Reconstructing the branching order of species that split in quick succession is a data problem. When two speciation events happen close together, few mutations accumulate between them, and different genes tell different stories about who is related to whom. Theory says how many genes are needed. For quartet-based methods the number grows as $\log n / f^2$, where $n$ is the number of species and $f$ is the length of the shortest internal branch. Every one of these guarantees assumes the genes are independent draws, each sorting through ancestral populations on its own. Biology breaks that assumption in a specific, mappable way. In hybrids between diverging lineages, meiosis fails: chromosomes do not pair, checkpoints arrest sperm production, and recombination is suppressed across megabase-scale intervals around incompatibility loci, inversions and shifted centromeres. Thousands of "loci" in such an interval share one history. Meanwhile the rest of the genome may keep exchanging genes between lineages.
This note proposes a sample-complexity theory for species-tree estimation in which genes come in linked blocks, and in which the genome splits into two kinds of region. "Islands" are regions protected by meiotic incompatibility, which carry the true species-tree signal. "Permeable" regions are diluted by gene flow. The theory produces a computable decision rule: given a recombination map and a set of annotated incompatibility intervals, should one feed the species-tree method all loci, one locus per independent block, or only loci inside the protected islands? We test the rule in simulation and on mouse inter-subspecific hybrids, where the incompatibility architecture has already been mapped.
Core Claim
The quantity that controls accuracy is not the number of loci $m$. It is the number of effectively independent loci, weighted by how strongly each region favors the true topology. Write $B$ for the number of independent linkage blocks, $\bar b$ for the effective block size (the variance-inflation factor of gene-tree topologies within a block), and $\delta$ for the margin by which the true quartet topology beats its strongest rival in a region. Under the standard coalescent, $\delta = 1 - e^{-f} \approx f$; gene flow at rate $\gamma$ shrinks this margin. The block-level concentration argument behind the ASTRAL bound then gives a requirement of the form
$$ m \;\gtrsim\; \frac{\bar b \,\log n}{\delta^2}. $$
The form follows directly. Concentration of quartet frequencies holds over blocks rather than loci, so the variance term picks up $\bar b$, while the margin enters squared as it does in the original $\Theta(\log n / f^2)$ result. Comparing two candidate locus sets $S$ (for example "islands only" versus "whole genome") reduces to comparing the score $|S|\,\delta_S^2/\bar b_S$. The prediction that is new here is a crossover. Below a critical gene-flow level, using the whole genome wins. Above it, restricting to the meiotically protected islands wins, despite discarding most of the data. The crossover point can be computed from quantities that are measurable before any tree is built.
Relation to Prior Work
The bounds that motivate and constrain this work all come from one line of theory. Species Tree Estimation Using ASTRAL: How Many Genes Are Enough? proves matching $\Theta(\log n/f^2)$ bounds for the exact quartet-maximizing method on independent, error-free gene trees. Data Requirement for Phylogenetic Inference from Multiple Loci: A New Distance Method introduces METAL, which pools raw sequence differences across loci into one distance and needs total data $mk = O(f^{-2})$. Distance-based species tree estimation: information-theoretic trade-off between number of loci and sequence length under the coalescent shows that loci and sequence length trade off as $m = \Theta(f^{-2}k^{-1/2})$, so longer genes cannot replace more genes. All three treat loci as unlinked and discordance as neutral noise around one tree. We keep their proof techniques and replace the independence assumption with block structure and region-specific margins. Phylogenetic estimation error can decrease the accuracy of species delimitation shows empirically that when speciation intervals are short, species-level and within-species branching become indistinguishable. This is the regime where our margin $\delta$ is small and the choice of loci matters most. The radiation datasets in Clarification of the Phylogenetic Framework of the Tribe Baorini and Phylogenies without Branch Bounds: Contracting the Short, Pruning the Deep document the practical consequence: unresolved near-polytomies.
The meiosis literature supplies the biological structure that the theory needs as input. X Chromosome Control of Meiotic Chromosome Synapsis in Mouse Inter-Subspecific Hybrids maps a 4.7 Mb X-linked sterility interval, Hstx2, which interacts with Prdm9 and causes failed chromosome pairing and arrest of about 95% of pachytene spermatocytes in male hybrids. This gives us a concrete candidate island with known coordinates. Efficient targeted recombination with CRISPR/Cas9 in hybrids of Caenorhabditis nematodes with suppressed recombination shows that established hybrid lines suppress recombination strongly enough to resist even induced breaks. It motivates treating suppressed intervals as long, near-perfectly linked blocks. Four near-complete genome assemblies reveal the landscape and evolution of centromeres in Salicaceae shows that centromere positions move rapidly between related species. It supplies a second, structural kind of island whose boundaries can be read off assemblies. Speedy A governs non-homologous XY chromosome desynapsis adds a mechanism by which sex-chromosome pairing is policed. None of these papers asks what their mapped intervals imply for phylogenetic data requirements.
The ontology-querying direction plays a smaller, methodological role. Small Datalog Query Rewritings for EL shows that flat rewritten queries blow up because shared subexpressions are copied rather than reused, and that sharing structure removes the blowup. Query Optimization for Ontology-Mediated Query Answering builds a small "quotient summary" of a database, a compressed copy that groups equivalent individuals. The summary can soundly certify that a subquery will return nothing before the subquery is run on the full data. We borrow both lessons. Linked loci are the phylogenetic analogue of duplicated subexpressions, and a summary of the genome that collapses each block into one weighted unit is the analogue of the quotient summary. Here, though, the property we need is that the summary preserves the species-tree method's optimum when blocks are perfectly linked. We deliberately do not build an ontology-mediated query engine over genomic databases (Gap 2 in the source material), because it would be a database engineering project with no payoff for the phylogenetic question.
Research Program
The theoretical part extends the ASTRAL* analysis to a model in which the genome is a sequence of blocks. Within a block, gene trees are perfectly linked, or correlated with a specified decay. Each block is labelled as island (pure multispecies-coalescent sorting on the species tree) or permeable (a mixture in which a fraction $\gamma$ of gene trees follow an introgression topology). We will derive upper and lower bounds in terms of $B$, $\bar b$ and $\delta$. We will also derive the crossover gene-flow level $\gamma^*$ at which restricting to islands beats using all loci. The same extension will be carried out for METAL's pooled distance, where linkage inflates the variance of the summed differences. Recovering the original $f^{-2}$ results when $\bar b = 1$ and $\gamma = 0$ is a required sanity check.
The estimation part makes the decision rule usable on real data. $\bar b$ will be estimated from how quickly gene-tree agreement between loci decays with physical distance. $\delta$ for each region will be estimated from quartet frequencies inside that region. Island boundaries will come from two sources: annotated incompatibility intervals (Hstx2, centromere positions), and a data-driven scan for regions where gene-tree discordance drops. The output is a weighted locus set, with one weight per block, which is then passed unchanged to ASTRAL or METAL. The method changes only the input, so any gain is attributable to locus selection and weighting, not to a new tree-search algorithm.
Baselines
The naive reference point is the standard practice: every locus is treated as an independent gene tree and given to ASTRAL, or every alignment is pooled into METAL. This is what the existing bounds license, and it is the method our theory predicts will be overconfident and, under strong gene flow, wrong. It will be run on exactly the same loci, gene trees and compute as every other condition.
The strongest practical competitor is physical-distance thinning: keep one locus per fixed genomic window, with the window size chosen by a grid search over the same tuning budget we give our method. It is strong because it already addresses linkage. Our method beats it only if it achieves higher topological accuracy at equal numbers of retained blocks, or equal accuracy with fewer blocks. The case where we expect a difference is suppressed intervals far longer than any sensible window.
The strongest existing method is exact ASTRAL on all loci. Its guarantee is $\Theta(\log n/f^2)$ with an empirically near-perfect linear relationship between required genes and $1/f^2$ across 401 replicate simulated datasets with up to $10^5$ gene trees, reported in Species Tree Estimation Using ASTRAL. No reported result was found on Lacuna for ASTRAL or METAL accuracy under linked loci or region-specific gene flow, so the comparison numbers will come from our own reruns of that simulation design.
Two ablations isolate the key ingredients. The first applies the block correction but no island selection. The second selects random genomic regions matched to the islands in size and locus count. If random regions do as well as annotated islands, the meiotic biology is doing no work and the result reduces to "fewer, less correlated loci."
Why not just…?
Why not just thin loci by physical distance, as everyone already does? This is the most serious threat. In simulations with short linkage blocks it may well match our method. The proposal claims an advantage only where suppression is long-range and specific to the hybrid background. The Caenorhabditis results in Efficient targeted recombination show that recombination can be blocked in hybrid lines, which suggests that a recombination map measured within one species will underestimate linkage across the species boundary. We will vary block length, and the agreement between the species-specific recombination map and the true block structure, independently. If thinning matches our method across this grid, we will report that the contribution is the theory and the crossover rule, not the locus-selection procedure.
Why not just sequence more loci? Linked loci do not add independent evidence, so by the core inequality, growing $m$ inside a fixed set of blocks leaves the bound unchanged. Under gene flow it can also add evidence for the wrong topology. The simulations will show accuracy plateauing as $m$ grows with $B$ fixed. That plateau is a direct test of the block-level claim.
Why not just use longer sequences per locus? Mossel and Roch prove that sequence length reduces the locus requirement only by $\sqrt{k}$. Longer sequences also do nothing about margins eroded by gene flow, because they sharpen the estimate of each gene tree without changing which tree that gene follows.
Why not just model gene flow explicitly with a network method? The research file contains no network-inference paper to compare against, so we cannot cite a reported result. Network methods also answer a different question: they ask for the full reticulate history, not a bifurcating tree from a minimal reliable locus set. Our decision rule is cheap, sits in front of standard tree methods, and can serve as a diagnostic of when a network analysis is warranted.
Proposed Experiments
Simulation comes first. We will reproduce the ASTRAL* design (caterpillar, balanced and double-quartet species trees, up to $10^5$ gene trees) and add three knobs: block size, island fraction, and gene-flow rate $\gamma$ in the permeable regions. The decisive checks are three. Required gene count should scale with $\bar b/\delta^2$ rather than $1/f^2$. The empirical crossover gene-flow level should match the predicted $\gamma^*$ within simulation error. And with $\bar b = 1$, $\gamma = 0$, the original linear relationship to $1/f^2$ should be recovered.
The empirical test uses house mice. We will take the musculus–domesticus subspecies pair with outgroups and treat the 4.7 Mb Hstx2 interval, and other intervals implicated in pairing failure, as annotated islands. Then we will compare tree support and topology stability, under resampling of blocks, across the all-loci, thinned, block-corrected and island-only locus sets. The structural-island version repeats this on the Salicaceae assemblies, using shifted centromeres as island annotations. The Baorini multi-locus data serve as a short-branch case in which the score $|S|\,\delta_S^2/\bar b_S$ can be computed and checked against the observed lack of resolution.
The result that would falsify the proposal is annotated islands performing no better than size-matched random regions while block-corrected all-loci analysis matches island selection everywhere. That outcome would mean meiotic incompatibility maps carry no phylogenetic information beyond what generic linkage correction already captures.
References
- Species Tree Estimation Using ASTRAL: How Many Genes Are Enough?
- Data Requirement for Phylogenetic Inference from Multiple Loci: A New Distance Method
- Distance-based Species Tree Estimation: Information-Theoretic Trade-off between Number of Loci and Sequence Length under the Coalescent
- Phylogenetic estimation error can decrease the accuracy of species delimitation: a Bayesian implementation of the general mixed Yule-coalescent model
- Clarification of the Phylogenetic Framework of the Tribe Baorini (Lepidoptera: Hesperiidae: Hesperiinae) Inferred from Multiple Gene Sequences
- Phylogenies without Branch Bounds: Contracting the Short, Pruning the Deep
- X Chromosome Control of Meiotic Chromosome Synapsis in Mouse Inter-Subspecific Hybrids
- Efficient targeted recombination with CRISPR/Cas9 in hybrids of Caenorhabditis nematodes with suppressed recombination
- Four near-complete genome assemblies reveal the landscape and evolution of centromeres in Salicaceae
- Speedy A governs non-homologous XY chromosome desynapsis as a unique prerequisite for XY loop-axis organization
- Small Datalog Query Rewritings for EL
- Query Optimization for Ontology-Mediated Query Answering