Open Access

This cumulative thesis describes contributions to the field of interpretable machine learning in the healthcare domain. Three research articles are presented that lie at the intersection of biomedical and machine learning research. They illustrate how incorporating latent structure can provide a valuable compression of the information hidden in complex healthcare data. Methodologically, this thesis gives an overview of interpretable machine learning and the discovery of latent structure, including clusters, latent factors, graph structure, and hierarchical structure. Different workflows are developed and applied to two main types of complex healthcare data (cohort study data and time-resolved molecular data). The core result builds on Bayesian networks, a type of probabilistic graphical model. On the application side, we provide accurate predictive or discriminative models focusing on relevant medical conditions, related biomarkers, and their interactions.

Knowledge on the occurrence of pathogenic tick-borne bacteria Anaplasma phagocytophilum and Anaplasma ovis is scarce in sheep from Germany. In 2020, owners from five flocks reported ill thrift lambs and ewes with tick infestation. Out of 67 affected sheep, 55 animals were clinically examined and hematological values, blood chemistry and fecal examinations were performed to investigate the underlying disease causes. Serological tests (cELISA, IFAT) and qPCR were applied to all affected sheep to rule out A. phagocytophilum and A. ovis as a differential diagnosis. Ticks were collected from selected pastures and tested by qPCR. Most animals (n = 43) suffered from selenium deficiency and endoparasites were detected in each flock. Anaplasma spp. antibodies were determined in 59% of examined sheep. Seventeen animals tested positive for A. phagocytophilum by qPCR from all flocks and A. phagocytophilum was also detected in eight pools of Ixodes ricinus. Anaplasma phagocytophilum isolates from sheep and ticks were genotyped using three genes (16S rRNA, msp4 and groEL). Anaplasma ovis DNA was identified in six animals from one flock. Clinical, hematological and biochemical changes were not significantly associated with Anaplasma spp. infection. The 16S rRNA analysis revealed known variants of A. phagocytophilum, whereas the msp4 and groEL showed new genotypes. Further investigations are necessary to evaluate the dissemination and health impact of both pathogens in the German sheep population particularly in case of comorbidities.

Abstract Cellular stress has been associated with inflammation, yet precise underlying mechanisms remain elusive. In this study, various unrelated stress inducers were employed to screen for sensors linking altered cellular homeostasis and inflammation. We identified the intracellular pattern recognition receptors NOD1/2, which sense bacterial peptidoglycans, as general stress sensors detecting perturbations of cellular homeostasis. NOD1/2 activation upon such perturbations required generation of the endogenous metabolite sphingosine‐1‐phosphate (S1P). Unlike peptidoglycan sensing via the leucine‐rich repeats domain, cytosolic S1P directly bound to the nucleotide binding domains of NOD1/2, triggering NF‐κB activation and inflammatory responses. In sum, we unveiled a hitherto unknown role of NOD1/2 in surveillance of cellular homeostasis through sensing of the cytosolic metabolite S1P. We propose S1P, an endogenous metabolite, as a novel NOD1/2 activator and NOD1/2 as molecular hubs integrating bacterial and metabolic cues.

Phytopathogenic Verticillia cause Verticillium wilt on numerous economically important crops. Plant infection begins at the roots, where the fungus is confronted with rhizosphere inhabiting bacteria. The effects of different fluorescent pseudomonads, including some known biocontrol agents of other plant pathogens, on fungal growth of the haploid Verticillium dahliae and/or the amphidiploid Verticillium longisporum were compared on pectin-rich medium, in microfluidic interaction channels, allowing visualization of single hyphae, or on Arabidopsis thaliana roots. We found that the potential for formation of bacterial lipopeptide syringomycin resulted in stronger growth reduction effects on saprophytic Aspergillus nidulans compared to Verticillium spp. A more detailed analyses on bacterial-fungal co-cultivation in narrow interaction channels of microfluidic devices revealed that the strongest inhibitory potential was found for Pseudomonas protegens CHA0, with its inhibitory potential depending on the presence of the GacS/GacA system controlling several bacterial metabolites. Hyphal tip polarity was altered when V. longisporum was confronted with pseudomonads in narrow interaction channels, resulting in a curly morphology instead of straight hyphal tip growth. These results support the hypothesis that the fungus attempts to evade the bacterial confrontation. Alterations due to co-cultivation with bacteria could not only be observed in fungal morphology but also in fungal transcriptome. P. protegens CHA0 alters transcriptional profiles of V. longisporum during 2 h liquid media co-cultivation in pectin-rich medium. Genes required for degradation of and growth on the carbon source pectin were down-regulated, whereas transcripts involved in redox processes were up-regulated. Thus, the secondary metabolite mediated effect of Pseudomonas isolates on Verticillium species results in a complex transcriptional response, leading to decreased growth with precautions for self-protection combined with the initiation of a change in fungal growth direction. This interplay of bacterial effects on the pathogen can be beneficial to protect plants from infection, as shown with A. thaliana root experiments. Treatment of the roots with bacteria prior to infection with V. dahliae resulted in a significant reduction of fungal root colonization. Taken together we demonstrate how pseudomonads interfere with the growth of Verticillium spp. and show that these bacteria could serve in plant protection.

Abstract The expected signature is an analogue of the Laplace transform for probability measures on rough paths. A key question in the area has been to identify a general condition to ensure that the expected signature uniquely determines the measures. A sufficient condition has recently been given by Chevyrev and Lyons and requires a strong upper bound on the expected signature. While the upper bound was verified for many well‐known processes up to a deterministic time, it was not known whether the required bound holds for random time. In fact, even the simplest case of Brownian motion up to the exit time of a planar disc was open. For this particular case we answer this question using a suitable hyperbolic projection of the expected signature. The projection satisfies a three‐dimensional system of linear PDEs, which (surprisingly) can be solved explicitly, and which allows us to show that the upper bound on the expected signature is not satisfied.

Abstract With the advent of molecular genetic methods, an increasing number of morphologically cryptic taxa has been discovered. The majority of them, however, remains formally undescribed and without a proper name although their importance in ecology and evolution is increasingly being acknowledged. Despite suggestions to complement traditional descriptions with genetic characters, the taxonomic community appears to be reluctant to adopt this proposition. As an incentive, we introduce QUIDDICH, a tool for the QUick IDentification of DIgnostic CHaracters, which automatically scans a DNA or amino acid alignment for those columns that allow to distinguish taxa and classifies them into four different types of diagnostic characters. QUIDDICH is a system‐independent, fast and user‐friendly tool that requires few manual steps and provides a comprehensive output, which can be included in formal taxonomic descriptions. Thus, cryptic taxa do not have to remain in taxonomic crypsis and, bearing a proper name, can readily be included in biodiversity assessments and ecological and evolutionary analyses. QUIDDICH can be obtained from the comprehensive R archive network (CRAN, https://cran.r-project.org/package=quiddich).

A common task in natural sciences is to describe, characterize, and infer relations between discrete objects. A set of relations E on a set of objects V can naturally be expressed as a graph G = (V, E). It is therefore often convenient to formalize problems in natural sciences as graph theoretical problems. In this thesis we will examine a number of problems found in life sciences in particular, and show how to use graph theoretical concepts to formalize and solve the presented problems. The content of the thesis is a collection of papers all solving separate problems that are relevant to biology or biochemistry. The first paper examines problems found in self-assembling protein design. Designing polypeptides, composed of concatenated coiled coil units, to fold into polyhedra turns out to be intimately related to the concept of 1-face embeddings in graph topology. We show that 1-face embeddings can be canonicalized in linear time and present algorithms to enumerate pairwise non-isomorphic 1-face embeddings in orientable surfaces. The second and third paper examine problems found in evolutionary biology. In particular, they focus on inferring gene and species trees directly from sequence data without any a priori knowledge of the trees topology. The second paper characterize when gene trees can be inferred from estimates of orthology, paralogy and xenology relations when only partial information is available. Using this characterization an algorithm is presented that constructs a gene tree consistent with the estimates in polynomial time, if one exists. The shown algorithm is used to experimentally show that gene trees can be accurately inferred even in the case that only 20$\%$ of the relations are known. The third paper explores how to reconcile a gene tree with a species tree in a biologically feasible way, when the events of the gene tree are known. Biologically feasible reconciliations are characterized using only the topology of the gene and species tree. Using this characterization an algorithm is shown that constructs a biologically feasible reconciliation in polynomial time, if one exists. The fourth and fifth paper are concerned with with the analysis of automatically generated reaction networks. The fourth paper introduces an algorithm to predict thermodynamic properties of compounds in a chemistry. The algorithm is based on the well known group contribution methods and will automatically infer functional groups based on common structural motifs found in a set of sampled compounds. It is shown experimentally that the algorithm can be used to accurately predict a variety of molecular properties such as normal boiling point, Gibbs free energy, and the minimum free energy of RNA secondary structures. The fifth and final paper presents a framework to track atoms through reaction networks generated by a graph grammar. Using concepts found in semigroup theory, the paper defines the characteristic monoid of a reaction network. It goes on to show how natural subsystems of a reaction network organically emerge from the right Cayley graph of said monoid. The applicability of the framework is proven by applying it to the design of isotopic labeling experiments as well as to the analysis of the TCA cycle.

In this thesis, we elaborate upon Bayesian changepoint analysis, whereby our focus is on three big topics: approximate sampling via MCMC, exact inference and uncertainty quantification. Besides, modeling matters are discussed in an ongoing fashion. Our findings are underpinned through several changepoint examples with a focus on a well-log drilling data.