Bridging the Knowledge Gap in Multigenic Pathogenicity Predictions: Leveraging integrated networks and machine learning to explain predictions in oligogenic diseases

Recent years have seen significant progress in medical genetics, with enhanced access to sequenced genomic data and the evolution of computational methods for analysing genetic variation. These advances have improved our understanding of Mendelian 'one gene - one phenotype' genetic models, yet the transition to the study of oligogenic diseases, where a small number of genes are involved, remains a considerable challenge.With the rising reports of clinical oligogenic cases, resources and machine learning tools have been developed to leverage this data. Nevertheless, despite their high accuracy, these predictors can be viewed as “black-boxes” due to their limited interpretability. This complexity poses challenges for medical professionals in validating these predictions and restricts their understanding of potential underlying disease mechanisms. Our research aimed to tackle these limitations using structured background knowledge to provide additional context to predictions.Our first key contribution is a web platform designed to allow geneticists to filter and analyse patient-level variant data with predictors designed for oligogenic diseases. This platform allows in-depth exploration of predictions, including feature contribution analyses, gene pathogenicity networks and gene module mapping to biological networks.Taking a step further towards enhanced explainability, we extended our research to the design of a whitebox machine learning approach that could provide both predictions and meaningful explanations based on background integrated knowledge. Our second contribution is a biological knowledge graph that blends data from known oligogenic diseases with multi-scale biological networks, emphasising the importance of diverse information for understanding oligogenic complexity.Our third contribution builds on this, presenting an interpretable model based on path semantics between gene pairs. This model, capable of learning and applying rules for oligogenic interactions, offers a novel method for geneticists to validate predictions and hypothesise about the causal mechanisms of oligogenic diseases.In conclusion, this research shows how background knowledge can enhance the explainability of predictions for pathogenic genetic interactions. Our analysis platform gives geneticists the necessary tools to understand oligogenic predictions more effectively. Additionally, our biological knowledge graph opens new avenues for investigating the intricate relationships within oligogenic diseases. Finally, using our rule-based approach in tandem with these resources can improve prediction validation and facilitate the generation of mechanistic hypotheses, advancing our understanding of oligogenic diseases.Doctorat en Sciences