Knowledge Representations for Scientific Discovery

Licentiatavhandling, 2025

Scientific discovery has a lot to gain from advances in artificial intelligence
and machine learning, and where progress can have major societal impact.
These techniques can help scientists uncover patterns in large datasets, generate
new hypotheses, and guide experimental design. Combined with robotic systems
they can also greatly increase the number of performed experiments. This thesis
investigates how structured knowledge representations can support scientific
discovery in systems biology, with a particular focus on the model organism
Saccharomyces cerevisiae (baker’s yeast).

The work introduces two ontologies designed as semantic schemas for
research databases. The first captures metadata and results from µ-chemostat
experiments, accommodating multiple measurement modalities. The second
ontology formalises revisions to computational models, with a focus on domains
where mechanistic models are updated iteratively and it is important to record
what was changed and why.

In the third contribution, information about S. cerevisiae from public data-
bases is integrated into a knowledge graph with well-defined class hierarchies.
Graph neural networks, in combination with box embeddings representing the
hierarchical structure, are used to predict growth outcomes of double gene
deletions. Furthermore, explainability techniques are applied to identify can-
didate biological interactions, forming hypotheses about traits in S. cerevisiae.
One such hypothesis is experimentally validated, illustrating how structured
representations can aid data-driven discovery from publicly available resources.

Taken together, this work introduces knowledge representations for emerging
domains, designed as tools to support scientific discovery, while also demon-
strating how rich, structured representations can enhance the interpretation of
existing data.