Deep learning for genomics using Janggu.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Deep learning shows promise for understanding biological processes, but existing tools are often specific to fixed datasets or model architectures.
• This limitation hinders broad application and requires significant repetitive programming for new genomics tasks.

Purpose of the Study:

• To introduce Janggu, a Python library designed to facilitate deep learning applications in genomics.
• To provide a unified and flexible framework for data acquisition, pre-processing, and model evaluation in genomics research.
• To reduce programming overhead and enable computational biologists to rapidly assess biological hypotheses.

Main Methods:

• Janggu provides specialized dataset objects with a NumPy-like interface, ensuring compatibility with deep learning libraries like Keras and PyTorch.
• The library includes utilities for visualizing predictions as genomic tracks or exporting to bigWig format.
• Functionality is demonstrated through diverse genomics applications, including transcription factor binding site prediction and chromatin effect modeling.

Main Results:

• Janggu streamlines the evaluation of different model topologies for predicting transcription factor binding sites (e.g., JunD).
• The framework successfully demonstrates published models for predicting chromatin effects.
• Promoter usage prediction (CAGE) is improved by incorporating high-order sequence features, a novel capability of Janggu.

Conclusions:

• Janggu significantly reduces repetitive programming efforts in deep learning for genomics.
• The library's flexible framework and reusable components accelerate the assessment of biological hypotheses.
• Janggu empowers computational biologists to more efficiently apply deep learning to complex genomics problems.