Abstract
Large genomes such as the human genome are pervasively transcribed yet encode relatively few unambiguously functional elements. This has led to debate over whether pervasive transcription is indicative of large suites of uncharacterized functional elements or is simply background noise. Here we used a deep-learning model to estimate background transcription in the human genome as a way of distinguishing between these two hypotheses. We applied the model to randomised (reversed or shuffled) versions of the human genome and found that transcription is predicted to be sparse across all randomisation methods, initiating with at least four-fold lower frequencies than in the native human genome. This relatively low level of background transcription from the human genome suggests that most transcription is not a consequence of background noise, thus it requires other explanations. We find that randomizing only interspersed repeats in human genome has little impact on predicted transcription, suggesting that transcription of mobile elements does not explain the excess transcription in the human genome. Instead, most transcriptional events may derive from functional noncoding RNA transcripts, some general requirement for extensive transcription initiation/elongation, and/or mutational biases leading to the frequent appearance of transcription initiation sites by chance.