Introduction

Antigenic variation is used by many pathogens as a means of staying one step ahead of the host’s adaptive immune response. Trypanosoma brucei has long been a paradigm for the study of antigenic variation, and the protein responsible, the variable surface glycoprotein (VSG) has been the focus of much research [1–3]. Each trypanosome in a population expresses a single species of protein, and an inherent, parasite-driven switching process causes a proportion of the population to replace their active VSG gene with a different VSG gene, resulting in the expression of a protein in those cells with different epitopes exposed to the host immune system (at a rate of up to 10⁻² switches per cell/generation [4]). The post-genomic era has revealed T. brucei’s antigenic variation system to be unrivalled in its elaboration, particularly in terms of the scale of the numbers of genes that comprise the VSG family. Sequencing the genome of T. brucei has uncovered a gene family much greater in numbers and complexity than was previously thought. Characterisation to date suggests that at least 2,000 VSG genes are in the genome of each trypanosome, providing a spectacularly large repertoire of potential antigens [5, 6], particularly when compared to other pathogens that undergo antigenic variation, such as Plasmodium Falciparum (60 genes in PfEMP1 family [7]), Anaplasma marginale (~10 members in the msp2 & msp3 gene families [8]), and Borrelia burgdorferi (15 members in the vls gene family[9]).

The scale of the gene family size is also reflected in the complexity of switching mechanisms employed to change the identity of the surface antigen. The VSGs are expressed from one of approximately 20 bloodstream expression sites (BES)[10], the active expression occurring in a dedicated sub-nuclear organelle, the expression site body (ESB)[11], with the remainder of BESs being transcriptionally silent. A minor mechanism of VSG switching, accounting for only approximately 10% of events in wild type trypanosomes [12], is to turn off the transcription of the active BES and activate one of the silent BESs (‘transcriptional switching’). However, the majority of switching is through replacing the gene sequence in the active BES via gene duplication, which involves the copying of variable amounts of sequence, ranging from within the gene to the whole telomere [13, 14]. Insights into mechanisms involved in switching suggest that replacing expressed VSG sequence is driven by DNA recombination, and DNA repair/homologous recombination pathways and proteins (e.g. RAD51) have been identified to be involved in the gene duplication process of VSG switching [15] (reviewed in [16]). A further layer of complexity is the construction of novel VSG sequences in the BES from multiple donor VSG sequences, a form of segmental gene conversion termed ‘mosaic’ gene formation [17, 18]. Mosaic gene formation was previously considered to be a rare and minor mechanistic component of overall VSG switching in an infection (e.g. [14]). However, the revelation upon the sequencing of the T. brucei genome that a significant proportion of the VSG repertoire (80–90%) consisted of pseudogenes [19] that cannot be expressed as functional proteins began to alter that perception [5, 20]. It has become clear from subsequent experimental work that mosaic gene formation is an integral component of VSG switching, particularly after the early stages of infection (i.e. beyond the first peak of parasitaemia in mouse infections)[5, 21].

One of the challenges of analysing VSG expression in vivo, and in particular gaining an accurate measurement of the level of expressed diversity given the extent of the VSG repertoire (i.e. to what extent is the repertoire actually used during infection), has been the relatively limited resolution of available techniques–in particular the manual cloning and sequencing of individual VSG cDNAs that has been undertaken in recent studies [5, 21]). While this approach clearly provides accurate data at the level of individual VSG transcripts, the limitations have undoubtedly resulted in a low estimate of the diversity and complexity of VSG expression at the population level, and particularly with respect to minor variant populations. Additionally, although transcriptomics potentially provides the ability to overcome the resolution limitations of manually cloning and sequencing transcripts, the application of RNAseq to VSG expression from in vivo samples has long been deemed challenging, due to the requirement for assembling multiple closely related gene variants from a mixed population using short reads of 100–200 base pairs (e.g. Illumina)–this has similarly been an issue when attempting to resolve, for example, the diversity of the mammalian immunoglobulin gene repertoire underpinning the antibody response (e.g. [22]). However, a recent study subjected in vivo samples to Illumina sequencing (100bp, single-end reads) and demonstrated the utility of transcriptomics in terms of increased resolution [23], and were able to detect minor variants (0.1% of population) and up to 79 variants at a time point, although they were not able to identify significant mosaic gene expression.

Long read sequencing potentially provides the ability to further increase our resolution, particularly as the length of reads commonly reached with such technologies (average read length in Pacbio, for example, is quoted as 10,000–20,000 bp; http://www.pacb.com/smrt-science/smrt-sequencing/read-lengths/) far exceeds the length of the VSG transcript (approximately 1600 bp), meaning that the issue of assembly of closely related VSGs from multiple reads should be bypassed. Here, we present analysis of VSG expression from replicate in vivo T. brucei TREU927 infections in mice at 4 time points over 12 days using almost 500,000 Pacbio Sequencing reads. We demonstrate that long read technologies provide significant advantages for analysing the diversity of VSG expression. Our data suggest that the VSG population comprises significantly more variants even at an early stage of infection (up to 190 variants at day 10 post-infection), that the pattern of VSG expression is surprisingly reproducible (using the novel application of ecological diversity indices), and that mosaic gene expression can be detected much earlier in infection than has been possible previously. Our data also provide insights into the nature of mutations introduced by Pacbio long-read sequencing technology, as the dataset includes significant coverage of one sequence (>140,000 reads).

Results

Long read sequencing maps the VSG transcriptome at unprecedented resolution

Using PacBio long read RNA sequencing of 20 blood samples enriched for VSG transcripts from replicate in vivo T. brucei TREU927 infections in mice at 3, 6, 10 and 12 days post infection, we obtained 486,343 reads with an average read length of insert of 1,569 bp (Table 1, Fig 1B). Reads were filtered by length (1400-2000bp) based upon both literature on VSG genes [21, 24] and the read distribution in our dataset (Fig 1B) to remove reads resulting from sequencing artefacts and shorter fragments (i.e. partial reads), and on the basis of similarity to known VSGs (blastn ≥60% alignment against TriTrypDB-v26 [25]–note that the reads include both N-Terminal and C-Terminal domain sequences) (Fig 1C), resulting in a dataset of 296,937 ‘VSG’ reads. Of the reads that were of the appropriate length (1400-2000bp) but did not have ≥60% match to VSGs in the reference database (n = 102,940), 90,810 (88.2%) mapped partially to VSGs, 3,513 (3.4%) mapped to non-VSGs, and 8,617 (8.3%) did not produce any match to the TREU927 reference genome. Within the dataset of 296,937 VSG reads, each read on average represented the consensus sequence from 6.50 passes of the full length fragment by the DNA polymerase (‘full passes per read’; summarised in Table 1; full data in S1 Table), and for each of these reads there was robust identification of a donor gene for the N-Terminal domain (NTD); therefore, for these 296,937 reads we have high confidence that they contain all of the features necessary to be consistent with being full length VSG transcripts. The 296,937 reads represent a total of 449 VSGs (74.77% of VSG a-type and 25.22% VSG b-type [24]) across 20 samples, with the number of reads per VSG following a power-law distribution (Fig 1D), and provide a unique insight into the in vivo VSG transcriptome across time and animal replicate.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

(A) Parasitaemia measurements for the batches of mice sacrificed at days 3, 6, 10 and 12 days post-infection. A summary of analysis of Pacbio data generated from these samples is shown, including main decision steps (black boxes represent data filtering steps; data within black boxes was included), (B) read length distribution, (C) alignment of reads to VSG, (D) correlation of read number with number of VSGs identified.

https://doi.org/10.1371/journal.pntd.0007262.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Summary of data per sample.

https://doi.org/10.1371/journal.pntd.0007262.t001

PacBio sequences contain random errors even at high number of full passes

ORFs were identified in the 296,937 reads with a conservative minimum nucleotide size of 1200 nucleotides (reported size ranges of VSG NTDs and C-Terminal domains [CTDs] are approximately 300–350 and 100 amino acids, respectively [5, 21, 24]). Surprisingly, only 33,234 reads (11%) resulted in predicted ORFs. Although the percentage of reads with predicted ORF increased with increasing number of full passes, it remained well below 50% even for reads having 10 full passes or more (Fig 2A). Since the distribution of the number of reads with a detectable ORF over all VSGs was similar to total expression level distribution (Table 2), we hypothesize that the lack of identified ORFs was due to random sequencing errors rather than any systematic biases in the data, despite PacBio claiming an accuracy of more than 99% for reads with 15-fold coverage [26]. To investigate this hypothesis in more detail, we focused on the most abundant VSG (Tb08.27P2.380, 1551bp, 141,822 high-confidence reads) and annotated each discrepant base pair of each aligned read as either an insertion, deletion or mismatch with respect to the Tb08.27P2.380 reference genome sequence. All reads had an alignment score greater than 90% over the first 1266bp (the N-Terminal domain) (Fig 2B). The distribution of sequence errors showed a clear bimodal pattern across the N-Terminal domain, with 145 nucleotide positions having a consistent mismatch (131), deletion (10) or insertion (4) across more than 80% of the reads, and 1,112 nucleotide positions having errors in at least one but fewer than 2% of reads (Fig 2B). This suggests that the former represent genuine mutations already present in our inoculum (with respect to the reference genome sequence), whereas the latter represent either random sequencing errors introduced by Pacbio or low level genuine mutations that we cannot currently distinguish from Pacbio error. Previous studies have indicated accumulation of mutations over time in expressed VSGs, and we examined this in our data for reads aligning to Tb08.27P2.380 (for the N-terminal domain) by assessing the error rate for mismatches, insertions and deletions (S1 Fig). While these data indicated statistical support for differences in the data distribution across time points for all 3 mutation classes, due to the skewed nature of the data distribution (most bases have an error rate close to zero) this conclusion must be treated with a degree of caution. The assertion that the errors present in >80% of reads were ‘genuine’ mutations was further supported by these 145 mutations being consistently present in PCR amplicons sequenced by Sanger sequencing. These PCR amplicons had been generated from cDNA extracted from multiple samples (n = 7 for Tb08.27P2.380; representing sequences independently cloned and sequenced from 4 mice on days 3 and 10, S2 Fig). Insertions were the most common Pacbio-introduced error (average per-base error rate of 0.79% across the N-terminal domain sequence), followed by deletions (0.73%) and mismatches (0.33%) (Fig 2C), in agreement with what has been reported before [27]. Consistent with the ORF prediction pattern (Fig 2A), the overall error percentage was lower for reads with higher number of passes, but introduced sequencing errors (i.e. interpreted as mutations not present in the genome of the inoculated trypanosomes) remained present at more than 1000 nucleotide positions even for reads with 10 passes (Fig 2D). The nature of our data, comprising >141,000 reads of the same sequence, therefore provides an unusually robust insight into the nature of Pacbio errors and the caveats that must be placed upon interpretation of such data, as most studies involve much less coverage per single base pair.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

Analysis of Pacbio sequence data: (A) Number of reads aligning to VSGs per number of full passes, with proportion of those reads comprising predicted ORF in yellow; red line shows percentage of reads at each threshold of full pass number that contained a predicted VSG ORF. (B) Percentage of reads at each position in the N-Terminal domain that contained a mutation with respect to the reference genome sequence for VSG Tb08.27P2.380; alignment coverage is shown by the black line, insertions by green dot, deletions by red dot and mismatch by blue dot. (C) Focused representation of data in 2B, with only mutations with respect to the genome reference sequence <2% at each position in the N-Terminal domain VSG Tb08.27P2.380 shown; insertions shown by green dot, deletions by red dot and mismatch by blue dot. (D) Percentage error rate plotted against number of full passes; red line indicates number of nucleotide positions for those aligning to VSG Tb08.27P2.380 that contained an error with respect to genome reference sequence against number of full passes.

https://doi.org/10.1371/journal.pntd.0007262.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Top 20 variants; number of reads across dataset as measured by (i) mapping to reference VSG database and by (ii) sequence clustering; the relevant TREU927 reference VSG is indicated in the first column.

https://doi.org/10.1371/journal.pntd.0007262.t002

VSG population comprises more variants than expected and the pattern of expression is highly reproducible across replicates

Our data demonstrate that we can detect multiple VSGs in each sample, and that we can identify changes in VSG expression and diversity over time. We identified a median of 27 unique VSGs per sample at day 3 post infection (p.i.), which progressed to 82 VSGs at day 6 p.i., peaking at 187 VSGs at day 10 p.i. and reducing to 132 VSGs by day 12 p.i. (Fig 3A). When identified VSGs that mapped to single reads from single samples were removed, this resulted in an identification of 334 VSGs (median of 27, 81, 170 and 126 VSGs per sample at 3, 6, 10, and 12 days p.i., respectively).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. VSGs per sample.

(A) Average number of donor VSGs mapped at each time point, plotted with and without VSGs identified from single reads of at least one full pass (‘singletons’), (B) Proportion of reads that map to identified donor VSGs for each mouse and each timepoint; reads that map to the 10 most abundant identified donors (across the whole dataset from 20 mice) are shown; reads that map to donor genes other than these ten are represented as ‘others’.

https://doi.org/10.1371/journal.pntd.0007262.g003

Not only were the number of distinct VSGs consistent across samples for the same time point, but the expression pattern (proportion of reads per sample mapping to particular VSGs) was also highly reproducible between samples and over time (Fig 3B), albeit bearing in mind that these analyses are of batches of mice at four different time points rather than longitudinal samples of the same mice. The VSG that is dominant at day 3 (Tb08.27P2.380), presumably introduced as the dominant VSG in the inoculum, remains dominant in all mice at day 6, but is the single VSG with the most reads aligned in only two of five mice at day 10. Interestingly, by day 12, the VSG with the most reads per sample is the same in all five mice (Tb09.v4.0077) and this VSG was also most common at similar timepoints in previous analyses [21]. Additionally, the other eight VSGs that reads map to in mice at days 10 and 12 (Tb927.4.5730, Tb927.10.10, Tb11.v5.0932, Tb927.9.300, Tb09.v4.0088, Tb05.5K5.330, Tb927.9.16490 and Tb927.3.480; Fig 3B) are present in all ten mice suggesting a degree of conservation in the sequential expression of VSGs in independent infections, consistent with previous observations [21, 28, 29]. However, in all mice there were reads that mapped to VSGs distinct to these most favoured 10 VSGs (‘others’ in Fig 3B, which account for 10.36% of all VSG-mapped reads), and in some mice this proportion was particularly high (e.g. mice 3.5, 6.1 and 10.4; Fig 3B). This is particularly evident at day 6, where although the dominant VSG (Tb08.27P2.380) makes up most reads, the majority of reads that do not map to Tb08.27P2.380 map to VSGs other than the other top 9 VSGs in all mice. Additionally, at Day 10 we observe both the greatest number of VSGs and the least domination by any single VSG, but the proportion of ‘others’ either reduces or remains stable. These analyses combine to indicate that while there is a broad predictability in expression, with dominant VSGs at the beginning and end of infections, in between these timepoints there is a degree of stochasticity in the system–although eight VSGs comprise the majority of reads that do not map to either of the two dominant VSGs, the relative proportion of these ‘minority’ VSGs is not consistent, and there are furthermore many VSGs that are expressed at very low levels in all mice.

The analysis described thus far (Fig 3) has not taken into account any sequence similarity between VSGs, but relied on mapping reads to identified VSGs in the reference database. To analyse the population diversity of VSGs within and across samples using a method that is independent of mapping to existing databases (which are likely to be incomplete), we applied information theoretic measures more commonly used to quantify the biodiversity of ecosystems [30]. This approach initially applied a clustering algorithm to a proportion of reads (n = 33,205; comprising reads with predicted ORF) in order to enable identification of the reads that clustered on the basis of sequence similarity, as putatively distinct VSGs (Fig 4A). These data showed significant congruity with those described for the VSG mapping approach described above (Table 2). The top 10 clusters comprised 89.34% of all reads, compared to 89.68% for the VSG mapping approach, and the relative proportion of reads that either map to VSGs or cluster by sequence similarity is very comparable for the 10 most abundant VSGs (Table 2). These data indicate that the clustering algorithm applied was robust in terms of identifying individual VSGs, and therefore indicated a very similar pattern of a dominant early VSG, followed by an intermediate period of significant greater VSG diversity, ending up with a second dominant VSG by day 12.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. VSGs are expressed in a semi-predictable order.

(A) Expressed VSG sequences were clustered by sequence similarity (Materials and Methods and S1 Appendix) and ordered on the x-axis according to the cluster they were assigned to. The average similarity of each sequence to others from the same population was calculated (‘ordinariness’, a measure of how common that sequence is) and is plotted for each sequence on the y-axis. Grey lines indicate the profiles of individual mice, while coloured lines indicate the average for that particular cluster, coloured according to the cluster. (B) Diversity analysis showing the effective number of distinct VSG profiles found on day 3, 6, 10 and 12 for each mouse (dots) on that day, with the average across the days represented by the solid line.

https://doi.org/10.1371/journal.pntd.0007262.g004

The sequence similarity data also allowed the analysis of variability between mice using a new measure of population differentiation called normalised beta diversity [30] (Fig 4B). When looking at a single day, beta diversity is the effective number of distinct VSG profiles present on that day, giving information on the differentiation between the animals. This analysis indicates (similar to the VSG mapping data) the greatest beta diversity across individuals is at day 10 (Fig 4B solid line).

Further exploring each time point and variation between mice (Fig 4B dots), we can see that although the mice at day 3 show some distinct VSG profiles (albeit with overexpressed VSGs in individual mice common to all mice, SI Fig 3), at day 6 most mice (except for mouse 6.5) are broadly consistent with respect to which VSGs are present and how common they are. The effective number of VSG profiles increases further on day 10 with maximal divergence between mice at any time point, (Fig 4B, solid line). This value then decreases on day 12 (though mouse 12.2 is distinct), as the mice begin to express similar profiles again. These analyses again indicate that there is stochasticity in the process of VSG expression considered as a progression over 12 days, and there is semi-predictability rather than strict hierarchical progression through VSG expression, as has been described previously [21, 29, 31, 32]

Mosaic VSG genes are present early in infection

Mosaic genes were considered identified where BLAST hits for a particular read demonstrated non-overlapping homology to more than one distinct VSG in the reference database. This was commonly seen in the C-Terminal domain, where the same N-Terminal domain was in many instances observed with different C-Terminal domains (“3’ donation” in [21]). Using pairwise alignments of all reads that mapped to Tb08.27P2.380, based on the alignment coverage over the gene, donors were filtered based on the region representing the C-Terminal domain (the 3’ region approximating to 30% of the gene shown in Fig 5A). Donors were selected based on at least 80% alignment coverage to the CTD. These data show that the reads aligning to Tb08.27P2.380 consists of three subgroups based on their CTD donors, which are derived from either the reference gene Tb08.27P2.380 (43% of all reads), but also from Tb10.v4.0158 (29%) or Tb927.6.5210 (28%). The proportion of the three donor CTDs varies across time points, with the proportion of reads deriving from the donor Tb08.27P2.380 gene decreasing by days 10 and 12 (reducing from 46.55% at day 3 to day 26.19% at day 12, although the number of reads in total aligning to Tb08.27P2.380 is low by days 10 and 12). The frequent nature of this recombination has been observed previously [21]. We detected N-Terminal domain mosaics (within the constraints of our stringent selection criteria) at a much lower frequency (n = 45 over all 20 mice; three sequences at day 3, five at day 6, 13 at day 10 and 23 and day 12 –S2 Table), and in most cases these are single read examples, and so must be treated with some caution (albeit 12 of the putative mosaic reads have coverage of at least 7 full passes, a coverage level at which our analysis–Fig 2D–suggests should effectively remove sequencing-derived error). However, we have two examples where we have more than one read indicating N-Terminal domain mosaicism, with the additional support for one of these sequences that it is only detected in one mouse–given the complex nature of previously identified mosaic N-Terminal domains [5, 21], it is unlikely that identical mosaics would emerge in separate individual infections. Nevertheless, we do also have one putative mosaic sequence that occurs in two separate mice (balbc_6_0/100673/ccs5 and balbc_12_1/30571/ccs9 in mice 6.1 and 12.1, respectively; S2 Table)–this may either represent a gene currently not annotated in the TREU927 genome or be a true mosaic gene that was present in the initial inoculum and has remained at low levels throughout infection. The N-Terminal domain mosaic examples we have detected are mostly relatively simple mosaic genes (e.g. Fig 5B). Although we cannot formally rule out that at least a proportion of these mosaic genes were present in the original inoculum, the increased frequency over time is consistent with expectations that this process is rarest early in infection but becomes more prevalent as infections progress.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Mosaic gene identification.

(A) Average number of donor VSGs (y-axis) per position across scaled VSG reads (gene length scaled to 100 (x-axis); data derived from 296,937 VSG amplicons)–blue line shows data for all reads, red line shows data for all data without reads that map to Tb08.27P2.380. (B) Example mosaic gene candidates, represented by read identifier, and displayed to scale from 5’ to 3’ (N-terminal domain only), with colour of each segment indicating most likely donor VSG from the TREU927 VSG repertoire (VSG gene shown in key; where relevant sequences are identical for more than one potential reference gene donor; all possibilities are shown in key).

https://doi.org/10.1371/journal.pntd.0007262.g005

Discussion

The results illustrate the power of long read sequencing when applied to expressed gene diversity–we identified 449 VSGs across 20 individual samples, covering four time points post-infection (3, 6, 10 and 12 days). The identification of the VSGs was achieved by two approaches; mapping reads to a reference VSG database, and secondly clustering read sequences to identify distinct variants–importantly without reliance upon a reference genome or sequence database. These independent approaches were highly congruent in the number of VSGs and the proportion of reads that were attributed to individual VSGs (Table 2), meaning that the clustering approach may be particularly valuable for analysis of long read data generated from infections with trypanosome strains (or species) where a genome is either not available or is incomplete. When compared with previous approaches, such as manual cloning (801 VSG sequences that comprised 93 distinct VSGs or ‘sets’ across 11 mice across 19 days of sampling each [21]) or short read Illumina sequencing (289 VSGs for 4 mice– 3 mice sampled 9 times over 30 days and one mouse sampled 13 times over 105 days [23]), the Pacbio approach gives significantly higher resolution per sample. It must be acknowledged that in the present study the starting volume of infected blood for each sample was higher (200 μl versus 50–100 μl in [23] and approximately 15 μl in [21]), and additionally the inoculum in the current study was significantly greater and not clonal, meaning the study design may predispose to more expressed variants being detectable. The TREU927 clone used was also highly virulent, giving rise to a high parasitaemia early that was maintained for the 12 days of infection. This is not representative of the classical fluctuating profile of less virulent strains (or clones of this strain, e.g. [33]); however, for the purposes of assessing the utility of Pacbio this was advantageous. A proportion of the identified VSGs (115/449; 25.6%) derive from single reads in single samples, and therefore a degree of caution must be employed with these variants. However, when the singleton VSGs are removed, we can confidently conclude that we have identified 334 VSGs across our datasets–this ranges from a median of 27 VSGs in day 3 samples to 170 in day 10 samples. Therefore, despite these caveats, we can still conclude that the resolution in terms of diversity is significant for the long read approach, and likely to be of great utility for studies incorporating VSG diversity going forward.

Despite the limitations of the study design, where we have analysed batches of mice at four time points rather than longitudinal surveys of individual mice, our data across 20 mice and four time points are very consistent with a highly reproducible pattern of VSG expression over time (Fig 3 & Table 2). There was a remarkable degree of consistency in identity of dominant VSGs across independent infections–particularly as the inoculum used was not a single cell or a cloned inoculum (this is very distinct from, for example, Borrelia, where Pacbio analysis has indicated very little overlap in expressed antigen diversity across replicates from the same starting inoculum [34]). The data demonstrated a consistent emergence of the two sequentially dominant variants at the beginning and end of the infection period (Tb08.27P2.380 and Tb09.v4.0077), although during the period in between the dominant VSGs there was significant diversity in expressed VSGs that was consistent with an inherent degree of stochasticity in the system. This was reinforced by the application of biodiversity analysis (Fig 4), which illustrated the semi-predictable nature of the variant progression across the mice and timepoints. This chimes with previous work that described the semi-predictable expression of VSGs in T. brucei [21, 28, 29], and modelling approaches that have also reflected semi-predictable use of the VSG repertoire [31, 32, 35].

When analysing our data set and comparing with that of Hall et al, 2013, who used the same TREU927 strain, we have significant overlap in detected expressed variants. 90% of our reads correspond with a VSG detected in Hall et al. The dominant early VSG is different (corresponding to ‘Set_23’ in the Hall data, Table 2), although the Tb09.v4.0077 which becomes dominant by day 12 was similarly dominant by ~day 20 in Hall et al; differences are presumably due to the use of either a stabilate with a distinct passage history, or the use of a larger inoculum rather than single trypanosomes (i.e. inoculation of a population from a previous infection presumably expressing the dominant VSG at that particular stage). The dominant VSG in our dataset (Tb08.27P2.380) was annotated as a pseudogene in the reference genome (predicted to be truncated due to insertion of a stop codon). The annotation as a pseudogene is not consistent with our data as a dominant early VSG, as it would suggest mosaic gene formation providing a dominant early gene–indeed, recent reannotation has classified this gene as intact, which would be more in keeping with early expression favouring intact over pseudogene or mosaic VSGs [5, 21]. However, given the 1 × 10⁵ inoculum used in this study, it is also feasible that the transfer of Tb08.27P2.380 as the dominant expressed VSG from the donor mouse infection may have given rise to it being the dominant expressed VSG in the infections analysed.

We have identified mosaic genes (classified as reads demonstrated non-overlapping homology to more than one distinct VSG N-Terminal domain in the reference database) earlier in infection than has previously been identified, although we cannot formally rule out that at least some of these were introduced in the original inoculum. The rate of mosaic gene detection was very low in our study, mostly either single or very few reads, which probably reflects our timeframe being only 12 days post-infection; this also pertains if the trypanosomes expressing mosaic VSGs derived from the inoculum, which was also generated over a short duration (5–7 days) in the donor mouse. However, these data do indicate that the nature of the long read sequencing is highly beneficial in terms of mosaic gene identification; even low frequency expressed genes (within the limitation of the four orders of magnitude of coverage that the read number per sample provides) can be identified with some confidence due to the acquisition of the whole gene sequence–in order to achieve this with short read approaches a reasonable degree of read coverage would be required to identify and confirm putative mosaic genes. This has potential implications for the application of long read sequencing to significantly further our understanding of infection dynamics and the role of mosaic genes as infections progress. This is likely to be important in terms of ability to gain insights into the mechanisms of mosaic gene formation because of consequent increased ability to resolve defects in switching rate (e.g. analysis of DNA recombination gene mutants such as RAD51 that have been implicated in DNA recombination-based VSG switching [15])–at present it is not known if mosaic gene formation involves a mechanistic switch in terms of pathways; the ability to detect low frequency mosaic gene expression should provide the ability to study this. Additionally, detection of low frequency VSGs would enhance the ability overall to more fully analyse the temporal kinetics of VSG switching–providing an avenue for improved quality of inputs into modelling dynamics of VSG expression. The clustering approach developed in this study that does not rely upon a reference database would also make analysing expressed VSG diversity in the animal trypanosomes, T. congolense & T. vivax, feasible—the reference genome (and therefore genomic VSG repertoire) is less well annotated in these species than in T. brucei. One challenge for taking a similar approach in these species is the lack of conserved 3’ UTR sequence in expressed VSGs to enrich transcripts. However, such analyses may be particularly enlightening given the different structure and content of VSG repertoires recently described between the three genomes [36, 37], as well as the strikingly different arrangement of VSG expression sites in T. congolense compared to T. brucei [38].

We detected indels consistently when comparing Pacbio transcripts to the reference gene (Fig 2). While these differences may indeed be real, with our protocol we have somewhat limited resolution for conclusively differentiating indels introduced by the trypanosome from those potentially introduced by PCR. However, PCR is unlikely to be the sole cause of the observed mutations, because in the dominant VSG in our dataset (Tb08.27P2.380), which represents 141,822 reads across all 20 samples–therefore, 20 independent PCR reactions—we observe a consistent set of variations from the reference genome sequence (145 nucleotide positions across the NTD having a consistent mismatch (131), deletion (10) or insertion (4) across more than 80% of reads with respect to the reference sequence) across all reads–these are consistently present across all reads for this variant, including those reads with high fold coverage (i.e. greater than 10 full passes per read) (Fig 2). These data, across technical and biological replicates, lead us to conclude that these differences were present in most likely the genome copy, but also potentially a distinct BES-resident copy of this VSG that has accumulated mutations distinct to the genome basic copy of the gene, and mutations were not introduced by PCR. One possible explanation for this is that there is very likely a significant (and unknown) divergence in passage history between the sequenced reference genome TREU927 trypanosomes and those used in this experiment. This would be consistent with data from many pathogens of the increased mutability of telomeric/subtelomeric gene families [39]. Previous data indicated accumulation of point mutations in expressed VSGs over time within infections [5, 21], and in our data we saw some support for this process, but the skewed nature of the data distribution limits our ability to conclude increased mutations over time as an important aspect of VSG expression (it should be noted that a timeframe of 12 days is relatively short and will have limited our resolution). However, our data indicate that application of long read analysis over longer infection timeframes is likely to be a useful means of characterising the nature and role of this mechanism.

However, the multiple mutations that were present across multiple VSG sequences in our data, did enable detailed analysis of the nature of mutations detected in Pacbio sequencing (Fig 2). Ideally, to enable clear differentiation of PCR bias and artefact, errors introduced by Pacbio, and mutations introduced by the trypanosome, unique molecular identifiers (UMIs) would be added prior to PCR amplification (e.g. [22, 40]). While we did not incorporate this step, we can draw some conclusions from analysis of our data. When data for Tb08.27P2.380, which represents 141,822 reads, is analysed across the range of fold coverage per read, it is clear that most of these mutations are removed as the coverage increases (Fig 2)–although notably even at a high number of passes some introduced mutations remain. This strongly suggests that most of these are errors that are introduced by the Pacbio process, and the proportion we observed across the dataset (insertion 0.79%, deletion 0.73%, mismatches 0.33% per base pair) is consistent with that reported in other studies (e.g. [41]). The mutations also directly influenced the ability to predict open reading frames in our data—ORFs only being detected in 11.22% of VSG reads (33,234 of 296,937). Clearly, with these reads being generated from cDNA one would have expected most if not all to have identifiable ORFs. Therefore, these data indicate some of the limitations when using Pacbio, even with data that comprises multiple passes–the introduction of mutations does provide a layer of complexity to the analysis that must be addressed with care. This is particularly pertinent when trying to analyse multiple closely related genes, as in the case of VSGs. We were able to draw conclusions on the basis of sufficient coverage of a highly expressed dominant gene, combined with the inclusion of multiple biological replicates; without these elements interpretation would have been very difficult without parallel short read sequencing to correct errors introduced by the technology.

A further issue for consideration for the application of long read technologies to the analysis of expressed gene diversity is the number of reads per sample. Our data provided coverage over four orders of magnitude—although significantly greater in resolution than previous manual and laborious methods, this contrasts relatively poorly with the numbers of reads that short read applications deliver (millions). However, it should be noted that with the short read approach many reads will be required to robustly identify full length single variants (in particular to enable differentiation of closely related transcripts, either similar genome-encoded variants or related lineages of mosaic genes [5, 21]), whereas in theory at least a single pacbio read should provide the ability to robustly identify a particular VSG transcript. While the coverage is being improved with the newer platforms (e.g. the Pacbio Sequel potentially delivers a further tenfold increase in data per run), this may limit resolution in terms of detecting minor variants, for example. We did detect significant expressed diversity, and this is partly explained by our use of a relatively large inoculum, which was not cloned, of a virulent isolate that resulted in high and sustained parasitaemia. Therefore, we started with what was probably a relatively diverse population (albeit dominated by expression of Tb08.27P2.380), reflected in the diversity of VSGs detected at day 3 post-infection, which would be significantly lower in the event of a clonal or smaller initial inoculum.

While our data indicate that long read sequencing provides increased resolution in terms of identifying VSG diversity, clearly questions still remain. For example, why the VSG repertoire is so evolved and large? Our data suggest an increased proportion of repertoire is involved, even at early stages, compared to previous studies, which indicates a bigger proportion of the repertoire may be utilised during the lifetime of an infection (which in cattle can be many hundreds of days) than previous data suggests. This is consistent with the data of Mugnier et al [23], where multiple minor variants were observed using an Illumina sequencing approach. However, that study and ours both have limitations, one with relatively few biological replicates (albeit one mouse was followed for ~120 days) and one that only ventured to 12 days post-infection. Therefore, assessing antigen dynamics in the chronic phase of infections with tools that give significant resolution of expressed antigen diversity will be critical to furthering our understanding of the mechanisms of trypanosome antigenic variation. Key to studying this will be analysing the picture in the truly chronic stages of infection (as was done by Mugnier et al in the context of mouse infections), but particularly doing so in relevant hosts (e.g. cattle [42]) where the total population of trypanosomes in the animal will be potentially 1,000 times greater at peak parasitaemia and where infections may last for 100s of days–this will have a profound influence on the usage of the repertoire (our data, for example, was representative of a total population of approximately 1 × 10⁸ parasites per mouse). Additionally, recent studies indicates T. brucei populations inhabit different niches in the mammalian host (e.g. skin and adipose [43, 44]), to the extent that some show evidence of local adaptation with respect to metabolism ([44])–how this population compartmentalisation interacts with antigenic variation and immunity is likely to be important for parasite maintenance and transmission. Therefore, understanding the dynamics in both the chronic stages of infection and in clinically relevant hosts will potentially provide ideas on the selective pressures that maintain such an elaborate system. Additionally, given the significant advantages described above in terms of identifying low frequency variants (including mosaic VSGs), it may be that a combined long and short read approach is likely to be the optimal way of holistically and accurately identifying expressed VSG diversity; the increased read number of short read technologies in combination with the better resolution of long read technologies would provide significant power to examine the complexity of VSG expression in trypanosomes.

Materials and methods

Supporting information