Novel Applications of Machine Learning and Data Science Methods in Evolutionary Genomics

The field of evolutionary genomics studies the evolutionary mechanisms that lead to the broad diversity of life on earth. Understanding these mechanisms not only helps to disentangle the origin of life, but also has practical applications such as for developing novel drugs, or tracking global pandemics. Phylogenetics and population genetics are two major subfields of evolutionary biology. While phylogenetics focuses on the evolutionary history between distinct species, population genetics studies the evolutionary mechanisms within populations of a single species, or among closely related species. In modern-day evolutionary genomics, analyses are typically performed using molecular sequence data and rely upon mathematical models that are implemented in scientific software tools.

Over the past decades, technological improvements have led to an avalanche of molecular sequence data. The amount of available data increases at a higher pace than the cost of compute power decreases, according to Moore's law. Consequently, data need to be selected systematically for analysis. In addition, analysis methods need to be as fast as possible while maintaining analytical accuracy. ... mehrFurthermore, quantifying our confidence in these analytical findings becomes increasingly important to prevent misleading conclusions based on potentially uncertain results.

In this thesis, I explore applications of machine learning and data science methods to contribute towards conducting more systematic data selection and faster data analyses, as well as towards better quantifying the inherent uncertainty of such analyses.

My first contribution is a machine learning-based framework that predicts the difficulty of phylogenetic analyses. Usually, phylogenetic trees that represent the hypothetical evolutionary history among a set of distinct species are inferred via complex mathematical models such as the Maximum Likelihood (ML) criterion. These inference procedures are time- and resource-intensive and rely on heuristic tree inference algorithm implementations. Performing multiple independent inferences is often required to at least partially explore the vast search space of possible tree topologies. Yet, these independent inferences do not necessarily converge to a single phylogeny or even topologically highly similar trees. I initially present a novel quantification of this behavior and validate that this quantification accurately represents the inherent dataset difficulty under ML phylogenetic analyses. I further present a machine learning model that can predict this difficulty with high accuracy while being substantially faster than even a single ML tree inference.

My second contribution is a collaborative study of sequence simulation realism conducted with Johanna Trost and Dimitri Höhler. In an extensive data analysis, we assessed the degree of realism for sequence simulation in phylogenetics. These data simulations rely on statistical models of evolution. A plethora of distinct models exists, comprising simple to highly complex models with numerous degrees of freedom. We simulate data under increasingly complex evolutionary models via state-of-the art sequence simulation software. Using two distinct machine learning-based classification methods, we demonstrate that, across all models, we can easily distinguish empirical from simulated data with high accuracy. This indicates a lack of realism in sequence simulation. We further conduct thorough analyses to explain why current state-of-the art simulations fail to replicate empirical data.


My final contribution is a software tool for uncertainty estimation in dimensionality reduction for population genetics data. High-dimensional genetic data often require the application of dimensionality reduction techniques to identify patterns and facilitate their interpretation. While there exist numerous software tools for performing dimensionality reduction, none of these tools deploy uncertainty estimation. With Pandora, I present a novel framework that seamlessly integrates dimensionality reduction of population genetics data and the respective uncertainty estimation into a single software tool. Pandora estimates the uncertainty based on a bootstrapping approach and calculates three distinct notions of stability. Through comprehensive evaluations on empirical and simulated data, I demonstrate the usage and utility of Pandora for studies that rely on dimensionality reduction techniques.