Interpretation of deep learning in genomics and epigenomics

Area of Science:

• Genomics and Epigenomics
• Bioinformatics
• Machine Learning

Background:

• Machine learning, particularly deep neural networks (DNNs), is increasingly used for big data analysis in genomics and epigenomics.
• While DNNs offer high accuracy in predictions, their 'black-box' nature hinders understanding of underlying biological mechanisms.
• Model interpretability is crucial for advancing genomic and epigenomic research.

Purpose of the Study:

• To review current developments in deep neural network (DNN) interpretation methods.
• To focus on the application of these interpretation methods in genomics and epigenomics research.
• To discuss biological discoveries enabled by DNN interpretation and the associated advantages and limitations.

Main Methods:

• Review of state-of-the-art DNN interpretation methods from machine learning.
• Summarization of DNN interpretation applications in recent genomics and epigenomics studies.
• Analysis of interpretation methods for sequence motif identification, genetic variations, gene expression, chromatin interactions, and non-coding RNAs.

Main Results:

• Identified and described various DNN interpretation techniques relevant to biological data.
• Cataloged applications of DNN interpretation in key areas of genomics and epigenomics.
• Presented biological insights derived from applying these interpretation methods.

Conclusions:

• DNN interpretation methods are vital for unlocking biological discoveries in genomics and epigenomics.
• Current approaches offer significant advantages but also have limitations that need addressing.
• Further development in DNN interpretability will enhance our understanding of molecular and cellular mechanisms.