Spatial Sampling Effort and Geographic Hotspot Confounding

Geographic hotspot analyses of environmental bacterial metal resistance can be confounded by uneven sampling effort, incomplete coordinates, and concentration of samples from particular studies or expeditions. The Global Biogeography of Environmental Bacterial Metal Resistance provides a preliminary case in which resistance prevalence varies across grid cells, but the report explicitly requires sampling-effort correction and study-origin checks before regional patterns are interpreted. ₁

Evidence from the Global Metal-Resistance Analysis

The study extracted 260,652 environmental MAGs (metagenome-assembled genomes, which are genomes reconstructed from metagenomic reads) from MGnify and retained 30,497 after filtering host-associated biomes. ₁ Of these, 22,356 MAGs had usable geospatial coordinates, corresponding to 73.3% coordinate coverage of the environmental MAG set. ₁ ENA batch API retrieval returned 24,511 sample records, of which 16,964 (69.2%) had valid latitude/longitude pairs, leaving a 30.8% per-sample spatial data gap. ₁

The report’s interpretation section first warned that the 30.8% ENA coordinate gap implies that any global map of metal-resistance distribution built from public archives is subject to substantial geographic blind spots. It left open which biomes are most underrepresented. ₁ A later result in the same report narrows that geographic reading. The coordinate gap does not indicate that entire geographic regions were absent from the archive: the ENA coordinate file covered every area represented in the MAG coordinate dataset, with 0 of 532 MAG grid cells lacking ENA coordinate coverage. ₁ The distinction is therefore between missing coordinates for individual samples and missing geographic grid cells, rather than evidence that the sampled map is geographically complete. ₁

Among the 22,356 MAGs with coordinates, 2.8% carried at least one metal-resistance type. ₁ A 5° grid contained 289 cells with at least 5 MAGs; Fisher’s exact tests with Benjamini–Hochberg false discovery rate (BH FDR), a multiple-testing correction, identified 11 significant hotspots with odds ratio (OR) OR>2 and q<0.05, and 3 coldspots with OR<0.5 and q<0.05. ₁ The report states that metal-resistant genomes are not uniformly distributed among the 22,356 MAGs with coordinates. It also states that hotspot identification and sampling-effort correction are required before regional patterns can be interpreted. ₁

The strongest apparent hotspot was the Atacama/Andean region of Chile at lat=-25°, lon=-70°, where prevalence was 21.8%, OR=9.83, q=7.6e-12, and n=101 MAGs. ₁ Eastern and central USA cells at lat=40°, lon=-80° and -90° had OR=7.9 and 6.3, while an East/Southeast Asia cluster at lat=25-30°, lon=105-120° had OR=4.4–5.9. ₁ These results identify statistical geographic contrasts, but they do not by themselves establish that the contrasts reflect regional environmental selection rather than differences in sampling composition, study representation, or sampling intensity. ₁

Biome Composition as a Confounding Dimension

Biome-stratified tests show that resistance prevalence differs strongly among the sampled environmental categories: soil contained 7,939 MAGs with 5.8% prevalence, OR=5.05, and q=1.8e-82; rhizosphere contained 422 MAGs with 1.9% prevalence, OR=0.66, and q=0.30 (NS); and marine samples contained 13,995 MAGs with 1.2% prevalence, OR=0.20, and q=9.2e-80. ₁

This biome structure supports environmental-resistome by showing soil enrichment and marine depletion in the coordinate-filtered environmental MAG dataset. ₁ It also means that a geographic hotspot can be partly produced by where particular biomes were sampled, unless geographic comparisons account for biome composition. ₁

Study-Origin and Sampling-Intensity Risks

The report identifies the Atacama/Andean and USA hotspots as possible single-study or expedition-level artefacts and requires checking distinct sample_accession prefixes within each hotspot. ₁ This expedition-level clustering check is still untested. The report’s list of pending analyses gives the Atacama hotspot as OR=9.8. That is the value stated in that passage, and it differs from the OR=9.83 given in the detailed hotspot result above. ₁ Concentration of many MAGs from one study, expedition, or accession family could make a grid cell appear unusually enriched while measuring study-specific sampling rather than an independently replicated regional pattern. ₁

The report proposes testing whether hotspots persist after normalising by log(n_MAGs), where n_MAGs is the number of MAGs represented in a grid cell. ₁ This proposed correction directly tests whether hotspot status is robust to sampling effort, but the corrected results were not reported in the preliminary analysis. ₁

The findings therefore refine environment-embedding-geography: coordinate availability can support geographic mapping, but geographic interpretation remains conditional on sampling effort, study replication, and environmental composition. ₁ They also connect to subsurface-hydrogeological-zonation, because spatial contrasts cannot be interpreted independently of how samples were distributed across locations and collection contexts. ₁

Coordinate Quality: Field Sites versus Institutional Addresses

A second project shows that having a coordinate is not the same as having a valid sampling location. AlphaEarth Embeddings, Geography & Environment Explorer identifies 11,765 unique coordinate locations across 83,286 genomes with latitude and longitude. ₂ Its coordinate quality control (QC) classifies 50,109 genomes (60.2%) as good, 30,469 (36.6%) as suspicious clusters, and 2,708 (3.3%) as low precision (integer degrees). ₂

The suspicious-cluster class comes from a heuristic. It flags the 30,469 genomes (36.6%) that sit at shared coordinates with >50 genomes and >10 species, treating these as likely institutional addresses rather than sampling sites. However, several of the flagged locations are legitimate field research sites. ₂

• Rifle, CO: coordinates 39.54, -107.78, 1,883 genomes; a DOE IFRC groundwater research site. ₂
• Saanich Inlet, BC: coordinates 48.36, -123.30, 1,529 genomes; an oceanographic O2-minimum zone. ₂
• Siberian soda lakes: coordinates 52.11, 79.17, 812 genomes; from extremophile sampling campaigns. ₂
• Pittsburgh, PA: coordinates 40.44, -79.97, 630 genomes; described as likely institutional (diverse clinical). ₂
• Lima, Peru: coordinates -12.0, -77.0, 1,641 genomes; integer coordinates, described as likely approximate. ₂

The report calls the coordinate QC heuristic crude and a rough first pass, because it falsely flags legitimate field sites such as Rifle and Saanich Inlet. It proposes checking whether the genomes at each location have homogeneous isolation sources, which would indicate a real site, or diverse unrelated sources, which would indicate an institutional address. ₂ The central discoveries digest records the same caveat. It reports that 30,469 genomes (36.6%) cluster at coordinates with >50 genomes of >10 species, and that several flagged locations are legitimate DOE or research field sites: Rifle, CO (1,883 genomes, DOE IFRC), Saanich Inlet (1,529, oceanographic time series) and the Siberian soda lakes (812, extremophile campaigns). The digest says the heuristic needs refinement and that checking isolation_source homogeneity at each location would better distinguish field sites from institutional addresses. ₃

This refines the sampling-effort problem described above. Dense coordinate clusters can come from two different processes: intensive campaigns at real field sites, or many unrelated genomes geocoded to an institution or an approximate point. ₂ This suggests the hypothesis that effort normalisation alone, such as the proposed log(n_MAGs) correction, cannot tell these cases apart without provenance checks on location and isolation source. Neither project has tested that hypothesis. _2,1

Sampling Intensity and Spatial Structure in Other Projects

The same confound appears in habitat-occurrence data, which supports the concern beyond MAG grid cells. MicrobeAtlas detection of OTUs (operational taxonomic units, which are clusters of similar marker-gene sequences) across 13 environment categories conflates genuine habitat use with sampling intensity, primer bias (V3–V4 or V4), and the geographic and temporal non-uniformity of SRA/ENA sample contributions. ₄

Spatial structure is also unresolved in soil genomics analyses. A soil frontier analysis that obtained negative R² still needs a re-run with spatial blocking. That re-run would separate distributional shift, outlier leverage and true unpredictability. ₅

In a soil metal functional genomics analysis, spatial autocorrelation was not tested. The planned check is Moran’s I, a spatial autocorrelation statistic, applied to model residuals. If that result is significant, the plan is to follow it with SEVM (spatial eigenvector mapping, which adds spatial structure terms to the model). ₆ In both projects, spatial non-independence is listed as an analysis still to be done, not one already corrected for. _5,6

Tensions

The analysis provides strong statistical evidence for geographic enrichment and depletion under its tested grid-cell model, including 11 hotspots and 3 coldspots after BH FDR correction. ₁ At the same time, the report states that hotspot identification requires sampling-effort correction and that regional patterns should not yet be interpreted without checking study or expedition artefacts. ₁ The tension is therefore between detected spatial association and unresolved representativeness of the samples contributing to each association. ₁

A second tension concerns coordinate filtering itself. 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. _2,3 A filter that treats high genome density as an artefact therefore risks discarding sites where sampling effort is high by design. Keeping every dense cluster, on the other hand, also keeps likely institutional or approximate locations such as Pittsburgh and Lima. ₂

Open Directions

• Use the MAG coordinate table and ENA sample records to classify distinct sample_accession prefixes within each reported hotspot, then test whether hotspot significance persists after removing or stratifying dominant studies; this would address the report’s single-study or expedition-level artefact concern. ₁
• Recompute the 5°-grid Fisher’s exact tests after normalising or modelling sampling effort with log(n_MAGs), and ask whether the 11 hotspots and 3 coldspots remain significant under the effort-adjusted analysis. ₁
• Combine MAG resistance status, biome labels, grid-cell counts, and coordinate completeness in a stratified or matched model, and ask whether the Atacama/Andean, USA, and East/Southeast Asia signals persist after controlling for biome composition. ₁
• Compare the 22,356 coordinate-bearing MAGs with the 30,497-MAG filtered set by biome and sample source, and ask whether the 30.8% per-sample coordinate gap disproportionately excludes particular environmental categories. ₁
• Repeat hotspot analysis using study-aware resampling across the 532 represented grid cells, and ask whether regional enrichment is reproducible when no single accession family or sampling campaign dominates a cell. ₁
• Apply the proposed isolation-source homogeneity check to each of the 11,765 coordinate locations, and ask how many of the 30,469 suspicious-cluster genomes move to field-site status; this would close the gap left by the crude density heuristic. ₂
• Cross-reference the metal-resistance hotspot grid cells against the coordinate QC classes (good, suspicious cluster, low precision), and ask whether any hotspot is driven by institutional or integer-degree coordinates. _2,1
• Run Moran’s I on the residuals of the hotspot and soil models, followed by SEVM if the result is significant. Also re-run the models with spatial blocking, and ask whether the associations and the negative R² hold once spatial non-independence is modelled. _6,5,1