Accurate annotation of plant genomes remains complex due to the presence of many pseudogenes arising from whole-genome duplication-generated redundancy or the capture and movement of gene fragments by transposable elements. Machine learning on genome-wide epigenetic marks, informed by transcriptomic and proteomic training data, could be used to improve annotations through classification of all putative protein-coding genes as either constitutively silent or able to be expressed. Expressed genes were subclassified as able to express both mRNAs and proteins or only RNAs, and CG gene body methylation was associated only with the former subclass. More than 60,000 protein-coding genes have been annotated in the reference genome of maize inbred B73. About two-thirds of these genes are transcribed and are designated the filtered gene set (FGS). Classification of genes by our trained random forest algorithm was accurate and relied only on histone modifications or DNA methylation patterns within the gene body; promoter methylation was unimportant. Other inbred lines are known to transcribe significantly different sets of genes, indicating that the FGS is specific to B73. We accurately classified the sets of transcribed genes in additional inbred lines, arising from inbred-specific DNA methylation patterns. This approach highlights the potential of using chromatin information to improve annotations of functional genes.
Keywords: epigenomics; genome annotation; machine learning; maize; proteomics.
Copyright © 2019 the Author(s). Published by PNAS.