Ancestral gene content estimates under gain–loss–duplication depend on the chosen observation threshold Authors: Gergely J. Szöllősi, Tom A. Williams Published: 2026 Csűrös (1) introduces a gain–loss–duplication (GLD) likelihood for reconstructing ancestral gene content from phylogenetic copy-number profiles. The likelihood and its gradient are obtained in closed form, and the method includes an analytic correction for the probability that a family’s summed copy count falls below the observation threshold. In the original analysis (1), GLD was applied to subsets of a 269-genome archaeal dataset corresponding to individual clades, but not to the domain as a whole. To enable a combined analysis of all 269 Archaea, and so to estimate the gene content of the last archaeal common ancestor (LACA), we reimplemented GLD in C, 20- to 80-fold faster, with added safeguards against boundary solutions in maximum-likelihood estimation (2). Fitting all genomes together, we found that the focal output of the analysis—reconstructed ancestral gene content—depends on a methodological choice whose effect should be reported explicitly: the minimum size threshold for gene families included in the analysis. The published analyses (1) retain families whose total copy count is at least Ωmin = 4. This threshold was chosen for comparison with phylogenetic reconciliation, which needs four copies for an informative unrooted topology (1). However, the GLD analysis can be run with any Ωmin, and there is no strong a priori reason to choose one value over another. Refitting the archaeal dataset at each Ωmin from 1 to 10, the inferred number of gene families at the root falls monotonically (Fig. 1A), from 3,847 at Ωmin = 1, 1,843 at the reported Ωmin = 4, to 1,506 at Ωmin = 10, a 2.6-fold range with no plateau that might suggest a stopping criterion (2). Discarding only the singletons (Ωmin = 1 → 2) already removes 36% of the inferred ancestral content. Throughout the sweep the inferred root stays larger than the median genome in the dataset. The biological conclusion—ongoing genome reduction vs. expansions and contractions from a moderately sized LACA—thus depends on a threshold the likelihood does not determine. Threshold dependence of the GLD ancestral genome reconstruction, on the published archaeal dataset (90,243 families across 269 genomes). (A) Inferred number of gene families (green) and gene copies (amber) at the root as the observation threshold Ωmin sweeps from 1 to 10. Violins show the family and copy counts of each of the 269 extant genomes. The Ωmin = 1 and Ωmin = 4 points (blue, orange) mark the two fits whose predictions are shown in (B). (B) Posterior-predictive check: predicted vs. observed family-size distributions, after removing singletons and normalizing over the common Ω ≥ 2 universe (log–log). Per-bin predicted/observed ratios are 0.89 to 1.11 for the Ωmin = 1 fit and 0.17 to 2.56 for the Ωmin = 4 fit. Shaded region: family size <4. One approach to setting Ωmin might be to select a value under which the model reproduces the observed family size distribution for extant genomes (Fig. 1B). However, this posterior-predictive check reveals a tension between the GLD model and the data: the Ωmin = 1 fit remains close to the empirical data (per-bin ratios 0.89 to 1.11) while estimating an implausibly large-genomed ancestor (3,847 families, ∼6,550 genes), whereas Ωmin = 4 supports an ancestor of moderate genome size, but fails to reproduce the extant data. We attribute this to inadequate HGT modeling in profile-based methods—a bias opposite to the conservatism of reconciliation methods (1, 3). The GLD framework is a valuable advance, but we suggest that inferred ancestral genome sizes be reported together with the Ωmin that produced them, and that conclusions resting on Ωmin = 4 be qualified, since at that threshold the model does not reproduce the observed family-size distribution (Fig. 1B).