Mapping split reads.

DIPScan begins by mapping the sequencing reads to the user-provided reference genome. This is achieved using a two-step approach that accounts for split reads spanning large deletions. In the first step, the reads are mapped to the reference genome using BWA-MEM2 [41]. From the resulting alignment, only the reads that map with at most 8 clipped nucleotides (unmatched read extremities) are kept. This produces a classical alignment, containing reads that map perfectly, as well as those containing SNPs or short insertions/deletions and less than 8 clipped positions.

In the second step, all unmapped reads and those that map with at least 8 clipped nucleotides are extracted, as they can define split regions in the case of DelVG. These reads then undergo STAR alignment [42] (Spliced Transcripts Alignment to a Reference, v2.7.11a). Although STAR was originally designed for aligning RNA-Seq reads across splicing sites, it is also highly effective at accurately mapping split reads flanking large deletions (with specific run parameters, see Text A in S1 Appendix), with read portions on either side of the deleted sequence, that are considered as “non-canonical splice sites”. This makes STAR well-suited to identifying defective genomes of interest, as these split reads often indicate large-scale structural variations within the genome, such as deletions, insertions, duplications, or translocations.

The outputs of these two steps are merged together in a single mapping file (BAM format) containing full reads mappings and split read mappings.

Extraction of deletion boundaries.

The mapping file is then used to identify potential deletion boundaries by extracting reads containing “skipped regions” (denoted by ‘N’ in the CIGAR string of the SAM format) of more than 150 nucleotides. These “skipped reads” each span deleted regions, defining start and end coordinates relative to the reference genome. The resulting position pairs are consolidated into a non-redundant set for downstream analysis.

Computing defective metrics.

Using this set of unique start-end position pairs, various metrics are calculated. For each pair, these metrics include:

• Supporting reads: The number of reads spanning the exact deletion coordinates.
• Split frequency: The proportion of supporting reads relative to the total number of split reads (at any deletion coordinate). This value indicates the relative frequency of this particular deletion compared to other deletions.
• Total frequency: The proportion of supporting reads relative to the total number of mapped reads. This value indicates the relative frequency of this particular deletion compared to the whole sequencing dataset.
• Expected minimum frequency: The expected Total frequency if the proportion of DelVG was 2%. This value corresponds to 2% of read length divided by the non-deleted region size (reference genome size minus deletion length).
• Deletion start ratio: The local deletion-start prevalence at a breakpoint is quantified as the ratio of mean coverage over the 5 nucleotides downstream to the mean coverage over the 5 nucleotides upstream of the breakpoint start.
• Deletion end ratio: The local deletion-end prevalence at a breakpoint is quantified as the ratio of mean coverage over the 5 nucleotides upstream to the mean coverage over the 5 nucleotides downstream of the breakpoint end.

Breakpoint selection.

In order to filter out lowly prevalent breakpoints that may add noise in the data, the following default filters (though adjustable depending on the sequencing depth) are applied to ensure the robustness of detected breakpoints:

• Supporting reads should be greater than 100 reads. This criterion eliminates background noise.
• Total frequency should be greater than the Expected minimum frequency. This removes breakpoints that are poorly represented relative to the entire dataset.
• Both Deletion start ratio and Deletion end ratio must be smaller than 1. This removes meaningless breakpoints having inner-coverage higher than outer-coverage.

The filtered list may include multiple breakpoints that delineate various deleted regions and, notably, several DelVG sequences.

DelVG proportion estimation.

The potential risk posed by DelVGs in viral consensus reconstruction is particularly significant when DelVGs are dominant relative to the full-length genome. In such cases, a potential mutation may become dominant and subsequently incorporated into the final consensus sequence. Estimating the proportion of each potential DelVG is therefore important.

For a given segment, the n previously selected unique breakpoint positions (either starts or ends) divide the segment into n + 1 distinct regions. The median nucleotide coverages, denoted as , are computed for all the regions . Because coverage frequently declines at the edges of the first and last regions, the median coverage is calculated after excluding the first (or last) 150 nucleotides of these regions, respectively. If the first (or last) region is shorter than 150 nucleotides, the median coverage of the last (or first) region is used instead. If both regions are shorter than 150 nucleotides, the higher coverage value between the two is selected.

Given that the n breakpoint positions represent d distinct DelVGs (some breakpoints starts or ends may be shared across multiple DelVGs), our objective is to estimate the relative abundances of each DelVG () and of the full-length segment (d + 1) in the sample. These estimates should best explain the observed coverages across the n + 1 regions.

This estimation is formulated as a system of n + 1 linear equations, where each equation models the coverage of a region r as a weighted sum of the contributions from all sequences (DelVGs and the full-length segment):

Here, if region r is present in sequence i, and 0 if deleted, while for the full length sequence (i = d + 1), for all regions.

Beyond coverage of all regions, we also leverage split-read counts at each breakpoint, where each split-read count represents the number of reads spanning the breakpoint present in DelVG i. Using these split-reads counts, we calculate the relative proportion of each DelVG compared to all DelVGs:

To ensure consistency, the estimated DelVG abundances should also satisfy:

This constraint can easily be reformulated as:

yelding d additional equations that are integrated into the system. An illustrative example with 3 DelVGs is given in Fig A in S1 Appendix. Because these equations have no exact solution (no abundance coefficients satisfy all constraints), we estimate the sequence abundances that best match the data using Non-Negative Least Squares (NNLS) [43], as implemented in Scipy (v1.15.3).

After solving, the abundance coefficients are normalized by their sum to obtain the proportions of each DelVG and the full-length segment.

A sample is flagged as containing relevant DelVGs if the sum of DelVG proportions (excluding the full length segment) exceeds 50%.

Consensus correction.

Consensus correction relies on a predefined list of candidate mutations. To generate this list, DIPScan first identifies potential variant positions relative to the reference genome using iVar (v1.3) [8], applied to a filtered BAM file from which noisy split reads (those not mapping to selected breakpoints) have been removed. This produces a tabulated output containing all candidate mutation sites, along with comprehensive details for each, such as the mutation type (SNP, INDEL, reference/alternative nucleotides), the number of reads supporting the alternative allele, the total read coverage at the position, the statistical confidence metrics for the mutation, etc. This information then feeds into the consensus correction pipeline, which comprises the following steps, as illustrated in Fig 3:

1. 1) Identification of regions to be corrected The breakpoints identified previously are used to define the maximal region that may harbor dominant DelVG mutations and therefore necessitates correction. This maximal region is composed of two parts:
   • The start region (orange sequence on Fig 3): spanning from position 0 to the right-most position among the start positions of the breakpoints.
   • The end region (blue sequence on Fig 3): extending from the left-most position among the end positions of the breakpoints to the end of the reference segment.
2. 2) Selection of mutated positions DIPScan selects positions located within either the start or the end region of the genome (in red in Fig 3) that may need correction. To do so, the initial consensus sequence is aligned against the reference genome using MAFFT [44] (v7.525 -auto) in order to match the positions of the user-provided consensus with the positions of the reference genome. Then, using the alignment along with the called mutations from iVAR output, DIPScan selects positions that are covered by at least 10 reads, and specifically for the positions of the consensus with SNPs, also those that are affected by a statistically significant mutation (with a ), either major or minor, as computed by iVAR.
3. 3) Choosing the right nucleotide At this final step, for a given mutated position, we aim at associating each observed nucleotide to one or several sequences (DelVGs and full length), and use this information to incorporate the nucleotide associated to the full length sequence in the consensus. This can be done by leveraging the estimated frequency of each sequence (DelVG and full length), and the frequency of each nucleotide at the given position. But first, we can already resolve straightforward cases:
   • If the estimated proportion of the full-length genome is below a mutation threshold (, by default ), then the genome is considered too low in frequency to confidently determine the correct nucleotide at any position called by iVAR, and an N is incorporated at all the mutated positions in the consensus.
   • If the mutation frequency falls within the ambiguous range (by default, ), the mutation cannot be reliably identified, and an N is incorporated in the consensus.
   • If (major mutation) and (the full length sequence is dominant), then the mutation is retained in the consensus.
   • If (minor mutation) and (the full length sequence in not dominant) and the frequencies of full-length () and DelVG () variants do not overlap (i.e., , with by default) and the frequency of the mutation, , is compatible with the full length frequency (i.e., , with by default), then the minor mutation is incorporated in the consensus sequence.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Schematic representation of the consensus sequence correction process.

A. Defining regions to be corrected. Based on two previously identified breakpoints (coordinates [s1,e1] and [s2,e2]), two regions are flagged for potential correction: the start region (in orange) spans from coordinate 0 to the rightmost breakpoint start (s2), and the end region (blue) spans from the leftmost breakpoint end (e1) to the end of the reference genome. The central sequence (in black), with coordinates [s2,e1] is protected from correction as it is presumed to come exclusively from reads originating from the full length genome. B. Mutation detection. The initial consensus sequence is aligned against the reference genome. Mutations that fall within the designated correction regions are identified (highlighted in red). C. Correction of the identified positions in the consensus. Each identified mutation is resolved through one of the following ways: (i) replacement with an ambiguous character ‘N’, (ii) retention of the mutation when it is presumed to originate from the full length, or (iii) substitution with the probable full-length nucleotide when the mutation is attributed to a DelVG, the latter determined by comparing its frequency to the coverage of the DelVG sequence.

https://doi.org/10.1371/journal.pcbi.1014115.g003

All other cases correspond to more complex scenarios, such as the presence of multiple DelVGs along with the full length sequence, where sequence frequencies must be combined to match to the frequency of a given nucleotide. In this scenario, for a given mutated position, we try to match the detected sequences (DelVGs and full-length) with the possible nucleotides found in the reads at this position, such as the sum of the sequence frequencies associated to a given nucleotide is approximately equal to the frequency of the nucleotide. This corresponds to a variant of a well-characterized problem, known as the “Multiple Subset Sum”, and can be formulated as “For each label (nucleotide), identify a subset of objects (sequences) whose combined frequencies match the label’s frequency”. While this is an NP-hard problem, heuristics (such as simulated annealing or MCMC) can be used to find solutions. Moreover, since the number of sequences and unique nucleotides is very small in our case, a solution is findable very quickly. The solution is then used to assess whether the full length sequence can be associated unambiguously with one nucleotide (the nucleotide is then incorporated in the consensus) or not (a N is incorporated in the consensus).

Additionally, to prevent misaligned reads from being mistaken for mutations, we compare the read depth at each potential mutation site with the expected depth accounting for the full-length genome and all defective genomes spanning the position). If the mutation is in a lowly covered region compared to the expectation (the ratio of the position depth and the expected depth is less than 80%), then the mutation is considered unreliable and an N is incorporated.

Simulated dataset.

To evaluate DIPScan in a controlled setting, we simulated sequencing datasets containing DelVGs (see Fig 4). We started with 11 influenza consensus genomes (including all segments) from the real influenza dataset described in the next section, representing various subtypes. These genomes are referred to as Full Length True Genomes. We then randomly removed internal regions from one segment (either PB1, PB2, or PA) of these genomes, to create DelVG True genomes. Next, we introduced between 1 and 10 new mutations into each DelVG True genome, making them distinct from their associated Full Length True Genome. These modified genomes are called DelVG mutated Genomes. For each Full Length True Genome and its corresponding DelVG Mutated Genome, we simulated 10 datasets using ReSeq [45] (command reseq illuminaPE –maxFragLen 500) with a full-length/DelVG ratio (based on average coverage per position) ranging from 100% to 10%. This approach allowed us to test both the sensitivity and the precision of DIPScan. This resulted in 110 datasets, further described on Text B and Table A in S1 Appendix.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Simulation workflow.

Schematic illustration of the simulation workflow for generating mixed full-length and defective genome samples. Starting from a known full-length genome (with 8 segments), an internal sequence of one segment is deleted to create a DelVG sequence, and 1 to 10 random mutations are introduced. Sequencing reads are then simulated from mixtures of the full-length and DelVG genomes at nine distinct ratios(from 100%:0% to 10%-90% full-length:DelVG), yielding ten simulated samples per starting consensus genome.

https://doi.org/10.1371/journal.pcbi.1014115.g004

In parallel, we built two low coverage simulated datasets by sub-sampling each of the 110 simulated samples with a reduced coverage of 10% (500k reads) and 1% (50k reads) of the original sample. These datasets are used to assess the performances of DIPScan on extreme cases (which do not reflect a common occurence, given the short length of the viral genomes, and which may not even be adequate to study viral consensus in the presence of a high coverage DelVG).

Real influenza virus dataset.

We applied DIPScan on 551 real Influenza virus sequencing dataset sequenced at Institut Pasteur’s NRC (see Text C in S1 Appendix). It comprises 551 samples collected between 24/02/2025 and 04/04/2025, and consisting of: 125 A/H1N1pdm (22.7%), 232 A/H3N2 (42.1%), and 194 B/Victoria (35.2%). All these samples were visually inspected, to search for DelVG patterns (high coverage on the extremities and low coverage in the central region). Among these, 55 A/H1N1pdm (44%), 74 A/H3N2 (31.9%) and 103 B/Victoria (53.1%) showed patterns typical of DelVGs resulting from large deletion.

Breakpoint detection and tool comparison.

The first aspect of DIPScan we assessed was its ability to precisely identify DelVG breakpoints within input datasets. Since ViReMa, DG-Seq, and VODKA2 are specifically designed to detect these breakpoints, rather than estimate DelVG proportions or refine consensus sequences, this presents an ideal opportunity to benchmark DIPScan’s performance against these tools. For this comparison, we applied DIPScan, ViReMa, VODKA2, and DG-Seq to each simulated dataset and assessed their performance by matching the identified breakpoints with the ground-truth breakpoints, i.e., those predefined in the simulations. We ran all tools with their default parameters and analyzed their outputs without further processing. Consequently, our analyses incorporated each tool’s built-in filtering criteria. For example, DIPScan by default only reports breakpoints separated by more than 150 nucleotides. Hereafter, breakpoints reported by the tools that coincide with simulated breakpoints are considered true positives; reported breakpoints without simulated matches are considered false positives; and simulated breakpoints absent from tool outputs are considered false negatives.

As shown in Fig 5, ViReMa, DG-Seq, and VODKA2 detected 1,060, 582, and 270 breakpoints, respectively, of which only 46 (4.3%), 99 (17%), and 64 (23.7%) were true positives. In contrast, DIPScan identified 93 breakpoints, all of which were correct, while missing 6 true breakpoints. Among these missed breakpoints, four originated from low-frequency DelVGs (10%), one from a 20% DelVG, and one from a 30% DelVG, likely due to inaccurate DelVG proportion estimates or insufficient breakpoint read support, below the detection threshold. These 6 breakpoints do not constitute a problem in terms of consensus correction.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Evaluation of DIPScan for breakpoint detection.

Performance evaluation of DIPScan, ViReMa, DG-Seq, and VODKA2 in detecting DelVG breakpoints. (A) UpSet plot illustrating the overlap of breakpoints detected by the three tools and the true ones. For instance, 895 breakpoints are uniquely identified by ViReMa, while 38 breakpoints are correctly detected by all tools. Columns shaded in gray represent false breakpoints. (B) Breakdown of detected breakpoints per tool along the true breakpoints, categorized as true positives (TP), false positives (FP), and false negatives (FN), accompanied by key performance metrics: sensitivity, precision, false discovery rate (FDR), and F1-score.

https://doi.org/10.1371/journal.pcbi.1014115.g005

DG-Seq achieved perfect sensitivity (100%), detecting all true junctions, but at the cost of 483 false positives, resulting in low precision (17%), a high false discovery rate (83%), and an F1-score of 29.1%. By comparison, DIPScan demonstrated superior performance across key metrics, with perfect precision (100%), no false discoveries (FDR: 0%), and a global F1-score of 96.9%.

It is worth noting that all methods except DIPScan incorrectly detected many breakpoints in the 0% DelVG simulation, which contained no true breakpoints. While specific filtering criteria could eliminate many false positive breakpoints, optimization of these thresholds falls outside the scope of this study.

DelVG proportion estimation.

On the same simulated data, DIPScan accurately estimates the proportion of DelVG sequences. Correlation between the known ratio and the estimated ratio is high (Fig 6A, Pearson correlation of 0.99 between average estimated ratios over the 11 samples at each known ratio and the known ratio).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Evaluation of DIPScan for DelVG proportion estimation and consensus correction.

A. Accuracy of DelVG proportion estimation. The mean distribution and confidence intervals of DelVG ratios estimated by DIPScan (y-axis) is shown for the 11 samples together across each simulated input proportion (x-axis). B. Correction performance for DelVG-specific mutations. The proportion of corrected DelVG-specific mutations (y-axis) as a function of the simulated input DelVG portion (x-axis) is represented for the three following outcomes: replacement with an ambiguous base (gray curve), correct reversion to (or retention of) the full-length genome nucleotide (blue curve), or erroneous retention of the DelVG-specific mutation (yellow curve). As expected, ambiguities peak around a 50% DelVG proportion, while correct retention or reversion peaks at both low DelVG proportion (where the full-length genome is dominant) and high DelVG proportion (where the DelVG is dominant, triggering a reversion to the full-length base).

https://doi.org/10.1371/journal.pcbi.1014115.g006

Except for two samples (see Fig B in S1 Appendix), the estimations were slightly under estimated compared to the known proportion, although close to the expectation.

Regarding the sensitivity to low coverage, the Pearson correlation is still good (99%) on the two poorly covered simulated datasets (500k and 50k reads), over the different ratios (see Fig C(a) in S1 Appendix).

Across all combined datasets, DIPScan demonstrated an overall good accuracy of 94.5% (104/110, six incorrectly classified samples were false negative with low DelVG proportions) in classifying samples as containing DelVGs or not.

Mutation detection and consensus correction.

In our simulation scenario, three kinds of mutations can in theory be identified relative to the reference genome: (i) Mutations that are shared between the DelVGs and the full length sequence, originating from the original consensus sequence), (ii) mutations that are specific to the DelVGs, randomly introduced, and (iii) mutations that are specific to the full length sequence, if the randomly introduced nucleotide reverts the nucleotide to the reference. The latter case did not happen in our simulations.

As shown in Fig 6B and Fig D in S1 Appendix (percent and absolute values, respectively) and Fig E in S1 Appendix, DIPScan demonstrates a strong performance in handling common and DelVG-specific mutations, respectively. This trend is confirmed in simulated samples with low-coverage, although results show a higher detection difficulty and a greater variability for low-proportion DelVGs (see Fig C(b) in S1 Appendix).

First, for shared mutations, where no consensus correction is expected, DIPScan demonstrated high fidelity, correctly retaining 99.6% (3976/3980) of the mutations across all samples. Among the 14 changed positions, 11 were masked (replaced by Ns) as these mutations fell just below the default 98% threshold for fixed mutations. The 3 incorrect changes involved samples where the mutations coincidentally displayed the same full-length/DelVG ratio between two nucleotides. This led DIPScan to misclassify them as DelVG-specific mutations.

Then, DIPScan demonstrated good performance in detecting all DelVG-specific mutations introduced into the test samples. Across all samples, it correctly retained or corrected 73.4% (426/580) of full-length mutations, introduced Ns (ambiguous bases) for 23.8% (138/580), and misassigned 2.8% (16/580). Performance varied across the expected DelVG ratio. In the [50%, 90%] range, where DelVG presence complicates consensus reconstruction, DIPScan correctly retained or corrected 59.3% (172/290) of full-length mutations, assigned Ns for 37.9% (110/290), and incorrectly changed 2.8% (8/290). In contrast, for the [0%, 50%[range, performance improved, with 87.6% (254/290) of mutations correctly retained or corrected, Ns introduced in 9.7% (28/290), and misassignments remaining at 2.8% (8/290).

Notably, all misassigned nucleotides occurred in the same three samples where DelVG proportions were poorly estimated.

Altogether, these results demonstrate the good performances of DIPScan on simulated datasets at detecting DelVGs, their proportion, and correcting the consensus sequences accordingly.

DIPScan detection accuracy.

To assess DIPScan detection performances, we classified the results into five categories, when compared to the “coarse-grained” manual dataset (Fig 7):

• Consistently defective: both manual detection and DIPScan identified DelVGs in the sample;
• Consistently not defective: neither manual detection nor DIPScan identified any DelVGs in the genome;
• Inconsistent, defective in manual: DelVGs were identified manually but not by DIPScan. These cases are detailed below, as they may represent either a limitation of DIPScan or manual detection errors.
• Inconsistent, Defective with DIPScan (<50%): DIPScan identified DelVGs representing less than 50% of the genomes, which manual detection might miss due to overall coverage similarities between and outside the deletion breakpoints. This category is not considered as errors but rather reflects different strategies in defining DelVGs.
• Inconsistent, Defective with DIPScan (): DIPScan identified DelVGs representing more than 50% of the genomes. This category is detailed below, as it may indicate DIPScan’s oversensitivity or manual detection errors.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Evaluation of DIPScan on real datasets.

Evaluation of DIPScan detection on real Influenza virus sequencing dataset: Manual curation versus DIPScan. Results are categorized based on agreement between methods: Consistent (both methods agree on the presence (light green) or absence (dark green) of a DelVG) or Inconsistent. Inconsistent results are further broken down into: (i) DelVG detected only in manual curation, (ii) DelVG detected only in DIPScan with an estimated proportion below 50%, and (iii) DelVG detected only by DIPScan with a proportion above 50%. The results are grouped by Influenza virus segment and subtype: A. A/H1N1pdm (125 samples), B. A/H3N2 (232 samples), and C. B/Victoria (194 samples). Overall, we observe a good consistency between DIPScan results and manual curation. DelVGs detected only by DIPScan, at a proportion below 50% were usually missed by manual curation.

https://doi.org/10.1371/journal.pcbi.1014115.g007

Across the 4,408 analyzed segments (551 samples), DIPScan and manual detection agreed on 92.8% of segment classifications (Table 1A). Agreement rates varied by segments: 418 consistent PB1 (75.9%), 438 PB2 (79.5%), 487 PA (88.4%), 549 HA (99.6%), 548 NA (99.5%), 549 NS (99.6%), 551 MP (100%), and 549 NP (99.6%). Discrepancies included 33 segments classified as defective only in the manual dataset and 286 detected solely by DIPScan (Table 1A). Notably, no inconsistencies were found in the HA, NA, NS, NP, or MP segments that could be considered potential errors. Using manual curation as ground truth, DIPScan achieved 91% sensitivity and 54% precision (92.8% accuracy). However, most disagreements involved DIPScan detecting low-frequency DelVGs (<50%) missed during manual inspection. When considering only high-frequency DelVGs (relevant for consensus reconstruction) as true positives (Table 1B), DIPScan’s performance improved to 99% sensitivity and 88% precision (98.9% accuracy). Further review showed most remaining discrepancies were either overlooked in manual curation or negligible low-frequency DelVGs, with 49 cases remaining unclassifiable (Table 1B). These unclassifiable cases presented a complex scenario where it was unclear whether manual review or DIPScan was correct.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Confusion matrix comparing DIPScan and manual classification of 4,408 segments from 551 samples.

https://doi.org/10.1371/journal.pcbi.1014115.t001

In particular, among the cases where DIPScan detected DelVGs with a proportion but manual review did not (Inconsistent, Defective with DIPScan ()), 12 samples belonged to the A/H1N1pdm subtype. Upon re-evaluation, all instances affecting the PB2, PB1, and/or PA segments (n = 15) were found to be potential manual errors, primarily due to the difficulty in visually evaluating the proportions of full-length genomes over DelVGs. For the A/H3N2 subtype, 11 samples were identified, with all cases affecting the PB2, PB1, and/or PA segments (n = 13) ultimately attributed to potential manual errors. For the B/Victoria subtype, 29 samples were identified, all affecting the PB2, PB1, and PA segments. In 12 cases, the inconsistency was due a complex scenario likely involving a defective genome with a single breakpoint, resulting in truncation of the entire 5’ part. One case was due to a manual report error since a DelVG was unambiguously visible, and 18 cases were found to be also potential manual errors.

In summary, despite using similar detection criteria (identification of breakpoints or sharp coverage declines greater than 50%), key differences between manual curation and DIPScan emerged:

• Sensitivity to low-abundance DelVGs: DIPScan detects minor low-proportion DelVGs much more effectively than manual curation. This limitation led us to focus, in a second step, on DelVGs present at more than 50% abundance.
• Subjectivity in manual assessment: while both methods use a 50% cutoff to flag problematic DelVGs (to be excluded or corrected), visual inspection is less robust, as the estimated ratio of full-length genomes to DelVGs is less accurate, and the curator bias plays a significant role.
• Complex cases remain challenging: neither method perfectly resolves ambiguous or highly complex cases

DelVG proportion estimation.

Direct assessment of DIPScan’s accuracy in estimating DelVG proportions in real viral samples is inherently challenging due to the absence of a known ground truth. However, an indirect proxy for the true DelVG proportion can be derived from samples containing DelVG-specific mutations and/or full-length genome-specific mutations. These mutations produce mixed nucleotide signals at specific genomic positions, which can serve as a quantitative indicator of the relative abundance of the corresponding viral genomes within the sample.

We identified candidate mutations by analyzing the real datasets for variable positions (with several possible nucleotides), while systematically excluding regions between breakpoints, and the 150-nucleotide terminal segments of the genome. Using this approach, we identified 233 samples containing a total of 8,117 variable positions, with proportions likely attributable to either DelVGs or full-length genomes. Fig 8 presents the comparison of the DelVG proportions estimated by DIPScan to the observed mixed-position frequencies serving as a proxy for true proportions. The resulting Pearson correlation coefficient (r = 0.96, ) indicates a high agreement between DIPScan’s estimates and the empirical proxy, supporting DIPScan’s accuracy in quantifying DelVG abundance.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Evaluation of DIPScan proportion estimation on real datasets.

Ratios between DelVG-specific coverage and full-length sequence coverage estimated by DIPScan (y-axis) are compared to the ratios derived from DelVG-specific (or full-length-specific) mutation frequencies at highly unambiguous positions (x-axis). Each point corresponds to the average over potentially multiple mutations for a given sample. The high correlation (0.96) between DIPScan estimates and ratios derived from unambiguous positions validates its use for downstream consensus correction.

https://doi.org/10.1371/journal.pcbi.1014115.g008

Consensus correction.

Across the whole dataset, 333/627 defective segments (53.1%) were subjected to correction of at least one position (41.4% for PB1, 25.5% for PB2, 32.1% for PA, 0.3% for HA, 0.3% for NA, and 0.3% for NP). Among the corrected positions, DIPScan changed the nucleotide in 17.7% of the cases. In the other cases (82.2%), DIPScan incorporated an ambiguous nucleotide.

To assess DIPScan’s accuracy in correcting the consensus sequences, a ground truth dataset was build consisting of 25 positions across 19 samples where nucleotides could be unambiguously (manually) attributed to either the full-length genome or the defective genome.

As depicted in Fig 9, 24/25 positions (96%) in the selected ground truth dataset were correclty handled: 4 positions correctly changed, 13 positions correctly retained (13 unchanged), 7 positions appropriately masked, and only one position incorrectly changed, corresponding to a difficult case of DelVG proportion estimation. It is worth noting that DelVGs falling in the 45–55% range were not the most frequent (they represented 5–14% of the PA, PB1 and PB2 DelVGs in A/H1N1 and A/H3N2 samples, and 9–21% in PA, PB1, and PB2 DelVGs in B/Vic samples, see Table B in S1 Appendix).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Validation of consensus correction on real datasets.

Validation of consensus correction using 25 unambiguous mutations from 19 real Influenza virus datasets. Each of the 25 positions is classified by: (i) The proportion (estimated by DIPScan) of the detected DelVG covering the position (x-axis) and (ii) the outcome of the consensus correction: correctly changed (green), correctly unchanged (blue), masked with an ‘N’ (orange), or incorrectly changed (red). The y-axis shows the number of position in each category. We observe that globally, most of the positions (96%) are correctly changed, unchanged or masked, and 1 position is incorrectly changed.

https://doi.org/10.1371/journal.pcbi.1014115.g009

Masked positions resulted from: (i) Poor alignment (5 positions from 5 samples; detected via comparison of expected depth and coverage of the mutated position), (ii) Inability to match estimated full-length proportions to nucleotide frequencies (1 position from 1 sample), (iii) Estimated full-length proportions below the 2% noise threshold (1 position from 1 sample).

For samples with >50% DelVG content, the unexpected cases of correctly unchanged positions occurred when the initial alignment (used to generate the consensus) was imperfect and resulted in the coincidental introduction of full-length-specific nucleotide into the consensus instead of the dominant DelVG-specific nucleotide.

After correcting the consensus sequences, we assessed the effects of these adjustments on two key aspects: i) Changes in phylogenetic tree topology, ii) Modifications at drug resistance mutation sites. Since viral lineage assignment is predominantly based on the HA segment, a segment that is rarely affected by DelVGs, DIPscan correction is not expected to significantly influence lineage determination.

Phylogenetic tree topology: We inferred phylogenetic trees for the DelVG-affected segments (PB1, PB2, and PA) from H1N1, H3N2, and B viruses, and compared trees generated from uncorrected versus DIPScan-corrected consensus sequences. While the observed topological differences were modest, they were consistent (see Fig F in S1 Appendix), confirming that uncorrected sequences may introduce inaccuracy in phylogenetic reconstructions.

Drug Resistance Mutations: We screened the corrected defective segments for residue known to be associated with resistance to antivirals (oseltamivir, zanamivir, and baloxavir). None of the corrected positions corresponded to established drug resistance markers. Additionally, we examined other samples beyond our dataset for the presence of resistance mutations but we found no overlap with defective sequences. However, it remains a possibility that, in other datasets or with emerging resistance mutations, DelVG-associated sequence could affect the estimated frequency of antiviral resistance markers, particularly since PB1, PB2, and PA all harbor known or potential resistance-associated sites.

Breakpoint location hotspots.

With DIPScan results in hand, we could further characterize the DelVG sequences. In particular, we used the results from the 551 real dataset samples to examine the breakpoint positions along the genome and identify regions enriched with breakpoints (“hotspots”).

To do so, we grouped the results by virus subtype and segment, as they may exhibit distinct patterns of genomic deletions (see Fig 10, Fig G in S1 Appendix).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. Analysis of potential breakpoint hotspots.

The distribution of detected breakpoint ends (y-axis) is shown across the genomic segment, with the x-axis representing the location as a percentage of the segment’s total length. This distribution is colored by Influenza virus subtype of the 551 analyzed samples: B/Victoria in red, A/H1N1pdm in green, and A/H3N2 in blue. Only the 3 main affected segments are represented: A. PB1, B. PB2 and C. PA. This confirms that clear breakpoint hotspots exist near the segment ends of the three analyzed subtypes, and the three mainly affected segments (PB1, PB2, and PA).

https://doi.org/10.1371/journal.pcbi.1014115.g010

Our analysis demonstrates that DelVG breakpoints are clearly not uniformly distributed across the viral genome (Fig 10), instead exhibiting distinct “hotspots” that vary by segment and subtype. Moreover, the breakpoint start-end locations seem to be dependent from each other (see Fig G in S1 Appendix). These observations align with previously reported patterns of DelVG breakpoint localization in influenza [33,46,47].

More specifically, breakpoint distribution patterns revealed subtype-specific variations in DelVG generation depending on the segments.

For the PB2 segment (Fig 10B), all three subtypes displayed 5’ breakpoints predominantly at around 9% of the genome length, and a second peak was observed around 12% for A/H1N1pdm and A/H3N2, while B/Victoria exhibited a shift to 19%. The 3’ breakpoints clustered around 89% for all three subtypes, and a second peak was observed around 85% for A/H1N1pdm and A/H3N2, while B/Victoria exhibited a shift to 80%.

For the PB1 segment (Fig 10A), A/H1N1pdm showed 5’ breakpoints at 10%, A/H3N2 at 8% and 13%, and B/Victoria at 15%. A/H1N1pdm and A/H3N2 showed 3’ breakpoints at 78% and 86%, B/Victoria also shared a peak at 86% and exhibited a second one with a shift to 88%.

In the PA segment (Fig 10C), A/H1N1pdm showed 5’ breakpoints between 16–22%, A/H3N2 between 7–16%, and B/Victoria exhibited two main peaks at 19% and 24%. All three subtypes displayed 3’ breakpoints at 87%, then A/H1N1pdm and A/H3N2 had other peaks at 77% and 82%, and B/Victoria at a second peak at 79%, highlighting subtype-specific variations in DelVG generation.

While confirming these breakpoint hotspots, these results further validate the ability of DIPScan to accurately detect DelVG sequences, and its usage in a routine setting.