Statistical Genomics Foundry
Paper
domain/gene-family-evolution

CAFE 5 — Mendes et al. 2020 (source note)

1. Citation

Mendes FK, Vanderpool D, Fulton B, Hahn MW. 2020. “CAFE 5 models variation in evolutionary rates among gene families.” Bioinformatics 36(22–23):5516–5518. DOI: 10.1093/bioinformatics/btaa1022. PMID: 33325502. URL: https://academic.oup.com/bioinformatics/article/36/22-23/5516/6039105 Software: CAFE 5, source/docs/manual at https://github.com/hahnlab/CAFE5/releases (paper’s stated availability; no specific version number given in the note). Access date: 2026-07-03.

2. Access note

Read via the Oxford Academic article page (abstract + body text as surfaced there). The ResearchGate full-text PDF returned HTTP 403 and could not be read. This is a short Applications Note; the captured content covers abstract, results, and the reported per-dataset parameter values. License mode: restrictive / all-rights-reserved → OWN-WORDS paraphrase for prose; functional strings (parameter names, numeric values) verbatim. No CC license present (see §7). Boundary: some fine body detail (exact supplementary figures, any full sentence-level nuance I could not surface) may be incomplete; where a fact was not found it is marked “not stated in paper.”

3. Thesis (1 sentence)

CAFE 5 is a rewritten gene-family gain/loss package whose central addition over prior CAFE versions is explicitly modeling rate variation among gene families using gamma-distributed rate categories, rather than assuming one shared rate for all families.

4. Problem & context

Genome sequencing shows frequent gene gains and losses between species. Earlier CAFE versions estimated gene gain/loss parameters across a phylogeny but assumed all gene families evolve at the same rate, despite evidence of large rate variation among families. CAFE models rates of change with a birth-death process having a mean rate λ (lambda) of gain and loss common to all families; the paper motivates relaxing that single-rate assumption.

5. Method / approach

  • Birth-death model of gene-family size change on a phylogeny; mean gain/loss rate λ estimated by maximum likelihood (own-words).
  • Rate variation among families modeled with a discrete approximation of the gamma (Γ) distribution, the same approach used by DupliPHY. Number of discrete rate categories K is specified a priori; each category is assumed equi-probable (1/K). The Γ distribution is scaled so the mean rate across categories is 1, with shape parameter α (with α = β) estimated from the data.
  • Empirical Bayes step: estimates the posterior probability that a given family belongs to each rate category, enabling downstream “slow” vs. “fast” family analyses.
  • Interface: strictly command-line; the earlier script-based paradigm (accessory scripts) was discarded. Command-line arguments now “preclude the use of accessory scripts” (own-words).
  • Output: reconfigured to minimize post-processing; trees written to a Nexus file for easy viewing.
  • Multithreaded; the paper reports noticeable performance increases up to at least 64 cores (Supplementary Fig. S1).
  • Exact default parameter values (e.g. default K) are not stated in the paper.

6. Key claims / findings (atomic)

  • CAFE 5 is a complete rewrite with performance and user-interface enhancements over prior versions.
  • New capability vs. prior CAFE: explicit modeling of rate variation among families via gamma-distributed rate categories (analogous to among-site rate models for nucleotides/ amino acids).
  • Number of rate categories K is user-specified a priori; categories equi-probable (1/K); Γ mean scaled to 1; shape α (=β) estimated from data.
  • Empirical Bayes gives posterior probability of a family’s rate category → identify slow/ fast families.
  • Per-dataset results reported (functional values, verbatim):
    • Primates: K = 4, λ = 0.00453, α = 0.62
    • Birds of paradise: K = 10, λ = 0.00226, α = 0.98
    • Hymenoptera: K = 6, λ = 0.00375, α = 0.373
  • The most rapidly evolving families (in gain/loss) flagged include sex/reproduction-related and immunity families.
  • Multithreading scales with noticeable gains up to at least 64 cores.
  • Improved output: Nexus trees; reduced post-processing.

7. Load-bearing statements (OWN-WORDS paraphrase — restrictive license; functional strings verbatim)

  • Copyright line, verbatim (functional/legal string): ”© The Author(s) 2020. Published by Oxford University Press. All rights reserved.” → confirms restrictive posture.
  • Prior CAFE assumed a single gain/loss rate λ common to all families; CAFE 5’s headline change is modeling among-family rate variation with gamma rate categories (own-words).
  • K (number of discrete rate categories) is specified a priori and each category is equi-probable at 1/K; the Γ is scaled to mean rate 1 with shape α = β estimated from data (own-words; the symbols K, α, β, λ and “1/K” are functional strings).
  • Reported parameter values per dataset are functional numeric strings, reproduced verbatim in §6.
  • The script-based paradigm of earlier versions was discarded for a strictly command-line interface (own-words).

8. Stated scope, assumptions, limitations

  • Model assumes a birth-death process of gene-family size change on a given phylogeny.
  • K must be chosen a priori by the user (paper does not give a rule or default for choosing K).
  • Rate categories assumed equi-probable (1/K) and the Γ scaled to mean 1 — modeling assumptions carried by the gamma-categories approach.
  • The paper does not report the total number of gene families per dataset, nor any minimum.

9. Failure modes / invalidity patterns

  • Not spelled out in this paper. The Applications Note names no explicit diagnostic, error message, invalidity condition, or value range at which the method breaks.
  • [summarizer-inferred] The single free knob highlighted is K (chosen a priori) and α (estimated); the paper gives no guidance on mis-specifying K, so failure/robustness under wrong K is not characterized here.

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

  • No claim of Type-I error control / false-positive control anywhere in the paper.
  • No per-family p-value and no explicit null-hypothesis test described; family-level inference here is the empirical-Bayes posterior rate-category probability, not a p-value.
  • No minimum number of gene families / orthogroups stated (no ”> 100”, no ”> 1000”).
  • No error/annotation-uncertainty model documented as a CAFE 5 feature. The paper’s only contact with error modeling is a citation to the CAFE 3 work (its title contains “in the presence of error in genome assembly and annotation”); the note surfaces no statement that CAFE 5 itself implements an error model.
  • No specific CLI flag names (e.g. -i, -t, -p, -k, -e, -o) appear in the surfaced text.
  • No specific output-file names (e.g. Base_results.txt, Base_clade_results.txt, Base_asr.tre, Base_change.tab) appear; only “Nexus file” is named.
  • No per-family or per-clade λ output is described; only global λ per dataset is reported.
  • No total gene-family counts per dataset; no runtime seconds/speedup ratio vs CAFE 4 in main text (only the ≤64-core scaling note).

11. Open questions / ambiguities

  • Default value of K and how to choose it: not given.
  • Whether CAFE 5 carries the CAFE 3 error model, and via what option: unresolved from this paper (would need the CAFE5 manual — a separate source; do not import here).
  • Exact CLI flags and output-file names: unresolved from this paper (manual-level).
  • What downstream test, if any, converts posterior rate-category probabilities into a significance call: not described.

12. Guidance answers

Q: What CAFE 5 adds over earlier versions (exact gamma mechanism). Discrete approximation of the gamma (Γ) distribution with K categories specified a priori, each equi-probable (1/K), Γ scaled to mean 1, shape α (=β) estimated from data (ML); empirical Bayes then gives each family’s posterior probability of belonging to a rate category, enabling slow/fast-family analysis. Same approach as DupliPHY. (Own-words; symbols verbatim.)

Q: Error-model feature / flags. Not documented as a CAFE 5 feature in this paper. The only error-modeling reference is a citation to CAFE 3 (Han et al.), whose title mentions “error in genome assembly and annotation.” No -e/error-model flag, and no flag-argument concatenation syntax, appears in the paper. → Treat error-model flags as manual-level, not paper-stated.

Q: CLI functional strings (-i, -t, -p, -k, -e, -o); is K=4 a default?; output artifact names. None of these flags are stated in the paper. K=4 appears only as the value used for the primates dataset, not as a documented default. Output file names (Base_results.txt, Base_clade_results.txt, Base_asr.tre, Base_change.tab) are not named in the paper — only a “Nexus file” for trees is mentioned. → All of these flags/ filenames are manual-only relative to this source.

Q1: “explicit Type-I error control.” The paper makes no such claim. No mention of Type-I error, false positives, or error-rate control anywhere. → Skill attribution to this paper is unsupported by the source.

Q2: “needs > 100 orthogroups (preferably > 1000)” / any minimum. The paper states no minimum number of gene families/orthogroups. → The >100/>1000 threshold is not grounded in this paper; treat as convention/manual, not Mendes 2020.

Q3: per-family p-value vs global-lambda-only; against what null. The paper reports a global λ (per dataset) and, for family-level inference, an empirical-Bayes posterior probability of a family’s rate category — not a per-family p-value and not an explicit null-model significance test. Reported λ values are global per dataset; no per-family or per-clade λ output is described.

Q: What lambda is and reported outputs. λ is the mean rate of gene gain and loss under the birth-death model, common to all families (in the base model), estimated by ML. Reported outputs surfaced: global λ per dataset (0.00453 / 0.00226 / 0.00375), shape α per dataset, and Nexus trees. Per-family / per-clade λ outputs are not described in this paper.