Alex Huebner Feb 03, 2023
In the past ten years, there has been a huge increase in the number of metagenome-assembled genomes (MAGs) due to improvement in de novo assembly tools, such as MEGAHIT (Li et al. 2015) and metaSPAdes (Nurk et al. 2017), and the decreasing costs for Illumina short-reads sequencing. Many of the pipelines available for de novo assembly of short-read sequencing data and subsequent binning of the yielded contigs, e.g. nf-core/mag (Krakau et al. 2022) or ATLAS (Kieser et al. 2020), are using the software Prokka (Seemann 2014) to rapidly annotate the MAGs.
Due to its functionality, Prokka has become the de facto standard for this type of bacterial genome annotation. However, it has not been developed further in the last years and recently the author of Prokka, Torsten Seeman, recommended on Twitter to use Bakta (Schwengers et al. 2021) as a successor of Prokka.
In contrast to Prokka, Bakta aims to increase the ability to assign the newly annotated coding sequences to genes that are available in reference databases and to improve the export of the annotations, e.g. by using JSON files. The authors of Bakta performed a single genome (E. coli) benchmark between Bakta and Prokka, and could show that Bakta had a higher number of proteins with a known function at a similar run time.
To validate these results on a larger set of genomes and to evaluate the suitability of the Bakta’s reference databases, I ran Bakta on the 29 MAGs obtained from EMN001 that have been previously annotated using Prokka and compared the annotation results with each other.
In total, metaWRAP produced 29 MAGs that comprised 11,266 contigs. These contigs were provided as input to both Bakta and Prokka and we obtained 8,926 and 9,192 contigs with any type of annotations, respectively.
Figure 1: Comparison of the total number of annotations per MAG per method.
Overall, we observed that the number of annotations per MAG were very similar between the two methods across all MAGs. There were 8 MAGs for which Bakta obtained more annotations, 19 for which Prokka obtained more annotations, and 2 MAGs with the identical number of annotations. The biggest difference in the number of elements annotated was 166.
However, the type of annotations varied between the two methods (Table 1). Next to the types that were shared between the methods (CDS, tRNA, tmRNA, rRNA), Bakta additionally provided annotations for non-coding RNA (regions), small open reading frames (sorf), origin of replication (oriC), and CRISPR.
Table 1: The number of annotations per annotation type.
| Type | Bakta | Prokka |
|---|---|---|
| CDS | 39,575 | 39,999 |
| tRNA | 613 | 595 |
| tmRNA | 15 | 15 |
| rRNA | 15 | 13 |
| ncRNA-region | 190 | 0 |
| ncRNA | 76 | 0 |
| sorf | 1 | 0 |
| oriC | 1 | 0 |
| crispr | 9 | 0 |
The vast majority of annotations were coding sequences (Table 1). These coding sequences are of particular interest to the group because they are used as input into different screening tools to identify molecule classes, such as antimicrobial peptides, antibiotic resistance genes, or biosynthetic gene clusters. Therefore, we combined the annotations of both Bakta and Prokka and compared which coding sequences were identified by each program (Figure 2).
Figure 2: The fraction of annotations for coding sequences (CDS) that were identified by each method.
Across the 29 MAGs, 85.8% of the CDS were identified by both methods, while 6.3% and 7.8% were only found by Bakta and Prokka, respectively. When we look further on whether the detected coding sequences could be assigned to a known annotations in reference databases, such as COG, GO, KEGG, PFAM, RFAM, EC, or RefSeq (Table 2).
Table 2: The number of reference annotations for the coding sequences.
| detection | no. of CDS | both | no ref. annotation | only Bakta | only Prokka |
|---|---|---|---|---|---|
| both | 36,701 | 19,300 | 7,365 | 8,976 | 1,060 |
| only Bakta | 2,874 | 0 | 1,684 | 1,190 | 0 |
| only Prokka | 3,297 | 0 | 3,297 | 0 | 0 |
From the coding sequences that could be detected by either program, more than half of the CDS could be assigned to a known annotation in a reference database and only about 20% could not be identified by either program. Another 25% could only be linked to a reference by Bakta, while this was only true for about 3% of the CDS for Prokka. This superiority of Bakta with respect of being able to link CDS to annotations in reference databases was also apparent when looking at the CDS that were only detected by one of the two programs. For Prokka, all additionally found CDS were hypothetical proteins with no link to a reference annotation. In contrast, Bakta could link more than half of the additional CDS to a reference.
Next to coding sequences, there was also the annotation of RNA sequences that is shared between Bakta and Prokka. There were three types of RNA sequences that were detected in both programs: tRNAs, tmRNAs, and rRNAs (Table 1). Since the detection of tmRNAs did not differ between the programs, we will investigate whether there were annotation differences between the other two categories.
Table 3: Comparison of the overlap of tRNA and rRNA genes between Bakta and Prokka.
| RNA type | both | only Bakta | only Prokka |
|---|---|---|---|
| tRNA | 428 | 185 | 167 |
| rRNA | 1 | 14 | 12 |
For tRNA genes, half of the annotations were shared between Prokka and Bakta (Table 3), but there were about 25% of the tRNA annotations that were only found by either program. On closer inspection, we could identified that from these tRNA detected only by one program 121 tRNA genes were identified to have a match to the annotation from the other program that fell within three bases (Table 4) This was different for rRNA genes that were almost never shared between the programs (Table 3) and only a small number of the annotations that were only detected by one program fell within close distance of each other (Table 4).
Table 4: The number of RNA annotations within less than 10 bp from each other when using Bakta and Prokka.
| distance btw. annotation [bp] | rRNA | tRNA |
|---|---|---|
| 1 | 0 | 108 |
| 2 | 0 | 11 |
| 3 | 2 | 2 |
| 4 | 1 | 0 |
| 5 | 2 | 0 |
Prokka (Seemann 2014) and Bakta (Schwengers et al. 2021) are two programs that combine a number of specialised tools for the annotation of microbial sequences. While both programs rely on the prediction of gene sequences by Prodigal (Hyatt et al. 2010), the individual tools differ slightly between them. Although both programs have a very similar number of predicted genetic elements per MAG of EMN001 and many of these elements were detected by both programs, Bakta was able to link to a known annotation from a reference database to the predicted sequences more often. All coding sequences that were only predicted by Prokka were sequences without any match in the reference database, while Bakta detected many sequences that were missed by Prokka for which it also found a match in the reference database. The prediction of tRNAs was very comparable between the programs, while rRNAs differed strongly.
As the author of Prokka, Torsten Seeman, wrote on Twitter, it seems that Bakta is the future way to go. Its main advantage that it is able assign many more genetic elements to reference database and report these results in a less convoluted way. Furthermore, the expensive conversion of the results from the GFF file format to the GenBank file format, which takes up to a few hours when running Prokka on all contigs of a sample, is optional in Bakta and will save a lot of time.
However, Bakta does not seem to be much more sensitive in detecting and annotating a higher number of genetic elements compared to Prokka. Therefore, it will be necessary to run additional sequence search tools, e.g. based on Hidden Markov Models, with specialised databases to detect all DNA sequence classes in the samples.
Hyatt, Doug, Gwo-Liang Chen, Philip F. LoCascio, Miriam L. Land, Frank W. Larimer, and Loren J. Hauser. 2010. “Prodigal: Prokaryotic Gene Recognition and Translation Initiation Site Identification.” BMC Bioinformatics 11 (1): 119. https://doi.org/10.1186/1471-2105-11-119.
Kieser, Silas, Joseph Brown, Evgeny M. Zdobnov, Mirko Trajkovski, and Lee Ann McCue. 2020. “ATLAS: A Snakemake Workflow for Assembly, Annotation, and Genomic Binning of Metagenome Sequence Data.” BMC Bioinformatics 21 (1): 257. https://doi.org/10.1186/s12859-020-03585-4.
Krakau, Sabrina, Daniel Straub, Hadrien Gourlé, Gisela Gabernet, and Sven Nahnsen. 2022. “Nf-Core/Mag: A Best-Practice Pipeline for Metagenome Hybrid Assembly and Binning.” NAR Genomics and Bioinformatics 4 (1): lqac007. https://doi.org/10.1093/nargab/lqac007.
Li, Dinghua, Chi-Man Liu, Ruibang Luo, Kunihiko Sadakane, and Tak-Wah Lam. 2015. “MEGAHIT: An Ultra-Fast Single-Node Solution for Large and Complex Metagenomics Assembly via Succinct de Bruijn Graph.” Bioinformatics 31 (10): 1674–76. https://doi.org/10.1093/bioinformatics/btv033.
Nurk, Sergey, Dmitry Meleshko, Anton Korobeynikov, and Pavel A. Pevzner. 2017. “metaSPAdes: A New Versatile Metagenomic Assembler.” Genome Research 27 (5): 824–34. https://doi.org/10.1101/gr.213959.116.
Schwengers, Oliver, Lukas Jelonek, Marius Alfred Dieckmann, Sebastian Beyvers, Jochen Blom, and AlexanderYR 2021 Goesmann. 2021. “Bakta: Rapid and Standardized Annotation of Bacterial Genomes via Alignment-Free Sequence Identification.” Microbial Genomics 7 (11): 000685. https://doi.org/10.1099/mgen.0.000685.
Seemann, Torsten. 2014. “Prokka: Rapid Prokaryotic Genome Annotation.” Bioinformatics 30 (14): 2068–69. https://doi.org/10.1093/bioinformatics/btu153.