Statistical Genomics Foundry
Paper
domain/molecular-evolutiontopic/positive-selectiontopic/dnds

Schneider et al. 2009 — Estimates of positive selection inflated by sequencing/annotation/alignment errors

1. Citation

Schneider A, Souvorov A, Sabath N, Landan G, Gonnet GH, Graur D. 2009. “Estimates of positive Darwinian selection are inflated by errors in sequencing, annotation, and alignment.” Genome Biology and Evolution 1:114–118. DOI: 10.1093/gbe/evp012. PMID 20333182; PMCID PMC2817407. Open access, Creative Commons Attribution Non-Commercial License (CC BY-NC 2.5). Full text: https://europepmc.org/article/MED/20333182 (Europe PMC). Access date: 2026-07-05.

2. Access note

Full text read (Europe PMC fullTextXML, PMC2817407): abstract, all body sections (Introduction, Data and Methods, Results with Tables 1–4, Discussion), and reference list. Supplementary alignments were not retrieved. No paywall boundary encountered.

3. Thesis

Errors in sequencing, annotation, and alignment inflate inferred rates of positive Darwinian selection, so some published estimates of the prevalence of positively selected genes (PSGs) may be too high and should be treated with caution.

4. Problem & context

Published estimates of the proportion of PSGs vary enormously — for human alone the source lists 0.02%, 0.07%, 1.1%, 1.5%, 5.2%, and 8.7% across six studies (a “435-fold range”), and cites a claim of positive selection in 77% of vertebrate genes (Studer et al. 2008) and a claim that the chimpanzee PSG fraction is nearly twice the human one (Bakewell et al. 2007). The dN/dS (nonsynonymous/synonymous) excess is the standard positive-selection signal. The authors ask quantitatively how much of this variation is an artifact of data-quality errors.

5. Method / approach

  • Data: 7 eutherian mammals (human, chimpanzee, macaque, dog, cow, rat, mouse) from Ensembl, converted to Darwin database format. Assembly/build versions recorded per species (e.g. NCBI 36 for human, PanTro2.1 chimp, Mmul 1.0 macaque, CanFam 2.0 dog, Btau 3.0 cow, RGSC 3.4 rat, NCBI m36 mouse). Species chosen because their phylogenetic position is unambiguous.
  • Orthologs: from OMA (Orthologs Matrix project); 9,942 OMA groups initially. Exclusions: 1,227 groups with a CDS < 200 codons; 5,089 groups where the longest sequence was > 5% longer than the shortest; 6 groups where a CDS had > 1% of bases missing/unknown. Proteins aligned with the Darwin multiple sequence alignment package; coding DNA mapped back onto aligned amino acids to build codon-wise alignments of the CDSs.
  • Selection inference: CODEML from PAML v3.14b. Branch lengths and per-branch dN, dS estimated with codon frequencies from the F3×4 model, one dN/dS per branch fixed across sites (model = 1, NSsites = 0). 758 groups had at least one zero-length branch and were excluded → final data set 2,980 orthologous trees. PSGs identified with the improved branch-site model (“test 2” from Zhang et al. 2005, under “model A”): null allows sites with ω < 1 and ω = 1; alternative allows ω > 1 on an a priori foreground branch. Each of the 7 terminal branches run as foreground separately, all others background. Per PAML recommendation, CODEML run 3× for the null and 3× per alternative with initial branches randomized to 0.8–1.2× prior estimates; highest likelihood kept for the likelihood ratio test (LRT). Multiple-testing: Rom (1990) correction, false discovery rate 5%; other corrections said to give essentially the same results.
  • Three quality axes (each binarized):
    • Sequencing coverage: each CDS BLASTed against its species trace archive (Trace Archive), only hits ≥ 95% identity counted; coverage = number of trace hits overlapping each position. Split into “all sequence covered ≥ 3×” vs. “parts covered < 3×.” Human trace data were not kept, so human is excluded from the coverage analysis.
    • Alignment quality: Heads-or-Tails (HoT) algorithm (Landan & Graur 2007) — realign the sequences in reversed residue order; agreement between head and tail alignments measures reliability. Split into unambiguous (100% head/tail agreement) vs. ambiguous (< 100%).
    • Annotation certainty: Ensembl “known” genes (from manually curated UniProt/Swiss-Prot, RefSeq, TrEMBL) vs. “novel”/“inferred” genes (evidence from related species).
  • “All good” vs. “all bad”: genes high on all three axes (high coverage, known, 100% HoT) vs. genes deficient on all three, compared directly (Table 4).

6. Key claims / findings (pooled figures unless noted)

  • Coverage (Table 1): PSG fraction in poorly covered (< 3×) sequences is 3.3× higher than in well-covered (≥ 3×) — pooled 17.5% (1,546/8,852) vs. 5.3% (477/9,028); P(χ²) = 2.3 × 10⁻¹⁶⁶. Effect significant in every species tested.
  • Annotation (Table 2): PSG fraction in “inferred”/novel genes 1.9× higher than in “known” genes — pooled 14.6% (969/6,649) vs. 7.6% (1,080/14,148); P = 1.4 × 10⁻⁶¹. One exception: chimp went in the opposite direction (known 5.3% > inferred 2.5%, P = 0.0031).
  • Alignment / HoT (Table 3): PSG fraction in ambiguously aligned (HoT < 100%) genes 1.6× higher than perfectly aligned — pooled 15.6% (191/1,225) vs. 9.5% (1,858/19,635); P = 8.2 × 10⁻¹³. (HoT scores the whole MSA, so ambiguous/unambiguous counts are identical across species: 175 ambiguous, 2,805 unambiguous per species.)
  • Combined (Table 4): PSG fraction in “all bad” genes 7.3× higher than “all good” (Discussion; abstract states 7.2×) — pooled all-bad 27.5% (88/320) vs. all-good 3.8% (348/9,186); P = 2.4 × 10⁻⁹⁰. Mouse had no “all bad” genes.
  • Raw inferred PSG range across all genes: 0.9% (human) through 17.5% (macaque) to 23.3% (dog) (abstract). Filtering to good-quality genes collapses this: macaque drops from 17.5% to 2.0% (Discussion; ~an order of magnitude).
  • Branch-length correlation: strong positive correlation between inferred PSG fraction and branch length — r = 0.67 (P = 0.098) using all genes; r = 0.79 (P = 0.053) using only good genes. Read as evidence of residual dS (synonymous-rate) underestimation from saturation. Example dog–cow pairwise: 7.3% of genes had dS > 0.7, 13.7% had dS > 0.6, 26.0% had dS > 0.5.

7. Load-bearing statements (own-words — CC BY-NC 2.5 is own-words-only per license-policy.yml; functional strings kept verbatim as facts)

  1. (Mechanism, Introduction) because sequencing, annotation, and alignment errors are independent of codon position and amino-acid effect, they are equally likely to hit synonymous and nonsynonymous sites.
  2. (Mechanism, Introduction) errors affecting both site classes equally add a noise component at dN/dS = 1; low levels of this “neutral” noise won’t generally raise the overall ratio above 1, but estimation models allowing several site categories can detect the noise as a separate category and thereby inflate the inferred fraction of positively selected genes (PSGs) in a genome.
  3. (Combined effect, Abstract) the inferred PSG fraction in sequences deficient in all three criteria — coverage, annotation, and alignment — is 7.2 times higher than in genes with high trace sequencing coverage, “known” annotation status, and perfect alignment scores.
  4. (HoT definition, Methods) the HoT method rests on the a priori expectation that alignment should be independent of input-sequence orientation: for fully unambiguous cases, reversing residue order before alignment should yield exactly the reversed alignment.
  5. (Conclusion) the authors conclude that estimates of the prevalence of positive Darwinian selection may be inflated and should be treated with caution.

8. Stated scope, assumptions, limitations (authors’ own)

  • The three factors are not independent — their effects do not simply add up — but all push in the same direction (inflation).
  • Authors name additional inflation sources they did not fully control: purifying selection on synonymous sites, or a failure to identify orthology correctly.
  • Residual dS underestimation from saturation persists even though sophisticated methods are used to compensate for saturation, implying their good-gene estimates may still be somewhat inflated.
  • Analysis restricted to 7 eutherian mammals with unambiguous phylogeny; adding taxa deemed phylogenetically contentious. Human excluded from coverage axis (no trace data kept).
  • Quality axes binarized (≥3× vs <3×; 100% vs <100% HoT; known vs inferred); 95% BLAST-identity threshold — authors say alternative thresholds did not change conclusions (“data not shown”).

9. Failure modes / invalidity patterns

  • Data-quality errors masquerade as positive selection. Low sequencing coverage (< 3× trace), novel/inferred (vs. known) annotation, and ambiguous alignment each independently and jointly raise the inferred PSG fraction; combined they inflate it ~7×. The failure is specific to site/branch-site models with multiple ω categories, which pick up quality-error noise sitting at dN/dS ≈ 1 as a spurious positive-selection category.
  • Named diagnostics (detectors) the source uses:
    • Alignment: HoT score < 100% (head/tail disagreement) flags ambiguous alignment.
    • Sequencing: trace coverage < 3× (BLAST identity ≥ 95%) flags poor sequence quality.
    • Annotation: Ensembl “novel”/“inferred” status (vs. “known”) flags lower-confidence genes.
    • dS saturation symptom: high PSG–branch-length correlation and inflated pairwise dS (e.g. dS > 0.5–0.7 in a nontrivial fraction of dog–cow genes) signal underestimated synonymous rates and hence inflated dN/dS.
  • Zero-length branches make a tree unusable for ratio calculation (758 groups excluded on this basis) — a mechanical invalidity condition for per-branch dN/dS.

10. What the source does NOT address (confident silences)

  • No aligner comparison. A single aligner (the Darwin MSA package) is used; alignment quality is assessed post hoc via HoT. The paper does not benchmark aligners against each other and does not name a “worst” aligner — it partitions genes by HoT ambiguity, not by which program produced the alignment.
  • Does not prescribe a specific corrective tool/pipeline; it demonstrates that quality-filtering reduces spurious PSGs rather than recommending one named fix.
  • Does not analyze non-mammalian taxa, does not vary the substitution/codon model, and does not quantify the individual contribution of sequencing-error mechanisms at the base level.
  • No treatment of alignment-error effects on methods other than dN/dS codon models.

11. Open questions / ambiguities the source leaves unresolved

  • Because the three factors are non-independent and their effects are not cumulative, the marginal contribution of each (esp. alignment alone) to the combined 7× inflation is not decomposed.
  • The good-gene estimates may still be inflated by unaddressed dS saturation; the true PSG fraction is bounded but not pinned.
  • Whether the same inflation magnitudes hold for other selection tests, other clades, or other aligners is left open.

12. Guidance answers

  • Mechanism (in the paper’s terms): Sequencing/annotation/alignment errors are independent of codon position, so they hit synonymous and nonsynonymous sites equally, adding “noise” centered at dN/dS = 1. This noise does not by itself push the overall ratio above 1, but multi-category site models (branch-site / site models) can classify the noise as a separate site category with ω > 1, thereby inflating the inferred PSG fraction. (Verbatim: §7 quotes 1–2.)
  • Quantified false-positive rate attributable to alignment error, and worst aligners/conditions: For alignment specifically (HoT axis, Table 3), ambiguously aligned genes show a PSG fraction 1.6× higher than perfectly aligned genes (pooled 15.6% vs. 9.5%, P = 8.2 × 10⁻¹³). Combined with poor coverage and inferred annotation, the “all bad” vs. “all good” inflation is 7.2× (abstract) / 7.3× (Discussion), pooled 27.5% vs. 3.8%. The paper does not compare aligners or name a worst aligner — it uses one aligner (Darwin MSA) and flags ambiguity via HoT, so “which aligner was worst” is unanswered by this source (confident silence, §10).
  • Recommended fix (quote): The paper does not endorse a single tool. Its remedy is quality-filtering before/around codon-model inference — restricting to well-covered, “known”- annotated, unambiguously aligned genes (HoT = 100%) — plus interpretive caution. Closing recommendation, verbatim: the authors conclude that estimates of the prevalence of positive Darwinian selection may be inflated and should be treated with caution. (§7 quote 5.) It uses HoT as the alignment-reliability check and codon-wise (codon-aware) alignment in its own pipeline, but does not prescribe codon-aware alignment as the named fix.
  • License / must-quote: Confirmed CC BY-NC 2.5 (open access) from the article permissions statement — resolves to own-words-only in license-policy.yml (NC kept out of casts); §7 paraphrased, functional strings kept verbatim as facts.