QoALa workflow and example datasets

The goals of the QoALa workflow are to standardize and comprehensively compare viral quasispecies and OTU profiles across multiple samples based on user customization. For each comparison, equal length read alignments from the same gene or genomic region of interest (Fig 1a) are separately imported from any genome assembly pipeline’s output (see examples in Methods) and noise-minimized using position-wise nucleotide base replacement in the “vqsassess” function (Fig 1b). The noise-minimization step replaces potential erroneous nucleotides, defined as SNVs with a frequency lower than n% of the total read depth (cut-off percentage), with either the majority base or the dominant haplotype’s base at that position. This cut-off percentage can be determined using prior information such as the sequencing error rate, or by observing the change in the percentage of singleton haplotypes at different percentage cut-offs (see Methods and Fig 1a). Additionally, when comparing multiple samples, all alignments must be randomly down-sampled by the “vqsassess” function to achieve equivalent read depth (the shallowest depth among all samples is recommended) to reduce bias in diversity measurement caused by sample size disparity. SNV profiles of the samples visualized by the “snvcompare” function should be inspected to evaluate the result of noise-minimization (Fig 1c). Finally, prepared samples are pooled by the “vqscompare” function to (1) identify common haplotypes, (2) reclassify haplotypes into OTUs, (3) visualize diversity profiles and genetic relationships among samples (Fig 1d), and (4) summarize quantitative diversity metrics.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. QoALa workflow and main functions of longreadvqs package.

a, Sequencing read alignment that completely covers a gene of interest is used as input for subsequent analyses in the longreadvqs package. The percentage cut-off, which is used to define low-frequency SNVs potentially as noise, can be estimated from prior knowledge or the percentage of singleton haplotypes in the alignment. b, Sample preparation involves noise-minimization, where nucleotide bases with frequencies lower than the percentage cut-off are replaced with the majority base (recommended) or the base of the dominant haplotype at the respective position, followed by read down-sampling for sample size standardization. The percentage cut-off can also be customized for error-rich regions. c, The SNV profile after sample preparation should be compared between samples to visually check for remaining potential noise in the alignment. d, All prepared samples are pooled to identify shared haplotypes and reclassify them into OTUs based on genetic distance. Green texts indicate key functions (italicized) and arguments in the longreadvqs package. Solid boxes represent main steps in the QoALa workflow, while dashed boxes denote optional steps.

https://doi.org/10.1371/journal.pcbi.1014208.g001

To benchmark the QoALa workflow under controlled conditions, we generated simulated long-read viral quasispecies datasets for SARS-CoV-2 and HIV-1 with known ground-truth haplotype composition. Simulations comprised five haplotypes with fixed frequency distributions across five independent replicates and sequences were generated in silico using representative ONT and PacBio error models. These datasets enabled direct, quantitative comparison of QoALa with alternative correction strategies—including Strainline [33], Racon [44], and uncorrected reads—by assessing accuracy in recovering true haplotype and OTU structures across sequencing technologies.

We further evaluated the QoALa workflow using read alignments assembled from publicly available long-read datasets representing five of the most extensively studied viruses in the viral quasispecies field: HCV, HBV, HIV-1, SARS-CoV-2, and influenza A virus (IAV). These datasets spanned diverse sequencing technologies (PacBio and ONT), sampling designs (cross-sectional and short- or long-term longitudinal), and depths of coverage per gene of interest, ranging from 130 reads in the IAV M gene segment (S6 Table) to 854,147 reads in the SARS-CoV-2 ORF3a gene (S5 Table). In this setting, QoALa was used to assess sensitivity to read downsampling and to re-examine quasispecies structure at standardized depths, enabling direct comparison across samples. Estimated sequencing error levels, inferred from singleton haplotype proportions prior to noise minimization, varied widely—from ~25% in the SARS-CoV-2 ORF3a sample to nearly 100% in the HCV and HIV-1 samples (S1 Fig). This reanalysis revealed quasispecies-level diversity patterns that were not revealed in the original studies and underscores the need for customized long-read preprocessing, particularly with respect to cut-off percentage and sample size.

QoALa improves recovery of viral quasispecies diversity and composition across sequencing platforms

Across simulated SARS-CoV-2 (S gene) and HIV-1 (env gene) quasispecies datasets, error-correction approaches showed metric-dependent performance patterns that were largely consistent across sequencing platforms (Fig 2c–2e). Based on Bray–Curtis dissimilarity, which captures agreement in both composition and relative abundance compared to the ground truth, QoALa achieved the lowest or near-lowest dissimilarity at the OTU level for both SARS-CoV-2 S and HIV-1 env datasets, generally reducing Bray–Curtis deviation by approximately 2–4-fold relative to uncorrected reads. For the most recent platforms (ONT 2023 and PacBio HiFi), QoALa and Strainline, and in some cases Racon, showed comparable OTU-level Bray–Curtis performance with no significant differences. However, all OTU analysis (regardless of which pipeline was used for error-correction) relied on the QoAla workflow, as OTU analysis is not a functionality of Strainline or Racon. At the haplotype level (i.e., no grouping of reads into OTUs), Bray–Curtis values were uniformly high and close to one for both viruses across all tools, indicating that none of the approaches reliably recovered true haplotype compositions (Fig 2c).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Benchmarking performance on simulated datasets.

The left half of the figure shows results for SARS-CoV-2 (S gene) and the right half for HIV-1 (env gene); within each half, haplotype-level results are shown on the left and OTU-level results on the right. a, Multidimensional scaling (MDS) plots based on pairwise SNV distances among haplotypes (points), with ellipses indicating the five largest OTUs in one replicate. b, Stacked bar charts showing the relative abundance of haplotypes and OTUs, using the same color scheme as in panel a, across the ground truth, uncorrected reads, and reads corrected by different methods in the same replicate as a. c, Median Bray–Curtis dissimilarity, d, Median deviation in the Gini–Simpson diversity index (H_GS), and e, Median deviation in nucleotide diversity (π_m), each calculated across five replicates between correction methods and sequencing platforms relative to the ground truth. Letters in panels c–e denote statistically significant differences between methods (FDR < 0.05). Sequencing platform abbreviations are as follows: ON18, Oxford Nanopore Technologies 2018; ON20, Oxford Nanopore Technologies 2020; ON23, Oxford Nanopore Technologies 2023; PBCL, PacBio continuous long reads (CLR); PBHF, PacBio HiFi.

https://doi.org/10.1371/journal.pcbi.1014208.g002

Evaluation using the Gini–Simpson index (H_GS), which summarizes within-sample diversity and evenness independent of quasispecies composition, also revealed contrasting trends between taxonomic resolutions. At the haplotype level, Strainline and Racon generally showed smaller deviations from ground-truth H_GS than QoALa for both SARS-CoV-2 S and HIV-1 env datasets, reflecting better preservation of population evenness, though the magnitude and statistical significance of these differences varied by sequencing platform. In contrast, at the OTU level—where genetically related haplotypes are grouped—QoALa consistently yielded the smallest H_GS deviations across both viruses and most sequencing platforms. While Strainline and Racon occasionally showed numerically smaller deviations than uncorrected reads at the OTU level, these differences sometimes was statistically indistinguishable from uncorrected reads (Fig 2d). These platform-dependent patterns indicate that the advantages of Strainline and Racon at the haplotype level do not consistently translate to OTU-level analyses.

Nucleotide diversity (π_m), a sequence-based measure of underlying genetic and functional heterogeneity, showed the clearest and most consistent separation among tools. Across both SARS-CoV-2 S and HIV-1 env datasets, QoALa exhibited significantly smaller deviations from true π_m than all other approaches at both haplotype and OTU levels, improving π_m recovery by two to three orders of magnitude relative to uncorrected reads, i.e., from ~3 × 10^-2–10^-1 in uncorrected reads to ~10^-4–10^-3. While Strainline and Racon significantly reduced π_m deviation at the haplotype level, their OTU-level π_m deviations remained similar to uncorrected data (Fig 2e), indicating persistent loss of fine-scale genetic variation that was minimized by the QoALa workflow.

Effect of down-sampling on diversity metrics

While read down-sampling is necessary for standardizing sample size before comparison, it may alter viral quasispecies diversity measures, especially at very low sample sizes. To assess this impact, we sub-sampled the real dataset reads with replacement and computed nine diversity metrics. Downsampling ranged from 10,000 down to 25 reads, depending on the original depth of each read alignment sample (Methods). This process was repeated 100 times per sub-sample size. This assessment was conducted twice: once for the alignments that were sampled after noise-minimization, and another for the alignments that were sampled before noise-minimization.

The Gini–Simpson index (H_GS) and nucleotide diversity (π_m) were the most robust diversity metrics, remaining relatively stable across all sample sizes regardless of the original alignment depth (Fig 3). In contrast, diversity metrics directly reliant on read depth [number of haplotypes (H) and mutation frequency at the molecular level (Mfm)], indirectly reliant on read depth [Shannon entropy (H_S), mutation frequency at the molecular level (Mfe), functional attribute diversity (FAD), and nucleotide diversity at the entity level (π_e)], or normalized by read depth [normalized Shannon entropy (H_SN)] were impacted to varying degrees by sample size (S2–S11 Figs). Down-sampling post noise-minimization from high-depth alignments (>10,000 reads) minimally affected H_S and Mfm at sample sizes >1,000 reads (S2, S4, and S8 Figs), whereas sampling from low-depth alignments (<1,000 reads) resulted in abrupt changes in these metrics at any sample size (S6 and S10 Figs). Computing metrics from alignments sampled after noise-minimization generally yielded less varied values compared to those sampled before noise-minimization in most instances (S3, S5, S7, S9, and S11 Figs).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Effect of down-sampling on key diversity metrics for five viruses (virus:gene).

a, Distribution of the Gini-Simpson index (H_GS) for each viral species: gene samples computed at different sample sizes (100 repeated random samplings with replacement for each size) after (top) and before (bottom) noise-minimization. b, Distribution of nucleotide diversity (π_m) with a layout similar to that of panel a. Bottom legends indicate the names of samples for each virus dataset. For all plots, horizontal lines indicate medians, boxes the 25th to 75th percentile, whiskers the lowest and highest values excluding outliers, and dots the outliers. Each dot indicates the mean of one target, boxes the 25th to 75th percentile, lines medians and whiskers extend from minimum to maximum values.

https://doi.org/10.1371/journal.pcbi.1014208.g003

Regarding inter-sample comparisons, variations in metrics due to sample size and sampling strategy marginally affected comparisons in scenarios where the actual diversity measures of samples were either highly similar (Figs 3 and S6–S11) or markedly different (Figs 3 and S2–S5). However, down-sampling to very low sample sizes (<100 reads) or down-sampling before noise-minimization likely resulted in false equivalences or differences because of the high variation in computed diversity metrics (Figs 3 and S2–S11). Thus, our analysis suggests that diversity metrics from samples with read depths of <100 should be interpreted with caution.

Example scenarios of diversity comparison

Since a single mutation can differentiate one haplotype from another, and such mutations may introduce noise and obscure the true frequency of different major genetic variants within the quasispecies, solely using haplotype-based diversity metrics to summarize the viral quasispecies profile might hinder our ability to observe the continuous dynamics of the virus across samples. To illustrate this, we compared related viral samples in various scenarios using both traditional haplotyping based on a strict mutational profile and a novel OTU assignment based on genetic distance (Methods). In the latter, haplotypes are clustered by relatedness into larger operational taxonomic units (OTUs). This allows the user to better comprehend the structure of the quasispecies, i.e., the evolving genetic diversity surrounding several main variants within the viral cloud.

The studies on HIV-1 and SARS-CoV-2 served as examples of short-term longitudinal sampling scenarios. Within a few hours of each other, env genes of four HIV-1 samples infecting T cells in vitro exhibited numerous haplotypes, among which a common dominant haplotype was not found (Fig 4a). When OTU clustering was applied to the same samples, the members and proportions of all four samples were relatively similar (Fig 4b), with the 186 haplotypes being optimally clustered into 10 OTUs (Fig 4c and 4d). Examining five SARS-CoV-2’s S gene samples taken within a nine-day period from a single patient, the dominant haplotype and OTU occupied over half of the virus population in each sample (Fig 4a and 4b), with minor haplotypes appearing as outliers (Fig 4c and 4d). It is also worth noting that the sporadic appearance of many minority variants in the day 13 and 17 SARS-CoV-2 samples (Fig 4a and 4b) was likely due to noise rising from incomparable read depths in the original alignments, which were over ten times lower than the other three samples (S5 Table).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Examples of viral quasispecies and OTU comparisons for five viruses (virus:gene). a, Stacked bar chart presenting the proportion of each unique haplotype (same color) identified across different samples within each viral species: gene dataset. b, Stacked bar chart presenting the proportion of new OTUs, each consisting of genetically closely related haplotypes. c, Multidimensional scaling (MDS) plot of haplotypes’ pairwise SNV distance (dot), showing the five largest OTUs (circle) with a color scheme corresponding to plot b. d, MDS plot of all OTUs.

https://doi.org/10.1371/journal.pcbi.1014208.g004

As a model for long-term longitudinal sample comparison, studies on HCV and HBV were utilized. Analysis of the NS4 coding region of HCV revealed that viral haplotypes and their proportions underwent sudden changes after treatment. Interestingly, certain major haplotypes detected pre-treatment reappeared 18 months post-treatment (Fig 4a). Concurrently, OTU clustering provided better insights into the dynamics of HCV across three sampling points by grouping genetically closely related minor haplotypes into common clusters observed at varying proportions across all time points (Fig 4b, 4c and 4d). The within-host virus population dynamics were also clearly illustrated by HBV’s S gene. Both haplotype and OTU classifications underscored the gradual turnover of predominant haplotypes from timepoint one to four over a 30-month period, and a completely different population emerged at timepoint five, occurring over 80 months after the initial sampling (Fig 4a and 4b).

Lastly, in the scenario of the outbreak represented by IAV’s M gene segment samples, viruses were collected cross-sectionally among patients in two wards (B and C) within the same hospital. At the haplotype level, only two mostly identical sets of virus populations were found in more than one patient (B1, B2 and C1, C3), without other prominent links between cases or wards (Fig 4a). However, upon regrouping them into new OTUs, we discovered that all patients in ward B were partly infected with the same genetically similar sub-population, while all patients in ward C were infected with a similar set of OTUs. Furthermore, the predominant OTU found in two ward B patients (B1, B2) was also present as a minority group in all ward C patients, suggesting potential inter-ward transmission of IAV (Fig 4b, 4c and 4d).

The comparative profiles of viral quasispecies, whether analyzed at the haplotype or OTU levels, based on other genes from the same viral samples (including HBV’s P gene, HIV-1’s gag gene, and SARS-CoV-2’s ORF3a gene), exhibited visual similarities to the chosen genes depicted in Fig 4 (S12-S14 Figs). This suggests the absence of linkage disequilibrium between the observed genes of these three viruses, indicating that they may adequately represent the population dynamics at the whole-genome level. Interestingly, one comparable result between the original and our studies is the number and frequency of haplotypes found in the day 11 and day 15 SARS-CoV-2 samples. We captured a similar quasispecies diversity in the S gene samples and discovered two SNV positions in the ORF3a gene that created two rare haplotypes not reported in the original study (S7 Table and Ko et al., 2021 [45]).

Simulated sequencing data

To benchmark our analytical pipeline under controlled conditions, we generated simulated viral long-read sequencing datasets for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) and human immunodeficiency virus (HIV-1). For each virus, we created five independent replicates, each containing five haplotypes with fixed frequency distributions of 50%, 20%, 15%, 10%, and 5%. Mutations were introduced into reference genomes (NCBI accessions listed in S1 Table at virus-specific rates. For SARS-CoV-2, we applied a single nucleotide polymorphism (SNP) rate of 8 × 10⁻⁴ and indel rate of 5 × 10⁻⁵ per site [51]. For HIV-1, we used a SNP rate of 2 × 10⁻³ and indel rate of 3 × 10⁻⁴ per site [52]. Each replicate was generated with a unique random seed to produce distinct mutation patterns while maintaining consistent haplotype proportions across replicates. For each replicate, we first generated error-free “true reads” at 500 × coverage to serve as ground truth. We then simulated platform-specific sequencing reads at 10 × coverage using Badread v0.4.1 [53], modeling five sequencing technologies with their characteristic error profiles: Oxford Nanopore Technologies (ONT) 2018, 2020, and 2023 models, and PacBio continuous long reads (CLR) and highly accurate long sequencing reads (HiFi). The simulated dataset comprised 50 sets of reads (2 viruses × 5 replicates × 5 sequencing platforms).

Real sequencing data

To evaluate our package and workflow, we gathered raw sequencing reads from five projects focusing on whole genome sequence (WGS) of five different viruses: SARS-CoV-2, HIV-1, hepatitis C virus (HCV), hepatitis B virus (HBV), and influenza A virus (IAV), which are among the most studied species in the field of viral quasispecies. These reads were generated using either PacBio circular consensus sequencing (CCS) or ONT and were sourced from the National Center for Biotechnology Information’s Sequence Read Archive (NCBI SRA). Specifically, we selected projects that obtained either longitudinal viral samples or cross-sectional samples from a single outbreak, where the genetic relationships were suitable for comparison at the quasispecies level. PacBio read samples were obtained from three projects: three HCV samples from a chronic hepatitis patient collected before and after treatment over 18 months [46]; five HBV samples collected at various time points within 100 months from an untreated patient [47]; and five longitudinal SARS-CoV-2 samples collected between 8–17 days after clinical onset from a single patient [45]. ONT data included four HIV-1-infected T cell samples collected up to 24 hours post-infection [48] and seven IAV samples from different patients during a nosocomial outbreak [49]. The NCBI accession numbers of all 24 raw read samples across five projects are listed in S1 Table.

Raw reads were processed using comparable workflows for both sequencing platforms. Adapter sequences were trimmed using HiFiAdapterFilt v3.0.1 [54] for PacBio reads and Porechop v0.2.4 [55] for ONT reads. Quality filtering was performed using Filtlong v0.2.1 [56] to remove PacBio reads shorter than 1,000 bp or ranking in the lowest 10% by composite read score (a weighted combination of read length and mean quality); since all reads across datasets had mean quality scores at or above Q30, the relative filter removed reads based on length rather than quality. NanoFilt v2.8.0 [57] was used to remove ONT reads with average quality scores < 7. Trimmed and filtered reads were aligned to NCBI reference genomic sequences (S1 Table) using minimap2 v2.26 [58] with technology-appropriate presets (“map-hifi” for PacBio CCS and “map-ont” for ONT). SAM files from read mapping were converted to FASTA format and fragmented into genes or coding regions based on genomic annotations using SeqKit v2.7.0 [59], retaining only reads covering the full length of each gene or region (Fig 1a). The depth and length of final alignments are documented in S2-S6 Tables.

The longreadvqs package and QoALa workflow

The longreadvqs package implements the QoALa workflow, a standardized framework for generating noise-minimized alignments, harmonizing sequencing depth, extracting haplotypes and operational taxonomic units (OTUs), and performing cross-sample viral quasispecies comparisons. The workflow begins with importing equal-length read alignments from any genome assembly or long-read processing pipeline. Noise minimization is performed using the “vqsassess” function, which replaces low-frequency single-nucleotide variants (SNVs) with either the majority nucleotide (“conbase” method) or, when applicable, the nucleotide from the dominant haplotype (“domhapbase” method) (Fig 1b). The frequency threshold can be set based on prior knowledge, such as the estimated sequencing error rate or virus-specific mutation rates, or optimized empirically using the “pctopt” function. “pctopt” shows the percentage of singleton haplotypes (haplotypes represented by a single read) across candidate thresholds, allowing identification of the optimal cut-off where this percentage plateaus (elbow point) (Fig 1a). When analyzing multiple samples, “vqsassess” also performs random downsampling to the lowest sequencing depth to reduce bias in diversity comparisons.

Haplotypes are defined as unique nucleotide sequences across the full-length alignment of the analyzed gene or genomic region. After noise minimization, reads with identical sequences are grouped into the same haplotype, and haplotype frequencies are calculated as the proportion of reads assigned to each haplotype relative to the total number of retained reads. These haplotypes and their frequencies are extracted for each sample and evaluated using the “snvcompare” function. For each sample, SNV frequencies across the alignment are summarized from the read data produced by “vqsassess”, and the resulting SNV frequency profiles are plotted and arranged across samples to enable visual comparison of variant patterns and positional consistency along the sequence (Fig 1c). Because sequencing noise can vary along the gene or genomic region—for example, in homopolymer regions [60,61] or at soft-clipped read ends [62,63]—the “vqscustompct” function allows region-specific adjustment of the noise-minimization cut-off to improve accuracy (Fig 1b).

Processed samples are then integrated using “vqscompare” (Fig 1d), which identifies haplotypes shared across samples and visualizes unique haplotypes per sample in color-coded bar plots. Haplotype sequences are reclassified into OTUs based on genetic distances calculated from the SNV alignment using dist.dna (ape v5.7.1 [64]). The distance matrix is converted into dissimilarity coordinates and clustered using classical multidimensional scaling (cmdscale) followed by k-means clustering (kmeans) in stats v4.3.1 [65]. The number of clusters is user-defined, and OTU composition is visualized in bar plots, with MDS plots illustrating clusters and major haplotypes within each OTU. All resulting plots are generated using ggplot2 v3.4.4 [66].

In addition to visualizations, “vqscompare” outputs noise-minimized, downsampled read and SNV alignments, classified haplotypes and OTUs, and quasispecies diversity metrics. Nine viral quasispecies diversity metrics are calculated per sample using QSutils v1.18.0 [67], with an additional set based on OTU consensus reads and frequencies for large-scale diversity assessment. Together, these functions constitute the QoALa workflow, providing a reproducible, standardized approach for viral quasispecies characterization, sample harmonization, and cross-sample comparison.

Benchmarking analysis using simulated sequencing data

For benchmarking, the performance of the QoALa workflow was compared against three alternative approaches: Strainline-corrected reads, Racon-corrected reads, and uncorrected raw reads. Strainline represents a platform specifically designed to correct and reconstruct haplotypes from viral quasispecies [33], whereas Racon is a general-purpose long-read error-correction tool [44]. Both tools are applicable to ONT and PacBio sequencing data. All 50 simulated read sets for each virus were corrected using either Strainline (with sequencing-platform-specific parameters) or Racon v1.5.0 (with default settings). All uncorrected (n = 50) and corrected (n = 100) simulated datasets were aligned to their respective reference viral genomes using minimap2 v2.26 [58] with technology-appropriate presets (“map-ont” for ONT; “map-pb” or “map-hifi” for PacBio). SAM files were converted to FASTA format, and gene-specific regions were extracted using SeqKit v2.7.0⁵⁹—the spike (S) gene for SARS-CoV-2 and the envelope (env) gene for HIV-1. Only reads spanning the full gene length were retained for downstream analyses.

For the aligned uncorrected S and env reads, noise-minimization thresholds for the QoALa workflow were estimated with “pctopt”, selecting the median percentage at which singleton haplotype proportions plateaued across replicates for each platform—HIV-1 env: ONT 2018 (12%), ONT 2020 (5%), ONT 2023 (6%), PacBio CLR (9%), PacBio HiFi (1%); SARS-CoV-2 S: ONT 2018 (13%), ONT 2020 (6%), ONT 2023 (8%), PacBio CLR (8%), PacBio HiFi (1%) (S15 Fig). These thresholds were then passed to “vqsassess”, which imported all corrected and uncorrected read sets and uniformly downsampled them to a depth of 500 to match the error-free ground-truth coverage. Noise minimization was applied only to the uncorrected reads designated for QoALa’s threshold-based correction, while corrected reads and comparison uncorrected sets were downsampled but left otherwise unmodified.

A total of 200 aligned read sets (2 viral genes × 5 replicates × 5 sequencing platforms × 4 correction approaches) were subsequently compared with the two corresponding ground-truth viral quasispecies (100 aligned reads per ground truth) using “vqscompare”. Comparisons were conducted at both the haplotype and OTU levels. For OTU-level analyses, haplotypes were reclustered into 30 OTUs per gene, and resulting OTU frequency profiles were used as the basis for evaluation.

Quasispecies reconstruction accuracy was evaluated using three metrics. Bray-Curtis dissimilarity [68] was calculated between each reconstructed sample and its corresponding ground truth using “vegdist” in vegan R package v2.6.8 [69] to provide a direct comparative measure incorporating both qualitative similarity (presence of shared haplotypes or OTUs) and quantitative agreement (their relative abundances); Bray-Curtis ranges from 0 (identical composition) to 1 (completely dissimilar), with lower values indicating more accurate reconstruction. Within-sample diversity was summarized using the Gini-Simpson index (H_GS), where larger values indicate greater diversity and evenness; deviations from the ground-truth H_GS therefore reflect structural differences rather than specific compositional mismatches. Nucleotide diversity (π_m) was also computed as a sequence-based measure of genetic/functional diversity, where higher values reflect more divergent viral populations; although π_m is not a direct pairwise-comparative index, its deviation from the true π_m indicates how well each method captures underlying heterogeneity.

Statistical comparisons were conducted on per-replicate deviations from the ground truth for each metric across correction methods (QoALa, Strainline, Racon, and uncorrected reads) and across sequencing technologies. Two-sided Wilcoxon rank-sum tests (wilcox.test) [65] were applied independently for each metric and for both haplotype- and OTU-level analyses to accommodate small replicate sizes and non-normal deviation distributions. To improve reproducibility, p-values were adjusted using the Benjamini-Hochberg false discovery rate (FDR) procedure via p.adjust [65]; significance was defined at FDR < 0.05. Because Bray-Curtis dissimilarity decreases with accuracy, and H_GS and π_m increase with diversity, smaller deviation values indicate better agreement with the ground-truth population structure. This statistical framework identified correction methods that yielded significantly smaller deviations—and therefore more accurate reconstructions—than uncorrected reads and other correction approaches across sequencing platforms. All visualizations of metric comparisons and statistical outcomes were generated using ggplot2 v3.4.4 [66].

Down-sampling sensitivity analysis and viral quasispecies comparison for real sequencing data

To assess how sequencing depth influences diversity estimation and quasispecies comparison in the QoALa workflow, we conducted a sensitivity analysis using real sequencing datasets. We extracted read alignments from specific genes or regions of five representative viral species. Alignments were selected based on having lengths exceeding or closely approximating 1,000 bases, coverage depths greater than 100 reads, and soft-clipped regions—when present—constituting less than 25% of the total alignment length. Using “pctopt”, we determined the optimal noise-minimization cut-off percentage as previously described (Fig 1a). The selected alignments and their corresponding cut-off thresholds were: NS4 (15%) for HCV; P (5%) and S (5%) for HBV; env (15%) and gag (19%) for HIV-1; S (1%) and ORF3a (1%) for SARS-CoV-2; and the M segment (10%) for IAV (S1 Fig).

For high-depth samples (HCV, HBV, SARS-CoV-2), down-sampling series through “vqsassess” began at 10,000 reads and were iteratively halved until reaching a minimum of 78 reads. For low-depth samples (HIV-1, IAV), fixed subsample sizes of 300, 150, 100, 50, and 25 reads were used. For each down-sample size, random resampling with replacement was performed 100 times after noise minimization; an identical repeated-sampling protocol was also applied before noise minimization to directly compare the impact of noise reduction on depth sensitivity. For every iteration, nine diversity metrics from QSutils—fully described by Gregori et al.⁴²—were computed: number of haplotypes (H), Shannon entropy (H_S), normalized Shannon entropy (H_SN), Gini-Simpson index (H_GS), functional attribute diversity (FAD), mutation frequency at the entity level (Mfe), nucleotide diversity at the entity level (π_e), mutation frequency at the molecular level (Mfm), and nucleotide diversity (π_m). Resulting metric distributions were visually compared across subsample sizes and before/after noise minimization to evaluate metric stability and the degree of information loss due to aggressive down-sampling.

Apart from the sensitivity analysis, we performed viral quasispecies comparisons among within-species/gene samples to demonstrate the application of the QoALa workflow to real experimental and clinical data. Noise-minimized, down-sampled alignments for each selected gene and viral species—using the pctopt-optimized thresholds described above and down-sampled to either the lowest observed depth for that gene or to 1,000 reads if the lowest depth exceeded 1,000 (S2–S6 Tables)—were generated using the “vqsassess” function and used as inputs for individual “vqscompare” analyses. The number of clusters (k) for OTU classification via k-means clustering was fixed at 10 for all runs. The resulting summary plots illustrate viral quasispecies dynamics across a range of biological scenarios, including short-term (hours to days in HIV-1 and SARS-CoV-2 datasets), long-term (months to years in HCV and HBV datasets) longitudinal sampling, and samples collected from the same outbreak cohort (IAV dataset).