A typical T. cruzi minicircle is approximately 1.4-kb and contains four conserved sequence regions, each followed by a variable region containing a gRNA. Each conserved region is composed of three individual conserved sequence blocks (CSB-1, CSB-2, and CSB-3) each of which are broadly conserved 23. Numerous minicircle sequences with CSBs were found in the CL Brener and Esmeraldo strain WGS reads. Multicopy sequences can have extensive variability in kinetoplastids 24, so to be certain of isolating the greatest number of minicircle sequences, the intragenomic diversity of individual CSB sequences was first assessed by examining the conserved regions of all extracted T. cruzi minicircle sequences using a two-tag method to capture native variability of the intermediate sequence (see Methods).

Remarkable conservation was observed among the dozens of conserved region sequences identified (Fig 1). Less than 1% of sequences contained variations at each position within the CSBs, and the conservation for each strain extended well beyond the basic consensus. The base CSB-1 sequence was expanded from 10 nt to 18 nt common to CL Brener and Esmeraldo minicircles, with a strong 24-nt stretch in CL Brener. The G-rich, 8-bp CSB-2 element is expanded into a 26-nt purine-rich region in both T. cruzi strains, with a 29 nt length in CL Brener. The 12-nt CSB-3 was extended to 27 nt in both strains. These extended CSBs likely reflect species-specificity and could prove useful for taxonomic distinctions.

The conserved regions of the minicircles represent tags that can be used to pull minicircle sequences from the WGS reads while being confident that few, if any, sequences are missed due to intragenomic variation.

The areas of complete CSB conservation were used as tags to extract as many minicircle reads as possible. A sorting script extracted 54 CL Brener and 53 Esmeraldo sequence reads. If a sequence was found to contain minicircle sequence, then the read from that same clone in the opposite direction was also extracted from the database, and if the 'mate pair' sequences overlapped they were joined. In this manner 32 contiguous sequences were assembled. Dozens of the reads contained several complete variable regions, with each variable region potentially carrying a gRNA. The structural linkages between variable regions were recorded to determine if recombination occurred between minicircle sequence classes and if any correlation existed between editing target and minicircle neighbours.

Each variable region was expected to encode a single gRNA, thus for the task of gRNA prediction the sequences were broken into individual units of the upstream CSB-3 plus one variable region plus its downstream CSB. A spectrogram alignment of unique units revealed striking regions of nucleotide preference in several places within the variable regions (Fig 2). An enrichment of G residues was evident adjacent to the CSBs, while the central portion of the variable region indicated a bias toward As and Cs. This pattern held true for both CL Brener and Esmeraldo strains.

The minicircle variable regions displayed specific nucleotide preferences over their length, potentially reflecting the sequence bias of the gRNA genes. We next sought to correlate the putative gRNA genes with RNA editing events, however as many of these sequences are undetermined, some creative sequence manipulation was required in order to proceed.

One of the major goals of this analysis is to predict gRNAs. The current local alignment method of predicting gRNAs 8 requires knowledge of the sequence for the fully edited target genes that for the most part are not yet known for T. cruzi, ATPase 6 25 and COII 26 being the only exceptions. We have initiated the experimental determination of T. cruzi editing events, but in the meantime used predictions to facilitate our minicircle study.

Predicted edited sequences (Additional File 1) were generated by manually inserting or deleting Us in the unedited message sequence following the known sequences of the corresponding T. brucei edited mRNAs, while preserving conservation of the resulting amino acid sequence. Predicted sequences were generated for: CyB, CR3, CR4, MURF2, ND3, ND4, ND7, ND8, ND9, and RPS12. Concurrently, the edited sequence of COIII was obtained by RT-PCR for CL Brener: When compared with our in silico prediction, the true COIII sequence differed from the predicted version such that two of the estimated 30 gRNAs would have been missed (data not shown). For this reason, sizeable portions of the other predicted sequences were expected to yield many useful gRNA predictions. Additional confidence was gained for segments of predicted sequence where high-scoring matches with putative gRNAs were found.

A combination of actual and predicted mRNA sequences was used for exploring the gRNA assignments to specific genes. This method is not perfect due to the potential for error in our predictions, but these areas will be clarified as the edited sequences are obtained at the bench.

Potential gRNAs were identified among minicircle sequences from the genome project and in GenBank using a three-part process (Fig. 3abc): the single greatest Smith-Waterman local alignment score for a variable region was most likely to represent the overlap of the gRNA and target mRNA region. A permutation test was performed to determine the probability of a given 'best overlap', and then a false discovery rate (FDR – see Methods) control was used to determine whether a hybridization with a given probability was significant (Fig. 3c). This statistic provided the criterion used to identify 108 potential gRNAs from 248 minicircle variable regions obtained from GenBank 2527, CL Brener and Esmeraldo WGS reads (Additional File 2). The predicted minicircle gRNAs are positioned consistently within the variable region, providing more evidence that these highly scoring alignments represent regions of gRNA hybridization to target mRNAs.

The predicted maps of RNA editing for T. cruzi (Additional File 3) showed that while the final edited mRNA was conserved due to the required amino acid composition of the resulting proteins the gRNAs that perform the editing were variable from one strain to the next; this variability was not restricted to transition mutations. The inexact line-up of gRNAs from different strains on certain regions of the COIII map (Fig 4, 5, 6), for example, demonstrated that the start and stop positions of hybridization can drift from strain to strain, and that gRNA heterogeneity may occur within isolates.

In addition to gRNAs identified from minicircle sequences, two gRNAs were predicted by local alignments with high confidence to reside within CL Brener maxicircles. Both maxicircle gRNAs are on the opposite strand of nearby protein-coding genes, with an ND7 gRNA placed 95-bp downstream of ND5 and approximately 500-bp upstream of the repetitive region, and a MURF 2 gRNA located 66-bp downstream of the CR4 gene overlapping the start codon of ND4. The previously-described COII gRNA lies immediately downstream of the COII gene on the same coding strand 2527.

The spectrograms of full minicircle units (CSB to CSB) revealed areas of conserved nucleotide preference within the variable region (Fig 2). To explore this further, the sequences with identified gRNAs were aligned as follows: regions upstream of the gRNAs were aligned to their 3' ends, gRNAs were all aligned at +1, 5' to 3' of predicted hybridization, and regions downstream of the gRNAs were aligned by their 5' ends.

The resulting spectrogram revealed nucleotide preferences with respect to these orientations (Fig 7). As might be expected to promote annealing to the target message, the 5'-anchor region of the gRNAs had a higher than background preference for Cs (blue), and the entire gRNA had a slightly higher T to A skew (red to purple) with a spike centred approximately 12 nt into the predicted gRNA hybridization. While the areas immediately around the gRNAs were more T-rich (red) than the gRNAs themselves, the most clear transition occurred approximately 50-nt downstream of the gRNA where the preference for As and Ts dropped sharply and the region became more G-rich (green).

Given that the precise switch was dictated more by distance from the end of predicted hybridization than relative to the CSBs, these observations may have implications for the transcription and maturation of gRNAs. Note that the 5' ends of these gRNAs have not been physically mapped, and that corrections in the predicted editing events may either extend or decrease the 5' and 3' ends of the actual gRNA.

Although dozens of gRNAs were predicted using local alignments, approximately half of the minicircle sequences still lacked a statistically significant gRNA-mRNA match. Because the nucleotide preferences observed for assigned gRNAs had a direct correlation to the position of the gene within the minicircle sequence, landmarks for independent gRNA predictions in the absence of known editing events were derived. To do so required a statistical model and a method of finding the optimum path of a sequence through that model. In this case the Viterbi algorithm was used to find the best path through a Hidden Markov Model (HMM) for each known T. cruzi minicircle sequence (Fig 8, 9).

The 108 minicircle reads carrying gRNAs predicted by local alignments were first used to test the ability of this method to correctly predict those gRNAs. Although the 5' and 3' ends show differences with those predicted by local alignments, 85% of HMM-predicted gRNAs overlapped the local alignment predictions by at least 20 nt (Additional File 4). The remaining 15% contained very little or no overlap with the local alignment prediction. This method was then applied to all known T. cruzi minicircle sequences to predict the gRNAs (Additional File 5) and their locations within each minicircle sequence (Fig 9).

Compared to the local alignment placements (Fig 7), the nucleotide preferences surrounding the HMM predictions were even more visually pronounced (Fig 9). While this was to be expected given the neutrality of the local alignment method to nucleotide usage and the reliance of the HMM on context-dependent usage, there was not enough empirical evidence regarding T. cruzi gRNAs to characterize the accuracy of each prediction method. The validity of these predictions will be determined with further experimental editing information. Some of these sequences may also represent pseudogenes.

Each variable region class was assigned a unique identifier. A single T. cruzi minicircle read could contain a string of up to four such class numbers. In general the functional target mRNA of a gRNA had no bearing on its structural location relative to other gRNAs on a minicircle sequence, similar to the situation in T. brucei. Recombination among minicircles had not been documented, so this linkage information inherent within the sequence reads was examined for apparently contradictory links between sequence classes.

Two examples of discontinuity were found. Reads CLARO12TF and TCGA393TF contain identical downstream variable regions, but the upstream variable regions, separated by the intermediate CSBs, are completely different (Fig 10). The same was true for clones CLAOS79 and CMCBV30. These alternate linkages were evidence of recombination among minicircles.

The minicircle sequences presented represent a fraction of minicircle sequence and as two of 54 CL Brener reads bore evidence of recombination extrapolated to a level of 4% recombinants in the overall minicircle population. However, the minicircle reads presented do not represent complete coverage, and few reads contain three, let alone all four, linked variable regions. This incomplete linkage information is likely to compound the already limited scope of the data, and for these reasons 4% represents a conservative and imprecise estimate for the percentage of minicircle recombinants.

In order to identify the intragenomic variability of CSB sequences within the conserved regions, a two-tag discovery strategy was employed using the WGS data from CL Brener and Esmeraldo. The first step was to search restrictively for two tags (CSB-1 and CSB-3 for example) while capturing all of the variability of the third tag between the two landmarks (CSB-2 in the previous example). Employed combinatorially, the intragenomic variability of CSB-1, CSB-2, and CSB-3 was determined. From this information, search parameters were designed to allow positive identification of any number of full CSB-CSB units within any query sequence, without excluding sequences on the basis of slight intragenomic CSB variation. This method was 'greedy' in that if a given sequence contained multiple units it captured all of them at once, thereby preserving any physical linkages between variable regions. To assemble mate pairs with overlapping sequence, a standard overlap alignment algorithm was implemented in the form of a perl script to assemble sequences. Once assembled, the contiguous sequences were each checked by eye using BioEdit. CSB conservation was depicted using Weblogo 38, and spectrograms were constructed using CLUSTALX (alignments), BioEdit (alignment editing and visualization), and Adobe Photoshop (building composites of BioEdit screenshots for spectrograms).

A primer complementary to the predicted 3' end of COIII (2 pmol) was added to dNTPs (1 μL of a 10 mM mix) and RNA (1 μg) from CL Brener strain cells in a 12 μL reaction and heated to 65°C for 5 minutes, then placed on ice. For a 20 μL reaction, 4 μL of 5× First Strand Buffer (Invitrogen) and 2 μL of 0.1 M DTT were added. The reaction was warmed to 42°C in a thermocycler for 2 min. and 1 μL of SuperScript™ II (Invitrogen) was added and the tube gently mixed. The reaction was incubated at 42°C for 50 min. PCR was then performed on the RT reaction and a no-RT control using a 55°C annealing temperature. A ~900-bp band seen with ethidium staining represented the full length edited COIII cDNA that was then cloned using a TOPO TA cloning kit (Invitrogen) and sequenced. No signal appeared in the control lane after 30 cycles. Primers used: COIII Forward, TATATTTGTTGGTGTTAGTGG; COIII Reverse, TTATACACACAAATACATAACG.

The alignment algorithm used was a perl implementation of the Smith-Waterman dynamic programming method of determining the optimal hybridization of a query sequence with a target sequence 39. Local alignments allow overhangs on the part of the target and query without penalty, ensuring that in silico hybridizations of the gRNA with the mRNA could be detected even though the gRNA makes up only a portion of the minicircle query sequence and would hybridize with only a small region of the target mRNA sequence. The only peculiarity of the algorithm used here was the scoring matrix: it was designed to detect complementary (not identity), it allowed for the fact that non-canonical G-U base pairs are common, and it used a large gap penalty, reflecting the expectation that the gRNA should be exactly complementary to the target sequence it edits. For each minicircle sequence, optimal local alignments were performed against the fully edited sequence (predicted or actual) of every edited maxicircle gene and the information from the best alignment was saved to determine its significance (Fig 3A). Although many matrices were tested, the weight matrix used here evaluated a match (A-U, G-U, C-G) and with a score of 4, a mismatch (A-A, C-C, G-G, U-U, A-C, A-G, U-C) with a score of -20, and assessed a gap penalty of 100.

In order to determine the probability of a given 'best' local alignment score, the distribution of such scores is required for sequences of a given length and composition against a standard mRNA library. Given these factors this distribution is neither standard nor easy to derive. The method used here to address this problem is the permutation test 404142 implemented in perl. If a minicircle sequence is permuted, losing all information while retaining the overall nucleotide frequencies, the resulting 'best' score against the mRNA library will be drawn at a certain probability from the unknown but inherent distribution of scores. As the number of permutations sampled approaches infinity, the approximation of that unknown distribution approaches the actual distribution. In this manner the probability of observing a given best local alignment score was empirically estimated using 1000 gRNA permutations (Fig. 3B). Although extremely computationally intensive, this remains the best method of approximating probability distributions that are difficult to derive mathematically.

When multiple tests are performed with each allowing a 5% chance of a false positive, the overall number of false positives received can be quite high. Many methods have been suggested to control for multiple testing and several have been broadly applied in biological research 4344. The false discovery rate (FDR) method of correcting for multiple sampling 45 was most appropriate for this study given the expectation of discovering exactly one appropriate gRNA match for each minicircle sequence. To perform FDR control (Fig. 3C), where 'X' is a data set including 'n' points, X is ranked in decreasing order of probability (from 0 to 1) by probability such that P(X_n) > P(X_n-1)>... X₁. Then for i = n → 1, where α = 0.05, the following stepwise algorithm is applied: if at any step i⋅αn>prob(Xi) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaadaWcaaqaaiabdMgaPjabgwSixJGaciab=f7aHbqaaiabd6gaUbaacqGH+aGpcqWGWbaCcqWGYbGCcqWGVbWBcqWGIbGycqGGOaakcqWGybawdaWgaaWcbaGaemyAaKgabeaakiabcMcaPaaa@3E7A@ then k = i. Reject the null hypotheses for all points from X₁ to X_k. This analysis was performed in perl, with the results visualized by Microsoft Excel (Fig 3D).

HMMs use the known 'emission' probabilities of an observed variable to predict a 'hidden' variable with defined 'transition' probabilities to and from all possible states of that hidden variable. In this case, the differing nucleotide preferences along the minicircle sequence are used to determine the probability that a given nucleotide is positioned within the gRNA state, the hidden state we wish to know. The HMM constructed here consists of four distinct states each with characteristic nucleotide preferences (Fig 8). The upstream (U) state is itself a collection of 40 individual states each with identical nucleotide probabilities and a definite transition onwards to the next sub-state with a probability of 1. After beginning in the 'start' state, a gRNA sequence will enter the upstream state and invariably transition to the gRNA state, the final sequence of which constitutes a gRNA prediction. A nucleotide in the gRNA state can transition back into a gRNA state with a high probability or can transition on to the downstream region which itself is made up of two states, D1 and D2. Like the upstream region the two downstream regions are themselves composed of multiple sub-states with identical nucleotide probabilities, with D1 containing 48 nucleotides and D2 containing 25. The sequences finally leave the model through an 'end' state. The nucleotide probabilities for each state were calculated using the predicted gRNA alignments generated by local alignments (Fig 7). The optimal chain of these states through each minicircle sequence can be recognized as the chain with the highest cumulative probability under the assumption that the probability of transition to a subsequent state depends only upon the nature of the current state and assuming that there is independence among each observed point in the chain. While neither of these assumptions is precisely true for nucleotide sequences, HMMs have been used successfully to predict genes and alternative splicing 3132. A perl implementation of the commonly used Viterbi algorithm was applied here to determine this optimal path through each minicircle sequence 46.