Explainable Transformer Models for Functional Genomics in Prokaryotes

Abstract

The annotation of transcription start sites with computational methods is an important and unsolved problem in genomics. In recent years, several novel experimental methodologies -named Cappable-seq, SMRT-Cappable-seq and SEnd-seq -have been introduced for the detection of transcription start sites and applied on E. coli. In this study, a comparison is made between these new methodologies and the curated transcription start site data set featured by RegulonDB. The analysis between these data sets is facilitated using deep learning techniques that cover both unsupervised and supervised learning, where we expand upon a framework that allows for interpretable deep learning in genomics. This study finds annotations of recent techniques to surpass the quality of annotations provided by Reg-ulonDB. Analysis of the transformer network trained for the detection of TSS in E. coli reveals its attention scores to pinpoint important promoter regions previously discussed in literature. Additionally, findings support the occurrence of a complex interaction between sense and antisense output probabilities, prevalent on key positions for interference of the transcription process.

Analysis of next- and third-generation RNA-Seq data reveals the structures of alternative transcription units in bacterial genomes

“…Several newly developed techniques have provided a comprehensive view of the E. coli transcriptome by identifying full-length primary transcripts (13)(14)(15)(16)(17). For example, SMRT-Cappable-seq (6) combines the isolation of the full-length bacterial primary transcriptome with PacBio SMRT (Single Molecule, Real-Time) sequencing (6), and simultaneous 5' and 3' end sequencing (SEnd-seq) (7) captures both transcription start sites (TSSs) and transcription termination sites (TTSs) via circularization of transcripts (17). Despite the great progress in experimental techniques, there are still some deficiencies.…”