STAR: Supervised Machine Learning to Improve Transposable Element Polymorphism Detection for Phylogenetic Inference

Understanding the relationships among living things is vital step in determining how traits evolve. Did the traits of interest evolve only once or multiple times? What changes have occurred to allow their evolution? Using molecular/genomic markers to infer relationships is standard practice but the markers chosen vary in their efficacy and the potential complications they introduce. For example, the most used marker, Single Nucleotide Polymorphisms, can introduce uncertainty because their state can be reversed over evolutionary time. A marker type that is less impacted by these problems is retrotransposon insertion sites. However, those sites require substantial human intervention to identify, making their development difficult and expensive. We will develop a set of computational tools that incorporate machine learning to identify and utilize a plentiful and particularly effective type of marker (SINEs) from the available genome assemblies of a variety of organisms. We will begin with bats, an interesting and complex group of interest to multiple biologists and then expand the utility of our tools to other organisms that have similar markers. The tools will be made available in a user-friendly package for other researchers, allowing them to identify and utilize these easily interpretable signals rapidly and effectively for their groups of interest. Insights into organismal phylogeny are critical when conducting modern biological analyses, especially when addressing questions regarding the origin of adaptive traits. However, methods using sequence-based concatenation and coalescence methods can be inaccurate if assumptions are violated, which is often the case when considering ancient and/or rapid radiations. The results of phylogenetic reconstructions can vary greatly when branch lengths are short and depend on sampling, markers used, and the methodology. Furthermore, homoplasy of sequence-based characters introduces ambiguity. New phylogenetic methods that make use of retrotransposons, a type of transposable element (TE), have proven useful. However, current retrophylogenomic methods that utilize large numbers of retrotransposon insertions and multiple genome assemblies are time-consuming and require substantial manual input. Machine learning (ML) is well suited for mitigating these problems. We will standardize the identification of informative SINE ORIL (Orthologous Retrotransposon Insertion Loci) presence/absence patterns for phylogenetic analysis using ML approaches, diverse genomic signals, multiple datasets, and scaling to large genomic studies, all while outperforming rigid rule-based approaches in dynamic regions of the genome. We will use our unique expertise to curate a large SINE ORIL data set and leverage the data set to train ML models. We will apply the data to various branches of the tree of life, evaluating their potential to generalize beyond the training settings. We will also develop our tools for general distribution to the scientific community. Our work will bridge the gap between manually curated biological analyses and automation via machine learning. Retrotransposons and other TEs are ubiquitous in eukaryotes. Thus, the methods to be developed will be broadly applicable. As part of the work, we will train undergraduate and graduate students as well as generate a museum-ready display on bats, especially Ecuadorian bats, for rotation at the Museum at Texas Tech University. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.