{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "8dffc528-5186-40e2-a022-932640a9437a",
   "metadata": {},
   "source": [
    "# TP1 : Prise en main"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9f0f2913-e1cd-45f8-a5c7-654c4fca953c",
   "metadata": {},
   "source": [
    "## Préambule"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "266fe37a-36dc-46af-aca6-e83538abf1f2",
   "metadata": {},
   "source": [
    "### Modules requis"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5de70c5a-0f5c-4dc5-924e-5262f5236ad7",
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install numpy scikit-learn torch torchvision gudhi cripser biopython torchmetrics tqdm pydicom pymol-open-source-whl"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "0f1f3c84-3b53-4bdd-88d2-7bd7f8b81d7d",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-11-17T10:03:17.610714Z",
     "iopub.status.busy": "2026-11-17T10:03:17.609991Z",
     "iopub.status.idle": "2026-11-17T10:03:24.653769Z",
     "shell.execute_reply": "2026-11-17T10:03:24.653457Z",
     "shell.execute_reply.started": "2026-11-17T10:03:17.610678Z"
    }
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/nardingennequin/.local/lib/python3.12/site-packages/matplotlib/projections/__init__.py:63: UserWarning: Unable to import Axes3D. This may be due to multiple versions of Matplotlib being installed (e.g. as a system package and as a pip package). As a result, the 3D projection is not available.\n",
      "  warnings.warn(\"Unable to import Axes3D. This may be due to multiple versions of \"\n"
     ]
    }
   ],
   "source": [
    "import numpy as np\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.nn.functional as F\n",
    "import torch.utils.data\n",
    "import matplotlib.pyplot as plt\n",
    "from tqdm import tqdm\n",
    "from torchvision import datasets, transforms\n",
    "import torchmetrics\n",
    "import gudhi\n",
    "import gudhi.representations\n",
    "import cripser\n",
    "import pydicom as dicom\n",
    "import Bio.PDB as pdb\n",
    "from scipy.ndimage import zoom\n",
    "import zipfile\n",
    "import urllib\n",
    "import os\n",
    "import pymol\n",
    "from Bio.SeqUtils import seq1\n",
    "\n",
    "from sklearn.model_selection import train_test_split"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "81ae0fe8-5616-4f37-95c0-3df778408fe9",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-11-17T10:03:24.654498Z",
     "iopub.status.busy": "2026-11-17T10:03:24.654168Z",
     "iopub.status.idle": "2026-11-17T10:03:24.820582Z",
     "shell.execute_reply": "2026-11-17T10:03:24.820089Z",
     "shell.execute_reply.started": "2026-11-17T10:03:24.654483Z"
    }
   },
   "outputs": [],
   "source": [
    "device = \"cuda\" if torch.cuda.is_available() else \"cpu\""
   ]
  },
  {
   "cell_type": "markdown",
   "id": "46c86a00-a457-4130-8402-2c155971bb91",
   "metadata": {},
   "source": [
    "## Images de chiffres"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b50a568c-8a8b-4bd8-b1ba-05d5b3b4090e",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "### Dataset\n",
    "On utilisera en préambule le jeu de données MNIST qui contient des images de chiffres en noir et blanc.\n",
    "\n",
    "Chaque image est encodée par un tableau de taille `(1,28,28)`, où la première coordonnée correspond au nombre de canaux de l'image (ici un seul, pour une image en couleur ce serait 3 canaux pour chaque couleur primaire Rouge, Vert, Bleu). Les images sont encodées en niveaux de gris avec un nombre flottant entre 0 et 1 pour chaque pixel.\n",
    "\n",
    "Le module `datasets` de la libraire `torchvision` permet de récupérer ce jeu de données au format `Tensor` de pyTorch."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "3829564e-f8ee-4706-8264-2a4d757f9f34",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-11-17T10:03:26.337618Z",
     "iopub.status.busy": "2026-11-17T10:03:26.337304Z",
     "iopub.status.idle": "2026-11-17T10:03:32.194619Z",
     "shell.execute_reply": "2026-11-17T10:03:32.193962Z",
     "shell.execute_reply.started": "2026-11-17T10:03:26.337592Z"
    }
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 9.91M/9.91M [00:01<00:00, 7.37MB/s]\n",
      "100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28.9k/28.9k [00:00<00:00, 289kB/s]\n",
      "100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1.65M/1.65M [00:00<00:00, 2.67MB/s]\n",
      "100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4.54k/4.54k [00:00<00:00, 7.80MB/s]\n"
     ]
    }
   ],
   "source": [
    "X_train = datasets.MNIST(\"data\", train=True, download=True, transform=transforms.ToTensor())\n",
    "dataloader_train = torch.utils.data.DataLoader(X_train, batch_size=60, shuffle=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "bd0a62df-8432-492f-892c-ba50242c7a1c",
   "metadata": {},
   "source": [
    "L'objet `X_train` est ici un itérateur qui renvoie, à chaque itération, l'image et sa classification (le chiffre représenté dans l'image). Le dataloader permet de mélanger aléatoirement le dataset et pourra être utilisé comme un itérateur au moment d'entraîner un modèle d'apprentissage afin de découper le jeu de données en plusieurs parts plus petites (\"batches\") pour diminuer la place requise en mémoire vive."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "d2b6bf72-e10f-4b85-a4e2-1d4bd22a5793",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-11-17T10:03:50.303551Z",
     "iopub.status.busy": "2026-11-17T10:03:50.303288Z",
     "iopub.status.idle": "2026-11-17T10:03:50.532398Z",
     "shell.execute_reply": "2026-11-17T10:03:50.531957Z",
     "shell.execute_reply.started": "2026-11-17T10:03:50.303534Z"
    },
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "5\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "4\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "9\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "for i in range(5):\n",
    "    plt.imshow(X_train[i][0][0], cmap=\"gray\")\n",
    "    print(X_train[i][1])\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "efbfb589-b37f-4d94-945d-435375559491",
   "metadata": {},
   "source": [
    "Rajoutez ci-dessous les définitions du dataset de test (`X_test` et `dataloader_test`) sur le modèle du jeu d'entraînement:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "52079fca-9661-459e-a27f-311d07a24340",
   "metadata": {},
   "outputs": [],
   "source": [
    "X_test           = None\n",
    "dataloader_test  = None"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "630b171a-f20d-4383-901c-a7f1a36e6a76",
   "metadata": {},
   "source": [
    "### Réseau de neurones convolutionnel"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "bb6e3808-cfcd-4d28-a33d-bead11f57228",
   "metadata": {},
   "source": [
    "Notre objectif est de construire un réseau de neurone simple capable de reconnaître un chiffre dans une image.\n",
    "On utilise la libraire pyTorch."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "59383397-0475-4510-9749-16a78d74a1a8",
   "metadata": {},
   "source": [
    "#### Formulation du réseau de neurones\n",
    "Le réseau de neurones ci-dessous est le réseau le plus simple possible que l'on puisse définir sur nos entrées.\n",
    "Il ne comprend qu'une seule couche linéaire, c'est à dire que le réseau va devoir ajuster des poids indépendamment pour chaque pixel de l'image lors de l'entraînement."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "e9229ef6-8fbc-44fd-8fbf-79e3924f3693",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-11-17T10:17:04.778856Z",
     "iopub.status.busy": "2026-11-17T10:17:04.778095Z",
     "iopub.status.idle": "2026-11-17T10:17:04.781557Z",
     "shell.execute_reply": "2026-11-17T10:17:04.780943Z",
     "shell.execute_reply.started": "2026-11-17T10:17:04.778838Z"
    }
   },
   "outputs": [],
   "source": [
    "class Net(nn.Module):\n",
    "    def __init__(self, in_channels, num_classes):\n",
    "        super().__init__()\n",
    "        self.fc1   = nn.Linear(in_channels*28*28, num_classes)\n",
    "\n",
    "    def forward(self, x):\n",
    "        x = self.fc1(x.reshape(x.shape[0], -1))\n",
    "        return x"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "dd743d0d-c771-4221-ada9-98381a1e3f2b",
   "metadata": {},
   "source": [
    "La cellule ci-dessous définit l'architecture du réseau de neurones. Pour l'utiliser, il faut d'abord l'instancier:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "1f76bf9d-fe08-4a5b-9dcf-27ccb02429ae",
   "metadata": {},
   "outputs": [],
   "source": [
    "model = Net(in_channels=1, num_classes=10).to(device)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b9afa715-cb8f-4d41-9a29-7a7a7d644fd3",
   "metadata": {},
   "source": [
    "Avec ces paramètres, le modèle doit ajuster $(28\\times 28 + 1)\\times 10 = 7850$ poids lors de l'entraînement."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "23f6df9b-b362-4742-84e3-a58fc68c7934",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Paramètres fc1.weight : 7840 poids\n",
      "Paramètres fc1.bias : 10 poids\n",
      "Nombre total de paramètres: 7850\n"
     ]
    }
   ],
   "source": [
    "num_weights = 0\n",
    "for name, values in model.named_parameters():\n",
    "    print(\"Paramètres\", name, \":\", np.prod(values.shape), \"poids\")\n",
    "    num_weights += np.prod(values.shape)\n",
    "print(\"Nombre total de paramètres:\", num_weights)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1cacefeb-ed6d-43f1-8865-e9531a960456",
   "metadata": {},
   "source": [
    "Plutôt que d'utiliser un réseau dense qui traite chaque pixel indépendamment, une stratégie possible est d'extraire l'information *localement* autour de chaque pixel en utilisant un *filtre* qui condense l'information autour du pixel en fonction de ses voisins, en parcourant l'image le long d'une fenêtre glissante. Ainsi, on n'a besoin d'apprendre que le filtre, généralement une matrice carrée de petite taille, pour déterminer comment l'information circule de pixel en pixel.\n",
    "L'opération fondamentale est appelée *convolution*. Pour $M$ une image de taille $h\\times w$, et $F$ une matrice carrée de taille $n\\times n$, la convolution produit une nouvelle image $M' = F * M$ de taille $(h-n)\\times(w-n)$, où\n",
    "$$\n",
    "(F*M)_{ij} = \\sum_{x=0}^{n}\\sum_{y=0}^{n} F_{xy}M_{i+x,j+y}.\n",
    "$$\n",
    "Pour obtenir plutôt une image de même dimension $h\\times w$, il est possible de considérer plutôt une image d'origine $M$ de taille $(h+n)\\times(w+n)$, en remplissant les coefficients supplémentaires avec des $0$. En appliquant plusieurs convolutions à la suite, on peut ainsi faire circuler l'information de proche en proche, et le réseau se charge d'apprendre un filtre $F$ pertinent pour la tâche considérée.\n",
    "\n",
    "Aux couches convolutionnelles peuvent également s'ajouter des phases de \"pooling\", qui réduisent la dimension de l'image en appliquant un principe similaire. Par exemple, un _max pooling_ de largeur $2$ divise par deux les dimensions de l'image:\n",
    "$$\n",
    "  M'_{ij} = \\max \\{M_{2i,2j}, M_{2i+1,2j}, M_{2i,2j+1}, M_{2i+1,2j+1}\\}.\n",
    "$$\n",
    "\n",
    "Une fois que les dimensions de l'image sont suffisamment réduites, on peut par exemple appliquer une couche dense qui nécessitera d'apprendre un bien moins grand nombre de poids. Le réseau convolutionnel est également particulièrement adapté pour certaines tâches qui nécessitent d'apprendre des raisonnements \"locaux\", comme par exemple le débruitage ou la localisation d'objets dans une image."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "45b96114-8add-487e-85e3-4b20f8e197a1",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-11-16T15:55:18.873874Z",
     "iopub.status.busy": "2026-11-16T15:55:18.873681Z",
     "iopub.status.idle": "2026-11-16T15:55:18.877504Z",
     "shell.execute_reply": "2026-11-16T15:55:18.876903Z",
     "shell.execute_reply.started": "2026-11-16T15:55:18.873861Z"
    }
   },
   "source": [
    "![convolutional neural network example](https://upload.wikimedia.org/wikipedia/commons/6/63/Typical_cnn.png)\n",
    "*Exemple de réseau de neurones alternant les couches convolutionnelles et les phases de pooling.*\n",
    "![segmentation example](https://miro.medium.com/v2/1*JtWbxWaF5tJfTJOvDr1GsA.png)\n",
    "*Exemple d'application à la localisation d'objets dans une image.*"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8288616d-57f1-4d6d-9526-6f91cec4caf6",
   "metadata": {},
   "source": [
    "**Exercice**: Remplacez l'architecture du réseau de neurone défini ci-dessus par une architecture de réseau de neurone convolutionnel constituée des couches successives suivantes:\n",
    "* Une couche convolutionnelle `Conv2d` avec un noyau 3 par 3, un remplissage (padding) de 1, et 8 canaux en sortie.\n",
    "* Une autre couche convolutionnelle avec les même paramètres et 4 canaux en sortie\n",
    "* Une couche linéaire finale qui produit un vecteur de 10 valeurs.\n",
    "\n",
    "Chaque application d'une couche convolutionnelle sera suivie d'une activation (`F.relu`) et d'un pooling (`MaxPool2d`) avec un noyau 2 par 2 et un pas (stride) de 2.\n",
    "\n",
    "Pour la couche linéaire, il est nécessaire de connaître la taille du vecteur en entrée. Celui-ci doit être de la forme `(4,7,7)` (chaque pooling divise la taille de l'image par 2)."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "336a96cf-3ccd-4ce6-bd5f-f766d9c26d42",
   "metadata": {},
   "source": [
    "#### Entraînement\n",
    "Pour l'entraînement, on utilise la fonction de pénalisation `CrossEntropyLoss`, et l'optimiseur `Adam`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "9bdc5b63-b146-41a6-881c-b593aaf13c8a",
   "metadata": {},
   "outputs": [],
   "source": [
    "criterion = nn.CrossEntropyLoss()\n",
    "optimizer = torch.optim.Adam(model.parameters(), lr=0.001)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8149ac2b-e63e-4442-822e-a28ac3a73eb7",
   "metadata": {},
   "source": [
    "Complétez la section suivante afin d'entraîner le réseau de neurones. Pour chaque batch, il s'agit de faire une prédiction avec le modèle, puis de calculer une `loss` avec `criterion` pour comparer la prédiction à la réalité, puis d'utiliser `loss.backward()` pour calculer les gradients, et `optimizer.step()` pour mettre à jour les paramètres du réseau."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "daf2aaa0-33fb-4d01-98dc-42dc18f944f1",
   "metadata": {},
   "outputs": [],
   "source": [
    "num_epochs = 10\n",
    "ls = []\n",
    "for epoch in range(num_epochs):\n",
    "    model.train() # Indique au modèle que ses paramètres peuvent être modifiés\n",
    "    print(f\"Epoch {epoch + 1} / {num_epochs}:\")\n",
    "    for xs, ys in tqdm(dataloader_train, ncols=80):\n",
    "        optimizer.zero_grad() # Remet les gradients à zéro à chaque itération. Le comportement par défaut est de les accumuler\n",
    "        loss = 0 # À modifier!\n",
    "    print(\"  Current loss:\", loss)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8d268815-4ecb-4610-81ba-c668effb640d",
   "metadata": {},
   "source": [
    "#### Test\n",
    "La case suivante calcule le pourcentage de prédictions réussies par le réseau:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "814c6bc9-1574-4c37-9168-cf9c81b9a7c4",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Test accuracy: 0.10270000249147415\n"
     ]
    }
   ],
   "source": [
    "acc = torchmetrics.Accuracy(task=\"multiclass\", num_classes=10).to(device)\n",
    "model.eval()\n",
    "with torch.no_grad():\n",
    "    for xs, ys in dataloader_test:\n",
    "        zs = model(xs.to(device))\n",
    "        _, preds = torch.max(zs, 1)\n",
    "        acc(preds, ys.to(device))\n",
    "test_accuracy = acc.compute()\n",
    "print(\"Test accuracy:\", test_accuracy.item())"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "894b2f30-b521-48cb-ad2e-981418958ebd",
   "metadata": {},
   "source": [
    "Pour illustrer, affichez, pour les 5 premières images du dataset, l'image elle même et la classification prédite par le réseau de neurones."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "dafeef62-aa9e-4082-a701-80b349c54f27",
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "pass"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4f8e1451-a399-4bc4-8826-ab65583ccad8",
   "metadata": {},
   "source": [
    "## Structures de protéines\n",
    "Les structures 3D de protéines résolues expérimentalement sont répertoriées dans la Protein Data Bank (PDB), et accessibles sous la forme de fichiers `.pdb` ou mmCIF (`.cif`). On travaillera ici sur des fichiers au format PDB. Il s'agit d'un format simple organisé ligne par ligne: chaque ligne est limitée à 80 caractères et contient les informations sur un atome de la structure, ou alors des informations supplémentaires sur la structure contenue dans le fichier (date de résolution publication associée, auteurs, etc).\n",
    "\n",
    "Une ligne sur un atome contient, par exemple, ses trois coordonnées dans l'espace aux emplacements 31 à 54. En pratique, on utilisera un parser fourni par la bibliothèque [Biopython](https://biopython.org/) ou pyMOL."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f4330884-4c31-46c4-ad40-6b0a6b0ed16e",
   "metadata": {},
   "source": [
    "### Chargement d'un fichier PDB"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "00419f91-6840-46fe-b532-6c3abb09905d",
   "metadata": {},
   "source": [
    "La bibliothèque [Biopython](https://biopython.org/docs/dev/Tutorial/chapter_pdb.html) peut être utilisée pour télécharger des fichiers PDB, dont par exemple la protéine d'insuline humaine enregistrée sous l'identifiant [3I40](https://www.rcsb.org/structure/3I40)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ad6af3f8-9a74-483e-a19e-953209954ff0",
   "metadata": {},
   "outputs": [],
   "source": [
    "from Bio.PDB import PDBList \n",
    "pdbl = PDBList()\n",
    "insuline_fp = pdbl.retrieve_pdb_file(\"1FAT\", pdir=\"data/\", file_format=\"pdb\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "660242c0-4833-44d1-a7ed-2baffc2778a9",
   "metadata": {},
   "source": [
    "### Extraction de structure secondaire\n",
    "Les structures de protéines peuvent être traitées avec pyMOL pour, par exemple, extraire l'information des *structures secondaires* de la protéine (hélices α et feuillets β). Cette information est encodée par une suite de lettres parmi $H,L,S,-$, où $H$ désigne une hélice, $S$ un feuillet (*sheet*), $L$ une boucle (*loop*) et $-$ un résidu n'ayant pas de structure secondaire particulière. Par exemple:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "6f276e6f-c99e-4545-b2ef-9378dd776710",
   "metadata": {},
   "outputs": [],
   "source": [
    "myspace = {\"residues\": []}\n",
    "\n",
    "# Chargement du premier modèle de protéine contenu dans le fichier\n",
    "pymol.cmd.load(insuline_fp, \"insulin\")\n",
    "pymol.cmd.create(\"model1\", \"insulin\", source_state=1, target_state=1)\n",
    "\n",
    "seq, ss = \"\", \"\"\n",
    "# Une protéine peut être constituée de plusieurs chaînes.\n",
    "# Ici on va les mettre bout à bout.\n",
    "for chain in pymol.cmd.get_chains(\"model1\"):\n",
    "    myspace[\"residues\"] = []\n",
    "\n",
    "    # On charge tous les atomes de carbone alpha dans la variable \"residues\"\n",
    "    pymol.cmd.iterate(\n",
    "        f\"model1 and chain {chain} and name CA\",\n",
    "        \"residues.append((resn, ss))\", # resn = nom du résidu, ss = structure secondaire\n",
    "        space=myspace\n",
    "    )\n",
    "\n",
    "    # On parcourt les résidus\n",
    "    for resn, secondary in myspace[\"residues\"]:\n",
    "        aa = seq1(resn) # Conversion du nom du résidu en une seule lettre\n",
    "        if aa:\n",
    "            seq += aa\n",
    "            ss += secondary if secondary else \"-\"\n",
    "\n",
    "print(\"Amino acid sequence:\", seq)\n",
    "print(\"Secondary structure:\", ss)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c8faed23-5729-4bad-b78a-080b44dac1e0",
   "metadata": {},
   "source": [
    "Pour un prochain TP, on travaillera sur un jeu de données de structures 3D qui contient environ 15'000 structures sous la forme de fichiers PDB."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5c9cc69c-f90b-4b72-a5c9-a38a66700be7",
   "metadata": {},
   "outputs": [],
   "source": [
    "f,_ = urllib.request.urlretrieve(\"http://math.univ-lyon1.fr/~nardingennequin/files/ens/structures.zip\")\n",
    "s3dfolder = zipfile.ZipFile(f, \"r\")\n",
    "print(\"Liste des fichiers présents dans l'archive:\")\n",
    "s3dfolder.namelist()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a2015eda-cd79-4c7d-9029-8cbe6e0f4b74",
   "metadata": {},
   "source": [
    "La cellule suivante extrait tout les fichiers PDB dans un dossier `data/pdb/`.\n",
    "Pour éviter les problèmes d'espace disque (environ 1.5GB pour le dataset entier), vous pouvez stopper l'extraction après un seuil de votre choix (idéalement au moins 2000 structures pour avoir assez de données)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "fcfea0e2-b2df-417e-a4ed-68f4b95d7426",
   "metadata": {},
   "outputs": [],
   "source": [
    "for file in tqdm(s3dfolder.namelist(), ncols=80):\n",
    "    if file.startswith('dompdb/'):\n",
    "        s3dfolder.extract(file, 'data/')\n",
    "        if os.path.isfile(os.path.join(\"data/\", file)):\n",
    "          os.rename(os.path.join(\"data/\", file), os.path.join(\"data/\", file + \".pdb\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3ae38f03-23a3-4fff-aafa-17921bb5201b",
   "metadata": {},
   "source": [
    "**Exercice:**\n",
    "À l'aide des fonctions de pyMOL montrées ci-dessus, extraire les séquences et les structures secondaires de chaque protéine dans le jeu de données. Placez les chaînes de caractère obtenues dans deux dictionnaires python respectivement `all_seq` et `all_ss` (les clés seront les identifiants de protéines donnés par les noms de fichiers, e.g. `all_ss[\"1zjcA01\"] = \"LHHH...\"` pour le fichier `data/dompdb/1zjcA01.pdb`)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "id": "a6969296-63e9-46da-bdda-4a3bc6ea6fa9",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-11-16T16:33:33.937386Z",
     "iopub.status.busy": "2026-11-16T16:33:33.936770Z",
     "iopub.status.idle": "2026-11-16T16:33:33.945048Z",
     "shell.execute_reply": "2026-11-16T16:33:33.943295Z",
     "shell.execute_reply.started": "2026-11-16T16:33:33.937342Z"
    }
   },
   "outputs": [],
   "source": [
    "pass"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "57695bcd-98d8-469f-a141-a2e84338f045",
   "metadata": {},
   "source": [
    "Dans certaines situations, les chaînes de caractères correspondant à la séquence de la protéine et à sa structure secondaire peuvent être de tailles différentes. La cellule suivante se charge d'éliminer ces cas particuliers du dataset:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "2ad4185c-5d15-4a09-8910-a6bf8e91e0ef",
   "metadata": {},
   "outputs": [],
   "source": [
    "count = 0\n",
    "deleted_items = []\n",
    "for k in list(all_seq):\n",
    "    if len(all_seq[k]) != len(all_ss[k]):\n",
    "        count += 1\n",
    "        deleted_items.append(k)\n",
    "        del(all_seq[k])\n",
    "        del(all_ss[k])\n",
    "print(\"Removed\", count, \"edge cases.\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e4f9ec2e-1f4b-415a-8ed2-c4fe11279abc",
   "metadata": {},
   "source": [
    "Fin du fichier."
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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
}
