Human DNA/RNA motif mining using deep-learning methods: a scoping review

The current study aims to develop robust contextual knowledge of deep-learning methodology for DNA/RNA motif sequence identification and recognition of correct transcription factor-binding sites (TFBS) for gene regulatory mechanisms in humans. Knowledge of the exact sequence specificities of DNA- and RNA-binding to particular transcriptional factors (TF) seems to be an excellent strategy to develop unique deep-learning models for gene regulatory processes. But uncertainty in the sequence specificity of genomic sequences to a particular TFBS is a big issue. It may be possible to resolve this issue using deep-learning techniques, and thus, it will be helpful to gain generalizable domain knowledge of deep-learning architectures, which offers researchers to know better, their performance to select a unified computational approach for the discovery of a selective kind of motif pattern. This scoping review serves to synthesize evidence for DNA/RNA motif sequences binding with transcriptional factor sites using the PRISMA-ScR guidelines (Preferred Reporting Items for Systematic reviews and Meta-Analyses of Scoping Reviews) to better understand and further assessment of the scope of literature on DNA or RNA motif mining using deep-learning methods. A deep-learning architecture literature survey for DNA and RNA sequence specificity for human ChIP-seq (Chromatin Immuno-Precipitation sequence), DNase-seq (DNase hypersensitive site sequence), CLIP-seq (Cross Linking Immuno-Precipitation sequence), ATAC-seq (Assay for Transposase-Accessible Chromatin sequence), etc. datasets, common motif pattern, and their corresponding TF-DNA/RNA-binding site affinities are included in this study. Deep-learning (DL) models have been used to find selective motifs and have been demonstrated to be more reproducible than traditional methods. As per our literature survey, 33 DL models exist to detect DNA/RNA motifs that have varied framework designs and implementation styles. Through literature survey and PRISMA-ScR reporting guidelines, it is easy to analytically evaluate the performances of each DL model in terms of model size, automatic calibration ability, tool selection, and training set, and it has been found that the DESSO (DEep Sequence and Shape mOtif), DeepFinder, and DeepBind are the selective DL models that are appropriate to study the true biological relationship, especially concerning gene expression patterns and sequence analysis. This study concludes that the application of existing deep-learning methods in the field of motif discovery is the faster way to process complex data relevant to genomic sequences. Through the PRISMA-ScR reporting guidelines and literature survey analysis, more than 30 existing deep-learning models are compared, and it is concluded that complex DL models are preferred over simpler DL models in terms of performance and scalability evaluation. Selective selection of a DL model architecture can be made to understand the complex behavior of motifs and their associated regulatory mechanism at the gene level.