Advanced Substitution Model Selection Methods in RAxML-NG

With next generation sequencing, the cost to sequence a genome has decreased exponentially, even outpacing Moore's law. We therefore expect the use of phylogenetic inference tools to rise even further, with the tools facing a perpetual scalability challenge. The largest extent of studies (as well as users) now commends for increasingly automated methods as we observe a shift from manual workflows to reusable, highly automated pipelines. An important step in such phylogenetic analysis pipelines is model selection, which ranks the plethora of available DNA and protein substitution models according to their relative fit to the data. In this thesis, we present a novel implementation of a model selection procedure, integrated into RAxML-NG, a widely used tool for phylogenetic inference via Maximum Likelihood. It implements a number of heuristics to speed up execution and employs a robust parallelization scheme that also supports partitioned datasets and distributed memory systems. For 86.7% of amino acid, and 72.9% of single-gene DNA datasets, our implementation selects models with a BIC score difference of less than 10 compared to IQTree's ModelFinder and ModelTest-NG. ... mehrWith 48 cores, we reach a mean speedup per dataset of 3.54 (AA) and 5.82 (DNA) over IQTree. Over all datasets, the accumulated speedup with 48 cores over IQTree is 1.72 (AA) and 4.05 (DNA). By providing a fast, easily-accessible implementation without external dependencies, we hope to lower the barrier to prepend a model selection to phylogenetic analyses. This ensures a good model-data fit and potentially saves runtime during the tree search by limiting the use of computationally expensive mixture models.
Furthermore, its integration into the inference tool allows future work to incorporate model selection results during the tree search, paving the way for more adaptive approaches and runtime optimizations.