Phylogenetic analyses using concatenation of genomic-scale data have been seen as the panacea for resolving the incongruences among inferences from few or single genes. However, phylogenomics may also suffer from systematic errors, due to the, perhaps cumulative, effects of saturation, among-taxa compositional (GC content) heterogeneity, or codon-usage bias plaguing the individual nucleotide loci that are concatenated. Here, we provide an example of how these factors affect the inferences of the phylogeny of early land plants based on mitochondrial genomic data. Mitochondrial sequences evolve slowly in plants and hence are thought to be suitable for resolving deep relationships. We newly assembled mitochondrial genomes from 20 bryophytes, complemented these with 40 other streptophytes (land plants plus algal outgroups), compiling a data matrix of 60 taxa and 41 mitochondrial genes. Homogeneous analyses of the concatenated nucleotide data resolve mosses as sister-group to the remaining land plants. However, the corresponding translated amino acid data support the liverwort lineage in this position. Both results receive weak to moderate support in maximum-likelihood analyses, but strong support in Bayesian inferences. Tests of alternative hypotheses using either nucleotide or amino acid data provide implicit support for their respective optimal topologies, and clearly reject the hypotheses that bryophytes are monophyletic, liverworts and mosses share a unique common ancestor, or hornworts are sister to the remaining land plants. We determined that land plant lineages differ in their nucleotide composition, and in their usage of synonymous codon variants. Composition heterogeneous Bayesian analyses employing a nonstationary model that accounts for variation in among-lineage composition, and inferences from degenerated nucleotide data that avoid the effects of synonymous substitutions that underlie codon-usage bias, again recovered liverworts being sister to the remaining land plants but without support. These analyses indicate that the inference of an early-branching moss lineage based on the nucleotide data is caused by convergent compositional biases. Accommodating among-site amino acid compositional heterogeneity (CAT-model) yields no support for the optimal resolution of liverwort as sister to the rest of land plants, suggesting that the robust inference of the liverwort position in homogeneous analyses may be due in part to compositional biases among sites. All analyses support a paraphyletic bryophytes with hornworts composing the sister-group to tracheophytes. We conclude that while genomic data may generate highly supported phylogenetic trees, these inferences may be artifacts. We suggest that phylogenomic analyses should assess the possible impact of potential biases through comparisons of protein-coding gene data and their amino acid translations by evaluating the impact of substitutional saturation, synonymous substitutions, and compositional biases through data deletion strategies and by analyzing the data using heterogeneous composition models. We caution against relying on any one presentation of the data (nucleotide or amino acid) or any one type of analysis even when analyzing large-scale data sets, no matter how well-supported, without fully exploring the effects of substitution models.
