Summer project
Bioinformatics Specialization · Cohort 2026
In silico stress-testing of transcriptomic biomarkers for deployable tuberculosis triage
Blood gene-expression signatures can distinguish active tuberculosis from latent infection, but they are discovered in well-resourced laboratories and deployed on low-cost RT-qPCR cartridges. This project built a selection rule that chooses gene panels for their worst-case performance under simulated instrument failure — assay noise, probe dropout, coarse quantisation, narrow dynamic range — and then applied the same stress-testing to the rule itself. The rule did not survive: no two panel sizes are distinguishable with 75 subjects, and separating adjacent sizes would require thousands. Panel size is instead taken from the device's redundancy requirement. A second, pre-registered panel selected for maximal reproducibility transferred significantly worse than the first, because reproducible selection concentrates on mutually correlated genes. The work delivers an identifiability result with a quantified sample-size requirement, a measured counterexample to standard advice on stability selection, and a full audit trail of four defects the method found in its own pipeline.
With KAUST Academy
- Cohort 2026
University of Cambridge · Bioinformatics Summer Program
Want your entry updated or removed? Contact us.