Did the 2016 global melioidosis risk map, which predicted that the disease is likely present but unrecognised in 34 countries that had never reported it, correctly anticipate the places that have recorded their first locally acquired cases or first soil detections since then (such as Mississippi's Gulf Coast in 2020 to 2022), and would a model refitted on newer climate and soil data do better on that forward test?
- Good it would do
- 3 / 5
- Can AI help
- 3 / 5
- How neglected
- 4 / 5
- Can we check it
- 4 / 5
Why it's on hold
Park, with the panel halted. Anthropic's safeguards stopped the Claude safety reviewer. The outside safety opinion gave a concern (it would veto publishing new collection locations for the organism), and the outside prior-art search failed. Only Neglected (4) is a panel score; the other three are triage first guesses.
Why it's stuck
Melioidosis kills many people for the attention it gets: the 2016 Nature Microbiology study by Limmathurotsakul and colleagues estimated 165,000 human cases a year (95% credible interval 68,000 to 412,000) and that '89,000 people globally will die from melioidosis in 2015', and concluded that the disease is 'likely endemic in a further 34 countries which have never reported the disease.' That map is still the main global estimate, and it was checked only by cross-validation on its own data ('Model accuracy was assessed by calculating the mean cross-validated area under the curve (AUC) statistic', AUC 0.81). Nobody appears to have tested it forward, against places that first reported the disease after it was published.
What we'd try
Rebuild the 2016 boosted-regression-tree model from its public occurrence database (the paper says 'The final occurrence database with geo-positioned data points and the data dictionary are publicly available online', on figshare) and public climate and soil layers (for example WorldClim and SoilGrids). Then build a held-out set of first locally acquired human cases and first environmental detections reported after 2015, from the published literature and public health notices.
We'd look again if
Aaron decides whether the lab should do distribution mapping for this organism at all, and if so how the safety review is to be done and what release limits apply (map resolution, no new precise detection coordinates, no collection guidance). Do not re-vet before then: each run would hit the same block.
Think we got it wrong?
Ruled-out problems can come back. Tell us what we missed.
Members can read the full record for this problem, check its sources and comment. Open the member view → Not a member? Here is how to join.