Reply to Springer and Gatesy: The impact of long branches and misalignments on phylogenetic analysis is minimal

Springer and Gatesy (1) suggested that our avian species tree, constructed using NJst (2), may be inaccurate due to observed disparities with the ASTRAL tree and potential misalignments. However, akin to their critiques of our other work, their criticisms stem from subjective opinions and misconceptions of phylogenetic analysis, rendering their arguments unfounded.

Springer and Gatesy argued that NJst is less accurate than ASTRAL, citing evidence from two simulation studies (3, 4). However, assessing phylogenetic methods solely through simulations cannot fully reflect real-world scenarios. Other studies have indeed suggested that NJst outperforms ASTRAL (5). Notably, NJst outcomes align closely with RAxML (SI Appendix, Figs. S1–S5 in ref. 2), indicating the robustness of NJst. In contrast, ASTRAL results differ significantly, warranting further validation.

Springer and Gatesy suggested that data misalignments could impact our phylogenetic results, citing 2,305 long branches (≥1.0 substitutions/site) across 1,824 gene trees: 1,626 CDS, 188 intergenic segments, and 10 introns. This finding indicates an average of only 1.26 long branches per gene tree. Our analysis revealed a random distribution of these long branches across gene trees. In CDS gene trees, 634 showed long branches exclusively linked to an outgroup; in the remaining 991, long branches were uniformly distributed across ingroup taxa (Fig. 1A). Among intergenic gene trees, eight displayed long branches solely linked to an outgroup: in the remaining 180, long branches were uniformly distributed across ingroup taxa (Fig. 1B). Only one out of the ten intron gene trees exhibited long branches involving nonoutgroup species. After excluding 991 CDS and 180 intergenic gene trees, we reconstructed NJst and ASTRAL trees. The resulting NJst tree remained identical to the tree including all gene trees (Fig. 2A), whereas the ASTRAL tree changed to support Aquaterraves (Fig. 2B). These results underscore NJst's robustness to data heterogeneity, contrasting sharply with ASTRAL's evident sensitivity.

Fig. 1.

(A) Random distribution of long branches across CDS gene trees; (B) Random distribution of long branches across intergenic gene trees.

Fig. 2.

(A) NJst species tree after excluding the 991 CDS and 180 intergenic gene trees with long branch (≥1 substitutions/site); (B) ASTRAL species tree after excluding the 991 CDS and 180 intergenic gene trees with long branch (≥1 substitutions/site). Numbers on nodes indicate the bootstrap support values, and nodes with maximum bootstrap support are not shown. Note that in the resulting ASTRAL tree, Columbiformes is classified within Aquaterraves (highlighted in red), diverging from Springer and Gatesy's ASTRAL tree, which placed Columbiformes within Telluraves. In contrast, the resulting NJst tree remains identical to the NJst tree presented in ref. 2.

Springer and Gatesy claim that our intergenic segments likely contain a mixture of orthologs and paralogs, citing the identification of CR1-LINE retrolements. However, they present no evidence to demonstrate that the presence of retroelements in intergenic segments adversely affects phylogenetic analysis. Recent studies indicate that even in the presence of paralogs, species tree inference can be as accurate as analyses based solely on orthologs (6, 7). This was demonstrated by NJst and RAxML trees built exclusively using intergenic segments (see SI Appendix, Figs. S1–S5 in ref. 2), which showed close alignment with inferred intraordinal phylogenies from other data types, maintaining congruence with previous studies (8–10). These findings suggest that the purported impact of paralogs on phylogenetic analysis is actually negligible.

Despite extensive filtering and screening, genome-scale data often contains a small number of long branches and paralogs due to its intrinsic heterogeneity (5). However, Springer and Gatesy overstate their influence on phylogenetic analysis. Their puriest approach, demanding "absolutely clean" data, is unrealistic and would drastically reduce usable data, thus hindering phylogenetic inference (11). Our analysis demonstrates that these factors have minimal impact on phylogenetic analysis and that Springer and Gatesy exaggerate their influence.