2_ADNI_GNNs_vs_nonGNNs.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "bigger-sheet",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/laura/anaconda3/envs/graphgym/lib/python3.7/site-packages/tqdm/auto.py:22: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
      "  from .autonotebook import tqdm as notebook_tqdm\n"
     ]
    }
   ],
   "source": [
    "import sys, os, pickle\n",
    "import pandas as pd\n",
    "import numpy as np\n",
    "\n",
    "from sklearn import preprocessing, metrics\n",
    "from sklearn.model_selection import train_test_split, GridSearchCV, StratifiedKFold\n",
    "from sklearn.linear_model import LogisticRegression\n",
    "from sklearn.ensemble import RandomForestClassifier\n",
    "from sklearn.svm import SVC\n",
    "\n",
    "import seaborn as sns\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "from ml_models import create_labels, models_stats\n",
    "from create_datasets import create_nx_datasets, create_table_datasets"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "reflected-mirror",
   "metadata": {},
   "source": [
    "### 1. Create table datasets\n",
    "\n",
    "Create table datasets (with and without APOE) for the targets and genes of interest to use as input for non-GNN models."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "killing-organ",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Input: type of variant, best network from previous analysis, missense data, diagnosis/metadata\n",
    "table_dataset = create_table_datasets.adni_data('missense',\n",
    "                                                  'data/bio_networks/AD_SNAP_PPI_brain.edgelist',\n",
    "                                                  'data/ADNI/AD_PPI_missense.csv',\n",
    "                                                  'data/ADNI/ADNIMERGE_metadata.csv')\n",
    "table_dataset.to_csv(f'data/table_datasets/AD_PPI_missense_ADNI_labeled.csv')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "dominant-conversion",
   "metadata": {},
   "source": [
    "### 2. Baseline models - Logistic Regression only with APOE\n",
    "Create baseline models, meaning Logistic Regression models for each label only using the number of missense variants in APOE as input."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "interesting-broadcast",
   "metadata": {},
   "outputs": [],
   "source": [
    "targets = ['PET', 'PETandDX']\n",
    "\n",
    "for target in targets:\n",
    "    \n",
    "    result_df = pd.DataFrame(columns = ['fold', 'target', 'dataset', 'model', 'acc', 'pre', 'rec', 'f1', 'auc'])\n",
    "\n",
    "    data = pd.read_csv('data/table_datasets/AD_PPI_missense_ADNI_labeled.csv', index_col = 0)\n",
    "    data_wclass = create_labels.label_ADNI(data, target)\n",
    "\n",
    "    x = data_wclass.drop(columns=['y'])\n",
    "    x = x['APOE']\n",
    "\n",
    "    y = data_wclass['y']\n",
    "    x.index = x.index.str.upper()\n",
    "\n",
    "    for i in range(10):\n",
    "\n",
    "        i += 1\n",
    "\n",
    "        f = open(f'data/splits/10Fold_CV_{target}/k{i}_{target}.pkl', 'rb')\n",
    "        split = pickle.load(f)\n",
    "        f.close()\n",
    "\n",
    "        tr_idx = split['train']\n",
    "        te_idx = split['valid']\n",
    "\n",
    "        x_train, x_test = x.iloc[tr_idx], x.iloc[te_idx]\n",
    "        y_train, y_test = y[tr_idx], y[te_idx]\n",
    "\n",
    "        x_train = np.array(x_train).reshape(-1, 1)\n",
    "        x_test = np.array(x_test).reshape(-1, 1)\n",
    "\n",
    "        scaler = preprocessing.StandardScaler().fit(x_train)\n",
    "        x_train_scaled = scaler.transform(x_train)\n",
    "        x_test_scaled = scaler.transform(x_test)\n",
    "\n",
    "        logreg = LogisticRegression()\n",
    "        logreg.fit(x_train_scaled, y_train)\n",
    "\n",
    "        y_prob = logreg.predict_proba(x_test_scaled)\n",
    "\n",
    "        y_pred = logreg.predict(x_test_scaled)\n",
    "        acc = metrics.accuracy_score(y_test, y_pred)\n",
    "        pre = metrics.precision_score(y_test, y_pred)\n",
    "        rec = metrics.recall_score(y_test, y_pred)\n",
    "        f1  = metrics.f1_score(y_test, y_pred)\n",
    "        auc = metrics.roc_auc_score(y_test, y_prob[:, 1])\n",
    "\n",
    "        result_df.loc[len(result_df)] = [i, target, 'Only APOE', 'Baseline model', acc, pre, rec, f1, auc]\n",
    "\n",
    "    result_df.to_csv(f'results/2022_03_baseline_models_{target}.csv', index=False)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "animal-proposal",
   "metadata": {},
   "source": [
    "### 3. Other non-GNN models\n",
    "Build non-GNN models (Logistic Regression, SVM with Linear and RBF kernels, and Random Forest) using as input the number of missense variants found in the genes present in AD PPT-Ohmnet network (see Section 2 of this notebook)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "adult-snapshot",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "PET\n",
      "Fold 1\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=0.01, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.01, gamma=1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=5000)\n",
      "\n",
      "Fold 2\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=10, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.01, gamma=1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n",
      "Fold 3\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=1000, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=10, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=5000)\n",
      "\n",
      "Fold 4\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=0.01, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.01, gamma=1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=5000)\n",
      "\n",
      "Fold 5\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=1000, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1000, gamma=0.001, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n",
      "Fold 6\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=10, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.01, gamma=1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=500)\n",
      "\n",
      "Fold 7\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=10, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.01, gamma=1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=500)\n",
      "\n",
      "Fold 8\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=10, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=10, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=500)\n",
      "\n",
      "Fold 9\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=1000, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.01, gamma=1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=5000)\n",
      "\n",
      "Fold 10\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=0.1, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n",
      "PETandDX\n",
      "Fold 1\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=1000, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=100, gamma=0.01, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=500)\n",
      "\n",
      "Fold 2\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=100, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=100, gamma=1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=500)\n",
      "\n",
      "Fold 3\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=1000, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=100, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=500)\n",
      "\n",
      "Fold 4\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=100, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1000, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n",
      "Fold 5\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=0.01, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.1, gamma=0.01, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=500)\n",
      "\n",
      "Fold 6\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=10, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=10, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=5000)\n",
      "\n",
      "Fold 7\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=100, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1000, gamma=0.01, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n",
      "Fold 8\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=100, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1000, gamma=0.01, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=5000)\n",
      "\n",
      "Fold 9\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=1000, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1000, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n",
      "Fold 10\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=10, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1000, gamma=0.01, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=5000)\n",
      "\n"
     ]
    }
   ],
   "source": [
    "for target in ['PET', 'PETandDX']:\n",
    "    \n",
    "    print(target)\n",
    "        \n",
    "    data = pd.read_csv('data/table_datasets/AD_PPI_missense_ADNI_labeled.csv', index_col = 0)\n",
    "    data_wclass = create_labels.label_ADNI(data, target) # label dataset with the corresponding label\n",
    "    data_wclass = data_wclass.loc[:, (data_wclass != 0).any(axis=0)] # delete genes that do not have mutations (all zeroes)\n",
    "    \n",
    "    x = data_wclass.drop(columns=['y'])\n",
    "    y = data_wclass['y']\n",
    "    x.index = x.index.str.upper()\n",
    "    \n",
    "    results = [] # list of lists that stores performance metrics\n",
    "    \n",
    "    for i in range(10):\n",
    "\n",
    "        i += 1\n",
    "        \n",
    "        print('Fold', i)\n",
    "        print('-------------------------------------------------')\n",
    "\n",
    "        f = open(f'data/splits/10Fold_CV_{target}/k{i}_{target}.pkl', 'rb')\n",
    "        split = pickle.load(f)\n",
    "        f.close()\n",
    "\n",
    "        tr_idx = split['train']\n",
    "        te_idx = split['valid']\n",
    "\n",
    "        x_train, x_test = x.iloc[tr_idx], x.iloc[te_idx]\n",
    "        y_train, y_test = y[tr_idx], y[te_idx]\n",
    "\n",
    "        scaler = preprocessing.StandardScaler().fit(x_train)\n",
    "        x_train_scaled = scaler.transform(x_train)\n",
    "        x_test_scaled = scaler.transform(x_test)\n",
    "\n",
    "        # Logistic Regression\n",
    "        logreg = LogisticRegression()\n",
    "        logreg.fit(x_train_scaled, y_train)\n",
    "        y_prob_logreg = logreg.predict_proba(x_test_scaled)\n",
    "        y_pred_logreg = logreg.predict(x_test_scaled)\n",
    "        results.append([i, target, 'AD PPT-Ohmnet', 'Logistic Regression',\n",
    "                        metrics.accuracy_score(y_test, y_pred_logreg),\n",
    "                        metrics.precision_score(y_test, y_pred_logreg),\n",
    "                        metrics.recall_score(y_test, y_pred_logreg),\n",
    "                        metrics.f1_score(y_test, y_pred_logreg),\n",
    "                        metrics.roc_auc_score(y_test, y_prob_logreg[:, 1])])\n",
    "\n",
    "        # SVM Linear\n",
    "        svm_lin = SVC(probability=True)\n",
    "        param_grid_svm_lin = [{'C': [0.01, 0.1, 1, 10, 100, 1000]}]\n",
    "        grid_search_svm_lin = GridSearchCV(svm_lin, param_grid_svm_lin, cv=10, scoring='roc_auc', n_jobs = -1, verbose = 3)\n",
    "        grid_search_svm_lin.fit(x_train_scaled, y_train)\n",
    "        print(grid_search_svm_lin.best_estimator_)\n",
    "        y_prob_svm_lin = grid_search_svm_lin.predict_proba(x_test_scaled)\n",
    "        y_pred_svm_lin = grid_search_svm_lin.predict(x_test_scaled)\n",
    "        results.append([i, target, 'AD PPT-Ohmnet', 'SVM Linear',\n",
    "                        metrics.accuracy_score(y_test, y_pred_svm_lin),\n",
    "                        metrics.precision_score(y_test, y_pred_svm_lin),\n",
    "                        metrics.recall_score(y_test, y_pred_svm_lin),\n",
    "                        metrics.f1_score(y_test, y_pred_svm_lin),\n",
    "                        metrics.roc_auc_score(y_test, y_prob_svm_lin[:, 1])])\n",
    "\n",
    "        # SVM RBF\n",
    "        svm_rbf = SVC(probability=True, kernel='rbf')\n",
    "        param_grid_svm_rbf = [{'C': [0.01, 0.1, 1, 10, 100, 1000], 'gamma': [0.001, 0.01, 0.1, 1]}]\n",
    "        grid_search_svm_rbf = GridSearchCV(svm_rbf, param_grid_svm_rbf, cv=10, scoring='roc_auc', n_jobs = -1, verbose = 3)\n",
    "        grid_search_svm_rbf.fit(x_train_scaled, y_train)\n",
    "        print(grid_search_svm_rbf.best_estimator_)\n",
    "        y_prob_svm_rbf = grid_search_svm_rbf.predict_proba(x_test_scaled)\n",
    "        y_pred_svm_rbf = grid_search_svm_rbf.predict(x_test_scaled)\n",
    "        results.append([i, target, 'AD PPT-Ohmnet', 'SVM RBF',\n",
    "                        metrics.accuracy_score(y_test, y_pred_svm_rbf),\n",
    "                        metrics.precision_score(y_test, y_pred_svm_rbf),\n",
    "                        metrics.recall_score(y_test, y_pred_svm_rbf),\n",
    "                        metrics.f1_score(y_test, y_pred_svm_rbf),\n",
    "                        metrics.roc_auc_score(y_test, y_prob_svm_rbf[:, 1])])\n",
    "\n",
    "        # Random Forest\n",
    "        rf_clf = RandomForestClassifier()\n",
    "        param_grid = [{'n_estimators': [50, 500, 5000]}]\n",
    "        grid_search_rf = GridSearchCV(rf_clf, param_grid, cv=10, scoring='roc_auc', n_jobs = -1, verbose = 3)\n",
    "        grid_search_rf.fit(x_train, y_train)\n",
    "        print(grid_search_rf.best_estimator_)\n",
    "        y_prob_rf = grid_search_rf.predict_proba(x_test)\n",
    "        y_pred_rf = grid_search_rf.predict(x_test)\n",
    "        results.append([i, target, 'AD PPT-Ohmnet', 'Random Forest',\n",
    "                        metrics.accuracy_score(y_test, y_pred_rf),\n",
    "                        metrics.precision_score(y_test, y_pred_rf),\n",
    "                        metrics.recall_score(y_test, y_pred_rf),\n",
    "                        metrics.f1_score(y_test, y_pred_rf),\n",
    "                        metrics.roc_auc_score(y_test, y_prob_rf[:, 1])])\n",
    "\n",
    "        print()\n",
    "        \n",
    "    results_df = pd.DataFrame(results, columns=['fold', 'target', 'dataset', 'model', 'acc', 'pre', 'rec', 'f1', 'auc'])\n",
    "    results_df.to_csv(f'results/2022_03_nonGNN_models_{target}.csv', index=False)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "better-singing",
   "metadata": {},
   "source": [
    "### 4. Other non-GNN models without APOE\n",
    "Build non-GNN models (Logistic Regression, SVM with Linear and RBF kernels, and Random Forest) using as input the number of missense variants found in the genes present in AD PPT-Ohmnet network (see Section 2 of this notebook) **but deleting APOE gene from the input**."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "polyphonic-radiation",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "PET\n",
      "Fold 1\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=0.01, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.01, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=500)\n",
      "\n",
      "Fold 2\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=0.1, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.01, gamma=0.001, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n",
      "Fold 3\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=1, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=5000)\n",
      "\n",
      "Fold 4\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=1000, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.01, gamma=0.001, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n",
      "Fold 5\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=0.01, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1000, gamma=0.001, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=5000)\n",
      "\n",
      "Fold 6\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=10, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1, gamma=0.001, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=5000)\n",
      "\n",
      "Fold 7\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=0.01, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.1, gamma=0.001, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=500)\n",
      "\n",
      "Fold 8\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=0.1, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=100, gamma=0.01, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n",
      "Fold 9\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=10, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=100, gamma=0.01, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=500)\n",
      "\n",
      "Fold 10\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=0.01, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.01, gamma=0.001, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=500)\n",
      "\n",
      "PETandDX\n",
      "Fold 1\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=1, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n",
      "Fold 2\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=100, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1, gamma=1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n",
      "Fold 3\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=10, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.01, gamma=1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=500)\n",
      "\n",
      "Fold 4\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=1, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.01, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=5000)\n",
      "\n",
      "Fold 5\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=1, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=100, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=500)\n",
      "\n",
      "Fold 6\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=0.01, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.1, gamma=0.01, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=5000)\n",
      "\n",
      "Fold 7\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=1, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n",
      "Fold 8\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=0.01, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=1, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n",
      "Fold 9\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=10, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=10, gamma=1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=5000)\n",
      "\n",
      "Fold 10\n",
      "-------------------------------------------------\n",
      "Fitting 10 folds for each of 6 candidates, totalling 60 fits\n",
      "SVC(C=1, probability=True)\n",
      "Fitting 10 folds for each of 24 candidates, totalling 240 fits\n",
      "SVC(C=0.01, gamma=0.1, probability=True)\n",
      "Fitting 10 folds for each of 3 candidates, totalling 30 fits\n",
      "RandomForestClassifier(n_estimators=50)\n",
      "\n"
     ]
    }
   ],
   "source": [
    "for target in ['PET', 'PETandDX']:\n",
    "    \n",
    "    print(target)\n",
    "        \n",
    "    data = pd.read_csv('data/table_datasets/AD_PPI_missense_ADNI_labeled.csv', index_col = 0)\n",
    "    data_wclass = create_labels.label_ADNI(data, target) # label dataset with the corresponding label\n",
    "    data_wclass = data_wclass.loc[:, (data_wclass != 0).any(axis=0)] # delete genes that do not have mutations (all zeroes)\n",
    "    \n",
    "    x = data_wclass.drop(columns=['y', 'APOE']) # DELETE APOE GENE !!\n",
    "    y = data_wclass['y']\n",
    "    x.index = x.index.str.upper()\n",
    "    \n",
    "    results = [] # list of lists that stores performance metrics\n",
    "    \n",
    "    for i in range(10):\n",
    "\n",
    "        i += 1\n",
    "        \n",
    "        print('Fold', i)\n",
    "        print('-------------------------------------------------')\n",
    "\n",
    "        f = open(f'data/splits/10Fold_CV_{target}/k{i}_{target}.pkl', 'rb')\n",
    "        split = pickle.load(f)\n",
    "        f.close()\n",
    "\n",
    "        tr_idx = split['train']\n",
    "        te_idx = split['valid']\n",
    "\n",
    "        x_train, x_test = x.iloc[tr_idx], x.iloc[te_idx]\n",
    "        y_train, y_test = y[tr_idx], y[te_idx]\n",
    "\n",
    "        scaler = preprocessing.StandardScaler().fit(x_train)\n",
    "        x_train_scaled = scaler.transform(x_train)\n",
    "        x_test_scaled = scaler.transform(x_test)\n",
    "\n",
    "        # Logistic Regression\n",
    "        logreg = LogisticRegression()\n",
    "        logreg.fit(x_train_scaled, y_train)\n",
    "        y_prob_logreg = logreg.predict_proba(x_test_scaled)\n",
    "        y_pred_logreg = logreg.predict(x_test_scaled)\n",
    "        results.append([i, target, 'AD PPT-Ohmnet no APOE', 'Logistic Regression',\n",
    "                        metrics.accuracy_score(y_test, y_pred_logreg),\n",
    "                        metrics.precision_score(y_test, y_pred_logreg),\n",
    "                        metrics.recall_score(y_test, y_pred_logreg),\n",
    "                        metrics.f1_score(y_test, y_pred_logreg),\n",
    "                        metrics.roc_auc_score(y_test, y_prob_logreg[:, 1])])\n",
    "\n",
    "        # SVM Linear\n",
    "        svm_lin = SVC(probability=True)\n",
    "        param_grid_svm_lin = [{'C': [0.01, 0.1, 1, 10, 100, 1000]}]\n",
    "        grid_search_svm_lin = GridSearchCV(svm_lin, param_grid_svm_lin, cv=10, scoring='roc_auc',\n",
    "                                           n_jobs = -1, verbose = 3)\n",
    "        grid_search_svm_lin.fit(x_train_scaled, y_train)\n",
    "        print(grid_search_svm_lin.best_estimator_)\n",
    "        y_prob_svm_lin = grid_search_svm_lin.predict_proba(x_test_scaled)\n",
    "        y_pred_svm_lin = grid_search_svm_lin.predict(x_test_scaled)\n",
    "        results.append([i, target, 'AD PPT-Ohmnet no APOE', 'SVM Linear',\n",
    "                        metrics.accuracy_score(y_test, y_pred_svm_lin),\n",
    "                        metrics.precision_score(y_test, y_pred_svm_lin),\n",
    "                        metrics.recall_score(y_test, y_pred_svm_lin),\n",
    "                        metrics.f1_score(y_test, y_pred_svm_lin),\n",
    "                        metrics.roc_auc_score(y_test, y_prob_svm_lin[:, 1])])\n",
    "\n",
    "        # SVM RBF\n",
    "        svm_rbf = SVC(probability=True, kernel='rbf')\n",
    "        param_grid_svm_rbf = [{'C': [0.01, 0.1, 1, 10, 100, 1000], 'gamma': [0.001, 0.01, 0.1, 1]}]\n",
    "        grid_search_svm_rbf = GridSearchCV(svm_rbf, param_grid_svm_rbf, cv=10, scoring='roc_auc',\n",
    "                                           n_jobs = -1, verbose = 3)\n",
    "        grid_search_svm_rbf.fit(x_train_scaled, y_train)\n",
    "        print(grid_search_svm_rbf.best_estimator_)\n",
    "        y_prob_svm_rbf = grid_search_svm_rbf.predict_proba(x_test_scaled)\n",
    "        y_pred_svm_rbf = grid_search_svm_rbf.predict(x_test_scaled)\n",
    "        results.append([i, target, 'AD PPT-Ohmnet no APOE', 'SVM RBF',\n",
    "                        metrics.accuracy_score(y_test, y_pred_svm_rbf),\n",
    "                        metrics.precision_score(y_test, y_pred_svm_rbf),\n",
    "                        metrics.recall_score(y_test, y_pred_svm_rbf),\n",
    "                        metrics.f1_score(y_test, y_pred_svm_rbf),\n",
    "                        metrics.roc_auc_score(y_test, y_prob_svm_rbf[:, 1])])\n",
    "\n",
    "        # Random Forest\n",
    "        rf_clf = RandomForestClassifier()\n",
    "        param_grid = [{'n_estimators': [50, 500, 5000]}]\n",
    "        grid_search_rf = GridSearchCV(rf_clf, param_grid, cv=10, scoring='roc_auc',\n",
    "                                      n_jobs = -1, verbose = 3)\n",
    "        grid_search_rf.fit(x_train, y_train)\n",
    "        print(grid_search_rf.best_estimator_)\n",
    "        y_prob_rf = grid_search_rf.predict_proba(x_test)\n",
    "        y_pred_rf = grid_search_rf.predict(x_test)\n",
    "        results.append([i, target, 'AD PPT-Ohmnet no APOE', 'Random Forest',\n",
    "                        metrics.accuracy_score(y_test, y_pred_rf),\n",
    "                        metrics.precision_score(y_test, y_pred_rf),\n",
    "                        metrics.recall_score(y_test, y_pred_rf),\n",
    "                        metrics.f1_score(y_test, y_pred_rf),\n",
    "                        metrics.roc_auc_score(y_test, y_prob_rf[:, 1])])\n",
    "\n",
    "        print()\n",
    "        \n",
    "    results_df = pd.DataFrame(results, columns=['fold', 'target', 'dataset', 'model', 'acc', 'pre', 'rec', 'f1', 'auc'])\n",
    "    results_df.to_csv(f'results/2022_03_nonGNN_models_{target}_noAPOE.csv', index=False)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ceramic-business",
   "metadata": {},
   "source": [
    "### 5. Create graph datasets without APOE & graph classification with GNNs\n",
    "\n",
    "Create the correspoding graph-datasets for each target/label without APOE gene.\n",
    "\n",
    "We then evaluated and tested the best GNNs obtained for each label as done with the original graph datasets.\n",
    "\n",
    "Configuration and grid files employed are in the subdirectory [graphgym_files](graphgym_files). The GNNs configuration in this case was the same as the best configuration found with AD PPT-Ohment graph datasets.\n",
    "\n",
    "We run GraphGym 10 times, with 10 different splits generated by 10-Fold Stratified Cross-Validation."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "complimentary-efficiency",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Network used: AD snap_brain_noAPOE\n",
      "# nodes = 28\n",
      "# edges = 46\n",
      "\n",
      "Dataset used: ADNI\n",
      "missense\n",
      "(28, 808)\n",
      "Creating samples graphs...\n",
      "Class: PET. Found 410 positive subjects out of 726\n",
      "Sample graph used: # nodes = 28 # edges = 46\n",
      "Density = 0.12169312169312169 Diameter = 6\n",
      "Resulting dataset saved at: data/graph_datasets/PET/AD_PPI_snap_brain_noAPOE_missense.pkl\n",
      "\n",
      "Network used: AD snap_brain_noAPOE\n",
      "# nodes = 28\n",
      "# edges = 46\n",
      "\n",
      "Dataset used: ADNI\n",
      "missense\n",
      "(28, 808)\n",
      "Creating samples graphs...\n",
      "Class: PETandDX. Found 182 positive subjects out of 317\n",
      "Sample graph used: # nodes = 28 # edges = 46\n",
      "Density = 0.12169312169312169 Diameter = 6\n",
      "Resulting dataset saved at: data/graph_datasets/PETandDX/AD_PPI_snap_brain_noAPOE_missense.pkl\n",
      "\n"
     ]
    }
   ],
   "source": [
    "dataset = 'ADNI'\n",
    "targets = ['PET', 'PETandDX']\n",
    "network = 'snap_brain_noAPOE'\n",
    "\n",
    "for target in targets:\n",
    "    outdir = f'data/graph_datasets/{target}'\n",
    "\n",
    "    result_nodes = create_nx_datasets.main('data', dataset, target, 'AD', network, 'missense', None)\n",
    "    outfile = f'{outdir}/AD_PPI_{network}_missense.pkl'\n",
    "    print('Resulting dataset saved at:', outfile)\n",
    "    print()\n",
    "\n",
    "    with open(outfile, 'wb') as f:\n",
    "        pickle.dump(result_nodes, f)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "37b4bfe0",
   "metadata": {},
   "source": [
    "### 6. Analyze results: tables and figures"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "closed-batman",
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "# PET\n",
    "pet_nonGNN        = pd.read_csv('results/2022_03_nonGNN_models_PET.csv', index_col='fold')\n",
    "pet_nonGNN_noAPOE = pd.read_csv('results/2022_04_nonGNN_models_PET_noAPOE.csv', index_col='fold')\n",
    "pet_GNN           = pd.read_csv('results/2022_03_networks_PET.csv', index_col='fold')\n",
    "pet_GNN_noAPOE    = pd.read_csv('results/2022_04_networks_PET_noAPOE.csv', index_col='fold')\n",
    "pet_bas           = pd.read_csv('results/2022_03_baseline_models_PET.csv',  index_col='fold')\n",
    "\n",
    "pet_all = pd.concat([pet_nonGNN, pet_nonGNN_noAPOE, pet_GNN, pet_GNN_noAPOE])\n",
    "pet_all = pet_all.loc[\n",
    "    (pet_all['dataset'] == 'Only APOE') |\n",
    "    (pet_all['dataset'] == 'AD PPT-Ohmnet') | \n",
    "    (pet_all['dataset'] == 'AD PPT-Ohmnet no APOE')]\n",
    "\n",
    "# PET&DX\n",
    "petdx_nonGNN        = pd.read_csv('results/2022_03_nonGNN_models_PETandDX.csv', index_col='fold')\n",
    "petdx_nonGNN_noAPOE = pd.read_csv('results/2022_04_nonGNN_models_PETandDX_noAPOE.csv', index_col='fold')\n",
    "petdx_GNN           = pd.read_csv('results/2022_03_networks_PETandDX.csv', index_col='fold')\n",
    "petdx_GNN_noAPOE    = pd.read_csv('results/2022_04_networks_PETandDX_noAPOE.csv', index_col='fold')\n",
    "petdx_bas           = pd.read_csv('results/2022_03_baseline_models_PETandDX.csv',  index_col='fold')\n",
    "\n",
    "petdx_all = pd.concat([petdx_nonGNN, petdx_nonGNN_noAPOE, petdx_GNN, petdx_GNN_noAPOE])\n",
    "petdx_all = petdx_all.loc[\n",
    "    (petdx_all['dataset'] == 'Only APOE') |\n",
    "    (petdx_all['dataset'] == 'AD PPT-Ohmnet') | \n",
    "    (petdx_all['dataset'] == 'AD PPT-Ohmnet no APOE')]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f29b33e2",
   "metadata": {},
   "source": [
    "#### Figure 3"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "7ca80ce5",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "The PostScript backend does not support transparency; partially transparent artists will be rendered opaque.\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1296x576 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "order = ['Only APOE', 'AD PPT-Ohmnet','AD PPT-Ohmnet no APOE']\n",
    "hue = ['Logistic Regression','SVM Linear', 'SVM RBF', 'Random Forest', 'GNN GraphGym']\n",
    "hue_bas = ['', '', 'Baseline model', '', '']\n",
    "\n",
    "fig, (ax1, ax2) = plt.subplots(ncols=2, figsize=(18, 8))\n",
    "sns.boxplot(x = 'dataset', y = 'auc', hue = 'model', data = pet_all, order = order, hue_order = hue, ax = ax1,\n",
    "           palette = sns.color_palette([\"#B79F00\", \"#00BA38\", \"#00BFC4\", \"#619CFF\", \"#F564E3\"]),\n",
    "           showmeans = True, meanprops = {\"markerfacecolor\":\"white\", \"markeredgecolor\":\"black\", \"markersize\":\"10\"})\n",
    "sns.boxplot(x = 'dataset', y = 'auc', hue = 'model', data = pet_bas, order = order, hue_order = hue_bas, ax = ax1,\n",
    "           palette = sns.color_palette(['#F8766D', \"#F8766D\", \"#F8766D\", \"#F8766D\", \"#F8766D\"]),\n",
    "           showmeans = True, meanprops = {\"markerfacecolor\":\"white\", \"markeredgecolor\":\"black\", \"markersize\":\"10\"})\n",
    "sns.boxplot(x = 'dataset', y = 'auc', hue = 'model', data = petdx_all, order = order, hue_order = hue, ax = ax2,\n",
    "           palette = sns.color_palette([\"#B79F00\", \"#00BA38\", \"#00BFC4\", \"#619CFF\", \"#F564E3\"]),\n",
    "           showmeans = True, meanprops = {\"markerfacecolor\":\"white\", \"markeredgecolor\":\"black\", \"markersize\":\"10\"})\n",
    "sns.boxplot(x = 'dataset', y = 'auc', hue = 'model', data = petdx_bas, order = order, hue_order = hue_bas, ax = ax2,\n",
    "           palette = sns.color_palette(['#F8766D', \"#F8766D\", \"#F8766D\", \"#F8766D\", \"#F8766D\"]),\n",
    "           showmeans = True, meanprops = {\"markerfacecolor\":\"white\", \"markeredgecolor\":\"black\", \"markersize\":\"10\"})\n",
    "\n",
    "ax1.set_ylim(0.3, 1.0)\n",
    "ax2.set_ylim(0.3, 1.0)\n",
    "\n",
    "ax1.set_ylabel('AUC', fontsize=16)\n",
    "ax2.set_ylabel('AUC', fontsize=16)\n",
    "\n",
    "ax1.set_xlabel('', fontsize=0)\n",
    "ax2.set_xlabel('', fontsize=0)\n",
    "\n",
    "ax1.tick_params(labelsize=16)\n",
    "ax2.tick_params(labelsize=16)\n",
    "\n",
    "ax1.set_title('(a) PET label - ADNI dataset', fontsize=16)\n",
    "ax2.set_title('(b) PET&DX label - ADNI dataset', fontsize=16)\n",
    "\n",
    "plt.tight_layout()\n",
    "\n",
    "# plt.savefig(f'figures/figure3.png', dpi=500)\n",
    "plt.savefig(f'figures/figure3.eps', dpi=500)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "dcf64f0d",
   "metadata": {},
   "source": [
    "#### p-values against baseline, GNNs and random classification"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "mounted-cologne",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Against baseline:\n",
      "( ) 7.1252e-01 AD PPT-Ohmnet - Logistic Regression\n",
      "( ) 6.8956e-01 AD PPT-Ohmnet - SVM Linear\n",
      "( ) 6.6811e-01 AD PPT-Ohmnet - SVM RBF\n",
      "( ) 6.3489e-01 AD PPT-Ohmnet - Random Forest\n",
      "( ) 1.0072e-01 AD PPT-Ohmnet - GNN GraphGym\n",
      "( ) 9.9997e-01 AD PPT-Ohmnet no APOE - Logistic Regression\n",
      "( ) 9.9999e-01 AD PPT-Ohmnet no APOE - SVM Linear\n",
      "( ) 9.9939e-01 AD PPT-Ohmnet no APOE - SVM RBF\n",
      "( ) 9.9992e-01 AD PPT-Ohmnet no APOE - Random Forest\n",
      "( ) 9.9871e-01 AD PPT-Ohmnet no APOE - GNN GraphGym\n",
      "\n",
      "Against GNN:\n",
      "(*) 3.5778e-02 AD PPT-Ohmnet - Logistic Regression\n",
      "(*) 2.6525e-02 AD PPT-Ohmnet - SVM Linear\n",
      "(*) 4.0317e-02 AD PPT-Ohmnet - SVM RBF\n",
      "( ) 6.2651e-02 AD PPT-Ohmnet - Random Forest\n",
      "( ) 5.0000e-01 AD PPT-Ohmnet - GNN GraphGym\n",
      "(*) 1.3357e-02 AD PPT-Ohmnet no APOE - Logistic Regression\n",
      "(*) 1.5148e-02 AD PPT-Ohmnet no APOE - SVM Linear\n",
      "( ) 1.8130e-01 AD PPT-Ohmnet no APOE - SVM RBF\n",
      "(*) 2.7457e-02 AD PPT-Ohmnet no APOE - Random Forest\n",
      "( ) 5.0000e-01 AD PPT-Ohmnet no APOE - GNN GraphGym\n",
      "\n",
      "Against random:\n",
      "( ) 9.9983e-01 AD PPT-Ohmnet - Logistic Regression\n",
      "( ) 9.9999e-01 AD PPT-Ohmnet - SVM Linear\n",
      "( ) 9.9994e-01 AD PPT-Ohmnet - SVM RBF\n",
      "( ) 9.9975e-01 AD PPT-Ohmnet - Random Forest\n",
      "( ) 1.0000e+00 AD PPT-Ohmnet - GNN GraphGym\n",
      "( ) 1.9665e-01 AD PPT-Ohmnet no APOE - Logistic Regression\n",
      "( ) 3.4927e-01 AD PPT-Ohmnet no APOE - SVM Linear\n",
      "( ) 7.8729e-01 AD PPT-Ohmnet no APOE - SVM RBF\n",
      "( ) 2.8052e-01 AD PPT-Ohmnet no APOE - Random Forest\n",
      "( ) 9.8960e-01 AD PPT-Ohmnet no APOE - GNN GraphGym\n",
      "\n",
      "With APOE vs. without APOE:\n",
      "(*) 9.8484e-05 Logistic Regression\n",
      "(*) 3.5104e-06 SVM Linear\n",
      "(*) 1.0920e-03 SVM RBF\n",
      "(*) 2.8652e-04 Random Forest\n",
      "(*) 3.1765e-05 GNN GraphGym\n"
     ]
    }
   ],
   "source": [
    "pet_all = pd.concat([pet_all, pet_bas])\n",
    "pvalues_bas_pet, pvalues_ran_pet = models_stats.pvalues_nonGNNs(pet_all)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "fitted-graduate",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Against baseline:\n",
      "( ) 5.7567e-01 AD PPT-Ohmnet - Logistic Regression\n",
      "( ) 5.2998e-01 AD PPT-Ohmnet - SVM Linear\n",
      "( ) 3.7087e-01 AD PPT-Ohmnet - SVM RBF\n",
      "( ) 4.9994e-01 AD PPT-Ohmnet - Random Forest\n",
      "( ) 5.1717e-02 AD PPT-Ohmnet - GNN GraphGym\n",
      "( ) 1.0000e+00 AD PPT-Ohmnet no APOE - Logistic Regression\n",
      "( ) 9.9997e-01 AD PPT-Ohmnet no APOE - SVM Linear\n",
      "( ) 9.9942e-01 AD PPT-Ohmnet no APOE - SVM RBF\n",
      "( ) 9.9973e-01 AD PPT-Ohmnet no APOE - Random Forest\n",
      "( ) 9.9947e-01 AD PPT-Ohmnet no APOE - GNN GraphGym\n",
      "\n",
      "Against GNN:\n",
      "(*) 1.7280e-02 AD PPT-Ohmnet - Logistic Regression\n",
      "(*) 1.5860e-02 AD PPT-Ohmnet - SVM Linear\n",
      "(*) 4.3754e-02 AD PPT-Ohmnet - SVM RBF\n",
      "(*) 2.7398e-02 AD PPT-Ohmnet - Random Forest\n",
      "( ) 5.0000e-01 AD PPT-Ohmnet - GNN GraphGym\n",
      "(*) 3.4273e-04 AD PPT-Ohmnet no APOE - Logistic Regression\n",
      "(*) 8.2945e-03 AD PPT-Ohmnet no APOE - SVM Linear\n",
      "( ) 5.3144e-01 AD PPT-Ohmnet no APOE - SVM RBF\n",
      "( ) 1.0200e-01 AD PPT-Ohmnet no APOE - Random Forest\n",
      "( ) 5.0000e-01 AD PPT-Ohmnet no APOE - GNN GraphGym\n",
      "\n",
      "Against random:\n",
      "( ) 9.9998e-01 AD PPT-Ohmnet - Logistic Regression\n",
      "( ) 1.0000e+00 AD PPT-Ohmnet - SVM Linear\n",
      "( ) 1.0000e+00 AD PPT-Ohmnet - SVM RBF\n",
      "( ) 9.9999e-01 AD PPT-Ohmnet - Random Forest\n",
      "( ) 1.0000e+00 AD PPT-Ohmnet - GNN GraphGym\n",
      "(*) 1.6727e-02 AD PPT-Ohmnet no APOE - Logistic Regression\n",
      "( ) 1.5417e-01 AD PPT-Ohmnet no APOE - SVM Linear\n",
      "( ) 9.9773e-01 AD PPT-Ohmnet no APOE - SVM RBF\n",
      "( ) 6.5781e-01 AD PPT-Ohmnet no APOE - Random Forest\n",
      "( ) 9.9717e-01 AD PPT-Ohmnet no APOE - GNN GraphGym\n",
      "\n",
      "With APOE vs. without APOE:\n",
      "(*) 5.0849e-07 Logistic Regression\n",
      "(*) 4.2853e-06 SVM Linear\n",
      "(*) 9.3743e-06 SVM RBF\n",
      "(*) 7.9737e-05 Random Forest\n",
      "(*) 9.4311e-08 GNN GraphGym\n"
     ]
    }
   ],
   "source": [
    "petdx_all = pd.concat([petdx_all, petdx_bas])\n",
    "pvalues_bas_petdx, pvalues_ran_petdx = models_stats.pvalues_nonGNNs(petdx_all)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "eb62fe29",
   "metadata": {},
   "source": [
    "#### Supplementary Tables"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "3040dc56",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>dataset</th>\n",
       "      <th>model</th>\n",
       "      <th>acc_format</th>\n",
       "      <th>pre_format</th>\n",
       "      <th>rec_format</th>\n",
       "      <th>f1_format</th>\n",
       "      <th>auc_format</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>fold</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet</td>\n",
       "      <td>Logistic Regression</td>\n",
       "      <td>0.599 ± 0.0534</td>\n",
       "      <td>0.6331 ± 0.0515</td>\n",
       "      <td>0.722 ± 0.156</td>\n",
       "      <td>0.6644 ± 0.066</td>\n",
       "      <td>0.6229 ± 0.0693</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet</td>\n",
       "      <td>SVM Linear</td>\n",
       "      <td>0.6074 ± 0.0553</td>\n",
       "      <td>0.6549 ± 0.0646</td>\n",
       "      <td>0.7 ± 0.1879</td>\n",
       "      <td>0.6604 ± 0.0688</td>\n",
       "      <td>0.627 ± 0.0493</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet</td>\n",
       "      <td>SVM RBF</td>\n",
       "      <td>0.6006 ± 0.061</td>\n",
       "      <td>0.6206 ± 0.079</td>\n",
       "      <td>0.8683 ± 0.1858</td>\n",
       "      <td>0.7062 ± 0.0506</td>\n",
       "      <td>0.6275 ± 0.0628</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet</td>\n",
       "      <td>Random Forest</td>\n",
       "      <td>0.6197 ± 0.0622</td>\n",
       "      <td>0.6799 ± 0.0612</td>\n",
       "      <td>0.6244 ± 0.1017</td>\n",
       "      <td>0.647 ± 0.0673</td>\n",
       "      <td>0.6291 ± 0.0771</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet no APOE</td>\n",
       "      <td>Logistic Regression</td>\n",
       "      <td>0.5538 ± 0.0176</td>\n",
       "      <td>0.5598 ± 0.0095</td>\n",
       "      <td>0.9805 ± 0.0277</td>\n",
       "      <td>0.7127 ± 0.0147</td>\n",
       "      <td>0.481 ± 0.067</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet no APOE</td>\n",
       "      <td>SVM Linear</td>\n",
       "      <td>0.5607 ± 0.016</td>\n",
       "      <td>0.5631 ± 0.0087</td>\n",
       "      <td>0.9902 ± 0.0236</td>\n",
       "      <td>0.7179 ± 0.0129</td>\n",
       "      <td>0.4942 ± 0.0459</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet no APOE</td>\n",
       "      <td>SVM RBF</td>\n",
       "      <td>0.5593 ± 0.0151</td>\n",
       "      <td>0.5625 ± 0.0086</td>\n",
       "      <td>0.9878 ± 0.0237</td>\n",
       "      <td>0.7168 ± 0.0124</td>\n",
       "      <td>0.5192 ± 0.0726</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet no APOE</td>\n",
       "      <td>Random Forest</td>\n",
       "      <td>0.5524 ± 0.027</td>\n",
       "      <td>0.5602 ± 0.0144</td>\n",
       "      <td>0.9634 ± 0.035</td>\n",
       "      <td>0.7084 ± 0.0202</td>\n",
       "      <td>0.4853 ± 0.0771</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet</td>\n",
       "      <td>GNN GraphGym</td>\n",
       "      <td>0.6362 ± 0.0748</td>\n",
       "      <td>0.6739 ± 0.1139</td>\n",
       "      <td>0.6488 ± 0.1169</td>\n",
       "      <td>0.654 ± 0.1031</td>\n",
       "      <td>0.6801 ± 0.0643</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet no APOE</td>\n",
       "      <td>GNN GraphGym</td>\n",
       "      <td>0.5372 ± 0.0321</td>\n",
       "      <td>0.5314 ± 0.1073</td>\n",
       "      <td>0.7821 ± 0.2263</td>\n",
       "      <td>0.586 ± 0.1563</td>\n",
       "      <td>0.5454 ± 0.0513</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>Only APOE</td>\n",
       "      <td>Baseline model</td>\n",
       "      <td>0.5992 ± 0.0399</td>\n",
       "      <td>0.6345 ± 0.0686</td>\n",
       "      <td>0.7659 ± 0.2127</td>\n",
       "      <td>0.6725 ± 0.0719</td>\n",
       "      <td>0.6406 ± 0.0691</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                    dataset                model       acc_format  \\\n",
       "fold                                                                \n",
       "1             AD PPT-Ohmnet  Logistic Regression   0.599 ± 0.0534   \n",
       "1             AD PPT-Ohmnet           SVM Linear  0.6074 ± 0.0553   \n",
       "1             AD PPT-Ohmnet              SVM RBF   0.6006 ± 0.061   \n",
       "1             AD PPT-Ohmnet        Random Forest  0.6197 ± 0.0622   \n",
       "1     AD PPT-Ohmnet no APOE  Logistic Regression  0.5538 ± 0.0176   \n",
       "1     AD PPT-Ohmnet no APOE           SVM Linear   0.5607 ± 0.016   \n",
       "1     AD PPT-Ohmnet no APOE              SVM RBF  0.5593 ± 0.0151   \n",
       "1     AD PPT-Ohmnet no APOE        Random Forest   0.5524 ± 0.027   \n",
       "1             AD PPT-Ohmnet         GNN GraphGym  0.6362 ± 0.0748   \n",
       "1     AD PPT-Ohmnet no APOE         GNN GraphGym  0.5372 ± 0.0321   \n",
       "1                 Only APOE       Baseline model  0.5992 ± 0.0399   \n",
       "\n",
       "           pre_format       rec_format        f1_format       auc_format  \n",
       "fold                                                                      \n",
       "1     0.6331 ± 0.0515    0.722 ± 0.156   0.6644 ± 0.066  0.6229 ± 0.0693  \n",
       "1     0.6549 ± 0.0646     0.7 ± 0.1879  0.6604 ± 0.0688   0.627 ± 0.0493  \n",
       "1      0.6206 ± 0.079  0.8683 ± 0.1858  0.7062 ± 0.0506  0.6275 ± 0.0628  \n",
       "1     0.6799 ± 0.0612  0.6244 ± 0.1017   0.647 ± 0.0673  0.6291 ± 0.0771  \n",
       "1     0.5598 ± 0.0095  0.9805 ± 0.0277  0.7127 ± 0.0147    0.481 ± 0.067  \n",
       "1     0.5631 ± 0.0087  0.9902 ± 0.0236  0.7179 ± 0.0129  0.4942 ± 0.0459  \n",
       "1     0.5625 ± 0.0086  0.9878 ± 0.0237  0.7168 ± 0.0124  0.5192 ± 0.0726  \n",
       "1     0.5602 ± 0.0144   0.9634 ± 0.035  0.7084 ± 0.0202  0.4853 ± 0.0771  \n",
       "1     0.6739 ± 0.1139  0.6488 ± 0.1169   0.654 ± 0.1031  0.6801 ± 0.0643  \n",
       "1     0.5314 ± 0.1073  0.7821 ± 0.2263   0.586 ± 0.1563  0.5454 ± 0.0513  \n",
       "1     0.6345 ± 0.0686  0.7659 ± 0.2127  0.6725 ± 0.0719  0.6406 ± 0.0691  "
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "mean_pet = pd.DataFrame()\n",
    "mean_pet['dataset'] = pet_all['dataset']\n",
    "mean_pet['model'] = pet_all['model']\n",
    "mean_pet['mean_acc'] = pet_all.groupby(['dataset', 'model'])['acc'].transform('mean')\n",
    "mean_pet['std_acc'] = pet_all.groupby(['dataset', 'model'])['acc'].transform('std')\n",
    "mean_pet['mean_pre'] = pet_all.groupby(['dataset', 'model'])['pre'].transform('mean')\n",
    "mean_pet['std_pre'] = pet_all.groupby(['dataset', 'model'])['pre'].transform('std')\n",
    "mean_pet['mean_rec'] = pet_all.groupby(['dataset', 'model'])['rec'].transform('mean')\n",
    "mean_pet['std_rec'] = pet_all.groupby(['dataset', 'model'])['rec'].transform('std')\n",
    "mean_pet['mean_f1']  = pet_all.groupby(['dataset', 'model'])['f1'].transform('mean')\n",
    "mean_pet['std_f1']  = pet_all.groupby(['dataset', 'model'])['f1'].transform('std')\n",
    "mean_pet['mean_auc'] = pet_all.groupby(['dataset', 'model'])['auc'].transform('mean')\n",
    "mean_pet['std_auc'] = pet_all.groupby(['dataset', 'model'])['auc'].transform('std')\n",
    "mean_pet = mean_pet.drop_duplicates()\n",
    "\n",
    "mean_pet['acc_format'] = round(mean_pet[\"mean_acc\"], 4).map(str) + ' ± ' + round(mean_pet[\"std_acc\"], 4).map(str)\n",
    "mean_pet['pre_format'] = round(mean_pet[\"mean_pre\"], 4).map(str) + ' ± ' + round(mean_pet[\"std_pre\"], 4).map(str)\n",
    "mean_pet['rec_format'] = round(mean_pet[\"mean_rec\"], 4).map(str) + ' ± ' + round(mean_pet[\"std_rec\"], 4).map(str)\n",
    "mean_pet['f1_format']  = round(mean_pet[\"mean_f1\"], 4).map(str) + ' ± ' + round(mean_pet[\"std_f1\"], 4).map(str)\n",
    "mean_pet['auc_format'] = round(mean_pet[\"mean_auc\"], 4).map(str) + ' ± ' + round(mean_pet[\"std_auc\"], 4).map(str)\n",
    "mean_pet[['dataset', 'model', 'acc_format', 'pre_format', 'rec_format', 'f1_format', 'auc_format']]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "c8a912d8",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>dataset</th>\n",
       "      <th>model</th>\n",
       "      <th>acc_format</th>\n",
       "      <th>pre_format</th>\n",
       "      <th>rec_format</th>\n",
       "      <th>f1_format</th>\n",
       "      <th>auc_format</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>fold</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet</td>\n",
       "      <td>Logistic Regression</td>\n",
       "      <td>0.6627 ± 0.0798</td>\n",
       "      <td>0.6977 ± 0.0777</td>\n",
       "      <td>0.7424 ± 0.109</td>\n",
       "      <td>0.715 ± 0.0707</td>\n",
       "      <td>0.6825 ± 0.076</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet</td>\n",
       "      <td>SVM Linear</td>\n",
       "      <td>0.6531 ± 0.0541</td>\n",
       "      <td>0.6987 ± 0.0651</td>\n",
       "      <td>0.7319 ± 0.1307</td>\n",
       "      <td>0.7053 ± 0.0463</td>\n",
       "      <td>0.6872 ± 0.0654</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet</td>\n",
       "      <td>SVM RBF</td>\n",
       "      <td>0.669 ± 0.0637</td>\n",
       "      <td>0.7172 ± 0.0779</td>\n",
       "      <td>0.7371 ± 0.1248</td>\n",
       "      <td>0.7174 ± 0.0491</td>\n",
       "      <td>0.7025 ± 0.0638</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet</td>\n",
       "      <td>Random Forest</td>\n",
       "      <td>0.6906 ± 0.0579</td>\n",
       "      <td>0.7414 ± 0.0473</td>\n",
       "      <td>0.7149 ± 0.1135</td>\n",
       "      <td>0.7231 ± 0.0638</td>\n",
       "      <td>0.6901 ± 0.075</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet no APOE</td>\n",
       "      <td>Logistic Regression</td>\n",
       "      <td>0.568 ± 0.0294</td>\n",
       "      <td>0.5737 ± 0.0192</td>\n",
       "      <td>0.9617 ± 0.0455</td>\n",
       "      <td>0.7184 ± 0.0247</td>\n",
       "      <td>0.4425 ± 0.0725</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet no APOE</td>\n",
       "      <td>SVM Linear</td>\n",
       "      <td>0.5586 ± 0.0204</td>\n",
       "      <td>0.5701 ± 0.0158</td>\n",
       "      <td>0.9398 ± 0.0607</td>\n",
       "      <td>0.709 ± 0.0224</td>\n",
       "      <td>0.4687 ± 0.0918</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet no APOE</td>\n",
       "      <td>SVM RBF</td>\n",
       "      <td>0.5711 ± 0.0188</td>\n",
       "      <td>0.5758 ± 0.0145</td>\n",
       "      <td>0.9617 ± 0.0587</td>\n",
       "      <td>0.7197 ± 0.0205</td>\n",
       "      <td>0.5571 ± 0.0481</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet no APOE</td>\n",
       "      <td>Random Forest</td>\n",
       "      <td>0.5679 ± 0.0307</td>\n",
       "      <td>0.576 ± 0.019</td>\n",
       "      <td>0.9339 ± 0.0684</td>\n",
       "      <td>0.7119 ± 0.0309</td>\n",
       "      <td>0.5122 ± 0.0916</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet</td>\n",
       "      <td>GNN GraphGym</td>\n",
       "      <td>0.6992 ± 0.0683</td>\n",
       "      <td>0.7381 ± 0.0688</td>\n",
       "      <td>0.7552 ± 0.1105</td>\n",
       "      <td>0.7408 ± 0.0611</td>\n",
       "      <td>0.7521 ± 0.0589</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>AD PPT-Ohmnet no APOE</td>\n",
       "      <td>GNN GraphGym</td>\n",
       "      <td>0.5502 ± 0.033</td>\n",
       "      <td>0.5741 ± 0.0413</td>\n",
       "      <td>0.8139 ± 0.1811</td>\n",
       "      <td>0.6416 ± 0.0979</td>\n",
       "      <td>0.5554 ± 0.0485</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>Only APOE</td>\n",
       "      <td>Baseline model</td>\n",
       "      <td>0.7066 ± 0.0682</td>\n",
       "      <td>0.7556 ± 0.0714</td>\n",
       "      <td>0.7368 ± 0.1018</td>\n",
       "      <td>0.7412 ± 0.0615</td>\n",
       "      <td>0.6901 ± 0.0979</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                    dataset                model       acc_format  \\\n",
       "fold                                                                \n",
       "1             AD PPT-Ohmnet  Logistic Regression  0.6627 ± 0.0798   \n",
       "1             AD PPT-Ohmnet           SVM Linear  0.6531 ± 0.0541   \n",
       "1             AD PPT-Ohmnet              SVM RBF   0.669 ± 0.0637   \n",
       "1             AD PPT-Ohmnet        Random Forest  0.6906 ± 0.0579   \n",
       "1     AD PPT-Ohmnet no APOE  Logistic Regression   0.568 ± 0.0294   \n",
       "1     AD PPT-Ohmnet no APOE           SVM Linear  0.5586 ± 0.0204   \n",
       "1     AD PPT-Ohmnet no APOE              SVM RBF  0.5711 ± 0.0188   \n",
       "1     AD PPT-Ohmnet no APOE        Random Forest  0.5679 ± 0.0307   \n",
       "1             AD PPT-Ohmnet         GNN GraphGym  0.6992 ± 0.0683   \n",
       "1     AD PPT-Ohmnet no APOE         GNN GraphGym   0.5502 ± 0.033   \n",
       "1                 Only APOE       Baseline model  0.7066 ± 0.0682   \n",
       "\n",
       "           pre_format       rec_format        f1_format       auc_format  \n",
       "fold                                                                      \n",
       "1     0.6977 ± 0.0777   0.7424 ± 0.109   0.715 ± 0.0707   0.6825 ± 0.076  \n",
       "1     0.6987 ± 0.0651  0.7319 ± 0.1307  0.7053 ± 0.0463  0.6872 ± 0.0654  \n",
       "1     0.7172 ± 0.0779  0.7371 ± 0.1248  0.7174 ± 0.0491  0.7025 ± 0.0638  \n",
       "1     0.7414 ± 0.0473  0.7149 ± 0.1135  0.7231 ± 0.0638   0.6901 ± 0.075  \n",
       "1     0.5737 ± 0.0192  0.9617 ± 0.0455  0.7184 ± 0.0247  0.4425 ± 0.0725  \n",
       "1     0.5701 ± 0.0158  0.9398 ± 0.0607   0.709 ± 0.0224  0.4687 ± 0.0918  \n",
       "1     0.5758 ± 0.0145  0.9617 ± 0.0587  0.7197 ± 0.0205  0.5571 ± 0.0481  \n",
       "1       0.576 ± 0.019  0.9339 ± 0.0684  0.7119 ± 0.0309  0.5122 ± 0.0916  \n",
       "1     0.7381 ± 0.0688  0.7552 ± 0.1105  0.7408 ± 0.0611  0.7521 ± 0.0589  \n",
       "1     0.5741 ± 0.0413  0.8139 ± 0.1811  0.6416 ± 0.0979  0.5554 ± 0.0485  \n",
       "1     0.7556 ± 0.0714  0.7368 ± 0.1018  0.7412 ± 0.0615  0.6901 ± 0.0979  "
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "mean_petdx = pd.DataFrame()\n",
    "mean_petdx['dataset'] = petdx_all['dataset']\n",
    "mean_petdx['model'] = petdx_all['model']\n",
    "mean_petdx['mean_acc'] = petdx_all.groupby(['dataset', 'model'])['acc'].transform('mean')\n",
    "mean_petdx['std_acc'] = petdx_all.groupby(['dataset', 'model'])['acc'].transform('std')\n",
    "mean_petdx['mean_pre'] = petdx_all.groupby(['dataset', 'model'])['pre'].transform('mean')\n",
    "mean_petdx['std_pre'] = petdx_all.groupby(['dataset', 'model'])['pre'].transform('std')\n",
    "mean_petdx['mean_rec'] = petdx_all.groupby(['dataset', 'model'])['rec'].transform('mean')\n",
    "mean_petdx['std_rec'] = petdx_all.groupby(['dataset', 'model'])['rec'].transform('std')\n",
    "mean_petdx['mean_f1']  = petdx_all.groupby(['dataset', 'model'])['f1'].transform('mean')\n",
    "mean_petdx['std_f1']  = petdx_all.groupby(['dataset', 'model'])['f1'].transform('std')\n",
    "mean_petdx['mean_auc'] = petdx_all.groupby(['dataset', 'model'])['auc'].transform('mean')\n",
    "mean_petdx['std_auc'] = petdx_all.groupby(['dataset', 'model'])['auc'].transform('std')\n",
    "mean_petdx = mean_petdx.drop_duplicates()\n",
    "\n",
    "mean_petdx['acc_format'] = round(mean_petdx[\"mean_acc\"], 4).map(str) + ' ± ' + round(mean_petdx[\"std_acc\"], 4).map(str)\n",
    "mean_petdx['pre_format'] = round(mean_petdx[\"mean_pre\"], 4).map(str) + ' ± ' + round(mean_petdx[\"std_pre\"], 4).map(str)\n",
    "mean_petdx['rec_format'] = round(mean_petdx[\"mean_rec\"], 4).map(str) + ' ± ' + round(mean_petdx[\"std_rec\"], 4).map(str)\n",
    "mean_petdx['f1_format']  = round(mean_petdx[\"mean_f1\"], 4).map(str) + ' ± ' + round(mean_petdx[\"std_f1\"], 4).map(str)\n",
    "mean_petdx['auc_format'] = round(mean_petdx[\"mean_auc\"], 4).map(str) + ' ± ' + round(mean_petdx[\"std_auc\"], 4).map(str)\n",
    "mean_petdx[['dataset', 'model', 'acc_format', 'pre_format', 'rec_format', 'f1_format', 'auc_format']]"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.7.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}