Independent exploration · 2026–present
The question
How can biomedical knowledge graphs help investigate gene relationships, and what evidence and representation choices influence a model’s predictions?
My starting point is NLRP3 and LGALS3 (Galectin-3), connecting this computational work with my earlier research experience in neuroinflammation.
What I have worked on
- Prepared a graph-based link-prediction workflow using OptimusKG and BioPathNet.
- Separated background graph relationships from supervised links and evaluated rankings on withheld recorded interactions.
- Ranked 732 candidate NLRP3 partners and compared model results with gene-degree and shared-neighbour measures.
- Inspected graph paths to understand what recorded context might inform a prediction and where uncertainty remains.
Why the data choices matter
An observed relationship depends on the source, coverage, and representation of the underlying evidence. Missing graph edges do not automatically establish biological negatives. Simple baselines help reveal whether a model is mainly reflecting well-connected genes.
What the results mean
This is exploratory computational work. Predicting a recorded graph relationship does not establish physical binding, causal biology, or therapeutic relevance. Those questions require additional evidence and appropriate experimental validation.
Current direction
I am exploring how literature evidence can support the revision and expansion of a seed graph across successive cycles. This remains a developing research direction, separate from the link-prediction work described above.