Semantic Multi-Classifier Systems Identify Predictive Processes in Heart Failure Models across Species

Abstract

Genetic model organisms have the potential of removing blind spots from the underlying gene regulatory networks of human diseases. Allowing analyses under experimental conditions they complement the insights gained from observational data. An inevitable requirement for a successful trans-species transfer is an abstract but precise high-level characterization of experimental findings. In this work, we provide a large-scale analysis of seven weak contractility/heart failure genotypes of the model organism zebrafish which all share a weak contractility phenotype. In supervised classification experiments, we screen for discriminative patterns that distinguish between observable phenotypes (homozygous mutant individuals) as well as wild-type (homozygous wild-types) and carriers (heterozygous individuals). As the method of choice we use semantic multi-classifier systems, a knowledge-based approach which constructs hypotheses from a predefined vocabulary of high-level terms (e.g., Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways or Gene Ontology (GO) terms). Evaluating these models leads to a compact description of the underlying processes and guides the screening for new molecular markers of heart failure. Furthermore, we were able to independently corroborate the identified processes in Wistar rats.

Keywords: Heart failure phenotypes; Wistar rat; semantic multi-classifier systems; zebrafish.

Figure 1

Heatmap of the two most frequently selected GO terms (RNA polymerase II transcription factor binding (55.0%) and transaminase activity (42.0%)) in zebrafish. The figure provides the gene expression levels of the genes these terms comprise. Within each term, the genes are ordered according their Spearman correlation to the class label (mutant/control). Each gene was z-transformed individually. Within each class, hierarchical clustering (average linkage) was used to organize the samples.

Figure 2

Heatmap of the two most frequently selected KEGG terms arginine biosynthesis (75%) and biosynthesis of unsaturated fatty acids (55%) in zebrafish. The figure provides the gene expression levels of the genes these terms comprise. Within each term, the genes are ordered according their Spearman correlation to the class label (mutant/control). Each gene was z-transformed individually. Within each class, hierarchical clustering (average linkage) was used to organize the samples.

Figure 3

Leave one out cross-validation on rat data using the terms selected in the zebrafish experiments. Additionally, an experiment based on the whole gene expression profiles is shown. Each row indicates a separate experiment. The cell color denotes the predicted stretch time of a sample. The axis on the right gives the overall accuracies (%) of the experiments.

Figure 4

Ventricular fraction shortening of weak contractility/heart failure genotypes (zebrafish). For all genotypes, the differences between mutants and controls were tested ($n=7$, Wilcoxon-Rank-Sum tests, Bonferroni correction for multiple testing). Significance levels are indicated as *: p < 0.05, **: p < 0.01, ***: p < 0.001.

Figure 5

Schematic drawing of a semantic multi-classifier system (S-MCS): The training of a S-MCS uses prior domain knowledge in form of a vocabulary of selected semantic terms (e.g., KEGG pathways of GO terms) for the analysis of incoming gene expression profiles. Each term is analyzed independently by a separate expert (classifier) which focuses on the chosen subset of measurements. The selection of these experts is aggregated via a fusion architecture, leading to a “mixture of experts”.