Avocado: a multi-scale deep tensor factorization method learns a latent representation of the human epigenome

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• The human epigenome contains vast amounts of data, with measurements for every basepair.
• Existing methods for analyzing epigenomic data can be computationally intensive and may not capture all relevant information.

Purpose of the Study:

• To develop a novel deep neural network tensor factorization method called Avocado.
• To compress large-scale human epigenomic data into an information-rich, dense representation.
• To improve the accuracy of epigenomic data imputation and enhance downstream genomics task performance.

Main Methods:

• Proposed Avocado, a deep neural network tensor factorization technique.
• Learned a dense, information-rich representation of human epigenomic data.
• Evaluated the imputation accuracy and machine learning model performance using the learned representation.

Main Results:

• Avocado successfully compresses epigenomic data into a dense representation.
• The learned representation enables more accurate epigenomic data imputation compared to previous methods.
• Machine learning models utilizing the Avocado representation outperform those trained on raw epigenomic data.
• Improved prediction accuracy for gene expression, promoter-enhancer interactions, replication timing, and 3D chromatin architecture.

Conclusions:

• Avocado provides an effective method for representing and utilizing complex human epigenomic data.
• The learned representation facilitates improved performance in diverse genomics applications.
• This approach has the potential to advance epigenomic data analysis and interpretation.