Does the embedded genome subset favour clinical or environmental isolates?

Projects disagree on which isolates make up the genome subset that carries environmental embeddings (satellite-derived vectors assigned from a genome’s sampling coordinates). One central digest says clinical isolates often lack valid coordinates. Another says clinical isolates are disproportionately embedded. The reports describe the bias in different directions again _1,2,3,4. The size of the subset is also stated inconsistently: 83,227 in one place and 83,287 in another ₁. These questions matter because any environment–genome association drawn from embeddings, or from harmonized metadata, inherits the composition of the subset. That evidence is synthesized in environment-embedding-geography.

Evidence Sides

Side A: clinical isolates often lack valid coordinates, and coverage is described as environment-biased The pitfalls digest says valid coordinates are often missing, especially for clinical isolates. The SNIPE report describes coverage as biased toward environmental isolates with latitude/longitude data _1,2,3,4.

Side B: the subset over-represents clinical isolates The discoveries digest says clinical isolates have good geographic metadata and disproportionately have embeddings. The prophage report describes a bias toward clinically and environmentally well-sampled lineages, which sits between the two framings _1,2,3,4.

Harmonized environment categories showing human clinical and gut dominance. From env_embedding_explorer.

Side C: metadata and coordinate caveats that shape either reading

• Field coverage. The structured env_broad_scale field (an ontology-coded broad environment label) is cleaner but covers only 42% of genomes. Keyword harmonization of the free-text isolation_source field covers 71% of genomes with a label but leaves 17% Other and 12.5% Unknown ₅.
• Distance-decay. Distance-decay is the decline in genomic similarity with geographic distance. It may reflect environmental context, epidemiological structure, institutional clustering, or approximate coordinates _5,6.
• Subset size. The pitfalls digest gives 83,227 embedded genomes in one section and 83,287 in another ₁.

Possible Reconciliations

• Hypothesis 1: The statements refer to different denominators. “Often missing for clinical isolates” may describe clinical isolates as a whole. “Disproportionately embedded” may describe the embedded subset’s composition. If so, both could hold at once.
• Hypothesis 2: The project-level descriptions reflect each project’s own species or lineage selection rather than the shared embedded subset.
• Hypothesis 3: The two subset counts come from different query snapshots or filters, so the digest sections describe different tables.

The supplied evidence does not test these hypotheses.

Resolving Work

• Clinical coordinate rate: Use pangenome (the combined gene content of a species’ sequenced genomes) genome metadata joined to the embeddings table. Compute the fraction of clinical vs environmental isolates that have valid coordinates, and their shares within the embedded subset. Question: is the clinical bias a property of the source population, the embedded subset, or both?
• Project re-runs: Use the SNIPE and prophage genome sets. Tabulate each set’s isolation categories against the full embedded subset. Question: do the project-level bias statements describe project selection rather than platform coverage?
• Subset count: Use the embeddings table. Re-count rows and record the query and filter behind each digest figure. Question: which count is current, and why do the two differ?
• Field agreement: Use genomes that have both env_broad_scale and harmonized isolation_source labels. Cross-tabulate the two labels. Question: do the Other and Unknown bins skew clinical or environmental?
• Distance-decay controls: Use embedded genomes. Fit distance-decay models with and without stratifying by submitting institution and by coordinate precision. Question: does the geographic signal survive these controls?