DFAST: Automated Bacterial Genome Annotation

DFAST accepts a FASTA-formatted file as a minimum required input, and users can customize parameters, tools and reference databases by providing command line options or defining an original configuration file (see Supplementary Notes for more details). The workflow is mainly composed of two annotation phases, i.e. structural annotation for predicting biological features such as CDSs, RNAs and CRISPRs, and functional annotation for inferring protein functions of predicted CDSs. Figure 1 shows a schematic depiction of the pipeline. Each annotation process is implemented as a module with common interfaces, allowing both flexible annotation workflows and extensions for new functions in the future.
In the default configuration, functional annotation will be processed in the following order:

Orthologous assignment (optional) All-against-all pairwise protein alignments are conducted between a query and each reference genome. Orthologous genes are identified based on a Reciprocal-Best-Hit approach. It also conducts self-to-self alignments within a query genome, in which genes scoring higher than their corresponding orthologs are considered in-paralogs and assigned with the same protein function. This process is effective in transferring annotations from closely related organisms and in reducing running time.

Homology search against the default reference database DFAST uses GHOSTX as a default aligner, which runs tens to hundred times faster than BLASTP with similar levels of sensitivity where E-values are less than 10⁻⁶ (Suzuki et al., 2014 (link)). Users can also choose BLASTP. For accurate annotation, we constructed a reference database from 124 well-curated prokaryotic genomes from public databases. See Supplementary Data for the breakdown of the database.

Pseudogene detection CDSs and their flanking regions are re-aligned to their subject protein sequences using LAST, which allows frameshift alignment (Kiełbasa et al., 2011 (link)). When stop codons or frameshifts are found in the flanking regions, the query is marked as a possible pseudogene. This also detects translation exceptions such as selenocysteine and pyrrolysine.

Profile HMM database search against TIGRFAM (Haft et al., 2013 (link)) It uses hmmscan of the HMMer software package.

Assignment of COG functional categories RPS-BLAST and the rpsbproc utility are used to search against the Clusters of Orthologous Groups (COG) database provided by the NCBI Conserved Domain Database (Marchler-Bauer et al., 2017 (link)).

DFAST output files include INSDC submission files as well as standard GFF3, GenBank and FASTA files. For GenBank submission, two input files for the tbl2asn program are generated, a feature table (.tbl) and a sequence file (.fsa). For DDBJ submission, DFAST generates submission files required for DDBJ Mass Submission System (MSS) (Mashima et al., 2017 (link)). In particular, if additional metadata such as contact and reference information are supplied, it can generate fully qualified files that are ready for submission to MSS.
While the workflow described above is fully customizable in the stand-alone version, only limited features are currently available in the web version, e.g. orthologous assignment is not available. As a merit of the web version, users can curate the assigned protein names by using an on-line annotation editor with an easy access to the NCBI BLAST web service. We also offer optional databases for specific organism groups (Escherichia coli, lactic acid bacteria, bifidobacteria and cyanobacteria). They are downloadable from our web site and can be used in the stand-alone version. We are updating reference databases to cover more diverse organisms.