Anomaly detection in genomic catalogues using unsupervised multi-view autoencoders.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate identification of Transcriptional Regulator binding locations is crucial for genomic region analysis, including Cis Regulatory Elements (CREs).
• Next-generation sequencing (NGS) methods like ChIP-Seq can suffer from data anomalies and biases, complicating unsupervised detection.
• Supervised methods are often required to address these challenges in ChIP-Seq data.

Purpose of the Study:

• To develop a novel deep learning method for identifying atypical peaks in ChIP-Seq data.
• To leverage correlations across multiple experimental series to flag unreliable transcriptional regulator binding sites.
• To improve the accuracy and reliability of genomic region analysis.

Main Methods:

• Utilized deep learning with multiview convolutions for lossy compression of genomic region representations.
• Developed a method to identify groups of correlating experimental series and evaluate Cis Regulatory Elements (CREs) based on group completeness.
• Applied the method to the ReMap database's extensive ChIP-seq data.

Main Results:

• Demonstrated successful identification of correlating experimental groups using artificial data.
• Showed that peaks lacking known biological correlators were identified as less confirmed in real ChIP-seq data.
• Proposed normalization techniques to aid in the interpretation of deep learning models.

Conclusions:

• The developed approach effectively detects ChIP-Seq peaks with less corroboration than average.
• The method can be extended to similar biological data analysis problems and aids in identifying correlation groups.
• The implementation is available as an open-source tool named atyPeak.