Merge branch 'main' of https://github.com/tahamv/DeepDRA

# Conflicts:
#	main.py

DeepDRA.py

@@ -22,8 +22,7 @@ class DeepDRA(nn.Module):
     - mlp_output_dim (int): Output dimension for the MLP.
     """
 
-    def __init__(self, cell_modality_sizes, drug_modality_sizes, cell_ae_latent_dim, drug_ae_latent_dim, mlp_input_dim,
-                 mlp_output_dim):
+    def __init__(self, cell_modality_sizes, drug_modality_sizes, cell_ae_latent_dim, drug_ae_latent_dim):
         super(DeepDRA, self).__init__()
 
         # Initialize cell and drug autoencoders
@@ -35,7 +34,7 @@ class DeepDRA(nn.Module):
         self.drug_modality_sizes = drug_modality_sizes
 
         # Initialize MLP
-        self.mlp = MLP(mlp_input_dim, mlp_output_dim)
+        self.mlp = MLP(cell_ae_latent_dim+drug_ae_latent_dim, 1)
 
     def forward(self, cell_x, drug_x):
         """
@@ -141,6 +140,8 @@ def train(model, train_loader, val_loader,  num_epochs,class_weights):
 
             # Backward pass and optimization
             total_loss.backward()
+            torch.nn.utils.clip_grad_norm_(model.parameters(), 1)
+
             mlp_optimizer.step()
             total_train_loss += total_loss.item()
 

data_loader.py

@@ -9,7 +9,7 @@ from utils import DRUG_DATA_FOLDER
 
 class RawDataLoader:
     @staticmethod
-    def load_data(data_modalities, raw_file_directory, screen_file_directory, sep):
+    def load_data(data_modalities, raw_file_directory, screen_file_directory, sep, drug_directory=DRUG_DATA_FOLDER):
         """
         Load raw data and screening data, perform intersection, and adjust screening data.
 
@@ -29,7 +29,7 @@ class RawDataLoader:
 
         # Step 2: Load drug data files for specified data modalities
         drug_data = RawDataLoader.load_raw_files(intersect=True, data_modalities=data_modalities,
-                                                 raw_file_directory=DRUG_DATA_FOLDER)
+                                                 raw_file_directory=drug_directory)
 
         # Step 3: Update the 'data' dictionary with drug data
         data.update(drug_data)
@@ -135,14 +135,11 @@ class RawDataLoader:
                 df.columns = df.columns.str.replace('_cell_CN', '')
                 df.columns = df.columns.str.replace('_cell_exp', '')
 
-                # Note that drug_comp raw table has some NA values so we should impute it
-                if any(df.isna()):
-                    df = pd.DataFrame(SimpleImputer(strategy='mean').fit_transform(df),
-                                      columns=df.columns).set_index(df.index)
-                if file.startswith('drug_comp'):  # We need to normalize the drug_data comp dataset
-                    df = ((df - df.mean()) / df.std()).fillna(0)
-                elif file.startswith('drug_desc'):  # We need to normalize the drug_data comp dataset
-                    df = ((df - df.mean()) / df.std()).fillna(0)
+                df = (df - df.min()) / (df.max() - df.min())
+                df = df.fillna(0)
+
+                print("has null:")
+                print(df.isnull().sum().sum())
                 if intersect:
                     if file.startswith('cell'):
                         if cell_line_names:

evaluation.py

@@ -53,7 +53,7 @@ class Evaluation:
 
 
     @staticmethod
-    def evaluate(all_targets, mlp_output, show_plot=True):
+    def evaluate(all_targets, mlp_output, show_plot=False):
         """
         Evaluate model performance based on predictions and targets.
 
@@ -172,3 +172,5 @@ class Evaluation:
             f'AVG: Accuracy: {avg_accuracy:.3f}, Precision: {avg_precision:.3f}, Recall: {avg_recal:.3f}, F1 score: {avg_f1score:.3f}, AUC: {avg_auc:.3f}, ,AUPRC: {avg_auprc:.3f}')
 
         print(" Average AUC: {:.3f} \t Average AUPRC: {:.3f} \t Std AUPRC: {:.3f}".format(avg_auc, avg_auprc, std_auprc))
+        return {'Accuracy': avg_accuracy, 'Precision': avg_precision, 'Recall': avg_recal, 'F1 score': avg_f1score, 'AUC': avg_auc,
+                'AUPRC': avg_auprc}

main.py

@@ -3,6 +3,7 @@ from sklearn.model_selection import train_test_split
 from sklearn.utils.class_weight import compute_class_weight
 
 from torch.utils.data import TensorDataset, DataLoader, SubsetRandomSampler
+from sklearn.model_selection import KFold
 
 from DeepDRA import DeepDRA, train, test
 from data_loader import RawDataLoader
@@ -13,6 +14,12 @@ import torch
 import numpy as np
 import pandas as pd
 
+# Step 1: Define the batch size for training
+batch_size = 64
+
+# Step 2: Instantiate the combined model
+ae_latent_dim = 50
+num_epochs = 25
 
 def train_DeepDRA(x_cell_train, x_cell_test, x_drug_train, x_drug_test, y_train, y_test, cell_sizes, drug_sizes,device):
     """
@@ -33,16 +40,10 @@ def train_DeepDRA(x_cell_train, x_cell_test, x_drug_train, x_drug_test, y_train,
     - result: Evaluation result on the test set.
     """
 
-    # Step 1: Define the batch size for training
-    batch_size = 64
 
-    # Step 2: Instantiate the combined model
-    ae_latent_dim = 50
-    mlp_input_dim = 2 * ae_latent_dim
-    mlp_output_dim = 1
-    num_epochs = 25
     model = DeepDRA(cell_sizes, drug_sizes, ae_latent_dim, ae_latent_dim, mlp_input_dim, mlp_output_dim)
     model= model.to(device)
+
     # Step 3: Convert your training data to PyTorch tensors
     x_cell_train_tensor = torch.Tensor(x_cell_train.values)
     x_drug_train_tensor = torch.Tensor(x_drug_train.values)
@@ -96,8 +97,53 @@ def train_DeepDRA(x_cell_train, x_cell_test, x_drug_train, x_drug_test, y_train,
     # Step 11: Test the model
     return test(model, test_loader)
 
+def cv_train(x_cell_train, x_drug_train, y_train, cell_sizes,
+                                    drug_sizes, device, k=5, ):
+
+
+    splits = KFold(n_splits=k, shuffle=True, random_state=RANDOM_SEED)
+    history = {'AUC': [], 'AUPRC': [], "Accuracy": [], "Precision": [], "Recall": [], "F1 score": []}
+
+    for fold, (train_idx, val_idx) in enumerate(splits.split(np.arange(len(x_cell_train)))):
+        print('Fold {}'.format(fold + 1))
+
+        train_sampler = SubsetRandomSampler(train_idx)
+        test_sampler = SubsetRandomSampler(val_idx)
+        model = DeepDRA(cell_sizes, drug_sizes, ae_latent_dim, ae_latent_dim)
+        # Convert your training data to PyTorch tensors
+        x_cell_train_tensor = torch.Tensor(x_cell_train.values)
+        x_drug_train_tensor = torch.Tensor(x_drug_train.values)
+
+        y_train_tensor = torch.Tensor(y_train)
+        y_train_tensor = y_train_tensor.unsqueeze(1)
+
+        # Compute class weights
+        classes = [0, 1]  # Assuming binary classification
+        class_weights = torch.tensor(compute_class_weight(class_weight='balanced', classes=classes, y=y_train),
+                                     dtype=torch.float32)
+
+        # Create a TensorDataset with the input features and target labels
+        train_dataset = TensorDataset(x_cell_train_tensor, x_drug_train_tensor, y_train_tensor)
+
+        # Create the train_loader
+        train_loader = DataLoader(train_dataset, batch_size=batch_size, sampler=train_sampler)
+        # Train the model
+        train(model, train_loader,train_loader, num_epochs, class_weights)
+
+
+        # Create a TensorDataset with the input features and target labels
+        test_loader = DataLoader(train_dataset, batch_size=len(x_cell_train), sampler=test_sampler)
+
+        # Test the model
+        results = test(model, test_loader)
 
-def run(k, is_test=False):
+        # Step 10: Add results to the history dictionary
+        Evaluation.add_results(history, results)
+
+
+    return Evaluation.show_final_results(history)
+
+def run(k, is_test=False ):
     """
     Run the training and evaluation process k times.
 
@@ -119,6 +165,7 @@ def run(k, is_test=False):
                                                             screen_file_directory=BOTH_SCREENING_DATA_FOLDER,
                                                             sep="\t")
 
+
     # Step 3: Load test data if applicable
     if is_test:
         test_data, test_drug_screen = RawDataLoader.load_data(data_modalities=DATA_MODALITIES,
@@ -126,19 +173,22 @@ def run(k, is_test=False):
                                                               screen_file_directory=CCLE_SCREENING_DATA_FOLDER,
                                                               sep="\t")
         train_data, test_data = RawDataLoader.data_features_intersect(train_data, test_data)
-        X_cell_test, X_drug_test, y_test, cell_sizes, drug_sizes = RawDataLoader.prepare_input_data(test_data,
-                                                                                                    test_drug_screen)
+
 
     # Step 4: Prepare input data for training
-    X_cell_train, X_drug_train, y_train, cell_sizes, drug_sizes = RawDataLoader.prepare_input_data(train_data,
+    x_cell_train, x_drug_train, y_train, cell_sizes, drug_sizes = RawDataLoader.prepare_input_data(train_data,
                                                                                                    train_drug_screen)
 
+    if is_test:
+        x_cell_test, x_drug_test, y_test, cell_sizes, drug_sizes = RawDataLoader.prepare_input_data(test_data,
+                                                                                                    test_drug_screen)
+
     rus = RandomUnderSampler(sampling_strategy="majority", random_state=RANDOM_SEED)
-    dataset = pd.concat([X_cell_train, X_drug_train], axis=1)
-    dataset.index = X_cell_train.index
+    dataset = pd.concat([x_cell_train, x_drug_train], axis=1)
+    dataset.index = x_cell_train.index
     dataset, y_train = rus.fit_resample(dataset, y_train)
-    X_cell_train = dataset.iloc[:, :sum(cell_sizes)]
-    X_drug_train = dataset.iloc[:, sum(cell_sizes):]
+    x_cell_train = dataset.iloc[:, :sum(cell_sizes)]
+    x_drug_train = dataset.iloc[:, sum(cell_sizes):]
 
     # Step 5: Loop over k runs
     for i in range(k):
@@ -148,18 +198,21 @@ def run(k, is_test=False):
         if is_test:
 
             # Step 7: Train and evaluate the DeepDRA model on test data
-            results = train_DeepDRA(X_cell_train, X_cell_test, X_drug_train, X_drug_test, y_train, y_test, cell_sizes,
+            results = train_DeepDRA(x_cell_train, x_cell_test, x_drug_train, x_drug_test, y_train, y_test, cell_sizes,
                                     drug_sizes, device)
+
         else:
-            # Step 8: Split the data into training and validation sets
-            X_cell_train, X_cell_test, X_drug_train, X_drug_test, y_train, y_test = train_test_split(X_cell_train,
-                                                                                                     X_drug_train, y_train,
-                                                                                                     test_size=0.2,
-                                                                                                     random_state=RANDOM_SEED,
-                                                                                                     shuffle=True)
-            # Step 9: Train and evaluate the DeepDRA model on the split data
-            results = train_DeepDRA(X_cell_train, X_cell_test, X_drug_train, X_drug_test, y_train, y_test, cell_sizes,
-                                    drug_sizes, device)
+            # # Step 8: Split the data into training and validation sets
+            # X_cell_train, X_cell_test, X_drug_train, X_drug_test, y_train, y_test = train_test_split(X_cell_train,
+            #                                                                                          X_drug_train, y_train,
+            #                                                                                          test_size=0.2,
+            #                                                                                          random_state=RANDOM_SEED,
+            #                                                                                          shuffle=True)
+            # # Step 9: Train and evaluate the DeepDRA model on the split data
+            # results = train_DeepDRA(X_cell_train, X_cell_test, X_drug_train, X_drug_test, y_train, y_test, cell_sizes,
+            #                         drug_sizes, device)
+
+            results = cv_train(x_cell_train, x_drug_train, y_train, cell_sizes, drug_sizes, device, k=5)
 
         # Step 10: Add results to the history dictionary
         Evaluation.add_results(history, results)

mlp.py

@@ -27,7 +27,7 @@ class MLP(nn.Module):
             nn.Linear(input_dim, 128),
             nn.ReLU(inplace=True),
             nn.Linear(128, output_dim),
-            nn.Hardsigmoid(),
+            nn.Sigmoid(),
         )
 
     def forward(self, x):

utils.py

@@ -8,6 +8,7 @@ DRUG_DATA_FOLDER = os.path.join(DATA_FOLDER, 'drug_data')
 GDSC_RAW_DATA_FOLDER = os.path.join(DATA_FOLDER, 'GDSC_data')
 CCLE_RAW_DATA_FOLDER = os.path.join(DATA_FOLDER, 'CCLE_data')
 CTRP_RAW_DATA_FOLDER = os.path.join(DATA_FOLDER, 'CTRP_data')
+SIM_DATA_FOLDER = os.path.join(DATA_FOLDER, 'similarity_data')
 
 GDSC_SCREENING_DATA_FOLDER = os.path.join(GDSC_RAW_DATA_FOLDER, 'drug_screening_matrix_GDSC.tsv')
 CCLE_SCREENING_DATA_FOLDER = os.path.join(CCLE_RAW_DATA_FOLDER, 'drug_screening_matrix_ccle.tsv')
@@ -30,7 +31,7 @@ SIM_KERNEL = {'cell_CN': ('euclidean', 0.001), 'cell_exp': ('euclidean', 0.01),
 SAVE_MODEL = False  # Change it to True to save the trained model
 VARIATIONAL_AUTOENCODERS = False
 # DATA_MODALITIES=['cell_CN','cell_exp','cell_methy','cell_mut','drug_comp','drug_DT'] # Change this list to only consider specific data modalities
-DATA_MODALITIES = ['cell_CN','cell_exp','cell_mut', 'drug_desc']
+DATA_MODALITIES = ['cell_exp', 'drug_desc']
 RANDOM_SEED = 42  # Must be used wherever can be used