I am training a RandomForestClassifier from sklearn.ensemble with the following code: adata = ad.read_h5ad(f'{data_dir}{ct}_clean_log1p_normalized.h5ad') adata = adata[:, adata.var.highly_variable] print(f'AnnData for {ct}: {adata}') # Extract feature matrix (X) and target vector (y) X = adata.X y = adata.obs['clinical_dx'] # Convert sparse matrix to dense for [insert reason] if issparse(X): X = X.toarray() # Encode the target variable for [insert reason] le = LabelEncoder() y_encoded = le.fit_transform(y) X_train, X_test, y_train, y_test = train_test_split(X, y_encoded, test_size=0.2, random_state=42) # Initialize the classifier rf_classifier = RandomForestClassifier(n_estimators=100, random_state=42) # Train the classifier rf_classifier.fit(X_train, y_train) # Validate on the test set y_pred_rf = rf_classifier.predict(X_test) # View validation report validation_report = classification_report(y_test, y_pred_rf, target_names=le.classes_) print(validation_report) with open(f'{rfc_dir}validation_report.txt', "w") as report_file: report_file.write(validation_report) # Generate the confusion matrix cm = confusion_matrix(y_test, y_pred_rf, labels=rf_classifier.classes_) # Make percentage cm_normalized = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis] plt.figure(figsize=(10, 9)) sns.heatmap(cm_normalized, annot=True, fmt='.1%', cmap='Blues', xticklabels=le.inverse_transform(rf_classifier.classes_), yticklabels=le.inverse_transform(rf_classifier.classes_)) plt.xlabel('Predicted Label') plt.ylabel('True Label') plt.title('Confusion Matrix (Random Forest)') plt.savefig(f"{rfc_dir}confusion_matrix", bbox_inches='tight') # Get feature importances feature_importances_rf = rf_classifier.feature_importances_ number_of_features = 200 # Create a DataFrame for better visualization feature_importance_rf_df = pd.DataFrame({ 'Ensembl': adata.var_names, 'Importance': feature_importances_rf }) top_features = feature_importance_rf_df.sort_values( by='Importance', ascending=False ).head(number_of_features) top_features.to_csv(f'{rfc_dir}markers.csv', index=False) Unfortunately, I cannot share the data since it is HIPAA-protected. For some reason, regardless of what celltype (there is a different anndata for each of 8 celltype) I train a classifier on, the importance scores in the CSV file are all incredibly low. The highest importance score in any case is around 0.01. Is this a red flag or just something to do with my datasets? Has anyone experienced this before? Thanks Continue reading...
