DNA Methylation Can be Used to Predict Whether Genes Express Transcripts, Proteins, or Both

Genes exhibit a bimodal distribution based on mRNA abundance levels; they are in either a high abundance or low abundance population. The purpose of these two populations is unclear. No mechanisms for this bifurcation of transcript abundance have been described. We previously reported that proteins in maize come almost exclusively from genes in the high-abundance mRNA population, explaining at least partially why this population is enriched for functional genes. To explore potential mechanisms for these expression patterns, we used the random forest machine learning classifier trained on our published gene expression atlas data (mRNA and protein) and on published whole-genome bisulfite sequencing data. Based solely on DNA methylation patterns, the trained classifier predicts with high accuracy which genes are in the high abundance mRNA population (ROC AUC 0.97). The classifier predicts with even greater accuracy which genes express proteins (ROC AUC 0.99). Much of the predictive power comes from exon methylation in the CpG and CHG contexts. The classifier was used to derive a set of 33,180 maize genes that are predicted to express proteins. The predicted set was compared to the maize filtered gene set and shown to share 29,833 genes, to exclude 14,202 genes, and to possess 3,345 genes not in the filtered set. The predicted set was more enriched than the filtered set for classical genes, syntenic genes, and named genes. The classifier derived 43,163 genes predicted to express high-abundance mRNA; 34,211 of these genes overlapped with the 44,035 genes of the filtered set.