Background
Most statistical methods for phylogenetic estimation in use today treat a gap (generally representing an insertion or deletion, i.e., indel) within the input sequence alignment as missing data. However, the statistical properties of this treatment of indels have not been fully investigated.

Results
We prove that maximum likelihood phylogeny estimation, treating indels as missing data, can be statistically inconsistent for a general (and rather simple) model of sequence evolution, even when given the true alignment. Therefore, accurate phylogeny estimation cannot be guaranteed for maximum likelihood analyses, even given arbitrarily long sequences, when indels are present and treated as missing data.

Conclusions
Our result shows that the standard statistical techniques used to estimate phylogenies from sequence alignments may have unfavorable statistical properties, even when the sequence alignment is accurate and the assumed substitution model matches the generation model. This suggests that the recent research focus on developing statistical methods that treat indel events properly is an important direction for phylogeny estimation.

Over the last decade, dramatic advances have been made in developing methods for large-scale phylogeny estimation, so that it is now feasible for investigators with moderate computational resources to obtain reasonable solutions to maximum likelihood and maximum parsimony, even for datasets with a few thousand sequences. There has also been progress on developing methods for multiple sequence alignment, so that greater alignment accuracy (and subsequent improvement in phylogenetic accuracy) is now possible through automated methods. However, these methods have not been tested under conditions that reflect properties of datasets confronted by large-scale phylogenetic estimation projects. In this paper we report on a study that compares several alignment methods on a benchmark collection of nucleotide sequence datasets of up to 78,132 sequences. We show that as the number of sequences increases, the number of alignment methods that can analyze the datasets decreases. Furthermore, the most accurate alignment methods are unable to analyze the very largest datasets we studied, so that only moderately accurate alignment methods can be used on the largest datasets. As a result, alignments computed for large datasets have relatively large error rates, and maximum likelihood phylogenies computed on these alignments also have high error rates. Therefore, the estimation of highly accurate multiple sequence alignments is a major challenge for Tree of Life projects, and more generally for large-scale systematics studies.

We have assembled a collection of web pages that contain benchmark datasets and software tools to enable the evaluation of the accuracy and scalability of computational methods for estimating evolutionary relationships. They provide a resource to the scientific community for development of new alignment and tree inference methods on very difficult datasets. The datasets are intended to help address three problems: multiple sequence alignment, phylogeny estimation given aligned sequences, and supertree estimation. Datasets from our work include empirical datasets with carefully curated alignments suitable for testing alignment and phylogenetic methods for large-scale systematics studies. Links to other empirical datasets, lacking curated alignments, are also provided. We also include simulated datasets with properties typical of large-scale systematics studies, including high rates of substitutions and indels, and we include the true alignment and tree for each simulated dataset. Finally, we provide links to software tools for generating simulated datasets, and for evaluating the accuracy of alignments and trees estimated on these datasets. We welcome contributions to the benchmark datasets from other researchers.