CAFE: a computational tool for the study of gene family evolution (De Bie et al. 2006)

1. Citation

De Bie T, Cristianini N, Demuth JP, Hahn MW. 2006. “CAFE: a computational tool for the study of gene family evolution.” Bioinformatics 22(10):1269–1271. DOI 10.1093/bioinformatics/btl097. URL: https://academic.oup.com/bioinformatics/article/22/10/1269/237347 No open-access CC license; page carries ”© The Author 2006. Published by Oxford University Press. All rights reserved.” Access date: 2026-07-03.

2. Access note

Applications Note (short, ~3 pages). Abstract and main-text body text were readable via the article web page; the typeset PDF is subscriber-only. Content below was reconstructed from the readable web text and cross-checked against an independently hosted copy and the paper’s own downstream description; the three branch-attribution method names and their descriptions were corroborated by a second source. No paywall boundary blocked the load-bearing content the guidance targets. License mode: restrictive/all-rights-reserved → §7 rendered as own-words paraphrase; functional strings (term/method/parameter names, equation-role symbols) reproduced verbatim.

3. Thesis (1 sentence)

CAFE is a tool that fits a stochastic birth-and-death process to gene-family sizes over a given phylogeny, and from that fit estimates rates, infers ancestral family sizes, flags families with unusual (accelerated) rates of gain and loss via a p-value, and points to the branches responsible for a family’s significance.

4. Problem & context

Gene families change in size across a phylogeny through gene gain (duplication) and loss. The paper’s premise is that these size changes can be modeled statistically as a birth-and-death process along a species tree, letting one ask which families evolve unusually fast and where on the tree the unusual change occurred. CAFE packages this analysis for a specified tree plus a matrix of observed family sizes in extant species.

5. Method / approach

Model: a stochastic birth-and-death process for gene-family size evolution over the phylogeny, governed by a single global rate parameter λ — the per-gene probability of gain or loss per unit time. The model assumes gene birth and gene death are equally probable (one shared λ, not separate birth and death rates).

Given a tree and observed extant family sizes, CAFE performs:

Rate estimation — estimate the global birth-and-death rate λ by maximum likelihood, or the user may supply a fixed numerical λ.
Ancestral reconstruction — infer the most likely gene-family size at every internal (ancestral) node; described as the “Viterbi” (most-likely-state) assignment for each family.
Per-family significance — for each family, compute a p-value for the observed extant family sizes under the fitted model (see §5a).
Branch attribution — for significant families, identify which branches drive the small p-value, via three methods (see §5b).

5a. Per-family p-value (load-bearing)

For each family, CAFE computes a p-value for the observed extant-species family sizes given the fitted birth-and-death model (the null = the single-global-λ birth-death process fitted to the data). Small p-values indicate the observed configuration is unlikely under that model; families with large variance in size — especially among closely related species — tend to get low p-values. The paper states that computing these p-values exactly is not feasible, so Monte Carlo sampling (simulation under the fitted model) is used to obtain them.

5b. Branch-specific attribution (load-bearing) — three methods

Each method asks, for a family already flagged, which branch(es) make its overall p-value small; the reference/null in each case is the fitted single-global-λ model:

Viterbi — using the Viterbi (most-likely) ancestral-size assignments, compute a p-value for the parent→child transition along each branch; branches with low p-values represent unusually large changes (expansions or contractions).
Branch cutting — test whether the family’s overall p-value increases when a given branch is “cut,” i.e. the probabilistic coupling between that parent and child family size is removed; a large increase implicates that branch as responsible for the low overall p-value.
Likelihood ratio test — re-maximize the family’s likelihood allowing a separate λ on the branch under test, and compare (ratio) the two-parameter likelihood to the single-λ likelihood to assess whether an extra rate parameter is warranted on that branch.

5c. Inputs and assumptions

Tree: a Newick description of a rooted and bifurcating phylogenetic tree, including branch lengths in units of time.
Data: a family-size matrix giving gene-family sizes for the extant taxa; may hold one family up to thousands of families for the specified tree.
λ: user-supplied single numeric value, or found by maximum likelihood.
Assumes gene birth and death are equally probable. Computation scales linearly with the number of species.

6. Key claims / findings

CAFE fits a birth-death model with a single global rate λ to family-size data over a fixed tree.
It estimates λ (ML) or accepts a fixed λ.
It infers most-likely ancestral family sizes at internal nodes (Viterbi assignment).
It assigns each family a p-value for having an accelerated / unusual rate of gain–loss, computed under the fitted model via Monte Carlo sampling.
For significant families it locates responsible branches by three methods: Viterbi per-branch transition p-values, branch cutting, and a per-branch likelihood ratio test.
Required tree is rooted, bifurcating, with branch lengths in time units.
Runtime scales linearly with number of species.
The paper reports no multiple-testing / FDR correction across families (see §9, §10, §12).

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

Abstract’s stated capabilities (paraphrased): for a specified tree and given extant family sizes, CAFE can estimate the global birth-and-death rate, infer the most likely family size at all internal nodes, identify families with accelerated rates of gain and loss quantified by a p-value, and identify which branches cause the p-value to be small for significant families.
Per-family p-value (paraphrased): for each family CAFE computes a p-value for the extant-species family sizes given the model; obtaining these p-values requires Monte Carlo sampling for computational feasibility.
Branch cutting (paraphrased): “cutting” a branch means removing the probabilistic coupling between the parent and child family sizes on that branch; if the family’s overall p-value rises substantially after cutting, that branch is implicated as the cause of the low overall p-value.
Verbatim functional strings: λ; p-value; method names “Viterbi”, “branch cutting”, “likelihood ratio test”; input format = Newick, tree “rooted and bifurcating” with “branch lengths in units of time”.

8. Stated scope, assumptions, limitations

Single global rate λ; birth and death equally probable (one shared rate, no separate gain vs. loss rates in the base model).
Requires a rooted, bifurcating tree with time-calibrated branch lengths supplied by the user.
p-values are obtained by Monte Carlo, i.e. approximate/simulation-based, not closed-form.
It is an Applications Note describing the tool; the underlying birth-death methodology is attributed to the associated primary methods work (Hahn et al.), not re-derived here.

9. Failure modes / invalidity patterns

Tree/branch-length dependence: results are conditional on the user-supplied rooted, bifurcating tree and its branch lengths in time units; a wrong topology or miscalibrated branch lengths invalidates the fit (source states the input requirement; the invalidation consequence is [summarizer-inferred]).
Single-λ assumption: the base null assumes one global rate with equal birth and death; a family or lineage that genuinely violates this is exactly what the p-value / branch tests are meant to surface, but the fitted global λ is itself the reference, so a poorly-fitting global rate biases every family’s p-value [summarizer-inferred].
Monte Carlo approximation: p-values are simulation-estimated, so precision depends on sampling; the paper notes Monte Carlo is used for feasibility but does not state a sample-count default or precision bound in the readable text.
High-variance families: families with large size variance, especially among closely related species, tend toward low p-values — flagged by the source as the signal, i.e. the expected driver of significance.

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

Multiple testing across families: the readable text contains no mention of multiple-testing correction, false discovery rate (FDR), family-wise error, or any adjustment for testing thousands of families simultaneously. The paper presents the per-family p-value (and per-branch p-values) with no correction step and leaves any such adjustment unstated — i.e. to the user.
No stated default for the Monte Carlo sample size or a p-value significance threshold in the readable text.
No separate birth vs. death rate model (base model ties them).
No treatment of unrooted or multifurcating trees, or of branch lengths not in time units.

11. Open questions / ambiguities

How many Monte Carlo samples are used, and what precision that gives for very small p-values.
What significance threshold, if any, the authors intend for the per-family or per-branch p-values.
How the three branch-attribution methods should be reconciled when they disagree.
Whether any of the per-branch p-values are themselves corrected for testing every branch.

12. Guidance answers

What CAFE (v1) computes on top of the birth-death model: (1) estimates the global birth-death rate λ (ML or user-fixed); (2) infers most-likely ancestral family sizes at internal nodes (Viterbi assignment); (3) assigns each family a p-value for accelerated/unusual gain-loss, computed under the fitted single-global-λ birth-death model via Monte Carlo sampling (the null is that fitted model); (4) identifies responsible branches for significant families via three methods — Viterbi (per-branch parent→child transition p-value), branch cutting (does the overall p-value rise when the parent-child coupling on a branch is removed), and a per-branch likelihood ratio test (fit a separate λ on the branch and compare likelihoods). Null for the per-family and per-branch tests alike is the fitted global-λ model.
Multiple testing across families: Not addressed. The readable text has no mention of FDR / multiple-comparison correction; CAFE reports raw per-family (and per-branch) p-values with no correction step. This is a confident silence — the paper neither performs nor prescribes any across-family correction, effectively leaving it to the user. Supports the “no CAFE primary for multiple-testing” hypothesis in the guidance: this paper does not provide that primary.
Input requirements / tree assumptions: Newick, rooted and bifurcating tree with branch lengths in units of time; a family-size matrix for extant taxa (one to thousands of families); λ user-specified or ML-estimated. Assumes birth and death equally probable; runtime scales linearly with number of species.
Functional strings (verbatim, per guidance must-quote): per-family significance output = “p-value”; branch-attribution methods = “Viterbi”, “branch cutting”, “likelihood ratio test”; rate parameter = λ.