Identifying promoter sequence architectures via a chunking-based algorithm using non-negative matrix factorisation

Core promoters are stretches of DNA at the beginning of genes that contain information that facilitates the binding of transcription initiation complexes. Different functional subsets of genes have core promoters with distinct architectures and characteristic motifs. Some of these motifs inform the selection of transcription start sites (TSS). By discovering motifs with fixed distances from known TSS positions, we could in principle classify promoters into different functional groups. Due to the variability and overlap of architectures, promoter classification is a difficult task that requires new approaches. In this study, we present a new method based on non-negative matrix factorisation (NMF) and the associated software called seqArchR that clusters promoter sequences based on their motifs at near-fixed distances from a reference point, such as TSS. When combined with experimental data from CAGE, seqArchR can efficiently identify TSS-directing motifs, including known ones like TATA, DPE, and nucleosome positioning signal, as well as novel lineage-specific motifs and the function of genes associated with them. By using seqArchR on developmental time courses, we reveal how relative use of promoter architectures changes over time with stage-specific expression. seqArchR is a powerful tool for initial genome-wide classification and functional characterisation of promoters. Its use cases are more general: it can also be used to discover any motifs at near-fixed distances from a reference point, even if they are present in only a small subset of sequences.