What this is
Canonical name: AlphaEarth. 1
Known aliases: No aliases were reported in the source. 1
Stable external identifier: None was reported in the source. 1
AlphaEarth is an environmental-embedding dataset used in the Environmental Resistome at Pangenome Scale analysis to provide a continuous environmental representation for species in the KBase KE pangenome collection. 1 The KBase Data Lakehouse atlas supports this reusable environmental-layer interpretation by inventorying 83,287 AlphaEarth environment embeddings indexed to KBase genomes. 2 The AlphaEarth Embeddings, Geography & Environment Explorer characterizes these as 64-dimensional satellite-derived embeddings for 83,287 genomes, representing 28.4% of 293,059 genomes; 79,449 records have valid values across all dimensions and 3,838 contain at least one NaN. 3
The SNIPE analysis supports AlphaEarth as an ecological-context layer beyond AMR and gene-content studies: among species with embeddings, SNIPE-bearing species numbered n=1,069 and non-SNIPE species n=13,977, with significant differences in 22 of 64 dimensions after Welch’s t-tests with Bonferroni correction. 4 The largest effect was dimension A19, with Cohen’s d = 0.26 and p = 5.0e-14; effect sizes ranged from |d| = 0.11–0.26, indicating a modest environmental shift rather than strict habitat segregation. 4 The dimensions derive from satellite remote-sensing imagery at sample-collection latitude/longitude coordinates and may capture vegetation indices, temperature, precipitation, and land cover, but no named biological interpretation was assigned to A19 or the other dimensions. 4 This refines existing environmental-association claims because AlphaEarth coverage was 28.4% of genomes and biased toward environmental isolates with latitude/longitude metadata. 4
The pitfalls refines these coverage figures by reporting 83,227 of 293,059 genomes, or 28.4%, as covered in one quality check; a separate check found 3,838 of 83,287 genomes, or 4.6%, with at least one NaN among 64 dimensions, leaving 79,449 after complete-value filtering. 5 Coverage depends on valid latitude and longitude and is especially sparse for clinical isolates. 5
The Lanthanide Methylotrophy Atlas — Summary refines the coverage limitation by reporting AlphaEarth coverage for 1,457 of 3,690 xoxF-bearing genomes (39.5%), compared with a 28% pangenome baseline, leaving 60.5% without environmental coordinates. 6 It proposes coverage-restricted PCA (principal component analysis) and UMAP (Uniform Manifold Approximation and Projection), stratified by phylum, to test whether xoxF carriers form distinct environmental or biogeographic clusters. 6
The prophage ecology analysis supports AlphaEarth as a continuous environmental-gradient layer beyond AMR and marker-gene applications: among 2,008 species with at least 5 embedded genomes, environmental niche breadth was partially correlated with prophage module count after controlling for genome size (partial Spearman rho=0.468, p=8.41e-110). 7 This result refines interpretation of AlphaEarth-derived ecological breadth by showing a strong module-level association in a prophage analysis, while the study was limited to 28% of genomes and a biased subsample toward clinically and environmentally well-sampled lineages. 7
Evidence across source documents
The analysis included 2,659 species with both antimicrobial-resistance (AMR) data and AlphaEarth environmental embeddings. 1 A separate atlas analysis included 2,684 species with at least 3 genomes and AlphaEarth embeddings. 8 These are reported as distinct analysis cohorts; the sources do not reconcile their different counts. 1,8
Of 64 embedding dimensions, 52 significantly correlated with AMR diversity after false-discovery-rate correction at FDR < 0.05. 1 Embedding dimension A34 had the strongest reported correlation with AMR diversity, with Spearman rho = +0.24. 1
A Mantel test found that environmental distance, measured as cosine distance between AlphaEarth embeddings, predicted AMR-profile distance, measured as Bray–Curtis distance between resistance-mechanism fractions, with r = 0.098 and p = 0.001. 1 Stratified Mantel correlations were strongest for clinical species (r = 0.177), moderate for soil species (r = 0.129), and weakest for aquatic species (r = 0.061). 1
The atlas supports the continuous environmental-layer interpretation: environmental diversity predicted AMR count (Spearman rho=0.466, p=1.6e-144) and was negatively correlated with AMR core fraction (rho=-0.173, p=1.8e-19) among the 2,684-species cohort. 8 These associations refine the earlier embedding evidence but do not establish causation. 8
The ecotype analysis refines these AMR-specific associations by showing that, across 172 species, the median partial correlation was 0.0025 for environment versus 0.0143 for phylogeny, with no significant environmental effect in 156 species (90.7%). 9 Phylogeny dominated gene-content similarity in 60.5% of species, whereas environment dominated in 39.5%; this result supports a limited, context-dependent environmental signal rather than a uniformly strong whole-genome effect. 9 The analysis further suggests the hypothesis that AlphaEarth embeddings do not fully capture ecologically relevant environmental variation. 9
The environmental-versus-human-associated reanalysis supports that hypothesis and refines the clinical-bias interpretation: environmental species (n=37) had median partial correlation 0.051, compared with 0.084 for human-associated species (n=93) and 0.109 for Mixed/Other species (n=53); the one-sided Mann-Whitney U test for Environmental > Human-associated gave U=1536 and p=0.83. 10 The continuous analysis likewise found no relationship between the fraction of environmental genomes and partial-correlation strength (rho=-0.085, p=0.25), while the human-associated fraction gave rho=0.030 and p=0.69. 10 Thus, the reanalysis contradicts the hypothesis that environmental species have stronger environment–gene-content correlations, while confirming that clinical sampling bias alone does not explain the weak signal. 10
The new explorer refines this interpretation by showing that environmental samples have a stronger geographic embedding gradient than human-associated samples: environmental categories rise from mean cosine distance 0.27 within <100 km to 0.90 beyond 10,000 km (3.4x), whereas human-associated categories rise from 0.37 to 0.75 (2.0x). 3 Across 50,000 sampled genome pairs, mean cosine distance increased monotonically across the reported geographic bins—from 0.41 at <100 km to 0.82 at 10K–20K km—and plateaued at intercontinental distances, supporting spatially autocorrelated environmental context rather than random variation. 3 This supports AlphaEarth as a spatial environmental layer but refines the gene-content interpretation: the relationship to environmental context does not by itself establish a strong environment–gene-content relationship. 3 The pitfalls analysis supports stratification by environment category by reporting a pooled near-versus-far ratio of 2.0x, compared with 3.4x for environmental samples and 2.0x for human-associated samples. 5
The Gene Function Ecological Agora supports AlphaEarth as an ecological-context layer beyond AMR analyses: k = 10 clustering of embeddings placed 35.6% of 90 Cyanobacteriia species in a marine+sponge-dominant cluster, 34.0% of 50 Mycobacteriaceae species in a gut+sludge cluster and 20.0% in another gut cluster, while 729 Bacteroidota species were distributed across marine and gut-associated clusters. 11 These results refine the earlier geographic-distance evidence by showing focal-clade environmental concentration, but they establish ecological consistency rather than causal effects of environment on gene innovation. 11 The analysis is summarized in Gene Function Ecological Agora.
The functional-dark-matter analysis supports AlphaEarth as an auxiliary environmental layer for testing unknown gene functions: among 151 accessory dark-gene clusters tested across 31 species, 10 of 137 clusters with environmental-category tests showed significant enrichment at FDR < 0.05, while 1 of 67 showed significant AlphaEarth embedding divergence. 12 A pre-registered lab–field mapping found 29 of 47 testable clusters concordant (61.7%), but the one-sided binomial test gave p = 0.072; Fisher’s combined probability across 47 tests gave p = 0.031. 12 NMDC validation confirmed four pre-registered abiotic predictions, including pH-associated carriers correlated with pH (ρ = +0.157, n = 4,366, FDR = 7.4e-25) and anaerobic carriers correlated negatively with dissolved oxygen (ρ = -0.298, n = 272, FDR = 1.5e-6). 12 These results refine the earlier geographic-distance evidence by showing that AlphaEarth can contribute to gene-level environmental hypothesis testing, while leaving the strength of general lab–field concordance unresolved. 12
The PHB granule ecology analysis supports AlphaEarth as a niche-breadth layer while refining the interpretation of environmental associations. Among 2,008 species with sufficient embedding representation, PHB-positive species had median embedding variance of 0.3295 (n = 531) versus 0.2472 for PHB-negative species (n = 1,477), with Mann–Whitney U = 446,546 and p = 1.88 x 10^-6; the raw PHB–niche-breadth association was rho = 0.106 with p = 1.77 x 10^-06. 13 However, PHB-positive species had median genome size of 4.34 Mbp versus 2.44 Mbp for PHB-negative species, genome size correlated with niche breadth at rho = 0.302 with p = 1.5e-43, and the genome-size-controlled partial correlation was rho = -0.047 with p = 0.037, a 56.3% reduction and sign reversal. 13 This contradicts treating the raw PHB–embedding association as a PHB-specific effect and supports genome size as its primary correlate, while the analysis remains limited to 2,008 species, or 7.2% of total species diversity. 13
The PHB study further found environmental enrichment across genome-size strata: high- versus low-variability environments showed 10.7% versus 2.5% phaC positivity in Q1, 18.5% versus 4.0% in Q2, 34.3% versus 10.9% in Q3, and 50.9% versus 37.4% in Q4, with enrichments of 4.4x, 4.6x, 3.1x, and 1.4x and p = 1.18 x 10^-11, p = 3.25 x 10^-47, p = 4.17 x 10^-70, and p = 3.21 x 10^-14, respectively. 13 These results support environmental context as relevant to PHB distribution but do not directly validate temporal variability, because AlphaEarth and NMDC measurements do not themselves measure feast/famine dynamics. 13
The prophage study supports the geographic-gradient interpretation while adding a cross-modal ecological application: AlphaEarth niche breadth remained associated with prophage module count after genome-size control, and environmental effects on prophage composition were detected beyond genome size and family-level phylogeny. 7 In a complementary analysis of 1,773 species, PERMANOVA, a permutation-based multivariate analysis of variance, found environment significant at p=0.01, with F=30.04 versus F=6.17 for family-level phylogeny; the analysis therefore refines the earlier weak and context-dependent gene-content results by showing that environmental structure can be stronger for a specialized prophage-module phenotype than for whole gene-content similarity. 7 The result remains an association, and AlphaEarth coverage was only 28% of genomes. 7
The new dataset characterization also refines the sampling-bias limitation. Among AlphaEarth genomes, 38% were human-associated: Human clinical included 16,390 genomes (20%), Human gut 13,466 (16%), and Human other 1,669 (2%); Soil included 6,073 (7%), Marine 5,850 (7%), and Freshwater 5,840 (7%). 3 Coordinate quality was heterogeneous: among 83,286 genomes with latitude and longitude, 50,109 (60.2%) were Good, 30,469 (36.6%) Suspicious cluster, and 2,708 (3.3%) Low precision. 3 The explorer reports that suspicious-cluster flags include legitimate field sites such as Rifle and Saanich Inlet, so the geographic signal should not be treated as free of metadata artifacts. 3
The report treats the embedding analysis as supplementary confirmation of the discrete environment analysis, but notes that the embedding dimensions are opaque and that coverage was 28% of genomes. 1 The atlas likewise reports limited AlphaEarth coverage and geographic-metadata bias, so the new associations should not be generalized beyond the sampled genomes. 8 In the ecotype cohort, AlphaEarth embeddings covered 28.4% of genomes, which refines the earlier coverage estimate and reinforces the limitation on environmental inference. 9 The reanalysis further found that environmental species had a 21% NaN partial-correlation rate versus 7% for human-associated species, so environmental species were disproportionately filtered rather than preferentially retained. 10 The explorer independently reports NaN values in 3,838 of 83,287 records (4.6%), with their cause unknown. 3 The functional-dark-matter report similarly cautions that AlphaEarth embeddings cover only 28% of genomes (83K/293K), and that sparse environmental metadata limit interpretation. 12 The lanthanide methylotrophy atlas supports this caution: 60.5% of its xoxF genomes lacked environmental coordinates, limiting geographic inference for that marker-defined group. 6 The pitfalls analysis refines it further by noting that human-associated samples can dilute geographic patterns and that missing coordinates or metadata should not be interpreted as biological absence. 5
The reanalysis used genome-level harmonized environment classifications and the same within-analysis methodology for comparison groups, but used all genomes with embeddings rather than diversity-maximizing downsampling; its median partial correlation across 183 species was 0.081 versus 0.003 in the original analysis, a reported 27x difference. 10 The report states that absolute values are not comparable across methodologies, while the within-method environmental-versus-human-associated comparison remains valid. 10
The explorer further found uneven metadata coverage and harmonization: isolation_source covered 76,295 genomes (91.6%), host 53,091 (63.7%), env_broad_scale 34,800 (41.8%), env_local_scale 31,541 (37.9%), and env_medium 31,483 (37.8%). 3 Keyword matching mapped 5,774 unique isolation_source values to 12 broad categories, capturing 71% of genomes with a label; 17% remained Other and 12.5% were Unknown. 3 These limitations refine interpretation of both the earlier AMR and ecotype associations by showing that “environment” is partly inferred from incomplete, heterogeneous metadata.
The SNIPE environmental comparison likewise establishes statistical association rather than a specific causal habitat mechanism, and its AlphaEarth coverage was biased toward environmental isolates with geographic metadata. 4 Its modest effect sizes refine stronger interpretations of environmental niche separation and make the result compatible with the existing evidence for spatial structure without establishing that a particular environmental variable causes SNIPE distribution. 4
These findings extend environmental-resistome by supplying a continuous environmental layer, contribute to multi-omics-integration by integrating environmental embeddings with AMR mechanism profiles, and inform pangenome-integration by testing environmental versus phylogenetic structure in gene content. 1,8,9 They also support ecotype-environment-gene-content by providing a systematic reanalysis of the null environment–gene-content relationship and its methodological boundary conditions. 10 The new explorer supports environment-embedding-geography by establishing geographic distance-decay and environment-stratified embedding behavior, and refines ecotype-environment-gene-content by motivating an environmental-only reanalysis rather than treating a stronger relationship as established. 3 The Gene Function Ecological Agora supports environment-embedding-geography and refines ecotype-environment-gene-content through focal-clade environmental clustering, while integrating AlphaEarth with phylogenomic and functional evidence. 11 The functional-dark-matter analysis supports environment-embedding-geography and cross-tenant-data-bridging by using AlphaEarth alongside Fitness Browser, pangenome, and NMDC evidence to prioritize environmentally testable dark genes. 12 The lanthanide methylotrophy atlas supports environment-embedding-geography and refines pangenome-integration by applying the environmental layer to xoxF-bearing genomes while highlighting incomplete coordinate coverage and the need for phylum-stratified embedding analyses. 6 The PHB study supports environment-embedding-geography by using embedding variance to quantify environmental breadth, while refining interpretation of that breadth by showing genome-size confounding and limited coverage. 13 The prophage ecology analysis supports environment-embedding-geography and cross-tenant-data-bridging by using AlphaEarth niche breadth to test prophage-module ecology while controlling for genome size and linking pangenome patterns to NMDC validation. 7 The SNIPE analysis supports environment-embedding-geography by testing environmental association for a defense-system gene family and refines interpretation through explicit coverage and causal-mechanism caveats. 4 The pitfalls report supports provenance-aware-resource-discovery by documenting coverage, coordinate-quality, and missingness checks that should precede AlphaEarth joins and geographic interpretation. 5 They are summarized in Pan-Bacterial AMR Gene Landscape, BERDL Data Atlas — Inventory, Topic Map, and Cross-Reference Synergies, Ecotype Correlation Analysis, Ecotype Reanalysis — Environmental vs Human-Associated Species, AlphaEarth Embeddings, Geography & Environment Explorer, Functional Dark Matter — Experimentally Prioritized Novel Genetic Systems, Gene Function Ecological Agora, Lanthanide Methylotrophy Atlas — Summary, Polyhydroxybutyrate Granule Formation Pathways — Distribution Across Clades and Environmental Selection, pitfalls, Prophage Gene Modules and Terminase-Defined Lineages Across Bacterial Phylogeny and Environmental Gradients, and SNIPE Defense System in the KBase Data Lakehouse Pangenome. 8,2,9,10,3,12,11,6,13,5,7,4