AI Insight
This simulation study tested how two widely used phylogenetic network inference methods — find_graphs (ADMIXTOOLS 2) and SNaQ — respond to hidden paralogy and substitution rate variation when detecting hybridization from genomic data. Contrary to expectations, hidden paralogy had little effect on network inference, while lineage-specific substitution rate variation severely distorted find_graphs results, generating false hybridization signals by inflating f-statistic residuals beyond the standard acceptance threshold. SNaQ proved more robust overall, though it also showed reduced accuracy under lineage-specific rate variation, and the study further demonstrates that the commonly used threshold of 3 for worst residuals in find_graphs produces elevated false positive rates even under ideal conditions.
Why it matters
These findings have direct implications for evolutionary biologists using genomic data to detect hybridization and gene flow, as they suggest that rate variation — a common biological reality — can produce misleading results with widely adopted tools, calling for more careful calibration of acceptance thresholds in empirical studies.
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Phylogenetic network inference methods are increasingly used to detect hybridization and gene flow from genomic data, but their robustness to common sources of model violation remains poorly characterized. We conducted a simulation study to evaluate the effects of hidden paralogy and substitution rate variation on two widely used network inference methods: find_graphs from ADMIXTOOLS 2 and SNaQ. Using an eight-taxon species tree calibrated from an empirical reptile phylogeny, we simulated data under various levels of hidden paralogy (from none to strong) and three levels of rate variation (none, gene-specific, and lineage-specific). We found that hidden paralogy had limited impact on network inference under the conditions examined: both network methods correctly favored a tree without reticulation, and ASTRAL recovered the correct species tree every time. In contrast, lineage-specific rates severely biased find_graphs, inflating worst f-statistic residuals well beyond the standard acceptance threshold. SNaQ correctly selected a tree model almost always across all conditions, though its network with h = 1 reticulation displayed the true species tree with a lower probability under lineage-specific rates. We also show that the standard worst residuals threshold of 3 for find_graphs produces inflated type I error even without rate variation, and we recommend empirical calibration of this threshold within each study system.