DNA-Promoter-Classifier

A PyTorch-based deep learning classifier for DNA promoter sequence detection using the UCI Promoter Gene Sequences dataset. Implemented using a 1D Convolutional Neural Network (CNN).

Project Overview

Features

• Loads and parses the UCI Promoter dataset
• One-hot encodes DNA sequences (A, T, G, C)
• Configurable 1D CNN for binary classification

Dataset

• Source: UCI Promoter Gene Sequences Dataset
• Size: 106 sequences (57 nucleotides each)
• Labels: '+' for promoter, '-' for non-promoter

Getting Started

Installation

git clone https://github.com/Thanos-png/DNA-Promoter-Classifier.git
cd DNA-Promoter-Classifier

Requirements

Install dependencies:

pip install -r requirements.txt

Ensure GPU is Available

python -c "import torch; print(torch.cuda.is_available())"

If True your GPU is ready.If False check your CUDA installation or the CPU will be used automatically.

Running the Project

Train the PromoterCNN Model

cd src/
python train.py

This will:

• You can choose if you want the trainning to be done with Cross-Validation python train.py --method cv or using the default train/val/test split method python train.py.
• First try to load preprocessed data from ../data/processed/.
• If not found, automatically:
  • Load the raw data from the specified path
  • Preprocess it (one-hot encoding, train/test split)
  • Save the processed data for future use
  • Return the DataLoaders for training
• Loads processed data (DataLoaders).
• Instantiates PromoterCNN and Adam optimizer.
• Runs epochs of forward and backward passes with gradient updates.
• Logs training & validation loss/accuracy.
• Save the model and the preprocessed data.

Hyperparameters used:

Parameter	Value	Description
epochs	20	Number of epochs
batch_size	16	Number of samples per batch
lr	1e-3	Learning rate
val_size	0.25	Validation size for train/val/test split
test_size	0.2	Test size for train/val/test split
method	split	Training method (split/cv)
k_folds	5	Number of folds for cross-validation

Expected output:

Training with Train/Val/Test Split
Using device: cuda
Training samples: 63
Validation samples: 21
Test samples: 22
Total samples: 106

Starting training for 20 epochs...
Epoch 1/20 - Train Loss: 0.6921 - Val Loss: 0.6780 - Val Acc: 0.5238
New best model saved! Accuracy: 0.5238
Epoch 2/20 - Train Loss: 0.6528 - Val Loss: 0.6701 - Val Acc: 0.5714
New best model saved! Accuracy: 0.5714
Epoch 3/20 - Train Loss: 0.6030 - Val Loss: 0.6455 - Val Acc: 0.7619
New best model saved! Accuracy: 0.7619
Epoch 4/20 - Train Loss: 0.5755 - Val Loss: 0.6222 - Val Acc: 0.7619
Epoch 5/20 - Train Loss: 0.5476 - Val Loss: 0.6024 - Val Acc: 0.7143
Epoch 6/20 - Train Loss: 0.5238 - Val Loss: 0.5835 - Val Acc: 0.8095
New best model saved! Accuracy: 0.8095
Epoch 7/20 - Train Loss: 0.4392 - Val Loss: 0.5630 - Val Acc: 0.7619
Epoch 8/20 - Train Loss: 0.4716 - Val Loss: 0.5417 - Val Acc: 0.8571
New best model saved! Accuracy: 0.8571
Epoch 9/20 - Train Loss: 0.3845 - Val Loss: 0.5203 - Val Acc: 0.6667
Epoch 10/20 - Train Loss: 0.3189 - Val Loss: 0.4981 - Val Acc: 0.7619
Epoch 11/20 - Train Loss: 0.2973 - Val Loss: 0.4771 - Val Acc: 0.7619
Epoch 12/20 - Train Loss: 0.2997 - Val Loss: 0.4538 - Val Acc: 0.8571
Epoch 13/20 - Train Loss: 0.2287 - Val Loss: 0.4377 - Val Acc: 0.8571
Epoch 14/20 - Train Loss: 0.2143 - Val Loss: 0.4283 - Val Acc: 0.7619
Epoch 15/20 - Train Loss: 0.1571 - Val Loss: 0.4202 - Val Acc: 0.7143
Epoch 16/20 - Train Loss: 0.1462 - Val Loss: 0.4052 - Val Acc: 0.7143
Epoch 17/20 - Train Loss: 0.1459 - Val Loss: 0.3834 - Val Acc: 0.8095
Epoch 18/20 - Train Loss: 0.1119 - Val Loss: 0.3832 - Val Acc: 0.7619
Epoch 19/20 - Train Loss: 0.1301 - Val Loss: 0.3732 - Val Acc: 0.8095
Epoch 20/20 - Train Loss: 0.0833 - Val Loss: 0.3718 - Val Acc: 0.7619

Training completed! Best Accuracy: 0.8571
Best model saved as: ../results/models/promoter_cnn.pth

Test the PromoterCNN Model

cd src/
python test.py

This will:

• You can choose to test either the train/val/test split model python test.py or the cross-validation model python test.py --method cv.
• Load the saved CNN model from ../results/models/.
• Load the processed test data or create it from raw data if not found.
• Evaluate the model on the test set using BCEWithLogitsLoss.
• Generate and save comprehensive visualization plots:
  • Confusion matrix heatmap with percentages
  • ROC curve with AUC score
  • Precision-Recall curve with AUC score
  • Prediction score distribution by class
  • Metrics summary bar chart
• Display detailed classification metrics:
  • Test loss and accuracy
  • Precision, recall, F1-score for each class
  • Confusion matrix breakdown
  • ROC-AUC and PR-AUC scores
• Save all plots to ../results/plots/ directory.

Expected output:

Testing Train/Val/Test split model
Using device: cuda
Test samples: 22
Loaded model from: ../results/models/promoter_cnn.pth

==================
 GENERATING PLOTS
==================
Confusion matrix saved: ../results/plots/confusion_matrix_train-val-test_split.png
ROC curve saved: ../results/plots/roc_curve_train-val-test_split.png
Precision-Recall curve saved: ../results/plots/precision_recall_train-val-test_split.png
Prediction distribution saved: ../results/plots/prediction_distribution_train-val-test_split.png
Metrics summary saved: ../results/plots/metrics_summary_train-val-test_split.png

===============================
 FINAL TEST EVALUATION RESULTS
===============================
Test Loss: 0.4982
Test Accuracy: 0.9545

Classification Report:
              precision    recall  f1-score   support

Non-Promoter       1.00      0.91      0.95        11
    Promoter       0.92      1.00      0.96        11

    accuracy                           0.95        22
   macro avg       0.96      0.95      0.95        22
weighted avg       0.96      0.95      0.95        22


Confusion Matrix:
True Negatives: 10, False Positives: 1
False Negatives: 0, True Positives: 11

Additional Metrics:
Accuracy: 0.9545
Precision: 0.9167
Recall: 1.0000
F1-Score: 0.9565
ROC-AUC: 0.9835
PR-AUC: 0.9834

Final Test Accuracy: 0.9545

Key Implementations

Data Preprocessing

• Read promoters.data, parse labels (+ → 1, - → 0) and sequences.
• One-hot encode each nucleotide into a 4-dim vector.
• Split into train/val/test sets (60/20/20 stratified).
• Wrap data in a PyTorch Dataset and create DataLoaders.

CNN Model

A 1D Convolutional Neural Network (CNN) implemented with PyTorch:

Architecture:

• Input: One-hot encoded DNA sequences (57×4)
• Conv1D Layer: 4→32 channels, kernel_size=5, ReLU activation
• MaxPool1D: kernel_size=2 for downsampling
• Fully Connected: 832→64 neurons, ReLU activation
• Dropout: 50% regularization
• Output Layer: 64→1 neuron (raw logits)
• Loss Function: BCEWithLogitsLoss (combines sigmoid + BCE)

Parameters: ~27K trainable parameters

Results & Analysis

Test Performance

• Test Loss: 0.4982
• Test Accuracy: 0.9545

Classification Report

              precision    recall  f1-score   support

Non-Promoter       1.00      0.91      0.95        11
    Promoter       0.92      1.00      0.96        11

    accuracy                           0.95        22
   macro avg       0.96      0.95      0.95        22
weighted avg       0.96      0.95      0.95        22

Confusion Matrix

True Negatives: 10
False Positives: 1
False Negatives: 0
True Positives: 11

Additional Metrics

• Accuracy: 0.9545
• Precision: 0.9167
• Recall: 1.0000
• F1-Score: 0.9565
• ROC-AUC: 0.9835
• PR-AUC: 0.9834

Model Variance

Due to the small dataset size (106 samples), model performance can vary between training runs:

• Typical Range: 70-95% test accuracy
• Small Test Set: Only 22 samples, so 1-2 misclassifications significantly impact accuracy

Contact

For questions or feedback, feel free to reach me out:

• Email: [email protected]