Defining Intergenic Transcribed Regions As Junk DNA or Novel Genes Using a Machine Learning Approach

With advances in transcript profiling, the presence of transcriptional activities in intergenic regions has been well established. Intergenic transcripts have been readily identified in many model systems, but little is known about the prevalence of intergenic transcription across species with a wide range of genome sizes. More importantly, whether intergenic expression represents transcriptional noise or the activity of novel genes remains unclear. We identified intergenic transcribed fragments in 15 diverse plant species with 17-fold differences in genome sizes and found that hundreds to thousands of intergenic transcribed fragments can be identified in any plant species. However only 0.4-3.4% of all transcription originates from intergenic regions and the proportion of intergenic expression does not scale with the amount of intergenic space, indicating that intergenic expression does not increase proportionately with genome size. To assess the functionality of intergenic transcribed regions, we first built machine learning models in A. thaliana that can accurately classify functional (phenotype genes) and non-functional (pseudogene and random unexpressed intergenic) sequences using a wide range of biochemical, evolutionary, and sequence structural features. Next, by applying the models to intergenic regions, we found that 10-15% of intergenic transcribed regions exhibit similarity to phenotype genes and thus are likely functional. However, the rest were more similar to nonfunctional sequences and are likely transcriptional noise unless falsified with further experimental testing.