn_metrics, gcn_embeddings = gcn_dpmon.run()\n" ] }, { "cell_type": "code", "execution_count": null, "id": "fac7426c", "metadata": {}, "outputs": [], "source": [ "from bioneuralnet.downstream_task import DPMON\n", "\n", "output_dir_base = Path(\"/home/vicente/Github/BioNeuralNet/dpmon_gat_multiple_results/kipan\")\n", "all_results_gat = []\n", "\n", "# To test several graphs sequentially. We can loop over the graphs dynamically like this:\n", "\n", "comparison_runs = [\n", " {\"name\": \"similarity_22\", \"graph\": similarity_22, \"omics\": omics_kipan},\n", " {\"name\": \"spearman_12\", \"graph\": spearman_12, \"omics\": omics_kipan},\n", "]\n", "\n", "for run_config in comparison_runs:\n", " graph_name = run_config[\"name\"]\n", " A_full = run_config[\"graph\"]\n", " omics = run_config[\"omics\"]\n", " \n", " current_output_dir = output_dir_base / graph_name\n", " current_output_dir.mkdir(parents=True, exist_ok=True)\n", "\n", " # to test a the same two graphs but with a different GNN we only need to update the 'model' parameter\n", " dpmon_params_base = {\n", " \"adjacency_matrix\": A_full,\n", " \"omics_list\": omics,\n", " \"phenotype_data\": Y_labels,\n", " \"phenotype_col\": \"target\",\n", " \"clinical_data\": clinical_numeric,\n", " \"tune_trials\" : 20,\n", " \"model\": 'GAT', # change to GCN for testing different GNN\n", " \"tune\": True, \n", " \"cv\": True, \n", " \"n_folds\": 5,\n", " \"repeat_num\":5,\n", " \"gpu\": True,\n", " \"seed\": SEED,\n", " \"output_dir\": current_output_dir\n", " }\n", " \n", " dpmon_tunned = DPMON(**dpmon_params_base)\n", " predictions_df, metrics, embeddings = dpmon_tunned.run()\n", "\n", " all_results_gat.append({\n", " \"graph\": graph_name,\n", " \"predictions\": predictions_df,\n", " \"metrics\": metrics,\n", " \"embeddings\": embeddings\n", " })\n", "\n", "# Once all the experiments are done, we can review the results for each graph configuration:\n", "for res in all_results_gat:\n", " graph_name = res[\"graph\"]\n", " graph_metrics = res[\"metrics\"]\n", " \n", " acc_row = graph_metrics.loc[graph_metrics['Metric'] == 'Accuracy'].iloc[0]\n", " f1_macro_row = graph_metrics.loc[graph_metrics['Metric'] == 'F1 Macro'].iloc[0]\n", " f1_weighted_row = graph_metrics.loc[graph_metrics['Metric'] == 'F1 Weighted'].iloc[0]\n", " recall_row = graph_metrics.loc[graph_metrics['Metric'] == 'Recall'].iloc[0]\n", " precission_row = graph_metrics.loc[graph_metrics['Metric'] == 'Precision'].iloc[0]\n", " auc_row = graph_metrics.loc[graph_metrics['Metric'] == 'AUC'].iloc[0]\n", " aupr_row = graph_metrics.loc[graph_metrics['Metric'] == 'AUPR'].iloc[0]\n", " \n", " acc_avg, acc_std = acc_row['Average'], acc_row['StdDev']\n", " f1_macro_avg, f1_macro_std = f1_macro_row['Average'], f1_macro_row['StdDev']\n", " f1_weighted_avg, f1_weighted_std = f1_weighted_row['Average'], f1_weighted_row['StdDev']\n", " recall_avg, recall_std = recall_row['Average'], recall_row['StdDev']\n", " precision_avg, precision_std = precission_row['Average'], precission_row['StdDev']\n", " auc_avg, auc_std = auc_row['Average'], auc_row['StdDev']\n", " aupr_avg, aupr_std = aupr_row['Average'], aupr_row['StdDev']\n", "\n", " print(f\"\\nResults for: {graph_name}\")\n", " print(f\"Accuracy (Avg +/- Std): {acc_avg:.4f} +/- {acc_std:.4f}\")\n", " print(f\"F1 Macro (Avg +/- Std): {f1_macro_avg:.4f} +/- {f1_macro_std:.4f}\")\n", " print(f\"F1 Weighted (Avg +/- Std): {f1_weighted_avg:.4f} +/- {f1_weighted_std:.4f}\")\n", " print(f\"Recall: {recall_avg:.4f} +/- {recall_std:.4f}\")\n", " print(f\"Precision: {precision_avg:.4f} +/- {precision_std:.4f}\")\n", " print(f\"AUC: {auc_avg:.4f} +/- {auc_std:.4f}\")\n", " print(f\"AUPR: {aupr_avg:.4f} +/- {aupr_std:.4f}\")" ] }, { "cell_type": "code", "execution_count": null, "id": "98442c32", "metadata": {}, "outputs": [], "source": [ "from sklearn.model_selection import StratifiedKFold, ParameterSampler, RepeatedStratifiedKFold\n", "from sklearn.metrics import accuracy_score, f1_score, recall_score, roc_auc_score, average_precision_score, precision_score\n", "from sklearn.preprocessing import StandardScaler, label_binarize\n", "from sklearn.linear_model import LogisticRegression\n", "from sklearn.neural_network import MLPClassifier\n", "from sklearn.pipeline import Pipeline\n", "from sklearn.base import clone\n", "from scipy.stats import loguniform, randint\n", "from sklearn.ensemble import RandomForestClassifier\n", "from sklearn.tree import DecisionTreeClassifier\n", "from xgboost import XGBClassifier\n", "from sklearn.svm import SVC\n", "\n", "SEED = 8183\n", "y = Y_labels[\"target\"]\n", "X = pd.concat([omics_kipan,clinical_numeric],axis=1)\n", "\n", "print(f\"Successfully created X vector with shape: {X.shape}\")\n", "print(f\"Successfully created y vector with shape: {y.shape}\")\n", "\n", "pipe_lr = Pipeline([\n", " ('scaler', StandardScaler()),\n", " ('model', LogisticRegression(solver='lbfgs', max_iter=1000, penalty=None, random_state=SEED))\n", "])\n", "\n", "pipe_mlp = Pipeline([\n", " ('scaler', StandardScaler()),\n", " ('model', MLPClassifier(max_iter=500, early_stopping=True, n_iter_no_change=10, random_state=SEED))\n", "])\n", "\n", "pipe_xgb = Pipeline([\n", " ('scaler', StandardScaler()),\n", " ('model', XGBClassifier(eval_metric='logloss', tree_method='hist', max_bin=128, random_state=SEED))\n", "])\n", "pipe_rf = Pipeline([\n", " ('scaler', StandardScaler()),\n", " ('model', RandomForestClassifier(random_state=SEED))\n", "])\n", "\n", "pipe_svm = Pipeline([\n", " ('scaler', StandardScaler()),\n", " ('model', SVC(probability=True, random_state=SEED))\n", "])\n", "\n", "pipe_dt = Pipeline([\n", " ('scaler', StandardScaler()),\n", " ('model', DecisionTreeClassifier(random_state=SEED))\n", "])\n", "\n", "params_lr = {'model__penalty': ['l2'], 'model__C': loguniform(1e-4, 1e2)}\n", "\n", "params_mlp = {\n", " 'model__hidden_layer_sizes': [(100,), (100, 50), (50, 50)],\n", " 'model__activation': ['relu', 'tanh'],\n", " 'model__alpha': loguniform(1e-5, 1e-1),\n", " 'model__learning_rate_init': loguniform(1e-4, 1e-2)\n", "}\n", "params_xgb = {\n", " 'model__n_estimators': randint(50, 200),\n", " 'model__learning_rate': loguniform(0.01, 0.3),\n", " 'model__max_depth': randint(3, 7),\n", " 'model__subsample': [0.8, 1.0], \n", " 'model__colsample_bytree': [0.8, 1.0]\n", "}\n", "params_rf = {\n", " 'model__n_estimators': randint(100, 300),\n", " 'model__max_depth': [10, 20, 30, None],\n", " 'model__min_samples_split': randint(2, 10),\n", " 'model__min_samples_leaf': randint(1, 5),\n", " 'model__max_features': ['sqrt', 'log2']\n", "}\n", "params_svm = {\n", " 'model__C': loguniform(1e-2, 1e3),\n", " 'model__kernel': ['rbf', 'linear'],\n", " 'model__gamma': ['scale', 'auto']\n", "}\n", "\n", "params_dt = {\n", " 'model__max_depth': randint(3, 15),\n", " 'model__min_samples_split': randint(2, 20),\n", " 'model__criterion': ['gini', 'entropy']\n", "}\n", "\n", "models_to_tune = {\n", " \"LogisticRegression\": (pipe_lr, params_lr),\n", " \"SVM\": (pipe_svm, params_svm),\n", " \"MLP\": (pipe_mlp, params_mlp),\n", " \"XGBoost\": (pipe_xgb, params_xgb),\n", " \"RandomForest\": (pipe_rf, params_rf),\n", " \"DecisionTree\": (pipe_dt, params_dt),\n", "}\n", "\n", "all_results = {\n", " \"LogisticRegression\": {\"acc\": [], \"f1_w\": [], \"f1_m\": [], \"recall\": [], \"auc\": [], \"precision\": [], \"aupr\": []},\n", " \"MLP\": {\"acc\": [], \"f1_w\": [], \"f1_m\": [], \"recall\": [], \"auc\": [], \"precision\": [], \"aupr\": []},\n", " \"XGBoost\": {\"acc\": [], \"f1_w\": [], \"f1_m\": [], \"recall\": [], \"auc\": [], \"precision\": [], \"aupr\": []},\n", " \"RandomForest\": {\"acc\": [], \"f1_w\": [], \"f1_m\": [], \"recall\": [], \"auc\": [], \"precision\": [], \"aupr\": []},\n", " \"SVM\": {\"acc\": [], \"f1_w\": [], \"f1_m\": [], \"recall\": [], \"auc\": [], \"precision\": [], \"aupr\": []},\n", " \"DecisionTree\": {\"acc\": [], \"f1_w\": [], \"f1_m\": [], \"recall\": [], \"auc\": [], \"precision\": [], \"aupr\": []},\n", "}\n", "\n", "\n", "\n", "for model_name, (pipeline, param_dist) in models_to_tune.items():\n", " print(f\"Evaluating model with nested CV: {model_name}\")\n", " \n", " outer_cv = RepeatedStratifiedKFold(n_splits=5, n_repeats=5, random_state=SEED)\n", "\n", " # Inner fold loop is for finding the best hyperparameters per fold.\n", " for fold_idx, (train_idx, test_idx) in enumerate(outer_cv.split(X, y), start=1):\n", " X_train_outer, X_test_outer = X.iloc[train_idx], X.iloc[test_idx]\n", " y_train_outer, y_test_outer = y.iloc[train_idx], y.iloc[test_idx]\n", " inner_cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=SEED)\n", " \n", " best_score_fold = -np.inf\n", " best_params_fold = None\n", "\n", " # A fixed seed = same hyperparamters each fold.\n", " # No seed = different hyperparameters each fold. \n", " # This adds more randomness, and may yield better generalization.\n", " param_sampler = list(ParameterSampler(param_dist, n_iter=20))\n", " \n", " for params in param_sampler:\n", " inner_scores = []\n", " \n", " for inner_train_idx, inner_val_idx in inner_cv.split(X_train_outer, y_train_outer):\n", " X_train_inner = X_train_outer.iloc[inner_train_idx]\n", " X_val_inner = X_train_outer.iloc[inner_val_idx]\n", " y_train_inner = y_train_outer.iloc[inner_train_idx]\n", " y_val_inner = y_train_outer.iloc[inner_val_idx]\n", " \n", " inner_pipeline = clone(pipeline)\n", " inner_pipeline.set_params(**params)\n", " inner_pipeline.fit(X_train_inner, y_train_inner)\n", " \n", " y_val_pred = inner_pipeline.predict(X_val_inner)\n", " score = f1_score(y_val_inner, y_val_pred, average='macro', zero_division=0)\n", " inner_scores.append(score)\n", " \n", " mean_score = np.mean(inner_scores)\n", " if mean_score > best_score_fold:\n", " best_score_fold = mean_score\n", " best_params_fold = params\n", " \n", " print(f\"Outer fold {fold_idx}: best params (inner CV F1-M={best_score_fold:.4f})\")\n", " print(f\"{best_params_fold}\")\n", "\n", " final_pipeline = clone(pipeline)\n", " final_pipeline.set_params(**best_params_fold)\n", " final_pipeline.fit(X_train_outer, y_train_outer)\n", " \n", " preds = final_pipeline.predict(X_test_outer)\n", " \n", " if hasattr(final_pipeline, \"predict_proba\"):\n", " proba = final_pipeline.predict_proba(X_test_outer)\n", " else:\n", " proba = None\n", " \n", " acc = accuracy_score(y_test_outer, preds)\n", " f1_w = f1_score(y_test_outer, preds, average='weighted', zero_division=0)\n", " f1_m = f1_score(y_test_outer, preds, average='macro', zero_division=0)\n", " recall = recall_score(y_test_outer, preds, average='macro', zero_division=0)\n", " precision = precision_score(y_test_outer, preds, average='macro', zero_division=0)\n", " \n", " auc = np.nan\n", " aupr = np.nan\n", " \n", " if proba is not None:\n", " try:\n", " if len(np.unique(y)) == 2:\n", " auc = roc_auc_score(y_test_outer, proba[:, 1])\n", " aupr = average_precision_score(y_test_outer, proba[:, 1])\n", " else:\n", " auc = roc_auc_score(y_test_outer, proba, multi_class='ovr', average='macro')\n", " y_test_bin = label_binarize(y_test_outer, classes=np.unique(y))\n", " aupr = average_precision_score(y_test_bin, proba, average='macro')\n", " except Exception:\n", " auc = np.nan\n", " aupr = np.nan\n", "\n", " print(f\"Fold {fold_idx} results: Acc={acc:.4f}, F1-W={f1_w:.4f}, \"\n", " f\"F1-M={f1_m:.4f}, Recall={recall:.4f}, Precision={precision:.4f}, AUC={auc:.4f}, AUPR={aupr:.4f}\")\n", " \n", " all_results[model_name][\"acc\"].append(acc)\n", " all_results[model_name][\"f1_w\"].append(f1_w)\n", " all_results[model_name][\"f1_m\"].append(f1_m)\n", " all_results[model_name][\"recall\"].append(recall)\n", " all_results[model_name][\"precision\"].append(precision)\n", " all_results[model_name][\"auc\"].append(auc)\n", " all_results[model_name][\"aupr\"].append(aupr)\n", "\n", "print(\"\\nFINAL BASELINE RESULTS\\n\")\n", "for model_name, metrics in all_results.items():\n", " avg_acc = np.mean(metrics[\"acc\"])\n", " std_acc = np.std(metrics[\"acc\"])\n", " avg_f1_w = np.mean(metrics[\"f1_w\"])\n", " std_f1_w = np.std(metrics[\"f1_w\"])\n", " avg_f1_m = np.mean(metrics[\"f1_m\"])\n", " std_f1_m = np.std(metrics[\"f1_m\"])\n", " avg_recall = np.mean(metrics[\"recall\"])\n", " std_recall = np.std(metrics[\"recall\"])\n", " avg_prec = np.mean(metrics[\"precision\"])\n", " std_prec = np.std(metrics[\"precision\"])\n", " avg_auc = np.nanmean(metrics[\"auc\"])\n", " std_auc = np.nanstd(metrics[\"auc\"])\n", " avg_aupr = np.nanmean(metrics[\"aupr\"])\n", " std_aupr = np.nanstd(metrics[\"aupr\"])\n", " \n", " print(f\"\\n{model_name}:\")\n", " print(f\"Accuracy: {avg_acc:.4f} +/- {std_acc:.4f}\")\n", " print(f\"Precision: {avg_prec:.4f} +/- {std_prec:.4f}\")\n", " print(f\"F1 Weighted: {avg_f1_w:.4f} +/- {std_f1_w:.4f}\")\n", " print(f\"F1 Macro: {avg_f1_m:.4f} +/- {std_f1_m:.4f}\")\n", " print(f\"Recall: {avg_recall:.4f} +/- {std_recall:.4f}\")\n", " print(f\"AUC: {avg_auc:.4f} +/- {std_auc:.4f}\")\n", " print(f\"AUPR: {avg_aupr:.4f} +/- {std_aupr:.4f}\")\n" ] } ], "metadata": { "kernelspec": { "display_name": ".enviroment", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.12.3" } }, "nbformat": 4, "nbformat_minor": 5 }