Statistical Genomics Foundry
Paper
domain/gene-family-evolution

Denton et al. 2014 — Extensive Error in the Number of Genes Inferred from Draft Genome Assemblies

1. Citation

Denton JF, Lugo-Martinez J, Tucker AE, Schrider DR, Warren WC, Hahn MW. 2014. “Extensive Error in the Number of Genes Inferred from Draft Genome Assemblies.” PLoS Computational Biology 10(12): e1003998. DOI: 10.1371/journal.pcbi.1003998. PMC4256071. Open access: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1003998 License: CC BY (permissive). Access date: 2026-07-03 via PMC mirror (https://pmc.ncbi.nlm.nih.gov/articles/PMC4256071/).

2. Access note

Full text read via the PMC HTML mirror (abstract, results, discussion). The simulation, fosmid, and RNA-Seq figures were verified verbatim against the PMC source (17,941 contigs → 32,025 predictions; ~20,000 extracted vs ~36,000 predictions; Table 3: 11,931 collapsed). Tables/figures were read only as reported in prose; supplementary files not read. No paywall boundary (fully OA).

3. Thesis (1 sentence)

Draft genome assemblies produce extensive, systematic error in the inferred number of genes — both adding and subtracting genes — with over 40% of gene families assigned the wrong size relative to a higher-quality version of the same genome.

4. Problem & context

Sequencing produces large data volumes but assemblies built from them are often incomplete and error-filled; those errors propagate into annotation, especially into the count of genes (total gene number and copies per gene family). The paper’s motivating framing: gene-number variation between genomes is used to infer biology, yet some of that variation is an artifact of assembly, not real evolution. Historical cautionary examples cited: the initial draft human genome contained 223 bacterial genes attributed to horizontal gene transfer that were later shown to be contaminants; and the estimated human gene count fell steadily (from an early 30,000–40,000 range) as assemblies and annotation improved.

5. Method / approach

Core design: compare multiple draft assemblies against a higher-quality version of the SAME genome, so the discrepancy in gene count is a measurable error rather than a cross-species inference.

  • Chicken (main empirical system): five assemblies compared against reference v4.0 (Galgal4), spanning traditional and next-generation technologies. As reported:
    • 2X fosmid-based — 281,711 contigs — 14.1% full-length CEGMA
    • 12X 454-based — 45,554 contigs — 68.2% CEGMA
    • 6.6X Ref 2.1 (Galgal3) — 71,609 contigs — 66.5% CEGMA
    • 82X Illumina-based — 27,093 contigs — 74.6% CEGMA
    • 12X Ref 4.0 reference — 25,017 contigs — 80.7% CEGMA Sequencing technologies used across these: Sanger (fosmid, BAC-end), 454 (fragment + 3 kb and 20 kb mate-pair inserts), Illumina (100 bp paired-end, 2 kb inserts).
  • Chimpanzee: published draft Pan_troglodytes-1.0 vs. the updated Pan_troglodytes-2.1.
  • Simulation (mechanism test): simulated fragmented assemblies to isolate the cause of gene-number inflation. Genomes were broken into contigs whose lengths were drawn from the length distribution of a real fragmented genome (Daphnia pulex), then genes were predicted (GENSCAN). (Abstract frames this as “simulated genome assemblies of Drosophila melanogaster.”)
  • Correction test: RNA-Seq (paired-end) used to reconnect gene fragments split across contigs.
  • Completeness quantified with CEGMA (Core Eukaryotic Genes Mapping Approach) — % of core eukaryotic genes recovered full-length. (Paper predates BUSCO; CEGMA is the completeness metric it uses.)

6. Key claims / findings (atomic)

  • Headline: across the assemblies, no more than ~60% of gene families matched the reference family size, so 40% or more of gene families were inferred to have the wrong size in draft assemblies.
  • Errors go BOTH directions: incorrect assemblies both add and subtract genes (stated in the abstract). This bidirectionality is the load-bearing point for referee use.
  • Over-counting mechanism (dominant driver of gene-number inflation): gene FRAGMENTATION — one real gene split onto multiple contigs is predicted as multiple separate genes. The simulations confirm fragmentation as “the major cause of increased gene numbers.”
    • Simulation magnitude: at 17,941 contigs the simulated assembly yielded 32,025 GENSCAN-predicted genes with start+stop codons — “a handful more than are present in the published Daphnia pulex genome” — i.e. fragmentation raised the GENSCAN count from the ~20,000 predicted on the uncut D. melanogaster reference up to roughly the Daphnia pulex gene number.
    • Predicted gene number rose monotonically as the simulated genomes became more fragmented.
    • Chimpanzee: of gene families with incorrect counts (1,693 genes examined), 1,279 were cleaved genes and 414 were split alleles — both are over-counting mechanisms (cleaved = one gene split across contigs; split alleles = the two haplotype alleles assembled as separate genes).
  • Under-counting mechanism: missing / incomplete genes in the most fragmented assemblies — genes absent or too fragmented to be recovered, producing apparent loss.
    • The 2X fosmid assembly was a clear outlier, with more than half of all gene families missing gene copies relative to the reference; only ~20,000 complete genes could be extracted from it vs. ~36,000 predictions in the other assemblies.
  • Chimpanzee scale: the earlier chimp assembly predicted almost 1,800 more genes than the updated version; 74% of families had the same gene count in the two annotations, 26% differed.
  • Assembly-quality dependence (qualitative): error tracks with assembly quality/completeness — the least contiguous / lowest-CEGMA assembly (fosmid, 14.1% CEGMA) is the extreme outlier for missing genes, while high fragmentation drives inflation. Average gene length in the chicken annotation is 27.8 kb, so genes commonly exceed contig lengths and get truncated/fragmented. The paper reports these relationships descriptively/categorically; it gives NO correlation or regression statistic tying an assembly-contiguity metric (N50, contig count, coverage) to gene-number error. It uses N50 (e.g. citing Daphnia pulex N50 scaffold size < 400 kb) but does not define it.
  • Correction: RNA-Seq effectively improves draft annotations, largely by connecting genes fragmented across contigs; almost 12,000 predicted genes were removed by combining them with other genes when requiring only a single RNA-Seq read as evidence (Table 3: 11,931).

7. Load-bearing statements — VERBATIM (CC BY permits reproduction; each with location)

  1. Abstract: “We find that upwards of 40% of all gene families are inferred to have the wrong number of genes in draft assemblies, and that these incorrect assemblies both add and subtract genes.”
  2. Abstract: “Using simulated genome assemblies of Drosophila melanogaster, we find that the major cause of increased gene numbers in draft genomes is the fragmentation of genes onto multiple individual contigs.”
  3. Results: “For each assembly no more than 60% of all gene families were the same size as in the reference assembly, meaning that the remaining 40% or more of families were inferred to have the wrong size.”
  4. Results: “The fosmid assembly was a clear outlier, with more than half of all gene families missing gene copies relative to the reference.”
  5. Chimpanzee result: “Of these, 1,279 were cleaved genes and 414 were split alleles.”
  6. Discussion (the downstream-inference payoff): “The cascading effects of these errors may affect many downstream conclusions, from inferences about the evolutionary histories of genes to the ability to map genes involved in disease.”
  7. Discussion: “However, it is often the case that a great deal of research will be based upon the draft assembly before it has reached a finished state, and erroneous conclusions may result.”

(Also captured, gene-family-evolution framing — verbatim: “The observed variation in gene numbers may represent genetic diversity resulting from the evolution of gene families, but may also have been incorrectly inferred from sequencing and assembly artifacts.” And: “Variation in the number of genes may have important consequences for understanding differences between species, especially for key morphological, physiological, and behavioral traits.”)

8. Stated scope, assumptions, limitations

  • Scope is eukaryotic / vertebrate-anchored: empirical systems are chicken and chimpanzee, with an invertebrate simulation (Drosophila/Daphnia-derived fragmentation). Do not generalize beyond eukaryotes.
  • The “error” is defined relative to a higher-quality assembly of the same genome, which is itself a draft (chicken v4.0), not a truly finished genome — the reference is a proxy for truth.
  • Gene prediction is GENSCAN-based for simulations; annotation-pipeline differences between assembly versions contribute to some of the observed count differences.

9. Failure modes / invalidity patterns (referee-relevant)

  • Fragmented assembly → inflated gene count: a single gene split across contig boundaries is predicted as ≥2 partial genes → false apparent DUPLICATION / gene-family EXPANSION. Symptom/detector: many short/partial gene models; families larger than in related well-assembled genomes; low contiguity (large gene length relative to contig/N50 length; chicken mean gene 27.8 kb).
  • Incomplete / low-coverage assembly → missing genes: genes absent or unrecoverable → false apparent LOSS / gene-family CONTRACTION. Symptom/detector: low CEGMA/core-gene completeness (fosmid outlier at 14.1% full-length CEGMA; “more than half of all gene families missing gene copies”).
  • Heterozygosity / haplotype split (“split alleles”): the two alleles of a locus assembled as two separate genes → false duplication. (414 of 1,693 incorrect chimp genes.)
  • Contamination misassembled into the genome (e.g. the 223 “horizontally transferred” bacterial genes in the draft human genome that were contaminants) → false gene gains.
  • General detector the paper endorses: low assembly completeness (CEGMA) and low contiguity are the warning signs; RNA-Seq evidence that collapses fragments is both a corrector and a diagnostic of fragmentation.

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

  • No named birth-death / CAFE gene-family-size-evolution model is invoked or tested. The paper connects error to “the evolutionary histories of genes” and “the evolution of gene families” in general terms but does not run or reference a specific gene-gain/loss rate estimator (e.g. CAFE) — see §12.
  • No quantitative correlation/regression between a contiguity statistic (N50/coverage/contig count) and gene-number error; relationships are categorical.
  • No prokaryote/bacterial pangenome analysis (eukaryote-scoped).
  • No explicit BUSCO threshold (predates BUSCO); uses CEGMA without prescribing a numeric completeness cutoff for downstream use.

11. Open questions / ambiguities left unresolved

  • The exact split of error attributable to assembly vs. to annotation-pipeline differences across chicken assembly versions is not fully partitioned.
  • How much the reference-as-proxy-for-truth assumption understates total error (since the reference is itself imperfect) is not quantified.
  • No stated numeric threshold (completeness or contiguity) above which gene-count inference becomes “safe.”

12. Guidance answers

  • Study design / species / references: Confirmed. Draft assemblies compared against higher-quality versions of the SAME genome. Chicken: five assemblies (2X fosmid/Sanger, 12X 454, 6.6X Ref2.1/Galgal3, 82X Illumina, 12X Ref4.0) vs. reference v4.0 (Galgal4). Chimpanzee: Pan_troglodytes-1.0 vs. 2.1. Technologies: Sanger (fosmid/BAC-end), 454, Illumina. “High-quality reference” = the most complete same-genome assembly (chicken v4.0 / chimp 2.1), itself a draft used as truth proxy.
  • Headline fraction: >40% of gene families have the wrong number of genes (no more than ~60% match reference size). Chimp: 26% of families differ (74% match). See §7 quotes 1 and 3.
  • Both error directions: Confirmed and load-bearing. Over-count = gene fragmentation (major driver of inflation; simulations: 17,941 contigs → 32,025 predicted genes; chimp: 1,279 cleaved genes). Under-count = missing/incomplete genes (fosmid outlier, >half of families missing gene copies). Additional over-count mechanism: 414 split alleles (haplotype duplication). Direction that dominates: for gene-number inflation, fragmentation dominates; the most fragmented/incomplete assembly (fosmid) is dominated by missing genes — so the direction flips with assembly regime.
  • Assembly-quality dependence + completeness metric: Error scales with completeness/contiguity qualitatively; completeness metric = CEGMA (fosmid 14.1% → Ref4.0 80.7% full-length core genes). No BUSCO (predates it); no correlation/regression statistic provided.
  • Downstream-inference warning (the payoff): The paper DOES explicitly connect the error to evolutionary inference in general terms — see §7 quote 6 (“cascading effects… from inferences about the evolutionary histories of genes…”) and the gene-family framing (“…variation in gene numbers may represent genetic diversity resulting from the evolution of gene families, but may also have been incorrectly inferred from sequencing and assembly artifacts”). BUT it does NOT name gene-gain/loss RATE estimation, birth-death models, or CAFE specifically. So: the link to gene-family-size evolution is stated; the link to a specific birth-death/CAFE analysis is NOT. Do not infer the CAFE connection for the paper.
  • Recommendations: (1) use RNA-Seq to reconnect fragmented gene models before trusting counts; (2) use high-quality genomes from closely related species to improve assembly/gene models; (3) avoid strong evolutionary conclusions from draft assemblies, which can contain large numbers of both added and missing genes, until the assembly is improved. No numeric completeness threshold is prescribed.