Year: 2026
Type: preprint
Abstract: Integrating biological knowledge with automated experimentation enables the generation and testing of hypotheses for scientific discovery. However, identifying promising hypotheses from heterogeneous biological data remains challenging. Here, we present a knowledge graph-based approach to prioritising gene-chemical interactions associated with decreased chemical stress resistance in the yeast Saccharomyces cerevisiae. We combine information from biological databases and ontologies with a hierarchy-aware graph neural network to rank candidate interactions for experimental testing. We selected 35 gene-chemical hypotheses involving six chemical stressors and evaluated them using automated growth assays. Ten hypotheses were supported after correction for multiple testing, including nine of the 30 highest-ranked candidates for the six target chemicals and one involving a poorly characterised gene. Experimental support varied across the six chemical stressors, highlighting differences in the outcomes of the tested gene-chemical interactions. These findings demonstrate how knowledge graph predictions can be evaluated experimentally to identify gene-chemical interactions for further investigation. Finally, we extended the Revisions for Improvements of Models in Biology Ontology (RIMBO) to represent knowledge graph updates alongside experimental evidence and provenance, retaining records of both supported and unsupported hypotheses. This approach connects knowledge graph-based prediction, experimental evaluation, and traceable updates to biological knowledge.
Source: Research Square
Institutions: Chalmers University of Technology, University of Gothenburg
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225867486
