ENNGene

ENNGene applies deep learning to train convolutional neural network (CNN) and hybrid CNN-recurrent neural network (RNN) models on genomic data for prediction and interpretation of sequence-associated features.

Key Features:

• Flexible Model Architectures: Design and customization of CNN and hybrid CNN-RNN architectures with various layer types and configurations.
• Multi-Input Support: Supports multiple input branches including sequence data, evolutionary conservation scores, and secondary structure, with input specified via genomic coordinates.
• Preprocessing: Automates necessary preprocessing steps for genomic inputs.
• Evaluation Metrics Export: Exports multi-class Receiver Operating Characteristic (ROC) curves, precision-recall plots, and TensorBoard log files for performance tracking.
• Interpretation via Integrated Gradients: Implements Integrated Gradients to generate graphical attributions highlighting the contribution of each input position.

Scientific Applications:

• Genomic pattern discovery: Detect complex patterns in large genomic datasets using CNN and RNN-based deep learning models.
• Integrative feature modeling: Combine evolutionary conservation scores and secondary structure with sequence data to improve predictive performance.
• RNA-binding protein analysis: Applied to the RBP24 dataset, models trained with ENNGene achieved state-of-the-art results, improving performance on more than half of the proteins analyzed.

Methodology:

Training of custom CNN and hybrid CNN-RNN models on genomic data with multiple input branches (sequence, evolutionary conservation scores, secondary structure) accepting genomic coordinates; automated preprocessing; export of multi-class ROC curves, precision-recall plots and TensorBoard log files; and use of Integrated Gradients for attribution.