Shared-Coordinate Filtering: Removing Institutional Artefacts vs. Keeping Intensively Sampled Field Sites

The disagreement is about whether dense clusters of genomes that share coordinates should be filtered out of geographic analyses. A shared-coordinate heuristic is meant to flag institutional addresses, yet the flagged clusters include legitimate, intensively sampled field sites. _1,2 Keeping every dense cluster also keeps likely institutional or approximate locations. ₁ The choice matters for spatial-sampling-effort-confounding, where uneven sampling effort already threatens geographic hotspot inferences.

Evidence Sides

Side 1: Dense clusters include legitimate field sites, so filtering on density risks discarding them. The shared-coordinate heuristic is meant to flag institutional addresses. Yet the flagged clusters include legitimate, intensively sampled field sites such as Rifle, CO and Saanich Inlet. _1,2 A filter that treats high genome density as an artefact therefore risks discarding sites where sampling effort is high by design. _1,2 On this side, high density is potentially informative rather than diagnostic of error.

Side 2: Dense clusters also include likely institutional or approximate locations, so keeping them all risks retaining such placements. Keeping every dense cluster also keeps likely institutional or approximate locations, such as Pittsburgh and Lima. ₁ On this side, at least some high-density coordinates likely do not mark true sampling sites, so retaining them could let such coordinates stand in for real sampling locations. ₁

Neither side’s evidence establishes how many flagged clusters belong to each category. The TENSION text reports examples, not a classification rate.

Possible Reconciliations

• Hypothesis: Genome density alone cannot separate the two kinds of cluster. A second signal, such as the environment label or the metadata for the originating study, may distinguish a field campaign from an institutional address.
• Hypothesis: Field-site clusters and institutional clusters differ in their within-cluster environmental composition. Field sites may be dominated by one environment type, while institutional addresses may mix unrelated sources.
• Hypothesis: A curated allow-list of known long-term field sites, applied before the density filter, would keep sites like Rifle, CO and Saanich Inlet without also keeping locations like Pittsburgh and Lima. This depends on such a list being reasonably complete.

Resolving Work

• Data: genome metadata for every cluster flagged by the shared-coordinate heuristic. Method: manually annotate each cluster as field site, institutional address or approximate location. Question: what fraction of flagged clusters falls into each class?
• Data: environment labels within each flagged cluster. Method: compare label diversity between annotated field-site and institutional clusters. Question: does environmental homogeneity separate the two classes?
• Data: study and BioProject identifiers attached to clustered genomes. Method: test whether a cluster traces to one sampling campaign or to many unrelated submissions. Question: does study provenance predict which clusters are artefacts?
• Data: geographic analyses run with and without the filter. Method: rerun the analyses three ways: unfiltered, filtered, and filtered with a field-site allow-list. Question: do the conclusions change depending on which dense clusters are kept?