In a conventional view of the prokaryotic genome organization, promoters precede operons and ribosome binding sites (RBSs) with Shine-Dalgarno consensus precede genes. However, recent experimental research suggesting a more diverse view motivated us to develop an algorithm with improved gene-finding accuracy. We describe GeneMarkS-2, an ab initio algorithm that uses a model derived by self-training for finding species-specific (native) genes, along with an array of precomputed "heuristic" models designed to identify harder-to-detect genes (likely horizontally transferred). Importantly, we designed GeneMarkS-2 to identify several types of distinct sequence patterns (signals) involved in gene expression control, among them the patterns characteristic for leaderless transcription as well as noncanonical RBS patterns. To assess the accuracy of GeneMarkS-2, we used genes validated by COG (Clusters of Orthologous Groups) annotation, proteomics experiments, and N-terminal protein sequencing. We observed that GeneMarkS-2 performed better on average in all accuracy measures when compared with the current state-of-the-art gene prediction tools. Furthermore, the screening of ∼5000 representative prokaryotic genomes made by GeneMarkS-2 predicted frequent leaderless transcription in both archaea and bacteria. We also observed that the RBS sites in some species with leadered transcription did not necessarily exhibit the Shine-Dalgarno consensus. The modeling of different types of sequence motifs regulating gene expression prompted a division of prokaryotic genomes into five categories with distinct sequence patterns around the gene starts.

Abstract. We revisit the polytope method for factoring sparse bivariate polynomials over finite fields, and address the bottleneck arising from solving the Hensel lifting equations using the sparse distributed polynomial representation. We revise the analysis when polynomials are represented as such, which reveals how performing the polynomial multiplications and ensuing additions in separate (serialised) phases causes the Hensel lifting phase to suffer from poor work, space, and I/O complexity, and hinges on the size of the intermediary output, as size is defined in the sparse distributed representation. We propose to overlap all polynomial arithmetic in one Hensel lifting step using a MAX priority queue. The overlapping approach adapts not only to the growth in the degree of the input polynomial but also to irregularities in the sparsity of intermediary output. It also results in evading expression swell and reducing the overall work and space complexity by an order of magnitude. When the priority queue is implemented as a cache-oblivious data structure, the overlapping approach achieves an order of magnitude improvement in I/O over the serialised approach, even when the latter is using cache efficient structures to assist in polynomial multiplications and additions. We present empirical results for the polytope method using a max-heap implementation of the global priority queue, which demonstrate extremely superior performance, and specifically against Magma, for sufficiently sparse input polynomials of very high degrees.

Accurate prediction of protein-coding genes in metagenomic contigs presents a well-known challenge. Particularly difficult is to identify short and incomplete genes as well as positions of translation initiation sites. It is frequently assumed that initiation of translation in prokaryotes is controlled by a ribosome binding site (RBS), a sequence with the Shine-Dalgarno (SD) consensus situated in the 5' UTR. However, ~30% of the 5,007 genomes, representing the RefSeq collection of prokaryotic genomes, have either non-SD RBS sequences or no RBS site due to physical absence of the 5' UTR (the case of leaderless transcription). Predictions of the gene 3' ends are much more accurate; still, errors could occur due to the use of incorrect genetic code. Hence, an effective gene finding algorithm would identify true genetic code in a process of the sequence analysis. In this work prediction of gene starts was improved by inferring the GC content dependent generating functions for RBS sequences as well as for promoter sequences involved in leaderless transcription. An additional feature of the algorithm was the ability to identify alternative genetic code defined by a reassignment of the TGA stop codon (the only stop codon reassignment type known in prokaryotes). It was demonstrated that MetaGeneMark-2 made more accurate gene predictions in metagenomic sequences than several existing state-of-the-art tools.

While computational gene finders for prokaryotic genomes have reached a high level of accuracy, there is room for improvement. GeneMarkS-2, a new ab initio algorithm, aims to improve prediction of species-specific (native) genes, as well as difficult-to-detect genes that differ in composition from the native genes. We introduce an array of pre-computed heuristic models that compete with the iteratively learned native model for the best fit within genomic neighborhoods that deviate in nucleotide composition from the genomic mainstream. Also, in the process of self-training, GeneMarkS-2 identifies distinct sequence patterns controlling transcription and translation. We assessed the accuracy of current state-of-the-art gene prediction tools along with GeneMarkS-2 on test sets of genes validated by proteomics experiments, by COG annotation, as well as by protein N-terminal sequencing. We observed that, on average, GeneMarkS-2 shows a higher precision in all accuracy measures. Screening of ~5,000 representative prokaryotic genomes reveals frequent leaderless transcription, not only in archaea where it was originally discovered, but in bacteria as well. Furthermore, species with prevalent leadered transcription do not necessarily use RBS sites with the Shine-Dalgarno consensus. The effort to distinguish leaderless and leadered transcription, depending on prevalence of one or the other, leads to classifying prokaryotic genomes into five groups with distinct sequence patterns around gene starts. Some of the observed patterns are apparently related to poorly characterized mechanisms of translation initiation.[Supplemental material is available for this article].

State-of-the-art algorithms of ab initio gene prediction for prokaryotic genomes were shown to be sufficiently accurate. A pair of algorithms would agree on predictions of gene 3′ends. Nonetheless, predictions of gene starts would not match for 15–25% of genes in a genome. This discrepancy is a serious issue that is difficult to be resolved due to the absence of sufficiently large sets of genes with experimentally verified starts. We have introduced StartLink that infers gene starts from conservation patterns revealed by multiple alignments of homologous nucleotide sequences. We also have introduced StartLink+ combining both ab initio and alignment-based methods. The ability of StartLink to predict the start of a given gene is restricted by the availability of homologs in a database. We observed that StartLink made predictions for 85% of genes per genome on average. The StartLink+ accuracy was shown to be 98–99% on the sets of genes with experimentally verified starts. In comparison with database annotations, we observed that the annotated gene starts deviated from the StartLink+ predictions for ∼5% of genes in AT-rich genomes and for 10–15% of genes in GC-rich genomes on average. The use of StartLink+ has a potential to significantly improve gene start annotation in genomic databases.