Background and Design Choices¶

This document provides the biological and computational context for the choices made in RolyPoly's presets and processing steps.

[!WARNING]
This document is still UNDER HEAVY WORK - Some of details may be incomplete, wrong, and for now this is not for human consumption.

filter-reads and assemble¶

tl;dr preset choice¶

How your RNA was prepared for sequencing is usually the main thing that determines which preset to use. Three factors tend to matter most: 1. whether your library was poly-A selected, 2. whether you want to filter potential host cDNA/mRNA and other contaminants, and 3. whether your sample came from enriched/purified virions ("virome"), where you can often skip rRNA removal and sometimes use lighter quality filtering.

• If your library is poly-A selected: RolyPoly includes a poly-A trimming step in filter-reads.
• If you want host filtering: provide a host reference genome to filter-reads and assemble (RolyPoly will try to mask likely viral regions in that reference before using it for filtering).
• If your sample is virome-enriched: you can often skip rRNA and host removal, use lighter quality filtering, and sometimes enable deduplication steps (depending on your goals).

Note: "host" removal can also remove EVEs or integrated viral sequences, so it may not be ideal if you are interested in non-RNA viruses or endogenous viral elements related to your study.

From sample to reads: the wet-lab pipeline¶

Before we get "reads", the biological material goes through a series of wet-lab steps that strongly shape what the sequencing data look like. Understanding that chain helps with picking in-silico settings and avoiding unexpected read loss.

The steps and methods below are usually chosen based on study goals and budget. Many introduce bias, but also reduce cost or increase throughput. For example, rRNA depletion kits do not equally target all domains of life, but they can substantially reduce the sequencing depth needed.

In most non-targeted genomics or metagenomics studies, the starting material can be anything from a single host species and its infecting virus, to a complex mixture spanning multiple domains of life (table 1). The "meta" prefix (e.g. metagenomics, metatranscriptomics) generally implies a mixed community as the source.

Regardless of source type, material is usually processed through a mix of chemical, enzymatic, and physical steps depending on study focus. The tables below summarize common steps across the wet-lab pipeline: sample handling before nucleic acid extraction (table 2), extraction itself (table 3), post-extraction processing (table 4), and library preparation (table 5). For RNA virus-centric/transcriptomic work, this usually means RNA extraction, often with rRNA depletion or poly-A selection, followed by reverse transcription to cDNA. Direct RNA sequencing is also possible (e.g. Oxford Nanopore), though still less common in many workflows.

Note: this overview is intentionally practical and not exhaustive.

Table 1: Common sample sources for RNA virus discovery

Once the sample is collected, it usually goes through pre-extraction processing. These steps act on physical or chemical properties of the sample (size, density, protection status, membrane/protein integrity) and can strongly affect what ends up in the extracted material. For our use case, one key distinction is whether the sample was further enriched for viruses (e.g. VLP-focused workflows, often called "viromics"). Some steps also serve multiple practical purposes, such as buffer exchange, inhibitor removal, or transport/storage stabilization.

A practical note on filtration: the steps listed in table 2 can be used in combination, and it is entirely possible to use the retentate from one filter as the input to a second filtration step, or to collect the filtrate instead - the choice depends on which size fraction is of interest.

Table 2: Common pre-nucleic acid extraction processing steps

Many of these steps are combined within a single protocol or commercial kit. Some steps also serve multiple roles - for example, a buffer exchange step may simultaneously remove inhibitors while concentrating the sample.

Nucleic acid extraction is the step that liberates nucleic acids from biological material. The method you choose can introduce bias in yield, purity, and which nucleic acid species are recovered. For RNA virus discovery, RNA extraction is usually the default, though total nucleic acid (TNA) extraction is also used when both DNA and RNA are of interest (for example, mixed RNA/DNA virus surveys or integrated proviral sequence detection).

Table 3: Common nucleic acid extraction methods

After extraction, a second round of processing is often used to reshape the nucleic acid pool by removing unwanted species or adjusting yield/purity before library prep. These steps act on the molecules directly (not the physical sample) and can have a large effect on what gets sequenced.

Table 4: Common post-nucleic acid extraction processing steps

Library preparation converts processed nucleic acid into sequencing-ready molecules. For RNA libraries, this begins with reverse transcription to cDNA; after that, the workflow broadly resembles DNA library prep. Choices at this stage, especially fragmentation, priming strategy, and strand specificity, can strongly affect how viral sequences are represented and interpreted downstream.

When working with multiple samples (for example, time points, locations, or tissue types), it helps to plan sample-level barcoding/indexing before library prep starts. Each sample can get a unique index (or dual-index combination), so libraries can be pooled in one sequencing run and demultiplexed later. This is standard and usually lowers per-sample cost. But prep decisions still matter: if samples are pooled too early (for example at the RNA stage), sample identity is lost. Keeping a clear index-to-sample map is essential for meaningful downstream comparisons. In practice, indexing strategy is often constrained by the sequencing center to avoid collisions with other projects in the same run. Many users therefore receive already-demultiplexed data; if not, you may also see index FASTQs in addition to R1/R2 files.

Table 5: Common library preparation steps

From sample to FASTQ: the processing chain¶

flowchart TD
    A[Sample\ncells / retentate / VLP pellet] --> B[RNA extraction\ne.g. TRIzol, column-based]
    B --> C{Enrichment or depletion?}
    C -->|Total RNA + ribo-depletion\nRibo-Zero, RNase H-based| D[ribo-depleted RNA\n★ recommended for broad virus discovery]
    C -->|Poly-A selection\noligodT capture| E[poly-A-enriched mRNA]
    C -->|No treatment\ntotal RNA| F[total RNA\nhigh rRNA content]
    D & E & F --> G[RNA fragmentation\nenzymatic or physical shearing\ntarget ~150–400 bp]
    G --> H[Reverse transcription\nrandom hexamers or oligo-dT]
    H --> I[End repair + dA-tailing]
    I --> J[Adapter ligation\nP5 + P7 Illumina adapters]
    J --> K[PCR amplification\n6–15 cycles]
    K --> L[Size selection\nSPRI beads or gel\nsets insert size distribution]
    L --> M[Paired-end sequencing\ntypically 2×75 bp or 2×150 bp]
    M --> N[Raw FASTQ\nR1 + R2 per sample]

From sequencing to FASTQ: basecalling and demultiplexing¶

The flowchart above ends at the sequencer, but the files you receive involve two more steps:

1. Basecalling - the instrument converts raw signal (images or electrical current) into base calls with per-base quality scores. For most Illumina instruments this happens on-board in real time; for Oxford Nanopore it is a separate, model-based step (basecallers improve frequently, so re-basecalling old data with a current model can be worthwhile).
2. Demultiplexing - reads are sorted into per-sample FASTQ files using their index sequences. Most users receive already-demultiplexed R1/R2 FASTQs (see the indexing note in table 5); if you receive undemultiplexed data, you will also get index FASTQs (I1/I2) and must split samples yourself before any per-sample processing.

Everything downstream in RolyPoly assumes per-sample FASTQ files as input.

From digitized data to in-silico processing¶

Important: the goal of in-silico processing is not to remove as many reads as possible. The goal is to reduce technical noise and increase biological signal for your specific question.

This part is inherently goal-dependent. Trimming, filtering, masking, host subtraction, normalization, and deduplication are tools, not goals by themselves. If a step does not clearly improve signal for the biological target you care about, it is often better to keep it conservative or skip it.

Pipeline, tool, and parameter choices should follow the study target, but they could also be informed by everything that happened before digitization (sample source, enrichment strategy, extraction chemistry, library prep, expected contaminants, expected insert size, and sequencing platform).

Examples:

• If you also care about DNA viruses, avoid aggressive host/DNA removal.
• If you know exactly which adapter sequence was used, start with that specific adapter set instead of broad default databases (some of which may match real viral sequences).
• If your focus is differential expression (host), preserve host reads and avoid filtering choices that distort host transcript abundance.
• If your focus is low-level variation (for example SNPs or quasispecies/subspecies structure), avoid error correction, over-normalization, and aggressive deduplication that can flatten minor variants.
• If your focus is recovering major viruses and improving genome completeness, stronger cleanup and downsampling reads with over represented kmers may actually make assembly easier.

It is also completely fine to run multiple in-silico "branches" from the same sample, each with different pipeline choices. For example, you can use the original raw reads (or lightly processed reads) for quantification, while using a normalized subset for assembly-oriented recovery.

Many recent workflows also combine strategies, such as pooling selected samples, running multiple assemblers and merging/curating outputs, applying k-mer abundance-based read normalization, and in some cases adding haplotyping/strain-resolution analyses after the fact.

Library preparation and viral recovery¶

Many RNA viruses, including phages and most negative-sense, ambisense, and segmented RNA viruses, are not polyadenylated, so enrichment strategy can strongly affect what gets recovered. In a direct comparison, poly(A)-selected libraries yielded viral reads but were insufficient for complete recovery of a non-polyadenylated virus genome, while ribo-depleted total RNA performed substantially better (Visser et al., 2016; PMID: 27250973). In practice, total RNA plus ribo-depletion is often a good starting point for discovery-oriented workflows.

Ribo-depletion also has limits. Probe performance depends on sequence match, so depletion can be less effective in non-model or mixed communities (Kim et al., 2019; PMID: 31783730). The two main depletion strategies have different failure modes: probe-based methods (Ribo-Zero, NEBNext, QIAseq) hybridize rRNA with sequence-specific probes and pull it down, so they miss rRNAs whose sequence diverges from the probe set - a real concern in mixed microbial communities (He et al., 2010; PMID: 20852648). Enzymatic methods (e.g. RNase H with custom probes) can be retargeted to new organisms by adding probes for their rRNA, but require knowing what is in the sample. In either case, computational rRNA filtering is commonly used as a second layer (Zhou et al., 2018; PMID: 29444661), but very aggressive filtering can discard informative reads. Dovrolis et al. showed that rRNA-focused sorting can capture rRNA-virus chimeras and reduce viral support for assembly (PMID: 34370725).

Independent RNA-seq evidence also supports treating host-virus chimeras cautiously: host-virus chimeric events can be infrequent and largely artifactual, consistent with RT template switching during library construction (Yan et al., 2021; PMID: 33980601). For preprocessing, this argues for conservative thresholds and careful interpretation of putative chimeric reads instead of blanket removal based on single-feature matches.

Subjective note: I have also observed some rRNA-virus chimeric contigs in published virus discovery works and general metatranscriptomes (see Table S6, sheet "rRNA_summary" in Neri et al., 2022), and the full extent of host-virus chimera is probably underappreciated. I am not sure I can recommend keeping chimeric reads though - I assume these are not biological and often arise from template switching after nucleic acid extraction - and properly trimming only the host portion of chimeric reads is not trivial (IMO).

RolyPoly preprocessing currently focuses on Illumina short-read libraries (poly-A selected, ribo- depleted, or unselected total RNA). Long-read/direct-RNA support is a separate ongoing area.

RNA degradation and sample quality¶

RNA integrity shapes the read profile before any computational step. Intact eukaryotic RNA shows sharp 18S/28S peaks (RIN ≈ 8–10); degraded RNA shows a smear shifted toward short fragments. For virus discovery, degradation has mixed effects:

• Fragmentation is partly free. Sequencing libraries need ~150–400 bp fragments anyway, so moderately degraded RNA can still yield usable libraries - dedicated degraded/low-input methods can match or outperform standard protocols on degraded material (Adiconis et al., 2013; PMID: 23685885).
• Coverage becomes 3'-biased. RT priming and fragment availability skew coverage toward the 3' ends of transcripts, which can leave 5' regions of viral genomes under-covered and complicate genome completion.
• Small RNAs are enriched. Degradation products and small-RNA species (siRNAs, tRNA fragments) become a larger fraction of the library, adding background that assembly must overcome.
• Low-input samples amplify artifacts. Low-input and VLP-enriched libraries often require more PCR cycles or amplification-based methods, which increases duplicates and chimeras (see PCR duplicates vs optical duplicates and Virus–host chimeric reads).

Practical guidance: check per-base quality and insert-size distributions before choosing trim/filter stringency, and treat unusually low read retention in filter-reads as a signal to inspect the input library rather than to loosen thresholds blindly.

Sequencing artifacts and their computational signatures¶

Each step between nucleic acid extraction and the sequencer can introduce artifacts, and the sequencing process itself adds base-call errors on top of them.

Base-call errors and quality scores¶

Every base in a FASTQ file carries a Phred quality score (Q), an integer estimate of the probability that the base call is wrong: Q = -10 × log₁₀(P_error). Common values:

Two properties of Illumina sequencing-by-synthesis matter for preprocessing:

1. Quality decays toward the 3' end. Each additional cycle adds more phasing noise and signal decay, so error rates rise along the read. This is why quality trimming (quality_trim_unmerged) trims from the 3' end, and why the trim stringency (trimq) trades read length for per-base accuracy.
2. Errors are mostly random, not systematic. A random error in one read is unlikely to be repeated in the read that overlaps it. This is the basis of overlap-based error correction (error_correct_1, error_correct_2): when R1 and R2 overlap, the two independent observations of the same molecule can be compared, and disagreements are resolved in favor of the higher-quality base. Merging overlapping pairs (merge_reads) therefore both lengthens the effective read and reduces its error rate - but only when the insert is short enough for the reads to overlap (see Paired-end reads and insert size).

A practical consequence: error correction and merging are most valuable for short-insert paired-end libraries and largely irrelevant for single-end input (RolyPoly auto-skips these stages for single-end data).

Paired-end reads and insert size¶

The insert is the actual cDNA fragment between the two Illumina adapters. Both ends are sequenced, producing R1 (reads the forward strand) and R2 (reads the reverse complement strand).

Adapter-P5 ─── INSERT (cDNA fragment) ─── Adapter-P7
    │                                           │
    └──► R1 reads →                 ← R2 reads ◄┘

Typical layout (insert longer than read length, no overlap):

5'─[P5]─[R1 ████████████████]─ · · · gap · · · ─[████████████████ R2]─[P7]─3'
         read 1 (fwd)           unsequenced middle           read 2 (rev-comp)
         └────────────────────────────────────────────────────────────────────┘
                               insert size (e.g. 300 bp)

Insert size is set by the size-selection step in library prep. Typical Illumina runs use 2×75 bp or 2×150 bp reads; inserts are usually 150–400 bp. The ratio of insert to read length determines which artifacts appear.

Short inserts, overlapping reads, and adapter contamination¶

As insert size decreases relative to read length, two problems arise:

Case 1 - Overlapping reads (insert slightly shorter than 2× read length):

5'─[P5]─[R1 ████████████]─[████ overlap ████]─[████████████ R2]─[P7]─3'

R1: ████████████████████                        (reads left → right)
R2:             ████████████████████            (reads right → left, RC)
              ←── overlap region ──►

BBMerge can merge overlapping pairs into a single, longer, more accurate read.
This is common for short viral genomes where short inserts are preferred.

Case 2 - Adapter contamination (insert shorter than read length):

5'─[P5]─[fragment]─[P7]─3'   (very short insert)

R1: [fragment sequence ████][P7-adapter sequence ░░░░░░░░░░]
                               ↑ adapter "bleed-in"
R2: [fragment RC ████][P5-adapter RC ░░░░░░░░░░]

Without adapter trimming, these tails can look like spurious sequence and reduce assembly quality.
Adapter trimmers (e.g., BBDuk, cutadapt, Trimmomatic) typically remove these by matching known
adapter sequence k-mers, and most also let you provide custom adapter sequences.

Poly-G tails and low-complexity reads¶

Two-color chemistry instruments (NextSeq, NovaSeq) cannot distinguish "G" from "no signal": both channels are dark, so any cycle where the cluster signal fails is called as G. When a fragment ends and the cluster runs out of template, the remaining cycles of the read become a poly-G tail:

Fragment ends mid-read (two-color chemistry):

R1: [real sequence ████████][GGGGGGGGGGGGGGGGGGGG]
                            ↑ no template left → every cycle reads "G"

On one-color chemistry instruments (MiSeq, MiniSeq) the same failure mode produces low-quality bases instead, which quality trimming catches. Poly-G tails, however, can carry high Q-scores (the instrument is confident there is no signal - it just reports it as G), so they survive quality trimming and must be removed by a dedicated poly-G trimmer or by low-complexity / entropy filtering (entropy_filter), which discards reads whose sequence diversity within a window falls below a threshold. Left in place, poly-G tails inflate G content, create spurious k-mer spikes, and can produce homopolymer-rich contigs during assembly.

PCR duplicates vs optical duplicates¶

Both appear as multiple reads with identical sequence, but their causes differ:

PCR duplicates                    Optical duplicates
──────────────────────────────    ──────────────────────────────
Same molecule amplified           Single cluster on the flowcell
multiple times during             that spreads or is misread as
library prep PCR:                 two adjacent clusters:

Flowcell surface:                 Flowcell surface:

 ● cluster A  (seq: ACGTACGT)      ●●  cluster A+B touching
 ● cluster B  (seq: ACGTACGT)          (seq: ACGTACGT for both)
 ● cluster C  (seq: ACGTACGT)
 (scattered anywhere on tile)     ← physically adjacent, same tile

Cause: amplification bias         Cause: optical/diffusion artifact
Worse with: more PCR cycles,      Worse with: patterned flowcells
  low-input / VLP libraries         (NovaSeq, NextSeq 2000)
Removed by: sequence identity     Removed by: sequence identity
  (any duplicate-removal tool)      AND cluster proximity

Both can inflate abundance estimates and mislead assembly coverage. In metatranscriptomes, though, highly expressed transcripts are expected, so aggressive duplicate removal can also discard real biology. In RolyPoly, deduplication is done as exact-sequence removal (seqkit rmdup) rather than coordinate-based deduplication.

Index hopping and sample cross-talk¶

On patterned-flowcell instruments (NovaSeq, NextSeq 2000), a small fraction of reads can be attributed to the wrong sample during cluster generation or demultiplexing - commonly called index hopping or index cross-talk. A read from sample A carries sample B's index, so it is written to sample B's FASTQ:

Sample A molecule:  [insert]─[P5-A][P7-A]   →  sequenced with index A
                                          ↓ index hop (misassigned during clustering/extraction)
Demultiplexed to B: [insert]─[P5-A][P7-B]   →  written to sample B's FASTQ

Computational signature: low-level "background" contamination that appears in many samples at roughly similar rates, often including sequences that cannot plausibly come from that sample. For virus discovery this matters most when a strong positive sample can seed false positives in negative controls or low-biomass samples on the same run. Non-redundant dual indexing largely remediates this (Costello et al., 2018; PMID: 29739332); computationally, cross-talk is best handled by running negative controls and interpreting trace detections cautiously rather than by per-read filtering.

Virus–host chimeric reads¶

Template switching during reverse transcription, or ligation of fragmented RNA molecules during library prep, can join viral and host RNA into a single read:

Origin molecules:

  ─────[viral genomic RNA segment]─────3'
                                         ↕ RT jumps (template switching)
  5'─────[host mRNA ─────────────]─────

Resulting chimeric cDNA / read:

  5'─[viral sequence ███████████][host sequence ░░░░░░░░░░░]─3'
      maps to viral genome ─┘         maps to host ────┘

Consequence 1: chimeric read is filtered by host-removal step
                → viral portion is lost along with the host portion

Consequence 2: chimeric read enters assembly
                → chimeric contig spanning virus + host
                → false annotations or missed virus

Consequence 3: rRNA–virus chimeras (described in Dovrolis et al., 2021; PMID: 34370725)
                → rRNA-sorting tools bin the chimera as rRNA
                → virus reads lost; assembly of complete genome fails

This is one reason RolyPoly keeps a permissive rRNA filter by default: strict rRNA filtering can discard the viral component of mixed/chimeric reads (Dovrolis et al., 2021; PMID: 34370725). The same caveat applies to host subtraction, where very strict mapping thresholds may remove reads that only partially match host sequence.

Digital read depth normalization¶

Metatranscriptomes are extremely uneven: a handful of transcripts (rRNA remnants, highly expressed host genes, one dominant virus) can account for most of the data, while the virus you are looking for sits at a few-fold coverage. Digital normalization addresses this computationally: reads are kept or discarded based on the k-mer coverage they carry, so that no region exceeds a chosen coverage cap (max_cov), while low-coverage regions - including rare variants - are preserved.

The mechanism (as implemented by Trinity's in-silico normalization, BBNorm, and ORNA):

For each read (in a single pass):
  1. Count the read's k-mers in a running k-mer table
  2. Compute the median k-mer depth of the read
  3. If median depth > max_cov  → discard the read (region is already well covered)
     If median depth ≤ max_cov  → keep the read, increment k-mer counts

The effect on a mixed population (dominant strain at ~100×, minor variant at ~2×):

Before normalization:                After normalization (max_cov = 5):

main:    ████████████████████ 100×   main:    █████ 5×
main:    ████████████████████ 100×   main:    █████ 5×
main:    ████████████████████ 100×   variant: █ 2×  ← kept: novel k-mers
variant: ██ 2×                            (median depth below cap)
         ↑ most main reads discarded,
           variant reads preserved

Because a minor variant carries k-mers that the dominant strain lacks, its reads have low median k-mer depth and are preferentially retained, while redundant reads from the saturated dominant strain are discarded. This is why normalization can make assembly of uneven metatranscriptomes both cheaper and, in some cases, more complete.

The trade-off is real and matters for variant work. Normalization is not free of information loss: a rare SNP-bearing read that happens to sit inside a deeply covered hotspot can be discarded because its other k-mers are already saturated. In a controlled two-genome benchmark (synthetic genomes with designed minor haplotypes at 2× against 20–100× main coverage), Trinity normalization at max_cov=5 retained only 3 of 12 designed SNP sites in the hotspot scenario, and 3 of 12 in a targeted sparse-minor scenario - while cutting total reads to ~11%. The same benchmark showed the upside: normalized read sets assembled faster and produced cleaner graphs, and post-hoc variant rescue by mapping reads back to assemblies recovered some lost alleles. The full fixture, benchmark harness, and per-site results are available in the digital normalization demo repository.

Practical guidance:

• Use normalization for assembly-oriented recovery of major genomes from uneven libraries - it reduces compute and can improve graph cleanliness.
• Avoid normalization (or use a high max_cov) when rare-variant/quasispecies structure is the goal - it can erase low-frequency variants, especially inside high-coverage hotspots.
• Prefer coverage-based methods (e.g. BBNorm) for paired data - they keep mates together, unlike k-mer-based single-end normalization which can break pairing.
• If in doubt, run both branches: normalized reads for assembly, raw (or lightly filtered) reads for variant calling and quantification - RolyPoly's modular steps make this easy.

Practical implications for preset choice¶

• If broad RNA virus recovery is the priority, a practical starting point is total RNA with ribo-depletion and moderate rRNA filtering.
• If libraries are multiplexed on patterned-flowcell instruments, non-redundant dual indexing can help reduce index-swap cross-talk (Costello et al., 2018; PMID: 29739332).
• If inserts are short and read overlap is high, read merging often improves effective read quality and assembly input (Bushnell et al., 2017; PMID: 29073143).
• If low-input amplification was used, it is usually safer to handle duplicate/chimera filtering conservatively and validate candidates by remapping support (Kugelman et al., 2017; PMID: 28182717).
• If the goal is variant/quasispecies recovery, avoid aggressive normalization and deduplication; if the goal is major-genome assembly, both can help (see Digital read depth normalization).

Background references (PubMed-checked)¶

• Visser M, Bester R, Burger JT, Maree HJ. Next-generation sequencing for virus detection: covering all the bases. Virol J. 2016;13:85. doi:10.1186/s12985-016-0539-x. PMID: 27250973.
• Dovrolis N, Kassela K, Konstantinidis K, et al. ZWA: Viral genome assembly and characterization hindrances from virus-host chimeric reads; a refining approach. PLoS Comput Biol. 2021;17(8):e1009304. doi:10.1371/journal.pcbi.1009304. PMID: 34370725.
• Yan B, Chakravorty S, Mirabelli C, et al. Host-Virus Chimeric Events in SARS-CoV-2-Infected Cells Are Infrequent and Artifactual. J Virol. 2021;95(15):e00294-21. doi:10.1128/JVI.00294-21. PMID: 33980601.
• Neri U, Wolf YI, Roux S, et al. Expansion of the global RNA virome reveals diverse clades of bacteriophages. Cell. 2022;185(21):4023-4037.e18. doi:10.1016/j.cell.2022.08.023. PMID: 36174579.
• Kim IV, Ross EJ, Dietrich S, et al. Efficient depletion of ribosomal RNA for RNA sequencing in planarians. BMC Genomics. 2019;20(1):909. doi:10.1186/s12864-019-6292-y. PMID: 31783730.
• Zhou Q, Su X, Jing G, Chen S, Ning K. RNA-QC-chain: comprehensive and fast quality control for RNA-Seq data. BMC Genomics. 2018;19(1):144. doi:10.1186/s12864-018-4503-6. PMID: 29444661.
• Costello M, Fleharty M, Abreu J, et al. Characterization and remediation of sample index swaps by non-redundant dual indexing on massively parallel sequencing platforms. BMC Genomics. 2018;19(1):332. doi:10.1186/s12864-018-4703-0. PMID: 29739332.
• Bushnell B, Rood J, Singer E. BBMerge - Accurate paired shotgun read merging via overlap. PLoS One. 2017;12(10):e0185056. doi:10.1371/journal.pone.0185056. PMID: 29073143.
• Kugelman JR, Wiley MR, Nagle ER, et al. Error baseline rates of five sample preparation methods used to characterize RNA virus populations. PLoS One. 2017;12(2):e0171333. doi:10.1371/journal.pone.0171333. PMID: 28182717.
• Adiconis X, Borges-Rivera D, Satija R, et al. Comparative analysis of RNA sequencing methods for degraded or low-input samples. Nat Methods. 2013;10(7):623-629. doi:10.1038/nmeth.2483. PMID: 23685885.
• He S, Wurtzel O, Singh K, et al. Validation of two ribosomal RNA removal methods for microbial metatranscriptomics. Nat Methods. 2010;7(10):807-812. doi:10.1038/nmeth.1507. PMID: 20852648.

Similarity-based taxonomy and demarcation¶

A conserved RdRp match establishes evidence of homology; membership of a named virus taxon requires the appropriate rank- and lineage-specific criteria. Reference-database composition, incomplete assemblies and marker selection all affect similarity-based assignments. Neither weighted support nor a single aligned-region identity is a universal demarcation threshold. Genome organization, other proteins and phylogenetic relationships may also matter. The ICTV classification overview explains how criteria are established for taxonomic groups.

RolyPoly's mmtax caveats document its actual defaults and its use of rank-aware weighted similarity votes rather than formal biological demarcation. The report caveats explain why a chart of selected contigs, often detected through RdRp markers, is not an abundance profile or a census of all segments of the viruses in a sample.