The impact of automated filtering of BLAST-determined homologs in the phylogenetic detection of horizontal gene transfer from a transcriptome assembly

- Differing rates of sequence evolution among genes complicates phylogenomics.
- Imposing arbitrary sequence similarity thresholds in phylogenomics effects results.
- No one threshold is ideal for building trees for the majority of genes in a genome.
- Merging results from several pipeline iterations yields the most informative trees.
- Interpreting results from phylogenomic pipelines requires caution.

Phylomes (comprehensive sets of gene phylogenies for organisms) are built to investigate fundamental questions in genomics and evolutionary biology, such as those pertaining to the detection and characterization of horizontal gene transfer in microbes. To address these questions, phylome construction demands rigorous yet efficient phylogenetic methods. Currently, many sequence alignment and tree-building models can analyze several thousands of genes in a high-throughput manner. However, the phylogenetics is complicated by variability in sequence divergence and different taxon sampling among genes. In addition, homolog selection for automated approaches often relies on arbitrary sequence similarity thresholds that are likely inappropriate for all genes in a genome. To investigate the effects of automated homolog selection on the detection of horizontal gene transfer using phylogenomics, we constructed the phylome of a transcriptome assembly of Alexandrium tamarense, a microbial eukaryote with a history of horizontal and endosymbiotic gene transfer, using seven sequence similarity thresholds for selecting putative homologs to be included in phylogenetic analyses. We show that no single threshold recovered informative trees for the majority of A. tamarense unigenes compared to the pooled results from all pipeline iterations. As much as 29% of trees built could have misleading phylogenetic relationships that appear biased in favor of those otherwise indicative of horizontal gene transfer. Perhaps worse, nearly half of the unigenes were represented by a single tree built at just one threshold, making it difficult to assess the validity of phylogenetic relationships recovered in these cases. However, combining the results from several pipeline iterations maximizes the number of informative phylogenies. Moreover, when the same phylogenetic relationship for a given unigene is recovered in multiple pipeline iterations, conclusions regarding gene origin are more robust to methodological artifact. Using these methods, the majority of A. tamarense unigenes showed evolutionary relationships indicative of vertical inheritance. Nevertheless, many other unigenes revealed diverse phylogenetic associations, suggestive of possible gene transfer. This analysis suggests that caution should be used when interpreting the results from phylogenetic pipelines implementing a single similarity threshold. Our approach is a practical method to mitigate the problems associated with automated sequence selection in phylogenomics.