Gene-model artifacts behave as directional signal, not noise, in pangenome introgression tests

Question: How do annotation-transfer errors and shared gene-model artifacts bias phylogenetic signal and introgression tests in comparative pangenome analyses of closely related taxa?

Answer

Across these papers, annotation and gene-model artifacts do not behave like the zero-mean, independent noise that phylogenomic and introgression machinery is designed to tolerate. They are directional, because transfer anchors target gene models to a source reference; they are taxon-correlated, because samples processed with the same pipeline, protein database, or reference panel share the same distortions; and they are concentrated in exactly the axes — gene presence/absence, orthogroup membership, and per-lineage substitution rate — that site-pattern and quartet-based tests use to infer relationships and gene flow. The result is that annotation provenance becomes recoverable from the data as if it were phylogenetic structure. The strongest documented evidence for this chain is still indirect: pangenome presence-absence variation recapitulating annotation method, and rate/topology asymmetry driving false introgression calls. No retrieved study closes the loop by injecting empirically calibrated annotation-transfer error into a D-statistic or network-inference pipeline, so the magnitude of the downstream introgression bias remains inferred rather than measured.

Presence-absence variation computed from public pangenome annotations recovers annotation provenance, not biology

PAV calculated directly from genome annotations hosted on public repositories recapitulates structure in annotation methods and not biological sequence differences, and re-annotating all genomes with a single identical pipeline resolves the broadest-stroke issues while leaving substantial residual problems, including exactly identical sequences carrying different gene-model structural annotations [20]. Heterogeneous annotations produced with different tools and protein databases introduce artifactual differences when comparing gene sets or gene structures between species [10]. The lack of widely accepted standards produces great diversity in completeness and accuracy across annotations, making comparative analyses susceptible to errors that can be misinterpreted as biological variation [7]. Consistent with this, gene number and CDS length show only moderate phylogenetic heritability (40% and 9.7%) across 301 drosophilid species, implicating species-specific factors including assembly error in shaping observed differences [3].

Transfer is reference-anchored, so its error has a preferred direction and a preferred topology

Reconstructing loci repeatedly from different reference genomes reveals reference bias in more than a quarter of BUSCO loci, with reconstructed loci acquiring erroneous similarity to the reference; excluding putatively biased loci yields a species tree topology more consistent with prior expectations [1]. The transfer mechanism itself is explicitly source-anchored, projecting annotations from a source genome through a variation graph onto other embedded genomes [21]. Error need not be random even in principle: error-correction objectives for perfect phylogenies are statistically inconsistent at all positive error rates, and the winning incorrect trees are not arbitrary but systematically share topological features, with deeper, more imbalanced topologies favoured [8]. Misannotation also corrupts the metrics used to certify input quality, since BUSCO misannotations can misrepresent assembly quality, and curated ortholog sets yield up to 6.99% fewer false positives than standard searches [17].

Orthogroup definition changes which loci exist, and the change is large and locus-specific

Conflicting cluster assignments against a graph-based reference vary from approximately 14% in MADS-box to approximately 57% in MYB and associate with protein length differences, while core, shell, and cloud classifications shift substantially by method and Ka/Ks for core genes is highly method-sensitive [13]. In a polyploid-rich genus, phylogenetic outcomes varied with the paralog processing method used and no method performed best across criteria [12]. Gene tree estimation error is not a benign analogue of incomplete lineage sorting: at matched overall discordance, it has a stronger detrimental effect on species tree accuracy than ILS, and it generates uniform high-entropy noise that does not diminish with additional genes [23]. Lineage-specific rate variation and estimation inaccuracy degrade inference of hybrid node placement in four-node cycles, with accurate detection complicated by gene tree estimation error [28].

The asymmetry annotation artifacts generate is the same asymmetry that introgression tests read as gene flow

Among-lineage substitution rate heterogeneity generates spurious signals of introgression, and reanalysis with an explicit model overturned prior interpretations of recurrent non-sister introgression in Jaltomata in favour of rate-driven artefacts plus two ghost events [15]. At shallow timescales, weak (17%) and moderate (33%) rate variation inflate false-positive rates for D-statistic and HyDe up to 35% and 100% respectively from a 500 Mb genome, and a more distant outgroup intensifies the spurious signal [26]. Lineage-specific rates severely biased network inference, and the standard worst-residual threshold of 3 for find_graphs inflates type I error even without rate variation [27]. Input data error more broadly inflates network complexity and produces non-biological reticulations, and this is exacerbated when the number of allowed reticulations exceeds the truth [6]. Biallelic site patterns also fail under deeper divergences where multiple hits cause false positives [14].

Shared errors shrink effective sample size, so adding genomes does not average them away

Gene tree dependence reduces the effective sample size by a constant factor, which matters because correlated errors behave as dependence rather than as independent replicates [25]. Simple implicit methods can outperform explicit phylogenetic ones under error: gene family presence/absence patterns across species trees consistently outperformed gene tree–species tree reconciliation for transfer inference [5]. Distinguishing true gene absence from non-detection is a modelling problem amenable to phylogenetic occupancy models that borrow information across related genomes [19]. Calibrated information-content measures allow detection of dataset dissonance and of loci that carry no topological information [4].

Conflicts and gaps

Whether orthology error materially biases topology is genuinely disputed. Hidden paralogy had limited impact on network inference in an eight-taxon simulation, with both network methods correctly preferring a tree without reticulation [27], whereas paralog processing method changed topology in Packera [12] and orthogroup method changed Ka/Ks in rice [13]; these may be different downstream quantities, but the papers do not reconcile them. Site and taxon filtering is also contested: subsampling to the most informative loci can produce a better species tree sample when data are uninformative [2], and dynamic site removal improves accuracy [9], yet maximum-likelihood tools showed no systematic overfitting across thousands of alignments and excluding sites substantially reduced both signal and accuracy [11]. Ghost lineages are treated as both a major interpretive hazard [22] and a low-probability one for ordering gene acquisitions [24]. The central gap is that no retrieved paper quantifies annotation-transfer error propagation into introgression statistics: the introgression-bias literature models rate variation, ILS, and estimation error, while the annotation literature stops at PAV, Ka/Ks, and assembly-quality metrics. Pangenome graph construction error is benchmarked against known genealogies [18] and false duplications are detectable in assemblies [16], but neither is propagated to gene-flow inference. Nothing establishes an acceptable transfer-error threshold.

What would change the answer

An error-injection experiment with empirically calibrated parameters: estimate per-gene transfer discordance from paired pipeline-swap and reference-swap reannotations of the same genomes, following the logic of multi-reference locus reconstruction [1] and single-pipeline reannotation [20], then inject that error into genealogy-aware pangenome simulations with known introgression [18] and measure false-positive rates for D-statistics, fd, QuIBL, and network search. A cheaper diagnostic would be a pipeline-swap null: annotate one taxon set two ways and report the distribution of ΔD and Δtopology as the artifact floor. Reanalysis of published claims in systems with reannotation resources — 82 Fusarium genomes [7], 301 drosophilids [3], cotton and soybean pangene sets [20] — using rate-aware and ghost-aware introgression methods [15] [22] would test whether published gene-flow signals survive annotation harmonisation. Extending occupancy-style models of unobserved gene presence [19] to the introgression likelihood would provide a principled rather than filtering-based correction.

References

[1] Ivan, J., Lanfear, R.. "PhyloRBT: A Phylogenetic Approach to Detect Reference Bias in Phylogenomic Datasets". bioRxiv. 2026-07-26T00:00:00.000Z. URL: https://doi.org/10.64898/2026.07.24.740642

[2] Milkey, A., Chen, J., Lewis, P. O.. "How much information is there for inferring species trees?". bioRxiv. 2026-04-02T00:00:00.000Z. URL: https://doi.org/10.64898/2026.04.01.715836

[3] Dhakad, P., Kim, B., Petrov, D., Obbard, D.. "Comparative gene annotation and orthology assignments across 301 species of Drosophilidae". bioRxiv. 2026-02-03T00:00:00.000Z. URL: https://doi.org/10.1101/2025.04.14.648771

[4] Milkey, A., Lewis, P. O.. "Estimating Bayesian phylogenetic information content using geodesic distances". bioRxiv. 2026-04-01T00:00:00.000Z. URL: https://doi.org/10.64898/2026.03.31.715656

[5] Mishra, S., Lercher, M. J.. "Horizontal Gene Transfer Inference: Gene presence-absence outperforms gene trees". bioRxiv. 2025-05-28T00:00:00.000Z. URL: https://doi.org/10.1101/2024.12.27.630302

[6] Tamiji, M., Sapoval, N., Nakhleh, L.. "Impact of Data Error on Phylogenetic Network Inference from Gene Trees Under the Multispecies Network Coalescent". bioRxiv. 2025-05-23T00:00:00.000Z. URL: https://doi.org/10.1101/2025.05.18.654708

[7] Hill, R., Reynolds, G., Hall, N., Swarbreck, D.. "Leveraging existing data to maximise quality and consistency across gene model annotations: a Fusarium pan-annotation". bioRxiv. 2025-03-14T00:00:00.000Z. URL: https://doi.org/10.1101/2025.03.12.642647

[8] Satas, G., Myers, M. A., Shah, S. P.. "Statistical Inconsistency of Error-correction Objectives for Perfect Phylogenies". bioRxiv. 2026-07-23T00:00:00.000Z. URL: https://doi.org/10.64898/2026.07.20.739615

[9] Liu, Q., Minh, B. Q., Lanfear, R., Charleston, M. A., Richards, S. A., Holland, B. R.. "Robust Phylogenetics". bioRxiv. 2025-04-02T00:00:00.000Z. URL: https://doi.org/10.1101/2025.04.01.646540

[10] Saenko, S., Hoff, K. J., Stanke, M.. "Annotation of 200 Insect Genomes with BRAKER for Consistent Comparisons across Species". bioRxiv. 2025-04-20T00:00:00.000Z. URL: https://doi.org/10.1101/2025.04.17.649312

[11] Togkousidis, A., Gascuel, O., Stamatakis, A.. "A Systematic Investigation of Overfitting in Maximum Likelihood Phylogenetic Inference". bioRxiv. 2026-01-08T00:00:00.000Z. URL: https://doi.org/10.1101/2025.10.07.680876

[12] Moore-Pollard, E. R., Ellestad, P., Mandel, J.. "Phylogenomic challenges in polyploid-rich lineages: Insights from paralog processing and reticulation methods using the complex genus Packera (Asteraceae: Senecioneae)". bioRxiv. 2025-10-13T00:00:00.000Z. URL: https://doi.org/10.1101/2025.09.14.676143

[13] Liu, S., Zhang, W., Yu, P.. "Methodological pitfalls in plant pangenome gene family identification may lead to biased evolutionary inferences". bioRxiv. 2026-05-18T00:00:00.000Z. URL: https://doi.org/10.64898/2026.05.15.725319

[14] Zhang, C., Nielsen, R.. "Beyond infinite sites: Generalized ABBA-BABA statistic for deeper phylogenies". bioRxiv. 2026-07-08T00:00:00.000Z. URL: https://doi.org/10.64898/2026.07.06.736715

[15] Pang, X.-X., Zhang, D.-Y.. "D-plus: Beyond the D-statistic -- Accurate Detection and Characterization of Introgression Despite Among-Lineage Rate Variation". bioRxiv. 2026-10-08T00:00:00.000Z. URL: https://doi.org/10.64898/2026.10.01.755911

[16] Asri, M., Eizenga, J. M., Hebbar, P., Real, T. D., Lucas, J., Loucks, H., Calicchio, A., Diekhans, M., Eichler, E. E., Salama, S., Miga, K. H., Paten, B.. "Evaluating genome assemblies with HMM-Flagger". bioRxiv. 2026-03-02T00:00:00.000Z. URL: https://doi.org/10.64898/2026.02.27.708355

[17] Alam, M. N. U., Palacios, C. R., Copetti, D., Wing, R. A.. "Universal orthologs infer deep phylogenies and improve genome quality assessments". bioRxiv. 2025-02-21T00:00:00.000Z. URL: https://doi.org/10.1101/2025.02.17.638702

[18] Piat, L., Denni, S., Dubois, S., Linard, B., Duvaux, L.. "Simulating population pangenomes under coalescent demographic models with MSpangenome". bioRxiv. 2026-07-03T00:00:00.000Z. URL: https://doi.org/10.64898/2026.06.29.735168

[19] Mattick, J. S. A., DeMontigny, W. C., Delwiche, C. F.. "Inferring Gene Presence in Incomplete Data via Phylogenetic Occupancy Modeling". bioRxiv. 2026-05-20T00:00:00.000Z. URL: https://doi.org/10.64898/2026.02.27.708499

[20] Bruna, T., Sreedasyam, A., Harder, A. M., Lovell, J. T.. "Evolutionary and methodological considerations when interpreting gene presence-absence variation in pangenomes". bioRxiv. 2025-08-14T00:00:00.000Z. URL: https://doi.org/10.1101/2025.08.14.670405

[21] Marthe, N., Zytnicki, M., Sabot, F.. "GrAnnoT, a tool for efficient and reliable annotation transfer through pangenome graph". bioRxiv. 2025-10-29T00:00:00.000Z. URL: https://doi.org/10.1101/2025.02.26.640337

[22] Forsythe, E. S., Pappa, B. S., Clavette, D. A., Mendoza, D. Y.. "Detecting cryptic ghost lineage introgression in four-taxon genomic datasets". bioRxiv. 2025-04-29T00:00:00.000Z. URL: https://doi.org/10.1101/2025.04.28.651118

[23] Tahmid, N., Rhythm, S. I., Bayzid, M. S.. "Differential Effects of Incomplete Lineage Sorting and Gene Tree Estimation Error on Gene Tree Distributions and Species Tree Inference". bioRxiv. 2026-08-25T00:00:00.000Z. URL: https://doi.org/10.64898/2026.02.21.707162

[24] Bernabeu, M., Manzano-Morales, S., Gabaldon, T.. "Phylogeny-aware simulations suggest a low impact of unsampled lineages in the inference of gene flow during eukaryogenesis". bioRxiv. 2024-11-19T00:00:00.000Z. URL: https://doi.org/10.1101/2024.10.04.616067

[25] He, W., Scornavacca, C., Chan, Y.-b.. "The effect of gene tree dependence on summary methods for speciestree inference". bioRxiv. 2026-01-26T00:00:00.000Z. URL: https://doi.org/10.1101/2024.06.06.597697

[26] Pang, X.-X., Liu, J., Zhang, D.-Y.. "Detecting Introgression in Shallow Phylogenies: How Minor Molecular Clock Deviations Lead to Major Inference Errors". bioRxiv. 2025-03-28T00:00:00.000Z. URL: https://doi.org/10.1101/2025.03.25.645378

[27] Li, B., Ane, C.. "Substitution rate variation, not hidden paralogy, drives false hybridization signal in phylogenetic network inference". bioRxiv. 2026-05-18T00:00:00.000Z. URL: https://doi.org/10.64898/2026.05.11.723986

[28] Tiley, G. P., Liu, N., Solis-Lemus, C.. "Extracting diamonds: Identifiability of 4-node cycles in level-1 phylogenetic networks under a pseudolikelihood coalescent model". bioRxiv. 2025-03-13T00:00:00.000Z. URL: https://doi.org/10.1101/2023.10.25.564087

How this survey was made

This survey answers one question: “How do annotation-transfer errors and shared gene-model artifacts bias phylogenetic signal and introgression tests in comparative pangenome analyses of closely related taxa?”

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 28 papers whose work this text draws on, chosen from the 131 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 21, 2026 and covered work from the previous two years.