Janggu

Janggu provides a Python framework for applying deep learning to genomics, enabling unified data acquisition and preprocessing, integration with Keras and PyTorch, prediction visualization (including bigWig export), evaluation of Keras models, and incorporation of high-order sequence features for sequence- and epigenome-based analyses.

Key Features:

• Unified Data Acquisition and Pre-processing Framework: Specialized dataset objects provide a unified approach to genomics data acquisition and preprocessing with reusable workflow components.
• Compatibility with Deep Learning Libraries: A numpy-like interface enables integration with Keras and PyTorch for model training and inference.
• Visualization Tools: Utilities allow visualizing predictions as genomic tracks and exporting predictions to the bigWig file format.
• Model Evaluation: Tools are provided for evaluating Keras-based model performance.
• High-order Sequence Features: Support for inclusion of high-order sequence features to improve predictive accuracy on complex genomic signals.

Scientific Applications:

• Transcription Factor Binding Prediction: Evaluation of model topologies for predicting binding sites of the transcription factor JunD.
• Chromatin Effect Prediction: Application of published models to predict chromatin effects from genomic and epigenomic data.
• Promoter Usage Prediction: Integration of DNase hypersensitivity, histone modifications, and DNA sequence features to predict promoter usage measured by CAGE.

Methodology:

Uses specialized dataset objects for data acquisition and preprocessing, a numpy-like interface for integration with Keras and PyTorch, prediction visualization via genomic tracks and bigWig export, evaluation tools for Keras-based models, and support for high-order sequence features.