Nikolaos Gialitsis, George Giannakopoulos, Marina Athanasouli

Abstract:
The representation of DNA sequences has been an interesting topic of discussion for many years. Presently, given the usefulness of representations built upon embeddings for Natural Language Processing (NLP), there have been efforts to transfer such paradigms to the DNA world and related problems. In this paper, we study different DNA representations on the well-studied problem of Conserved Non-coding Elements (CNEs), trying to understand how well existing representations utilize the value of context, both in terms of local, near context, but also of long-distance interactions in genomic sequences. To this end, we apply a number of methods, including probabilistic models (LDA) and hybrid probabilistic-neural models (lda2vec) on appropriate datasets, compare the results to pre-existing methods and discuss the findings to better understand the value and challenges of different representations in the given domain.