Year: 2026
Type: dissertation
Abstract: Multi-omics integration methods are hard to evaluate because benchmark datasets with a known ground truth are scarce, as real experimental data never comes with its generating mechanism attached. One way to obtain such data is to simulate a mechanistic dynamic model, which is the ground truth by construction. However, building and parameterising such models by hand is impractical at scale. This thesis presents a fully automated framework that builds and simulates an ordinary differential equation model directly from a biological knowledge graph, with no manual intervention. It parses the network from SBML, estimates kinetic parameters by combining the BRENDA and SABIO-RK databases with machine-learning predictions and principled statistical fallbacks, assembles an Antimony model, and simulates it under any stress and knockout conditions specified by the user, with a web interface that gathers the whole workflow into one place. The framework is demonstrated end-to-end on the Plant Stress Signalling network, a curated map of plant stress signalling built mainly on emph{Arabidopsis thaliana}, where it produces a complete, simulatable model together with the multi-condition, multi-omics-style data it generates as candidate benchmark data. In this demonstration, each simulated stress recruits its own coherent, biologically recognisable and cleanly separable module, which is the property a multi-omics benchmark needs. A second finding is that making the network run turns the model into a demanding test of the network itself. The pipeline brings out points in the graph and its reference data that inspection alone misses, traces them from its own outputs, and can measure whether a refinement helps.
Author: Atanasoski, Radoslav
Institution: University of Ljubljana
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225067894
