The pan-genome describes which orthologs are present in a group of related genomes. The core genome is the orthologs found in all genomes, the dispensable genome is the orthologs found in some of the species, while the unique genome is specific to an individual. Differences between genes in each of these groups will identify the possible differences in lifestyle between strains encoded in the genome. The size of the pan-genome also suggests the degree of the variation between the species/strains.

To estimate the pan-genome for P. fluorescens I followed the methods used by Hogg et. al which are described in the supplementary material as being suitable for pan-genome estimation. I modified this method to include a clustering technique and the steps I took are described in brief below.

Ortholog identification

All P. fluorescens protein coding gene sequences from all four strains were compiled into a fasta nucleotide database. The protein sequence of each gene was then used to search all six-frame translated sequences using tfasty. Using six frame translations attempts to account for possible frame shift errors in pair-wise comparisons.

Positive gene matches were those identified as having at least 70% sequence identity over at least 70% the length of the shortest of the two sequences.

The tfasty matches between sequences was used to generate a network of sequence similarity. Each node in the network represents a P. fluorescens gene from any of four genomes and each arc represents sequence similarity between the two sequences.

Groups of orthologous genes were then identified in the this network using the Markov clustering software (MCL). This software simulates a random walk over a network and more similar nodes are more likely to appear together in the same walk. Groups of orthologous genes were identified as those clustering together.

Clustering results

I wanted to test this clustering approach to determine how effective this process was in identifying orthologs. I aligned each cluster of sequences using MAFFT and then determined the consensus sequence from this alignment. The distribution of consensus sequence length as a percentage of the total alignment length is shown in the histogram below.

This figure shows that the majority of ortholog clusters have a consensus sequence length greater than 50% of the alignment length. However what I found unusual was that 2.3% of gene clusters produced a consensus sequence less than this.

I compared the number of sequences in the alignment, and length of alignment to determine if these factors could be related to the consensus sequence length. The graphs for these are shown below and suggest that neither of these two factors have an obvious effect.

I also compared the cluster efficiency, a metric produced by MCL, to determine if the way in which the data is clustered and therefore which sequences are included in each ortholog group is related to the consensus estimation. This also appeared to be unrelated.

Finally I manually looked at the alignments that had a low consensus length. In many of these alignment there were small sequences included which were much shorter than the rest of the sequences in the alignment. Plotting the difference between the largest and smallest sequence in each ortholog cluster does suggest there is a trend.

The likely reason for this is that small proteins are matching a conserved domain in a larger sequence and therefore being considered a match when generating the network for clustering. It’s debatable at what length a shorter sequence may not be considered to encode an orthologous function, but I opted for an arbitrary cut-off of a 50% difference in size. Repeating the analysis using this cutoff generated the following distribution of consensus length where the number of sequences with consensus less than 50% dropped for 2.3% to 1.6%.

This does not completely eliminate the poorly aligned gene clusters nor does this appear to reduce the obvious bias related to differences in sequence size. However this did seem to eliminate about a third of the poor sequence alignments. The best way to deal with the remaining clusters may just be to remove them.

Submission and annotation process

Submitting our genome sequence data for annotation was a simple process which first requires an IMG account. The microbe to be annotated also requires a GOLD genome project identifier. I had previously created an NCBI genome project and these appear to automatically also be given GOLD ids, so I was able to use this. To submit a genome for annotation I went to the data submission page, selected “IMG ER Submissions”, then clicked “Submit Dataset to IMG ER” at the bottom of the page. The species name was enough to search and find our genome project so that I could then upload all of our scaffolds and non-scaffolded contigs for annotation. Our genome annotation was completed in around a day, and the annotation results were available a few days later.

Annotation results

An early look at the results of the genome annotation suggests nothing unusual. The number of genes, genes per megabase, and mean gene size for P. fluorescens R124 appears similar to the other P. fluorescens strains. One interesting point though was that no ribosomal RNA or CRISPR regions were found in the genome. I believe this is likely because these types of repetitive regions are those which cannot be easily assembled from the ~500bp length 454 reads, and are therefore gaps in the current assembly.

The likely order of scaffolds can be estimated by aligning our sequences to a closely related reference genome. This is easy to do using using nucmer, part of the mummer package. The figure below is an example of simply plotting the results of each scaffold match to three different reference strains.

This output is useful but I wanted to move further to generate a genome map of each scaffold. This is somewhat more difficult to do though because rearrangements and repeats produce hits for the same scaffold in more than one area. I found the most pragmatic solution was to manually curate the nucmer alignment results by hand and order the sequencing scaffolds and contigs based on the longest continuous set of matches to the reference genome. I wrote the likely order as a YAML file which looks something like this:

This allowed to me specify the order of the scaffolds and contigs, the probable size of the unresolved gaps, and also add comments for regions I’m unsure about. Going from a YAML file to circos output is then just a case of rearranging the numbers to the correct format. This produced the map below which shows where the gaps between contigs and also shows the unresolved repeat regions left as gaps by newbler (green highlights on the inside track).

I’ve since further tried to identify genomic islands based on a related approach examining local versus global variance in tetranucleotide usage. This analysis also suggested some possible genomic islands in the Pseudomonas fluorescens R124 genome as regions with a divergent localised variance in tetranucleotide usage compared with the rest of the genome. One interesting result was that a genome scaffold with no sequence similarity to any reference genome did not show any unusual tetranucleotide variance. This leads to further questions about the origin of this region and the possible functionality encoded within.

I’ve been trying to learn about genomic islands in microbes and recently I’ve been looking at identifying genomic islands through unusual frequencies in tetranucleotide usage. The theory is that horizontally transferred DNA will show differential usage of tetranucleotides compared with the rest of the genome. My analysis was based on that of Dick et al. who used self organising maps to look for genomic islands in whole community sequence data. My initial results suggest there may be some regions which are genomic islands however further work will be needed. Next I’m going to look at differences in local and global tetranucleotide variance which also seems useful for identifying genomic islands.

A megablast search showed scaffold 5 did align to reference P fluorescens genomes which was surprising since, as I wrote above, scaffold 5 did not appear to part of the assembly. After a closer look however scaffold 5 is only ~5Kb in size while the scaffold map I produced was on a megabase scale. Therefore scaffold 5 was just too small to be seen by eye when compared to the other much large scaffolds.

The blast search using scaffold 8 returned a more interesting result. The best hit was a plasmid in Pseudomonas syringae pv. phaseolicola. The alignment between scaffold 8 and the plasmid is shown below (click for the larger version) where the plasmid open reading frames are shown in red and the aligned scaffold 8 regions are shown in blue.

This result indicates the likely reason that scaffold 8 does not align to any of the reference genomes is because it is plasmid in origin rather than genomic. A further blastx search with this scaffold identified four regions with sequence similarity to known proteins which are as follows: conjugal transfer proteins involved in the tranfer of genetic material, topoisomerases involved in unwinding DNA, and relaxases and replicases which are likely to be involved in plasmid replication. There was a fifth type of protein may be be related to Type IV (DNA or protein) secretion however the functional annotation of these was less clear. The blastx image result is shown below.

I’m still learning microbial genomics and I suspect it’s unsurprising to discover a plasmid containing sequence similarity to genes involved in replication and transfer. What does spark my interested is that the above blast image shows the rest of the plasmid does not appear in first 100 results returned by blast. This might indicate there is relatively novel data with low sequence similarity known genes waiting to be analysed.

UPDATE: Morgan Langille has rightly pointed out in a comment below that scaffold 8 could have low sequence similarity and still be part of the R124 genome if it’s an inserted genomic island.

There are genomes available for other strains of the P. fluorescens species so these can therefore be used as a template to determine the order of our R124 scaffolds. I initially tried blasting the scaffolds against a reference genome and and plotting the density of blast hits. However when I posted these results on FriendFeed Max pointed out this plot was difficult to interpret and Rob Syme suggested using mummer.

Mummer is a software package for aligning genomes and so I used the nucmer part of the package to compare the R124 scaffolds against the reference genomes of three other P. fluorescens strains. The plot below visualises each nucmer alignment match. This figure indicates the possible order of the scaffolds and also suggests that scaffold 5 (last row) does not appear in any of the reference genomes.

I also visualised the nucmer results as a dotplot between the R124 scaffolds and the reference strains. This plot indicates the likely orientation of the scaffolds and also suggests possible rearrangements in scaffold 3 (purple) and scaffold 7 (yellow) versus the reference strains – a result which I find rather interesting.

The current genomic scaffolds are available on github for both R124 and KY485 strains. I’m going to update the repositories as gaps are closed and the genomes annotated. So far the initial results of the sequencing show the genomes of both isolates are larger than we expected. The predicted genome size and coverage from the Roche GS De Novo Assembler (newbler) run for each strain is illustrated in the chart below (See here for the R code and data). The figure includes the genome sizes of already sequenced P. fluorescens isolates as references.

Genome size

The graph shows both P. fluorescens genomes appear larger than those of existing genomes. The R124 strain is predicted to be marginally larger by ~0.3 MBp than the largest already sequenced P. fluorescens genome while the KY485 strain is much larger by >4 MBp. The sequence data however is relatively fresh and therefore we expect the estimated genome size will change as we try to generate a complete build. Furthermore I believe there is the possibility the current data contains sequences from plasmids which would inflate the size estimates.

Sequencing coverage

The unexpected large size of each genome resulted in less coverage than we hoped. The total genomic coverage in scaffolds is highlighted by the darker grey bars in the barchart above. The R124 assembly has a reasonable ~85% of the predicted genome at 22X coverage. However we have only ~44% of the KY485 genome at 17X coverage – less than half the genome. This therefore indicates a large portion of the KY485 genome is still unknown.

Next step

Over the next weeks we will be trying to bridge gaps in the smaller of the two genomes using PCR and traditional sequencing. I’ll also be trying to estimate size of the gaps in each genome assembly using other P. fluorescens genomes as a reference. I’ll also try to determine if any differences genomic GC content suggest the presence of plasmids in the sequencing data.

I’ve built a small command for boson that combines with bioruby that allows querying pubmed at the command line and also makes refining by year, author, or journal relatively simple. You can see the code on github or watch the short video below to see how it works.