Multilayer network approaches for integration of biological and clinical multimodal datasets

Thesis

High-dimensional assays now routinely profile multiple molecular and cellular modalities per patient, yet integrating heterogeneous and incomplete measurements into reproducible, clinically meaningful strata remains challenging. This thesis presents a multilayer framework for patient stratification. ParetoCOGENT selects per-modality patient similarity networks by jointly optimising resampling stability and informativeness against null models, and MLModNet integrates these networks as a weighted multiplex that accommodates missing assays. Multiplex community detection and stability diagnostics were then used to identify endotypes supported by concordant structure across modalities.

Applied to the COVID-19 Multiomics Blood Atlas (COMBAT), MLModNet recovered endotypes that refined clinical severity categories. These endotypes aligned with coherent immune programmes, including interferon response, myeloid activation, complement and coagulation, and humoral signatures. They also stratified ICU-free survival and clinical trajectories, while soft cluster probabilities behaved as continuous molecular state readouts. Compared with alternative multimodal clustering methods, MLModNet formed part of the most consistent hostresponse backbone and showed the strongest continuous clinical readout in the main benchmark. In downstream prediction, MLModNet cluster labels were also more molecularly informative than WHO clinical severity classes, highlighting immunestate signals that were less fully captured by routine severity scoring. Through external validation on an independent proteomics cohort, reduced shared-protein panels preserved the major severity axis and endotype probabilities explained broad proteome variation beyond clinical labels. The full framework was then applied to GAinS sepsis, where four endotypes were reproduced across discovery and validation cohorts. These endotypes aligned with SRSq and with FP/CAPstyle severity axes, while showing concordant pathway programmes and shared molecular drivers. Taken together, this thesis presents a reproducible framework for moving from heterogeneous multiomics data to interpretable patient stratification across acute infections.

Authors: Sliwa, PA

• DOI: 10.5287/ora-r8kve8yjz
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: COVID-19; unsupervised learning; patient stratification; molecular endotypes; precision medicine; sepsis; multiomics; multilayer networks; patient similarity networks
• Subjects: Statistics; Medicine; Genomics