Statistical Genomics Foundry
Paper
domain/whole-genome-alignment

Kent et al. 2003 — Evolution’s cauldron (chains & nets)

1. Citation

Kent WJ, Baertsch R, Hinrichs A, Miller W, Haussler D. 2003. “Evolution’s cauldron: Duplication, deletion, and rearrangement in the mouse and human genomes.” PNAS 100(20):11484–11489. doi:10.1073/pnas.1932072100. Open-access mirror: PMC208784 (https://pmc.ncbi.nlm.nih.gov/articles/PMC208784/). Accessed 2026-07-03.

2. Access note

Full text read via the PMC open-access mirror. License posture: PNAS, freely readable but all-rights-reserved → this note is written own-words; tool/format names and numeric thresholds are kept verbatim as functional strings (Step 1.5 restrictive mode).

Capitalization caveat: the tool names were extracted through a page-to-markdown fetch that appears to have lower-cased them (rendered as blastz, axtchain, chainnet, netfilter, netsynteny, netclass, lavtoaxt). The UCSC canonical forms are camelCase (axtChain, chainNet, netFilter, lavToAxt, netSyntenic, netClass); treat the exact casing as [re-check against the PDF]. The concepts/names themselves are firm.

3. Thesis

By aligning the whole mouse and human genomes, the paper catalogs duplication, deletion, and rearrangement across all scales — single base to whole chromosome — using chained/netted alignments to organize the events. The two genomes are similar enough to align at the DNA level yet diverged enough to expose substantial structural change.

4. Problem & context

Whole-genome pairwise alignment produces many local, gapless alignment blocks that, alone, don’t express large-scale structure (synteny, inversions, duplications, deletions). The paper introduces a chaining step (assemble co-linear blocks) and a netting step (pick the single best chain per region, hierarchically) to recover evolutionary events from raw alignments.

5. Method / approach (pipeline)

Ordered stages (tool names as printed; casing flagged above):

  1. Masking before alignment. Transposon-derived repeats masked with repeatmasker; simple repeats (period ≤ 12) masked with tandem repeat finder, prior to alignment.
  2. Local alignment. blastz produced gapless local alignments; lavtoaxt converted output. Initial blastz covered ~35.9% of the human genome.
  3. Chaining (axtchain). Assembles the gapless blocks into chains using a variation of a k-dimensional tree (kd-tree)-based algorithm. Uses the same nucleotide scoring matrix as blastz but a novel piecewise linear gap scoring scheme to form maximally scoring chained alignments from the gapless subsections. Block order within a chain must be consistent with genomic order in both species (co-linearity). Scoring permits simultaneous gaps in both species (to model overlapping deletions). After chaining, ~34.6% of human bases align to mouse.
  4. Netting (chainnet). Builds the net hierarchy (see §6 for algorithm).
  5. Post-net classification/filtering. netsynteny, netclass, netfilter add synteny/class annotation and produce a filtered “syntenic subset.”

6. Key claims / findings (atomic)

  • Chain definition. A chain is an ordered sequence of pairwise nucleotide alignment “blocks” separated by larger gaps, some of which may be simultaneous gaps in both species.
  • Net algorithm. All bases start “unused.” Chains are sorted highest-score-first. Iterating: take the next chain, discard the parts intersecting already-covered bases, mark the remainder covered. A chain covering bases that fall in a gap of a previously taken chain is marked a child of that chain — forming a hierarchy of chains (parent/child levels).
  • Reference-directionality (asymmetry). The net is not symmetric between species: the human net allows only the single best-aligning mouse DNA to align to any one place in the human genome; the mouse net is built the opposite way. → target-net vs query-net differ.
  • No numeric chain-length/score threshold is stated. The number 5000 does not appear in the paper; there is no stated minimum chain length or minimum chain score cutoff. (Confirms the “5000 bp” figure is UCSC convention, not attributable to Kent 2003.)
  • Syntenic subset rule. To be “syntenic,” a chain must either have a very high score itself, or be embedded in a larger chain on the same chromosome from the same region. This subset filters out processed pseudogenes while keeping tandem duplications. (Qualitative “very high score” — no numeric cutoff given in the text.)
  • Synteny scale. 344 large (>100-kb) blocks of conserved synteny; these 344 appear at the top level of the human net.
  • Chain length distribution. 579 long chains span 100,000 to ~115,000,000 (≈115 Mb) bases; many thousands more short chains at the top level.
  • Per-megabase event rates (excluding transposon insertions). Per Mb of genomic alignment, on average: 2 inversions, 17 duplications (5 tandem or nearly tandem), 7 transpositions, and 200 deletions of 100 bases or more.
  • Larger events within blocks (≥100 kb). ~160 inversions ≥100 kb; ~75 duplications/ transpositions ≥100 kb.
  • Nonsyntenic small events. Nonsyntenic duplications ≈11.9/Mb (mostly processed pseudogenes); nonsyntenic transpositions ≈5.0/Mb.
  • Gene loss estimate. ≈2% of genes present in the human/mouse common ancestor were deleted in the mouse lineage.
  • Coding coverage. The complete net covers 96.7% of RefSeq coding regions; the syntenic subset covers 93.0%.
  • Gap-size distribution. Gap sizes do not follow a geometric distribution; Alu insertions create a peak around ~300 bases.

7. Load-bearing statements (own-words paraphrase — restrictive license)

No verbatim prose reproduced; facts/terms/numbers preserved. Functional strings (tool names, numbers) verbatim per Step 1.5.

  • A chain = an ordered sequence of gapless alignment “blocks” separated by larger gaps, some of which may be simultaneous gaps in both species.
  • axtchain reuses the blastz nucleotide scoring matrix but adds a piecewise linear gap scoring scheme to build maximally scoring chains; block order must be consistent with genomic order in both species.
  • Net construction: bases begin unused; chains sorted highest-score-first; each chain is trimmed to the not-yet-covered bases, which are then marked covered; a chain landing in a prior chain’s gap becomes that chain’s child — yielding a chain hierarchy.
  • The net is asymmetric: the human net keeps only the single best mouse DNA per human position; the mouse net is built oppositely.

8. Stated scope, assumptions, limitations

  • Method is a pairwise human↔mouse comparison; the net is inherently directional/ reference-anchored (one genome is the axis the “best chain per position” is chosen against).
  • Event counts are derived from what aligns; the ~34.6% aligned fraction bounds the analysis (regions that don’t align aren’t classified).
  • “Very high score” for the syntenic subset is a qualitative rule as stated; the paper’s printed text (as read) gives no numeric score cutoff for it.

9. Failure modes / invalidity patterns

  • Reference-bias / asymmetry. Because the net keeps only one best chain per position on the reference side, human-net and mouse-net results differ; conclusions about “the” alignment depend on which genome is the target. Using the wrong-direction net is a validity error the asymmetry statement warns against.
  • Repeat-driven spurious alignment. Handled by masking transposons (repeatmasker) and simple repeats period ≤12 (tandem repeat finder) before alignment; skipping masking would inflate spurious chains. (Paper prescribes masking as the guard; it doesn’t quantify the failure if omitted — [summarizer-inferred] that omission causes spurious chains.)
  • Processed pseudogenes masquerading as duplications. The syntenic-subset filter (netfilter + score/embedding rule) is the named detector separating processed pseudogenes (nonsyntenic) from real tandem duplications (kept).

10. What the source does NOT address

  • No fixed numeric minimum chain length or chain score (no 5000).
  • Not a multi-species/graph alignment method — strictly pairwise, one reference at a time.
  • Exact blastz scoring-matrix values and the piecewise-linear gap penalty coefficients are named but not tabulated in the read text.

11. Open questions / ambiguities

  • Precise numeric thresholds behind “very high score” and top-level vs child assignment are not given in the read text (may live in supplementary/UCSC docs, not this paper).
  • Exact camelCase of the tool names (axtChain/chainNet/etc.) needs confirmation against the PDF; the fetched markdown lower-cased them.

12. Guidance answers

  • Chain/net algorithm ordering & definitions. Captured (§5, §6, §7): local blastz blocks → axtchain assembles co-linear chains (order consistent in both species; simultaneous gaps allowed; piecewise linear gap scoring) → chainnet selects best chain per region, sorting by score and marking gap-filling chains as children (hierarchy).
  • Reference asymmetry / directionality. Yes — explicitly stated and paraphrased (§6, §7): the net is not symmetric; the human net keeps only the single best mouse DNA per human position, the mouse net the opposite. Backs the reference-bias validity axis.
  • Minimum chain length / score threshold? Absent. No such number; 5000 does not appear. Keep “5000 bp” labeled UCSC convention, not cited to Kent 2003. (§6)
  • Repeat handling. Yes — repeatmasker for transposons and tandem repeat finder for simple repeats (period ≤ 12) mask sequence before blastz; the syntenic subset further filters processed pseudogenes. (§5, §9)