Fossil calibrations at speciation nodes do not date gene trees

Question: How do Bayesian relaxed-clock dating methods perform on single-locus trees when fossil calibrations are available only at speciation nodes, and what alternatives to concatenation exist for dating gene duplication or deep coalescence events?

Answer

Taken together, the retrieved literature supports a narrow and somewhat uncomfortable answer: relaxed-clock dating of single-locus trees is usable but its error is dominated by information content rather than by the clock model, and the calibration constraint named in the question is a structural mismatch rather than a data shortage. Gene-tree nodes need not correspond to speciation nodes at all — single-locus lineages can diverge millions of years before the species they belong to [19], and gene-based divergence estimates carry a coalescent offset of order 2Ne generations [20] — so a fossil calibration that is correct for a species node is being applied to a node with a different biological meaning. Worse, in node-dated analyses the posterior at the calibrated node is often governed by the effective prior, particularly the maximum-age bound, rather than by the sequences [7]. What has genuinely replaced concatenation is not a new dating method but coalescent-aware branch-length estimation: discordance-aware pipelines produce less biased dates than concatenation or unpartitioned Bayesian analyses [10] and now extend to multi-copy gene families under duplication and loss [17]. For duplication and deep-coalescence events themselves, no retrieved paper dates them directly from fossils; the operative strategy is reconciliation of gene trees onto an already-dated species tree, with ILS-aware reconciliation outperforming duplication-loss-only models [9], and with the caveat that parsimony-cost gene tree parsimony is statistically inconsistent for every linear combination of duplication, loss and deep coalescence scores [14].

Single-locus dating accuracy is set by alignment information content, not by the relaxed-clock prior

Benchmarking 5,205 primate gene alignments in BEAST2 under a relaxed clock, date estimates depart further from the median age as alignments shorten, between-branch rate heterogeneity rises, and mean substitution rate falls — the three quantities that determine the statistical power available for dating [3]. The smallest deviations are recovered for genes with core biological functions such as ATP binding, cellular organisation and anatomical development, categories expected to be under strong negative selection and therefore comparatively clock-like [3]. Simulated alignments built from primate parameter ranges confirm these precision determinants but additionally expose directional bias when branch rates are highly heterogeneous, meaning that low-information single-locus trees are not merely imprecise but potentially biased [3].

Gene-tree nodes are not speciation nodes, so speciation-anchored calibrations are mis-assigned

The problem with calibrating single-locus trees is stated explicitly in that same benchmark: for gene-specific events, fossil calibrations only inform speciation nodes, and concatenation does not apply to divergences other than speciations [3]. Empirical work makes the mismatch concrete: in Fagus, five deeply diverged chloroplast lineages show divergence decoupled from speciation, predating modern species lineages by up to 28 million years [19]. A general coalescent correction makes the same point quantitatively — species divergence is delayed relative to average gene divergence by roughly 2Ne generations, a bias that Ks-based dating ignores unless ancestral effective population size is estimated [20]. The consequence at the level of trait inference is that hemiplasy arising from gene tree–species tree discordance is widespread in the bird tree of life [18], which is only interpretable if gene lineages and species lineages have genuinely different histories.

The calibration prior, not the sequence data, frequently determines the age

Node dating of crown angiosperms shows the posterior strongly constrained by the effective prior at that node, dominated by the maximum age constraint, whereas a skyline fossilized birth-death model yields similar crown ages under different origin-time priors [7]. A direct comparison of calibration strategies in Crocodylia finds calibration strategy matters far more than datatype, with node calibration giving younger ages and tip calibration producing older ages biased toward unrealistically ancient times through the structure of the underlying priors [6]. The delay between a clade's origin and its oldest fossil is inflated both by extinct lineages and by limited morphological character data, especially when fossils are few [16]. Data-driven alternatives exist: fossil occurrence data can be converted through a Bayesian Brownian Bridge into calibration densities, here constraining 110 nodes from more than 25,000 occurrences [4], and a corrected unresolved fossilized birth-death likelihood permits FBD priors for molecular clocks without morphological data and scales to hundreds of fossils [8].

Clock-model choice is itself a source of distortion and can be averaged rather than selected

Conditioning a Bayesian time tree on a single clock model treats clock choice as fixed and, when wrong, distorts both timescales and their uncertainty; a hierarchical clock-mixture model implemented for BEAST 2 retains the generating clock family in the 95% posterior model set across stratified simulations while keeping model-averaged substitution rate, root age and tree length accurate [1]. Reviews of the clock literature converge on the same prescription — reliability depends on explicit model testing, transparent calibration choices and careful interpretation rather than on the sophistication of the rate model alone [5].

Coalescent-aware branch-length estimation is the demonstrated replacement for concatenation, but it still dates species nodes

The clearest alternative to concatenation addresses branch lengths rather than the dating step: concatenation does not accommodate gene tree heterogeneity, and although Bayesian dating can model incomplete lineage sorting it does not scale, whereas a pipeline using discordance-aware topology and branch-length estimation followed by maximum-likelihood dating produces more accurate and less biased dates than concatenation or unpartitioned Bayesian methods and handles thousands of species and genes [10]. That approach has been extended to multi-copy data: no branch-length estimation method previously existed for gene family trees evolving with duplication and loss, and CASTLES-Pro now accounts for both duplication-loss and ILS while eliminating the systematic overestimation of terminal branch lengths typical of concatenation [17]. This is the strongest available answer to the concatenation half of the question, but note what it fixes — substitution-rate and branch-length estimation on the species tree — not the absolute dating of duplication nodes.

Duplication and deep-coalescence events are placed by reconciliation on a dated species tree, not dated directly

Reconciliation that treats ILS as a series of nearest-neighbour-interchange events and compares the cost of duplication-plus-loss against interchange-plus-loss is much more accurate than models ignoring ILS and at least as accurate as leading ILS-aware methods while scaling to larger datasets, demonstrated on 23 primate genomes [9]. This matters because the obvious shortcut is unsound: gene tree parsimony estimators are statistically inconsistent under the multispecies coalescent for all linear combinations of duplication, loss and deep coalescence scores [14]. Where absolute ages are wanted for duplications, the retrieved example applies a Bayesian molecular clock to the gene tree and then reconciles: glutamine synthetase GS2 diverged approximately 560 million years ago with GS1a and GS1b diverging about 70 million years later, and reconciliation revealed hidden duplications in genes conventionally treated as single-copy orthologs [11]. At genome scale, phylogenomic reconciliation that abandons synteny reliance and models fractionation scales to tens of thousands of gene trees and shows that LCA mapping can yield incorrect whole-genome-duplication predictions after fractionation [12], and probabilistic reconciliation via a Boltzmann distribution over reconciliations with Gibbs-style sampling extends the framework to segmental duplications [13].

Conflicts and gaps

Two disagreements are unresolved. First, calibration strategy: tip-dating is described in review as a more principled approach than node-dating [2], while a direct empirical comparison finds tip-calibrated timescales biased toward unrealistically ancient ages and node calibration more credible, calling for hybrid strategies [6], and a third study finds the skyline FBD robust precisely where node dating is prior-dominated [7]. Second, concatenation: the branch-length critique holds that concatenation fails to address gene tree heterogeneity [10], yet distance-based concatenation trees are argued to be statistically consistent even in anomaly zones, unlike character-based concatenation [15]. The larger gap is that the single-locus dating benchmark deliberately analysed only genes with no observed duplication or loss [3]; consequently no retrieved paper validates the accuracy of a duplication-node or deep-coalescence-node age estimate against a known true age. Reconciliation papers report event placements and relative orderings against a given species tree [9] [12], not posterior distributions for duplication ages anchored by fossils, so the propagation of speciation-node calibration error into duplication ages is unexamined.

What would change the answer

The decisive experiment is a controlled simulation extending the existing single-gene benchmark design [3] to gene families generated with known duplication times under a duplication-loss coalescent process, with calibrations supplied only at speciation nodes and scored on duplication-node age error and interval coverage. A second, cheaper test is to push the CASTLES-Pro and ML-dating pipeline [17] [10] onto multi-copy families with simulated true duplication ages and measure whether the terminal-branch bias correction carries through to event ages. Third, fossil-derived calibration densities built from occurrence data [4] and corrected unresolved FBD priors [8] could be attached directly to gene-tree nodes rather than species nodes and compared against speciation-anchored calibration on the same alignments. Fourth, applying a coalescent offset correction for ancestral Ne [20] inside a relaxed clock would show how much of the single-locus error documented empirically is coalescent depth rather than rate variation. Finally, sensitivity analysis that perturbs species-tree node ages and re-runs ILS-aware reconciliation [9] would reveal whether duplication orderings are robust to the calibration prior that dominates node-dated analyses [7].

References

[1] Xu, Y., Douglas, J., Bouckaert, R., Drummond, A. J.. "A hierarchical clock-mixture model for Bayesian phylogenetic dating". biorxiv. 2026-07-21. URL: https://doi.org/10.64898/2026.07.15.738815

[2] Heckeberg NS, Capobianco A, Khakurel B, Darlim G, Hohna S. "Practical Guide and Review of Fossil Tip-Dating in Phylogenetics.". pubmed. 2026-01-15. URL: https://doi.org/10.1093/sysbio/syaf050

[3] Louvel, G., Roest Crollius, H.. "Factors influencing the accuracy and precision in dating single gene trees". biorxiv. 2024-11-30. URL: https://doi.org/10.1101/2020.08.24.264671

[4] Wu R, Alvarez-Carretero S, Tong Y, Wan S, Schneider H, Clark J, Pisani D, Silvestro D, Donoghue PCJ. "Integrated analysis of fossils and molecular divergence time estimates a latest Jurassic origin of angiosperms.". pubmed. 2026-01-01. URL: https://doi.org/10.1038/s41477-026-02311-x

[5] Soares P, Rito T. "The Evolution of Molecular Clocks: Concepts, Calibrations, and Challenges.". pubmed. 2026-01-01. URL: https://doi.org/10.3390/biology15151329

[6] Davies SEL, Tasca A, Alvarez-Carretero S, Brochu C, Pisani D, Donoghue PCJ. "Competing methods of divergence time estimation recover diverse evolutionary timescales.". pubmed. 2026-10-07. URL: https://doi.org/10.1098/rspb.2026.0635

[7] Ma X, Zhang C, Yang L, Hedges SB, Zhong B. "New insights on angiosperm crown age based on Bayesian node dating and skyline fossilized birth-death approaches.". pubmed. 2025-01-01. URL: https://doi.org/10.1038/s41467-025-57687-9

[8] Lim W, Raskin LY, Li JK, Huelsenbeck JP, Nielsen R. "Estimating Divergence Times and Diversification Rates with the Unresolved Fossilized Birth-Death Process.". pubmed. 2026-01-01. URL: https://doi.org/10.64898/2026.09.10.749957

[9] Mishra, S., Smith, M. L., Hahn, M. W.. "reconcILS: A gene tree-species tree reconciliation algorithm that allows for incomplete lineage sorting". biorxiv. 2025-12-08. URL: https://doi.org/10.1101/2023.11.03.565544

[10] Tabatabaee Y, Claramunt S, Mirarab S. "Coalescent-based branch length estimation improves dating of species trees.". pubmed. 2025-01-01. URL: https://doi.org/10.1101/2025.02.25.640207

[11] Aledo E, Canas RA, Canton FR, Aledo JC. "Phylogenetic reconciliation provides new insights into the evolutionary diversification of the glutamine synthetase gene family in seed plants.". pubmed. 2025-01-01. URL: https://doi.org/10.1111/febs.70276

[12] Kalhor R, Lafond M, Scornavacca C. "Beyond Synteny: A Scalable Phylogenomics Method for Whole-Genome Duplication Detection.". pubmed. 2026-01-01. URL: https://doi.org/10.1177/15578666251415567

[13] Chan YB, Scornavacca C, Charleston M. "A Probabilistic Algorithm for Gene-Species Reconciliation with Segmental Duplications.". pubmed. 2026-01-01. URL: https://doi.org/10.1177/15578666251392595

[14] Sapoval, N., Nakhleh, L.. "On the consistency of duplication, loss, and deep coalescence gene tree parsimony costs under the multispecies coalescent". biorxiv. 2026-04-30. URL: https://doi.org/10.64898/2026.02.20.707019

[15] de Jong, M. J., Janke, A.. "DIST: Distance-based Inference of Species Trees". biorxiv. 2025-05-08. URL: https://doi.org/10.1101/2025.05.02.651899

[16] Carruthers T. "Interactions between phylogenetic trees and fossils when estimating evolutionary timescales.". pubmed. 2026-09-16. URL: https://doi.org/10.1093/jeb/voag055

[17] Tabatabaee Y, Zhang C, Arasti S, Mirarab S. "Species tree branch length estimation despite incomplete lineage sorting, duplication, and loss.". pubmed. 2025-01-01. URL: https://doi.org/10.1101/2025.02.20.639320

[18] Braun EL. "Evolution: Polymorphic genes and false convergence.". pubmed. 2026-05-04. URL: https://doi.org/10.1016/j.cub.2026.03.062

[19] Worth, J. R. P., Grimm, G. W., Ihara-Udino, T., Li, P., Papageorgiou, A. C., Simeone, M. C., Shanjani, P. S., Lopez-Saez, J. A., Chiang, Y.-C., Kitamura, K., Tomaru, N., Denk, T.. "Whole chloroplast genomes reveal a complex genetic legacy of lost lineages, past radiations and secondary contacts in the dominant temperate deciduous tree genus Fagus". biorxiv. 2025-06-06. URL: https://doi.org/10.1101/2025.06.03.653586

[20] Li, M., Li, X., Zhang, B.. "Tspecies, Rapid Optimization for Estimating Species Divergence Time Using Ks Distribution". biorxiv. 2025-05-23. URL: https://doi.org/10.1101/2025.05.18.654758

How this survey was made

This survey answers one question: “How do Bayesian relaxed-clock dating methods perform on single-locus trees when fossil calibrations are available only at speciation nodes, and what alternatives to concatenation exist for dating gene duplication or deep coalescence events?”

We keep our own collection of recent preprints and refresh it every day from bioRxiv, medRxiv, arXiv, PsyArXiv, SocArXiv and ChemRxiv. For each survey, we search that collection for recent work and keep the strongest matches: papers that use the same terms as the question, and papers on the same topic in different words.

The text was written from the titles and abstracts of those papers; the full texts were not read. Statements that draw on a particular paper are marked with a number in brackets. The reference list just above is the 20 papers whose work this text draws on, chosen from the 96 papers those searches returned; the rest of the shortlist is not printed. It is not a complete review of everything published on the question.

Treat this as a starting point and a reading list, not a peer-reviewed review. The searches behind this survey ran on October 29, 2026 and covered work from the previous two years.