Multi-Omics Integration

Multi-omics integration combines measurements from multiple biological modalities to analyze the same genes, strains, or phenotypes. The Acinetobacter baylyi ADP1 Data Explorer provides an ADP1-centered case study connecting genome features, TnSeq essentiality, flux balance analysis (FBA), mutant growth fitness, proteomics, pangenome classification, and functional annotations in one database. ₁ The ADP1 Deletion Collection Phenotype Analysis refines this case by integrating deletion-collection growth measurements with TnSeq classifications, functional annotations, and pangenome status across 2,034 genes and 8 carbon sources. ₂ The Triple Essentiality Concordance Analysis for *Acinetobacter baylyi* ADP1 extends it to direct concordance analysis of FBA, RB-TnSeq (random barcode transposon sequencing), complete-gene knockout phenotypes, mutant growth rates, and proteomics. ₃ The AlphaFold MSA Depth as a Lens on the Bacterial Annotation Gap extends the framework with AlphaFold multiple-sequence-alignment (MSA) depth as a structural-representation and annotation-richness layer linked to pangenome gene clusters. ₄

The Condition-Specific Respiratory Chain Wiring in ADP1 extends and refines the ADP1 framework by integrating gene-phenotype measurements, FBA, theoretical stoichiometry, cross-species Fitness Browser data, and proteomics to explain condition-specific respiratory-chain wiring. ₅ In ADP1, carbon sources selected qualitatively different respiratory configurations: quinate required Complex I, acetate required Complex I, cytochrome bo3, ACIAD3522, and additional components, lactate specifically required cytochrome bo3, glucose had no specifically required respiratory component, and urea was broadly demanding. ₅ This supports the existing condition-aware interpretation of fitness while refining simple total-yield explanations: quinate produced 4 NADH, or 0.57 NADH per carbon, yet Complex I had a growth ratio of 0.37, whereas glucose produced 9 NADH, or 1.50 NADH per carbon, and Complex I had a growth ratio of 1.44. The report proposes that concentrated NADH production from quinate catabolism can exceed alternative dehydrogenase capacity, but this is a theoretical stoichiometric interpretation rather than a measured flux distribution. ₅ Similar protein levels for Complex I, NDH-2, and ACIAD3522 under standard conditions—27.6 at the 66th percentile, 27.0 at the 59th percentile, and 26.2 at the 48th percentile, respectively, against a genome median of 26.4—support passive, flux-based wiring rather than a transcriptional switch. ₅ The report notes that the spread is only 1.4 units, that NDH-2 is therefore not repressed but constitutively co-expressed with Complex I, and that which dehydrogenase becomes limiting should depend on the NADH flux rate from the carbon source being catabolized. ₅ These abundances were measured under standard growth conditions only, so condition-specific upregulation of Complex I on aromatic substrates or of NDH-2 on glucose is not established here; the report proposes reanalyzing Stuani et al. (2014) quinate-versus-succinate proteomics for respiratory-chain proteins as a test of the wiring model. ₅

The Pan-Bacterial Metal Fitness Atlas extends this framework from organism-centered multi-omics to cross-species fitness–pangenome–module integration. It combines 559 metal experiments across 31 organisms and 16 metals with 383,349 gene × metal fitness records, pangenome conservation, ortholog families, module activity, and genome-scale prediction. ₆ Across 22 organisms and 14 metals, metal-important genes were 87.4% core versus 76.9% for baseline genes (OR=2.08, p=4.3e-162), while essential-metal tolerance genes had a mean core-fraction delta of +0.148 versus +0.081 for toxic metals (U=39, p=0.015). ₆ This supports a two-tier model in which broadly required core stress functions coexist with accessory specialized resistance mechanisms, rather than treating all metal responses as one resistance modality. ₆ A 1,286-KO repertoire score did not predict metal fitness, refining simple gene-content inference and motivating regulatory and expression-based integration. ₆

Molecular, ecological, and metadata integration

The Environmental Resistome at Pangenome Scale extends the framework from organism-centered modalities to an ecological resistome layer: across 14,723 bacterial species and 293K genomes, ecological environment was associated with AMR diversity, core/accessory composition, and resistance-mechanism composition. ₇ The Annotation-Gap Discovery via Phenotype-Fitness-Pangenome-Gapfilling Integration refines it by using metabolic-model gapfilling as an integration target, combining sequence homology, gene annotations, pangenome conservation, GapMind pathway evidence, and carbon-source fitness across 14 organisms. ₈ The Aromatic Catabolism Support Network in ADP1 extends it with an ADP1 aromatic-catabolism support network that jointly interprets growth phenotypes, FBA, genomic organization, co-fitness, and cross-species ortholog-transferred fitness. ₉

The BERDL Data Atlas — Inventory, Topic Map, and Cross-Reference Synergies supports and broadens this framework from project-specific integrations to an observatory-scale data layer: the BERDL Data Atlas inventories 1,740 deduplicated tables across 119 databases, 17 tenants, 17 biological topics, and 66 BERIL projects, with 536 schema-level cross-tenant bridges defined by 29 canonical join keys. ₁₀ Among 66 audited projects, 51 (77%) span multiple tenants, showing that molecular, fitness, pangenome, and ecological integration is an infrastructure pattern rather than only an ADP1-specific design. ₁₀ Schema compatibility is not validated value-space overlap, and only UC1 of five bridge use cases was sample-executed. ₁₀ The pitfalls further refines this infrastructure claim: live namespace discovery, schema and type inspection, and provenance-preserving joins are prerequisites because KBase Data Lakehouse is migrating from Delta to Iceberg, migration remains incomplete, and archived underscore-form references may no longer be executable. ₁₁

The Gene Function Ecological Agora extends the framework to a GTDB-r214, phylogeny-centered evidence chain integrating KO presence and participation, Sankoff-parsimony gain events, pangenome and MGE context, environmental metadata, BacDive phenotypes, AlphaEarth embeddings, and gene neighborhoods across 18,989 species representatives. ₁₂ Its 13.74M rank × clade × KO producer/participation scores, 17.07M gain events, and 3.94M KO × genus MGE-machinery records refine cross-modal integration by separating phylogenetic signal, annotation bias, mobile-element context, and within-clade heterogeneity. ₁₂ BacDive phenotype anchors confirmed mycobacterial aerobic-rod and Bacteroidota saccharolytic profiles. AlphaEarth environmental clustering recovered ecological structure, with focal clades concentrated in expected env-clusters. ₁₂ The report frames atlas, ecology, and phenotype as three independent measurement substrates. Their convergence is evidence for the same biology but does not establish causation. ₁₂

The ENIGMA Carbon Census — Tiered Knowledge Census of 83 Enrichment Compounds extends this pattern to a tiered chemical-to-ecological census integrating compound structures, pathway and enzyme annotations, ENIGMA-isolate utilization predictions, GTDB placement, SSO occurrence, and NMDC and Planet Microbe abundance. ₁₃ All 83 enrichment compounds were structure-resolved, 54 linked to KEGG, and 9 callable under the project definition, while 74/83 (89%) were organism-dark. ₁₃ This refines the meaning of cross-modal coverage: an integration result can be an explicit knowledge gap rather than a biological conclusion. ₁₃

The Metabolic Consistency of Pseudomonas FW300-N2E3 Across Four KBase Data Lakehouse Databases supports integration when evidence is matched at metabolite and organism levels. For Pseudomonas FW300-N2E3, Web of Microbes exometabolomics, kescience-fitnessbrowser mutant fitness, BacDive utilization phenotypes, and GapMind pathway predictions were joined through a manual metabolite crosswalk; among 21 testable metabolites, 17/21 (81%) were fully concordant, 4/21 (19%) partially concordant, none fully discordant, and mean concordance was 0.94. ₁₄ Across 41 comparisons, 37 were concordant (90.2%), but only 21/58 WoM metabolites (36%) were testable and 37/58 (64%) occurred only in WoM. ₁₄ Fitness Browser and GapMind were each concordant for 21/21 and 13/13 metabolites, whereas BacDive utilization was 3/7 (43%). ₁₄ The new Web of Microbes Data Explorer refines this bridge by distinguishing WoM I actions (amplification of metabolites already present) from E actions (emergence of metabolites absent from the starting medium). In the 2018 snapshot, E and I were mutually exclusive across 10,744 observations, with no organism consumption action, so production cannot be treated as utilization or uptake. ₁₅ For the directly matched FW300 strain, 19 WoM-produced metabolites also had Fitness Browser carbon or nitrogen-source tests; 5 were E and 14 were I, providing a concrete metabolite-production-to-fitness bridge while leaving consumption unmeasured. ₁₅ The pitfalls evidence further refines this concordance because Web of Microbes action codes are context-dependent and BacDive utilization is four-valued; only explicit + and - should enter utilization percentages. ₁₁

The Community Metabolic Ecology via NMDC × Pangenome Integration extends this framework to community-scale integration of NMDC taxonomy, GTDB pangenome pathway potential, GapMind completeness, and metabolomics. Across 220 samples and 80 pathways, 11 of 13 amino-acid pathways had negative correlations with ambient metabolites by a sign test (p = 0.011); leucine and arginine remained significant after Benjamini-Hochberg false-discovery-rate correction (leucine r = −0.390, q = 0.022, n = 62; arginine r = −0.297, q = 0.049, n = 80). ₁₆ This supports using metabolite pools as an ecological readout of community metabolic potential, while remaining primarily a soil test because all 33 Freshwater samples lacked paired metabolomics. ₁₆ Soil and Freshwater communities showed strong pathway-potential separation, with 17 of 18 amino-acid pathways differing by ecosystem type after correction, but absent pH, temperature, and total-organic-carbon measurements leave physicochemical and taxonomic explanations unresolved. ₁₆ The pitfalls evidence further refines this interpretation: classifier and metabolomics file_id namespaces must be bridged through sample_id using omics_files_table, and wide taxonomy matrices, entity-attribute-value environment metadata, and abiotic zeros representing unmeasured variables require explicit handling. ₁₁ The Functional Dark Matter — Experimentally Prioritized Novel Genetic Systems supports metadata-level validation from an independent sample set. Taxa carrying dark genes with nitrogen-source lab phenotypes were more abundant in nmdc samples with higher nitrogen availability. Carriers of pH phenotypes tracked sample pH, and carriers of anaerobic phenotypes were enriched in low-oxygen samples. These metagenomic samples were entirely separate from the pangenome-based carrier analysis. ₁₇ These remain abiotic correlations rather than molecular readouts. NMDC proteomics (346K observations) and metabolomics (3.1M observations) were not used in that analysis and are proposed only as a future source of more direct functional evidence. ₁₇ The BERDL Data Atlas likewise describes NMDC metabolomics (3.1M), proteomics (346K), and lipidomics (1.4M) layers as largely untapped, and its UC4 and UC5 bridges depend on that layer. It attributes these counts only to agent memory. ₁₀

The NMDC project positions this community-scale integration against prior method families rather than claiming a new phenomenon: its hypothesis is consistent with Mallick et al. (2019) and Noecker et al. (2016), both of which showed that community metabolomics is partially predictable from genomic or metagenomic functional profiles, while its use of GapMind pathway completeness as the predictor is methodologically distinct from flux-based or PICRUSt-style models and arrives at comparable conclusions. ₁₆ The report states that, to its knowledge, this is the first application of GapMind community-weighted pathway completeness scores to predict environmental metabolomics across a large cross-habitat NMDC dataset, integrating 305M GapMind pathway records (27,690 GTDB species × 80 pathways) with NMDC multi-omics for 220 samples spanning two habitat types; the priority claim is explicitly qualified as “to our knowledge” and is not backed here by an independent survey of prior work. ₁₆

The Prophage Gene Modules and Terminase-Defined Lineages Across Bacterial Phylogeny and Environmental Gradients extends this ecological integration architecture to prophage-associated molecular modules, environmental gradients, host phylogeny, contig organization, and metagenomic validation. Across 27,702 bacterial species and 4,005,537 eggNOG-annotated gene clusters, packaging and lysogenic-regulation modules occurred in 100.0% of species, while 34.9% carried all seven modules; these prevalence estimates are annotation-based and may include domesticated remnants and bacterial homologs. ₁₈ PERMANOVA, a permutation-based multivariate analysis of variance, found significant effects of genome-size quartile, environment, and family-level phylogeny (each p=0.01), with genome size dominant (F=212.99), followed by environment (F=30.04) and phylogeny (F=6.17); genome size correlated with prophage cluster count at rho=0.717. ₁₈ Environment therefore supports the existing claim that metadata can explain molecular composition beyond phylogeny, while also refining it by showing that genome size must be modeled explicitly. ₁₈

The Contamination Gradient vs Functional Potential in ENIGMA Communities extends ecological integration from chemical occurrence and abundance to contamination-linked functional potential. It connected ENIGMA geochemistry and community composition to kbase-ke-pangenome clades and eggnog functional proxies across 108 samples, producing 324 site functional-score rows and 12 model-result rows. ₁₉ Predeclared genus-level contamination–defense tests were null, while coverage-adjusted exploratory associations were sensitive to mapping mode, covariates, fraction strata, and multiple-testing correction; the relaxed model had beta = 0.000751, 95% bootstrap CI [0.000224, 0.001779], p = 0.000398, FDR q = 0.0462, whereas the strict model had beta = 0.000640, 95% bootstrap CI [0.000169, 0.001538], p = 0.00354, FDR q = 0.130. ₁₉

The SSO Subsurface Community Ecology — Spatial Structure, Functional Gradients, and Hydrogeological Drivers extends the ecological layer to spatially resolved 16S amplicon sequence variant (ASV) communities, hydrogeological zones, inferred guilds, and short-term groundwater stability. Across 9 wells spanning approximately 6 m, hydrogeological zone explained 27.5% of sediment community variance (F = 4.05, p = 0.0001), well identity explained 19.2% and was not significant (F = 0.80, p = 0.979), and a Mantel test detected distance-decay (Spearman ρ = 0.323, p = 0.029, 9,999 permutations). ₂₀ Groundwater rankings were nearly unchanged over 9 days (Mantel ρ = 0.867, p = 0.001), but taxonomy-to-trait inference covered only 21% of reads and direct SSO geochemistry was unavailable. ₂₀

The Lignin Enrichment and Ecological Memory in Microbial Communities extends ecological integration to experimentally manipulated bacterial and fungal communities. Using paired 16S V3–V4 and ITS2 data from 21 samples across 7 groups, it integrated treatment history, current carbon regime, bacterial and fungal composition, diversity, and Bray–Curtis/PERMANOVA analyses. ₂₁ Lignin enrichment explained 97.9% of bacterial community variance (R²=0.979, p=0.001), while Shannon diversity fell from 6.46 to 3.16 and observed OTUs from 1,594 to 163; labile-carbon supplementation increased Acinetobacter from 25.2% to 41.7% and Aeromonas from 0.1% to 20.2%, while Pseudomonas fell from 39.3% to 23.0%. ₂₁ Round-1 history explained 58.9% of Round-2 variance (F=14.31, R²=0.589, p=0.002), versus 32.7% for current carbon (F=4.85, R²=0.327, p=0.018). ₂₁ This supports condition-aware ecological integration but refines functional interpretation because pathway activity was proposed rather than directly measured and fungal replicate distances reached 0.99–1.00 for several groups. ₂₁

The Metadata Correlates of Eukaryotic Contamination in NMDC Prokaryote-Targeted Metagenomes extends the ecological and metadata layer to eukaryotic signal in prokaryote-targeted metagenomes. Across 2,759 NMDC ReadbasedAnalysis runs from 9 studies, GOTTCHA2 detected eukaryotic reads in 77% of runs; median eukaryotic fraction was 2.7%, mean 13.3%, and 20% exceeded 20%, with plastids representing a median 100% of detected eukaryotic signal. ₂₂ Cross-study associations were confounded with study or batch, whereas local vegetation and geography remained predictive within the dominant NEON soil study. ₂₂

The Rational Design of Protective Microbiome Formulations for Competitive Exclusion of *Pseudomonas aeruginosa* in Cystic Fibrosis Airways extends the architecture to therapeutic microbiome formulation by combining planktonic inhibition, carbon utilization, growth kinetics, patient metagenomics and metatranscriptomics, pairwise interactions, and pangenome pathway conservation against pseudomonas-aeruginosa in cystic-fibrosis airways. It analyzed 4,949 isolates from 175 patient samples, including 220 inhibition-tested isolates, 430 substrate-profiled isolates, 32 with growth kinetics, and 499 pangenome genomes across six species. ₂₃ This refines the framework by separating measured inhibition from inferred engraftability and genomic prebiotic hypotheses. ₂₃ Patient metatranscriptomics adds an activity layer. Engraftability screening used paired metagenomic (DNA abundance) and metatranscriptomic (RNA activity) data from 175 patient samples. In vivo, 47 KEGG pathways showed a >2× commensal-to-PA expression ratio, dominated by PTS sugar transport systems for maltose, trehalose, and N-acetylmuramic acid that only commensals expressed. ₂₃ During acute exacerbation, PA downregulated 170 of 207 measured pathways, including cysteine and methionine biosynthesis (10× lower) and OprD-associated imipenem resistance (13× lower). Only phosphate transport was upregulated (64×). The report’s inference that this makes PA more vulnerable to competitive exclusion is a hypothesis, not a direct measurement. ₂₃

The Regulatory and proteomic architecture of Δfur-permitted lipid A loss in *Caulobacter crescentus* extends it to a genetically defined envelope-rescue mechanism integrating transcriptomics, single-replicate outer-membrane proteomics, Caulobacter RB-TnSeq fitness, published ChvI regulons, and comparative annotation. Fur derepression was supported by correlation with an independent Δfur signal across 93 DEGs (Spearman ρ = 0.315, p = 2.08e-03, 71% sign concordance), whereas the cbb3/cyd/fix-NOPQ respiratory program remained a hypothesis because its 26-gene set had 1.04× fitness enrichment and hypergeometric p = 0.515. ₂₄ Outer-membrane proteomics complicates the transcript-level reading of envelope transport. In the Δlpxc state, MsbA-like CCNA_00307 transcript rose +0.89 (FDR=0.01) and LptC-related CCNA_03716 rose +0.56 (FDR=0.005), consistent with the Uchendu 2026 shared-component model. In contrast, the two canonical Lpt proteins actually detected declined in the rescued strain (LptD log2(4672/4659)=−0.47; LptE=−0.78). Single-replicate proteomics cannot show whether this reflects substrate-limited transport or contradicts the model, and the apparatus’s functional status awaits replicated proteomics. ₂₄ MsbA-like protein was detected only in rescued strain 4672 (abundance 300, NaN in others), so its direction relative to wild type is uncomputable. ₂₄ The sphingolipid transporter lptC2 (CCNA_01226) showed the opposite discordance: transcript −0.60 (FDR 0.034) but protein log2(4672/4659) = +1.08. Because lptC2 protein was already down −0.42 log2 in the intermediate strain (4659/4580), the net change versus wild type is +0.66 log2 (≈1.58×), not +1.08. ₂₄ This is a suggestive pilot observation. It is consistent with post-transcriptional stabilization of sphingolipid-transport machinery but does not establish it statistically, and it requires the replicated proteomics scheduled for summer 2026. ₂₄ The integration also yielded a null result. CtpA/CCNA_03113 had transcript logFC +0.58, pvalue=0.048, FDR=0.109 in 4599-vs-4584 and was not detected in the outer-membrane proteome. The pre-registered induction claim was therefore rejected, and CtpA’s LpxF-substitute hypothesis (Zik 2022) remains untested. The cumulative 4599-vs-4580 contrast (FDR=0.035) confounds Δfur and Δlpxc effects. ₂₄

The Harvard Forest Long-Term Warming — DNA vs RNA Functional Response extends the framework to a long-term environmental perturbation with paired DNA, RNA, taxonomy, KO composition, and metabolite-richness evidence. In a 25-year +5°C Harvard Forest experiment, Actinobacteria increased from 0.249 to 0.315 (log2 FC +0.341, q=0.049) and Acidobacteria decreased from 0.035 to 0.024 (log2 FC −0.549, q=0.049) in organic soil. ₂₅ Treatment explained 7.6% of genus-level Bray–Curtis variance (p=0.069), horizon 30.6% (p=0.0002), and treatment × horizon 41% (p=0.0002). ₂₅ Functional-gene composition explained approximately 12% of DNA and 10–11% of RNA response variance; after accounting for horizon × incubation, DNA and RNA responses were comparable, so the RNA-greater-than-DNA hypothesis was not supported. ₂₅ A 62-KO carbon list was enriched among heated-up DNA KOs in organic soil (5 of 57 versus 428 of 12,806; Fisher odds ratio 2.78, p=0.042), while individual pmoa, pmob, aceb glcb, and acea icl signals did not survive FDR across 14K KOs. ₂₅

Direct-sample sensitivity analyses that drop the incubated samples, and with them the incubation confound, refine this null rather than overturning it: treatment explained 12.7% of variance in the mineral-soil DNA pool (n = 14, F = 1.75, p = 0.081), 11.4% in the mineral-soil RNA pool (n = 11, F = 1.15, p = 0.254), and 10.0% in the organic-soil RNA pool (n = 14, F = 1.33, p = 0.190). ₂₅ DNA and RNA pools therefore showed comparable treatment R² of 10-13%, and the premise that the transcript pool is the more warming-sensitive readout was not supported in this dataset. ₂₅ The paired DNA-versus-RNA KEGG-Ortholog (KO) pool PERMANOVA and Procrustes alignment test used n = 25. ₂₅

The explanations offered for this convergence are hypotheses, not measurements. The report proposes that 25 years of heating is long enough for genome content to reflect warming-selected lineages, so that once the DNA-level community has equilibrated the RNA / DNA ratio of warming sensitivity collapses toward 1. ₂₅ It also notes that all samples were collected on a single date, 2017-05-24, so transcript abundances carry diurnal and microspatial noise unrelated to warming while the DNA pool integrates over years. ₂₅ A third proposal is compositional dilution: the metatranscriptome assembly recovered 14,302 distinct KOs versus 12,863 in the metagenome, and the richer rare-organism repertoire in RNA increases per-KO variance and reduces power under a relative-abundance framework. ₂₅ A fourth is substrate depletion feeding back to RNA as a global down-shift across functional categories rather than a redistribution, which would leave no preferred treatment axis in the RNA pool. ₂₅ The comparable-R² verdict accordingly does not rule out an early transient in which RNA leads DNA; it argues that on multi-decadal timescales the two pools’ warming responses converge in magnitude. ₂₅

Two measurement limits bound this DNA/RNA comparison. Metatranscriptome KO counts derived from contig annotations are transcript-pool composition, not TPM-quantified expression, and are biased by assembly quality. ₂₅ The analysis also had no quantitative metabolomics, natural organic matter, or proteomics layer, because the project excludes nmdc_arkin tables, even though equivalent quantitative layers exist there and could be added if scope expands. ₂₅

The Metagenome-Prioritized Phage Cocktails for Crohn’s Disease and IBD extends integration to patient-stratified therapeutic ecology by joining metagenomic ecotypes, pathway and biosynthetic-gene-cluster evidence, metabolites, strain adaptation, serology, endogenous viromics, and phage susceptibility. Across 8,489 curatedMetagenomicData MetaPhlAn3 samples, four ecotypes were selected; the 23-patient UC Davis cohort occupied E0, E1, and E3, while metagenomic calls agreed with clinical-covariate classifiers in only 41% (9/22) or 36% (8/22) of cases. ₂₆ A two-modality CCA (canonical correlation analysis) pilot on 106 paired HMP2 subjects produced r = 0.964, cliff CD-versus-nonIBD = +0.498, and MW p = 4e-4, but omitted pathway abundance and did not establish flux or causality. ₂₆ Six Tier-A pathobionts reached total score ≥ 2.5; iron/heme pathway enrichment among CD-up pathways was OR = 8.11 with FDR = 7.4e-6, while E. coli alone carried the iron-plus-genotoxin signature with 54 iron BGCs and 25 genotoxin BGCs. ₂₆ PhageFoundry contained 96 phages, 188 E. coli strains, and 17,672 tested susceptibility pairs; five phages covered 94.7% and eight covered 98.4% of tested strains, not UC Davis isolates. ₂₆ The pitfalls evidence contradicts treating the ecotype-associated differential list as stable without independent validation: held-out-species Jaccard values were 0.230 for E1 and 0.064 for E3, and independent within-substudy evidence reduced the Tier-A list from 33 candidates to 3. ₁₁ The discoveries digest supports the project’s multi-granularity design. The project produced two independent six-line cross-corroboration narratives from a single dataset, each showing the same biological claim across multiple analytical granularities. The iron-acquisition narrative ran from per-actionable MIBiG lookup to sample-level pathway × species attribution, cohort-level pathway-class enrichment, sample-level species × pathway co-variation, genomic BGC content, and AIEC literature. The bile-acid 7α-dehydroxylation narrative ran from within-carrier pathway differential abundance to subject-level metabolite differential abundance, a paired sample-level direct substrate–product signature, an informative strain-level null, and mechanism literature. ₂₇ The iron/heme result converged with E. coli Yersiniabactin/Enterobactin MIBiG matches, heme-biosynthesis attribution to E. coli, and AIEC iron-acquisition literature. This convergence spans analytical granularities within one dataset and is not replication in an independent cohort. ₂₇

The IBD project refines this cross-modality claim quantitatively. The dominant structure in HMP2 metagenomics plus metabolomics is a canonical correlation pair (r = 0.96) on which all 6 actionable Tier-A core pathobionts (H. hathewayi, M. gnavus, E. coli, E. lenta, F. plautii, E. bolteae) load CD-positive, ecotype-defining commensals (R. bromii, A. putredinis, L. eligens) load CD-negative, urobilin and secondary bile acids load CD-negative, and polyamines, long-chain PUFAs, fatty-acid amides, sphingolipids, N-acetylputrescine, and E. coli-specific cadaverine load CD-positive; a single composite axis stratifies CD from non-IBD subjects at cliff δ = +0.50 (p = 4×10⁻⁴). ₂₆ All four canonical correlations on the paired-subject set were very strong (r = 0.964, 0.928, 0.911, 0.889), so taxonomy and metabolomics share substantial cross-modality structure on that set. ₂₆ This remains a pilot construction, not a robustness demonstration: the paired set was 106 HMP2 CSM* subjects with 130 species at ≥10% prevalence (CLR-transformed) and 582 named metabolites in log10-intensity at ≥30% non-NaN coverage, each modality was first reduced by per-modality PCA to 30 components before CCA produced the 4 canonical pairs, and the report reads the fact that all 4 canonical correlations are very strong (>0.88) as suggesting that the per-modality PCAs are over-aligned, noting that a proper MOFA+ would impose sparsity to gate factor count; n=106 subjects is also modest for a 4-factor decomposition. ₂₆ Only the first axis discriminated disease groups: CC1 had canonical r 0.964, mean scores CD +0.235, UC +0.123, and nonIBD −0.593, Cliff statistic +0.498, and Mann–Whitney p 4e-4. ₂₆ CC2 (r 0.928, Cliff +0.092, p 0.51), CC3 (r 0.911, Cliff −0.146, p 0.30), and CC4 (r 0.889, Cliff −0.274, p 0.05) gave little or at best borderline CD-versus-nonIBD discrimination, so a high canonical correlation is not by itself evidence of clinical signal. ₂₆

The pilot’s modality set was fixed by data availability rather than by design. HMP2 pathway abundance is not in the mart — fact_pathway_abundance contains only CMD_IBD_PATHWAYS — so the planned three-modality MOFA+ on HMP2 taxonomy, pathways, and metabolomics fell back to two-modality joint factor analysis on HMP2 taxonomy plus HMP2 metabolomics, using sklearn CCA on per-modality PC scores because mofapy2 was unavailable. ₂₆ The same constraint is recorded twice: the two-modality design is taxonomy plus metabolomics with pathway not in the mart under the no-raw-reads scope, and HMP2 pathway data is not paired with HMP2 metabolomics. ₂₆ Adding pathways by reprocessing cMD_IBD on cohort-aligned subjects would require sample-level pairing that the available slices lack, since the cMD_IBD pathway slice exists but is not paired with HMP2 metabolomics; the cMD_IBD pathway slice is, however, paired with cMD_IBD species at sample level, so a three-modality decomposition on cMD_IBD subjects is proposed rather than performed. ₂₆ A second method limit refines what this pilot can show, and it is the report’s own methodological expectation rather than a tested equivalence, because mofapy2 was unavailable and MOFA+ was never run: the report asserts that CCA on PC scores captures the canonical-correlation signal MOFA+ would recover in the two-modality case, while noting that CCA cannot model modality-specific factors that load on only one modality, so that structure remains inside the per-modality PCA components. ₂₆

The CC1 loadings make that axis interpretable in both modalities. All six actionable Tier-A core species loaded positive, in the CD direction — M. gnavus +0.195, E. coli +0.173, F. plautii +0.153, H. hathewayi +0.144, E. lenta +0.109, E. bolteae +0.103 — along with oral-gut Tier-B candidates V. parvula +0.194, A. intestini +0.161, and V. atypica +0.151, while ecotype-defining commensals loaded negative (R. bromii −0.173, R. bicirculans −0.170, A. putredinis −0.169, plus L. eligens and B. intestinihominis). ₂₆ The report treats this as the pathobiont module structure independently rediscovered as a single principal direction in joint species-metabolite space, with two analytical approaches converging on the same biology. ₂₆ That rediscovery — reported as a joint multi-omics factor at CC1 canonical r=0.96 with CD-versus-nonIBD Cliff’s delta +0.50 — is the report’s basis for claiming that the per-modality analyses converge on the same biology rather than measuring orthogonal phenomena. ₂₆ Ecotype is implicit in the same loading structure: E1-Bact2-transitional and E3-Bacteroides-expanded species both load positive on CC1 while E0-commensal species load negative, so the axis may encode ecotype as well as diagnosis; regressing factor scores on ecotype plus diagnosis to test whether residual factor variance carries ecotype-specific biology remains future work. ₂₆

On the metabolite side of CC1, the CD-depleted direction carried urobilin (3 instances, max −0.143), glycodeoxycholate −0.113, lithocholate −0.080, and caproate −0.113, consistent with the project’s cross-cohort urobilin CD-down and bile-acid-pool depletion results. ₂₆ The CD-elevated direction carried linoleoyl ethanolamide (+0.110 and +0.106) and palmitoylethanolamide +0.108, which the report ties to the CD-up ebf/ecf BGC family (p<1e-31). ₂₆ It also carried sphingosine isomers (+0.109, +0.105) and 7-methylguanine +0.121. ₂₆ Long-chain PUFAs loaded positive as well — docosapentaenoate twice (+0.092, +0.091), adrenate +0.081, and arachidonate +0.048 — which the report ties to its long-chain-PUFA class result at OR=7.9 and to 75 % cross-cohort concordance. ₂₆ ADMA/SDMA loaded +0.099, which the report reads as a uremic-toxin marker connecting arginine catabolism to its TMA/choline results. ₂₆ Four polyamines loaded positive — N-acetylputrescine +0.105, diacetylspermine +0.091, cadaverine +0.086, and putrescine +0.084 — which the report connects to the polyamine metabolite-class OR=14.6 and to a cadaverine × E. coli correlation of ρ=+0.45. ₂₆

Subject-level metabolomics supports that axis from an independently analyzed chemical layer. Across 50 CD and 26 nonIBD HMP2 subjects, 52 of 579 named metabolites passed differential abundance, and two themes were Fisher-significant: polyamines (putrescine, N1-acetylspermine, N-acetylputrescine, anserine, diacetylspermine all CD-up) at OR=14.6 with FDR 0.008, and long-chain PUFAs (arachidonate, adrenate, DHA, DPA, EPA all CD-up) at OR=7.9 with FDR 0.009; tauro-α/β-muricholate and free taurine were CD-up, which the report reads as corroborating an F. plautii 7α-dehydroxylation deficit at metabolite level. ₂₆ A lipid-class result supports the PUFA theme through a different measurement rather than repeating it: 19 cholesteryl esters were CD-up, including the C20:4, C20:5 and C18:2 signals, which the report treats as the storage form of esterified PUFAs. ₂₆

These metabolite classes are placed in prior literature rather than used to derive mechanism. The report cites Franzosa et al. (2019) as HMP2 metabolomics plus microbiome integration showing a canonical bile-acid 7α-dehydroxylation deficit together with polyamine and lipid signatures, Wang et al. (2018) plasma polyamine signatures in inflammatory bowel disease as polyamine CD-biomarker confirmation, and Pegg (2014) on the toxicity of polyamines and their metabolic products for polyamine biology in IBD. ₂₆ It also cites Timsina et al. (2025) proteomic and metabolomic analysis of quaternary amine metabolism, but that citation carries no specific quantitative result into this project’s evidence base. ₂₆

The report proposes two cross-corroborated six-line mechanism narratives sitting inside this one axis as orthogonal molecular sub-mechanisms: iron acquisition centred on E. coli AIEC, and bile-acid 7α-dehydroxylation centred on F. plautii, E. lenta, and E. bolteae; both are interpretations of converging correlational streams rather than demonstrated mechanisms. ₂₆ The iron narrative stacks five evidence streams — per-actionable MIBiG lookup, pathway × pathobiont attribution (ρ=0.640 heme↔E. coli), cohort pathway-class enrichment (OR=8.1), sample-level co-variation (ρ=0.45 E. coli × iron pathway), and genomic BGC content (OR=44.4, driven by E. coli’s 54 iron BGCs) — which the report reads as converging on AIEC iron acquisition as a central CD-pathobiont specialization mechanism. ₂₆ The iron-siderophore BGC odds ratio of 44.4 is one of the largest enrichments in the project and complements rather than replicates the pathway-class iron OR of 8.1, because the two come from different evidence streams: genomic gene-cluster content versus pathway-cohort co-occurrence. ₂₆ The pathway stream itself was built on 52 CD-up MetaCyc pathways (FDR<0.10, |effect|>0.5) against a permutation-null mean of 0.077. ₂₆ Within that stream, 137 pathway–pathobiont attribution pairs passed |ρ_meta|>0.4, with heme biosynthesis (PWY-5920) ↔ E. coli at ρ=0.640 as the top hit, and the sample-level co-variation concentrated on E. coli at mean ρ=0.45 versus 0.13 for A. caccae. ₂₆ The report’s own synthesis describes both narratives as six-line and counts six independent evidence streams for the iron case, quoting its headline figures as OR=8.1 pathway and OR=44 BGC, and treats iron acquisition and bile-acid 7α-dehydroxylation as independent multi-line convergence stories within this one dataset. ₂₆

The bile-acid narrative similarly spans layers that are not interchangeable: within-carrier CD-up F420-dependent bile-acid metabolism pathways in F. plautii; subject-level CD-up tauro-α/β-muricholate and free taurine, consistent with reduced microbial 7α-dehydroxylation in CD; and a paired sample-level direct substrate-product signature in which F. plautii, E. lenta, and E. bolteae correlate negatively with primary tauro-bile acids and positively with secondary unconjugated bile acids. ₂₆ The report’s stated interpretation is that bile-acid pool sizes vary across cohorts due to dietary and antibiotic confounding, while the mechanistic substrate-product signature for active 7α-dehydroxylation is preserved at the within-paired-sample level. ₂₆ It therefore anchors the narrative on the paired sample-level evidence plus mechanism literature rather than on cohort-aggregate cross-cohort differential abundance, and states the general rule that cohort-aggregate cross-cohort differential abundance and within-cohort paired-sample correlation are different evidence streams, so one can be weaker without invalidating the other. ₂₆

A serology layer adds a weaker modality over the same species set, and its site-stratified breakdown is largely null. The layer comprised 48 (assay × species) tests across 67 HMP2 subjects with site-adjusted partial Pearson r, and no pair cleared the planned |r|>0.40 and FDR<0.10 threshold; the top pair was ANCA × M. gnavus at +0.31 with FDR 0.40. ₂₆ The leading pairs were ANCA × M. gnavus (partial r +0.31, raw ρ +0.24, FDR 0.40; Harvard r=+0.456, p=0.008, CCHMC r=−0.03 NS), OmpC × E. lenta (partial r +0.29, raw ρ +0.10, FDR 0.40; CCHMC r=−0.02 NS, Harvard r=+0.14 NS), ANCA × H. hathewayi (partial r +0.23, raw ρ +0.16, FDR 0.59; Harvard r=+0.345, p=0.050, CCHMC r=+0.02 NS), IgA-ASCA × E. coli (partial r +0.23, raw ρ +0.22, FDR 0.59; Harvard r=+0.30, p=0.09, CCHMC r=+0.06 NS), CBir1 × E. bolteae (partial r +0.22, raw ρ +0.22, FDR 0.59; CCHMC r=+0.381, p=0.035, Harvard r=+0.04 NS), and ANCA × F. plautii (partial r +0.22, raw ρ +0.32, FDR 0.59; CCHMC r=+0.443, p=0.013, Harvard r=+0.25 NS). ₂₆ The report reads all top pairs being positive as biologically plausible — pANCA, canonically UC-associated, co-elevated with multiple CD-pathobiont species; CBir1 as anti-bacterial-flagellin matching flagellated E. bolteae; and IgA-ASCA × E. coli as the most mucosal-immunity-relevant axis because IgA is the gut-immunity isotype — but these are interpretations of correlations that did not pass FDR correction and that are significant at only one site at a time, not demonstrated mechanisms. ₂₆

Phage evidence adds a fourth modality set over the same targets. The report describes three independent layers — curated literature, experimental susceptibility from PhageFoundry (96 phages × 188 E. coli strains × 17,672 pairs), and the in-vivo HMP2 phageome (630 viromics samples) — and reads them as converging on the same per-target classification: where one layer shows phage availability for a Tier-A pathobiont the others tend to show it, as for E. coli, and where one layer shows a gap, as for H. hathewayi and F. plautii, the others confirm the gap. ₂₆ That convergence is directional rather than quantitative: phageome × ecotype × diagnosis effect sizes were modest (max |ρ|≤0.18), the endogenous phageome is a noisier signal channel than species-level taxonomy, and the three layers were not matched on effect size. ₂₆ The project’s closing thesis stacks these layers onto the joint axis — CD at the gut-microbiome level as a single principal-direction phenomenon (CC1 r=0.96) within which 6 actionable Tier-A pathobionts and two cross-corroborated mechanism narratives define a state-dependent, hybrid-cocktail design framework with concrete per-patient drafts for 14 of 23 UC Davis CD patients — which remains a design framework assembled from correlational layers rather than a validated clinical result. ₂₆

The project’s figures document these layers. ₂₆

• NB07d_mofa_pilot.png — 3-panel: CC1×CC2 sample scatter colored by diagnosis (CD/UC/nonIBD; 106 subjects), CC1 top 15 species loadings (Tier-A core CD-positive vs commensal-CD-negative), CC1 top 15 metabolite loadings (urobilin CD-DOWN; polyamines + fatty-acid-amides + sphingosine + ADMA CD-UP). ₂₆
• NB09a_metabolomics_cd_vs_nonibd.png — 3-panel: volcano (cliff_delta vs -log10 FDR with key metabolites annotated), 11-theme Fisher OR bar (polyamines + PUFAs supported), top 15 CD-up and top 15 CD-down metabolites. ₂₆
• NB11_serology_pathobiont.png — 2-panel: site-adjusted partial Pearson r heatmap (8 species × 6 serology axes) and top 10 (assay × species) pairs scatter with FDR annotations. ₂₆
• NB04g_pathway_vs_taxon_ecotype.png — pathway-ecotype × taxon-ecotype cross-table heatmap with per-ecotype agreement bars. ₂₆

Evidence from databases, fitness, and orthogonal measurements

The ADP1 database contains 15 tables, 461,522 rows, and 135 MB for acinetobacter-baylyi-adp1 and 13 related genomes. ₁ Its genome_features table contains 5,852 genes and 51 annotation columns; TnSeq covers 58% of genes, FBA 15%, mutant growth on 8 carbon sources 39%, proteomics across 7 strains 41%, pangenome classification 54%, and functional annotations 34–55%. No gene has all six modalities. ₁ The report describes pairwise overlaps as substantial, particularly between essentiality, pangenome, and proteomics. However, FBA flux is the sparsest modality (15% of genes), which limited the FBA–TnSeq concordance analysis to 866 genes. ₁ The database links its 15 tables through 27 shared-column relationship pairs. Its core data flow runs from genome to genome_features (annotations, essentiality, proteomics, growth), then to pan_genome_features, genome_reactions, gene_phenotypes, and growth_phenotypes_detailed (FBA predictions). ₁ Protein abundance was measured across 7 strains for 2,383 genes. The report reads the high cross-strain correlation as evidence that the engineered modifications had targeted rather than global effects on the proteome. This is an interpretation, not a per-gene test. ₁ The deletion matrix spans 2,034 genes, excluding 499 essential genes and 316 incomplete-data genes. ₂

The atlas supports the same coverage-aware interpretation at larger scale: it contains 383,349 gene × metal records and 12,838 broad metal-important genes (3.3%), including 5,667 strict genes (1.5%), across 24 organisms with fitness matrices. ₆ It identified 1,182 ortholog groups conserved across at least two organisms, 601 across at least three, and 149 novel candidates with conserved metal fitness phenotypes; 89 were truly unknown, 43 had DUF/UPF domains, and 17 had partial functional hints. ₆ Module-level integration identified 600 metal-responsive module records among 19,453 module × metal-experiment records (3.1%); 183 responsive modules with conservation data had mean core fraction 0.826 and median 0.929. ₆

The ADP1 deletion analysis found 625 genes (31% of the complete matrix) with condition-specificity score ≥ 1.0 across 8 carbon sources; five PCA components captured 82% of variance, PC1 36.7% and PC2 12.7%, and median correlation across 28 condition pairs was r = 0.25 (maximum r = 0.58). ₂ The FW300 study supports this condition-specific interpretation at metabolite scale: 601 unique genes and 4,764 significant gene-condition hits occurred across 21 matched metabolites, 231 genes were significant in at least 3 conditions, and the top 18 in all 21. ₁₄ The respiratory-chain study further supports condition-specific integration by showing that carbon source can change the required respiratory configuration even when total reducing-equivalent yield does not provide a simple explanation. ₅

The triple-essentiality analysis supports proteomics as an independent continuous evidence axis: among 2,288 genes, essential genes had mean log2 expression 28.43 ± 2.94 versus 25.73 ± 2.96 for dispensable genes, a 2.70-log2-unit or 6.5-fold difference (Mann-Whitney p = 9.91×10⁻⁵⁹; ROC AUC = 0.743). ₃ The minimal-medium comparison contained 464 essential and 1,824 dispensable genes. Expression also correlated with essentiality (Pearson r = 0.345, p = 5.32×10⁻⁶⁵; Spearman ρ = 0.338, p = 3.28×10⁻⁶²), and minimal-medium log2 proteomics was the best overall continuous predictor in the comparison. ₃ The report treats proteomics and FBA as complementary measures in minimal media, with proteomics at AUC=0.74 and FBA at F1=0.67. ₃ Two caveats limit this evidence. First, the proteomics input was log2 expression averaged across 7 Acinetobacter strains, with locus tags parsed from complex feature IDs, so it is not an ADP1-only measurement. Second, the association was added in the refined analysis as an independent measurement axis and does not show that higher expression causes essentiality. ₃ The explorer’s 2,383-gene proteomics set and the 2,288-gene minimal-medium comparison are reported separately, and this page does not reconcile them. _1,3 Inverted fitness predicted knockout essentiality with AUC = 0.700 in rich and 0.725 in minimal media, whereas essentiality fraction yielded 0.344 and 0.403. ₃ RB-TnSeq versus knockout essentiality had recall = 7.9%, precision = 5.8%, F1 = 0.067, and Cohen’s kappa = -0.081 across 1,933 genes at the 0.05 threshold. ₃ The pitfalls evidence further refines this result: Fitness Browser essential genes are absent from transposon fitness records, genefitness-only analyses miss approximately 14.3% of protein-coding genes, and absence from genefitness supplies only an upper bound on essentiality. ₁₁ Respiratory-chain proteomics supports the same use of protein measurements as an orthogonal axis, but its near-equal dehydrogenase levels specifically argue that abundance alone does not identify the active respiratory route. ₅

The annotation-gap study supports orthogonal evidence at reaction-assignment level: 44 high-confidence pairs (21.9%) had BLAST, fitness, and pangenome support; 19 medium-confidence pairs (9.5%) had partial additional evidence; 33 low-confidence pairs (16.4%) rested on one stream; and 105 (52.2%) remained unresolved. ₈ Baseline FBA achieved 42.5% accuracy, recall 86.5% (244/282 growth-positive conditions), and precision 42.5% (244/574 predictions), while conditional gapfilling added 219 reactions and resolved 96/201 enzymatic reaction-organism pairs (47.8%) versus 70 (34.8%) for BLAST alone. ₈

The new Web of Microbes bridge supports the use of metabolite observations as an additional orthogonal layer but refines what that layer measures. WoM contains 589 metabolites, 332 (56.4%) unidentified, and 257 identified compounds; 69 (26.8%) had definitive ModelSEED name links, while 107 (41.6%) had formula-only candidate links and 81 (31.5%) were unmatched. Formula-only matching expanded 107 compounds to 900 ModelSEED molecules, so compound identity is often a candidate set rather than a definitive reaction assignment. ₁₅ The directly matched FW300 strain had 5,854 genes and 211 Fitness Browser experiments, while GW456-L13 had 5,243 genes and 106 experiments; these matches make metabolite-to-fitness analysis actionable, but the 2018 WoM snapshot’s small organism set and missing consumption measurements prevent a stronger producer–consumer or essentiality inference. ₁₅

The Genotype × Condition to Phenotype Prediction from ENIGMA Growth Curves refines this metabolite layer by making the integration method itself a variable. For 6 Web of Microbes strains and 105 metabolites, multivariate gradient-boosted decision-tree (GBDT) prediction failed, with AUC=0.500, whereas univariate per-metabolite point-biserial correlation between KO presence and metabolite production identified 940 strong associations (|r| > 0.7) — 454 production and 486 consumption — across all 62 variable metabolites. ₂₈ The feature set was 156 KOs with significant fitness effects on rich media in Pseudomonas Fitness Browser anchor organisms that are also variable across the 6 WoM strains, so Fitness Browser data supplied the filter that made a small exometabolomic panel analyzable. ₂₈ Because the panel is 6 strains, these are associations rather than independent causal validation. ₂₈

The same project refines the broader framework with a scale rule: cross-genus growth prediction (n=46K pairs) used multivariate GBDT to identify condition-specific features, while within-genus metabolite prediction (n=6 strains) required per-metabolite univariate correlation to recover genuine signal that the multivariate model missed. ₂₈ Growth-predictive KOs and metabolite-production-associated KOs were different feature sets (Spearman rho=0.043), answering different biological questions — “can it grow?” versus “what does it produce?” — and the report’s position is that gene content explains both when the method matches the sample size and biological resolution. ₂₈ Its synthesis restates the pairing directly: growth-predictive and metabolite-production KOs showed ρ ≈ 0.04, and at n=6 multivariate GBDT failed while per-metabolite univariate correlation recovered 940 mechanistic gene-metabolite associations covering all 62 variable metabolites. ₂₈

The formulation study supports orthogonal evidence through positive residuals after metabolic prediction: S. salivarius ASMA-737 (+74.1%), G. sanguinis ASMA-3044 (+62.2%), and N. mucosa ASMA-3643 (+57.2%) were candidate dual-mechanism inhibitors, although five-fold CV R² was only 0.145 ± 0.142. ₂₃ The Caulobacter study supports treating transcript–protein discordance as an integration result rather than forcing agreement; its outer-membrane proteome had one replicate per strain. ₂₄ Because that proteome supports no per-protein statistics, only directions can be interpreted for the lptC2 protein induction, Pal-Tol upregulation, LptD/LptE decline, and CCNA_01217 protein increase. All of these need replicated proteomics before they can be claimed at publication rigor. ₂₄ The NMDC analysis extends orthogonal evidence to community-level pathway potential and ambient chemistry rather than gene-level fitness. GapMind completeness was genomic potential, not expression, and the analysis used 174 analysis-ready merged samples; therefore its negative leucine and arginine associations support a community-scale ecological relationship but do not establish pathway activity or flux. ₁₆

Cross-dataset integration and limitations

The ADP1 project connected to kbase-ke-pangenome, kbase-msd-biochemistry, and kescience-fitnessbrowser. Four of five tested connection types exceeded 90% matching: 13/13 genome IDs, 1,210/1,330 reactions, 230/230 compounds, and 4,891/4,891 cluster IDs through mapping. ₁ The pangenome bridge linked all 4,891 KBase Data Lakehouse clusters to 4,081 unique ADP1 clusters and yielded a 100% gene-level match across 43,754 genes, although the direct string match was 0%. ₁ The atlas supports this infrastructure interpretation at observatory scale through 536 schema-level bridges, but only UC1 was sample-validated and genome_id can represent different identifiers across KBase, NCBI, and MAG pipelines. ₁₀ The pitfalls evidence refines the apparent high bridge coverage: short strain names are not globally unique, one example produced 8,434 erroneous genome matches including 1,751 clinical genomes, and 12 of 32 pangenome linkages through ncbi_strain_identifiers were incorrect genus matches; genus checks and assembly accessions such as GCF_* are therefore required. ₁₁

The FW300 crosswalk matched 28/58 WoM metabolites to Fitness Browser, 8/58 to BacDive P. fluorescens utilization, and 13/58 to GapMind; malate, arginine, and valine had four-way coverage. ₁₄ The Web of Microbes Explorer extends this result with a larger direct FW300 metabolite bridge: 19 WoM-produced metabolites were matched to Fitness Browser carbon or nitrogen tests, including lactate as an example of de novo production paired with a fitness-utilization phenotype. ₁₅ These observations are related but non-equivalent: WoM used R2A rich medium, Fitness Browser used minimal media, and BacDive aggregated species-level assays. ₁₄ WoM action semantics further refine this limitation because E means emergence from an initially absent medium, I means amplification of a present metabolite, and the snapshot contains no organism consumption action. ₁₅ The carbon census linked PubChem, KEGG, ModelSEED, Fitness Browser, ENIGMA genome-depot, GTDB, SSO, NMDC, and Planet Microbe, but 74/83 compounds remained organism-dark. ₁₃

The contamination bridge integrated a (108, 49) geochemistry matrix, 41,711 community rows, 212 communities, 1,392 genera, and 27,690 GTDB species rows; 862 genera were unmapped, while 380 mapped genera expanded to multiple clades. ₁₉ The SSO study found 65 of 1,038 genera functionally annotated, genus-level taxonomy covering 44% of sediment reads, and 56% outside genus-level inference. ₂₀ BacDive exact species-name matching accounted for 33,535 strains (34.5%), suffix removal added 8,692 (8.9%), and 55,107 (56.6%) remained unmatched. ₂₉ AlphaFold bridging reached 38,051,842 of 132,531,501 gene clusters (28.7%), while 70.7% lacked a bridge or AlphaFold entry. ₄

The atlas’s UC1 structural-fitness cohort contained 55,454 genes across 48 organisms with both fitness data and an AlphaFold model, including 22,303 distinct models; 6,635 genes were essential under min_fit ≤ −4. ₁₀ This supports structural–fitness integration but does not establish complete modality overlap. Similarly, deletion data were missing for 272 (10.5%) of 2,593 TnSeq-dispensable genes, and missing genes were less often pangenome-core than present genes (76.5% versus 93.3%; p = 1.4×10⁻²⁰). ₂

The Harvard Forest analysis used nmdc_metadata and nmdc_results for 42 biosamples, but abiotic_features was all zeros because of a parsing artifact; the +5°C label was therefore the only usable environmental contrast. ₂₅ The IBD study connected curatedMetagenomicData, HMP2, FRANZOSA, PhageFoundry, UC Davis profiles, and endogenous viromics, but HMP2-to-FRANZOSA bridging yielded only 9 strict replications, pooled m/z clustering had LOSO ARI = 0.000, and HMP2 viromics had an 80% Unknown fraction. ₂₆

The NMDC integration refines the interpretation of community-scale joins by showing both high bridge coverage and residual ambiguity: mean taxonomy-bridge coverage was 94.6%, 92% of samples mapped at least 85% of community abundance to GTDB pangenome species, and approximately 6.5% of mapped abundance involved genus-proxy-ambiguous taxa resolved by an alphabetical tiebreak. ₁₆ Its metabolite matching also exposed an isoleucine substring collision with leucine, while KEGG compound IDs had only a 2% annotation rate. ₁₆ The discoveries digest restates this finding from the same project rather than corroborating it independently. Only ~2% of compounds in nmdc_arkin.metabolomics_gold have KEGG compound IDs populated, so pathway matching must rely on compound-name substring patterns rather than KEGG-based ontologies. _27,16 The Web of Microbes result supports this identifier-quality warning: exact name links were high-confidence, but formula-only links generated an average of 8.4 ModelSEED molecules per WoM compound. ₁₅ The pitfalls evidence supports this caution: pangenome gene-cluster identifiers are species-specific, taxonomy joins must use genome_id, GapMind clade_name uses the full species-clade identifier, and annotation tables join at gene-cluster rather than gene level. ₁₁

The prophage NMDC bridge adds a related limitation: its 87.2% median matching coverage depended on indirect taxonomy-based prophage-burden inference and assumed genus-level conservation of prophage content, an assumption that may fail for recently acquired or lost prophages. ₁₈ Dedicated prophage detection was not used, the environmental categories were coarse, 35% of species were other_unknown, and only 28% of genomes had AlphaEarth embeddings; these limitations refine rather than overturn the broader claim that cross-modal joins can independently validate ecological associations. ₁₈

The respiratory-chain analysis further refines the interpretation of model-based integration. FBA predicted zero flux through NDH-2 and ACIAD3522 on all standard media because growth optimization preferentially routed NADH through ATP-favorable Complex I, even though gene-phenotype data indicated condition-specific requirements and ACIAD3522 had a 0.013 growth ratio on acetate. ₅ NDH-2 had no growth data, and ACIAD3522 may not be a respiratory NADH dehydrogenase in the strict sense; thus the integrated result identifies a testable model limitation rather than establishing alternative-pathway flux. ₅

Tensions and evidence hierarchy

The atlas contradicts a simple accessory-genome expectation for broad metal fitness: metal-important genes were 87.4% core, whereas a prior DvH condition-specific result found 71.2% core for heavy-metal genes. ₆ This tension is explained as a difference in scope rather than resolved: the atlas’s genome-wide definition captures general stress, envelope, DNA-repair, metabolism, and protein-quality functions, while condition-specific analyses more clearly expose specialized accessory resistance. ₆ The atlas also reports that excluding four duplicate P. fluorescens FW300 strains left the result robust (OR=2.065, p=5.9e-141 versus OR=2.083, p=4.3e-162), but rare-metal patterns remain dominated by one or two organisms. ₆

This distinction refines existing tensions between strong matched-set concordance and incomplete semantic coverage. The simple metal repertoire score failed to predict fitness, just as FBA class did not associate with quantitative growth-defect status among 478 genes (chi-squared = 0.93, p = 0.63; Kruskal-Wallis H = 1.67, p = 0.43). _6,3 Pangenome status also did not explain FBA discordance, with genes 93–100% core across discordance classes (OR = 0.89, p = 0.80). ₃ These findings support context-dependent essentiality and contradict inferring phenotype from presence alone.

The respiratory-chain result adds a tension between optimized model flux and measured condition-specific phenotype. FBA routed NADH through Complex I and predicted zero NDH-2 flux on all standard media, whereas Complex I fitness varied from a growth ratio of 0.37 on quinate to 1.44 on glucose, and the report proposed that capacity constraints can force use of suboptimal alternatives. ₅ This does not resolve whether NDH-2 compensates in vivo: NDH-2 had no growth data, the cross-species comparison retained only 5 of 14 organisms with validated NDH-2 and had only 4 organisms lacking it, with p = 0.52 for the opposite-direction difference. ₅ The tension therefore requires direct deletion, NADH/NAD⁺, flux, and ortholog analyses rather than preference for either the model or the phenotype.

The Web of Microbes bridge adds a related tension between metabolite production and utilization. WoM recorded organism production or amplification, whereas Fitness Browser tested growth on supplied substrates; the 19 FW300 matches therefore support a testable cross-modal association but do not show that the organism consumes the metabolites it produces. ₁₅ The absence of organism consumption data in this 2018 snapshot contrasts with the valid “decrease” action described for WoM in later or other versions, so the source version may underrepresent producer–consumer relationships. ₁₅ This tension is not resolved by the high FW300 concordance reported elsewhere because medium, action semantics, and assay direction differ. _14,15 The Genotype × Condition to Phenotype Prediction from ENIGMA Growth Curves refines this gap from the genotype side rather than closing it: per-metabolite correlation split its 940 associations into 454 production relationships (KO present → metabolite produced, for example K01048 PAPS synthase → taurine, K05710 thymidine phosphorylase → thymine) and 486 consumption relationships (KO present → metabolite consumed or degraded, for example K02613 lactate permease → lactate consumed, K07334 xanthine oxidase → hypoxanthine consumed), which the report calls mechanistically correct gene-function relationships; they are nonetheless inferred from KO presence across 6 strains, not measured consumption. ₂₈

The prophage study introduces a related tension between strong module-level environmental associations and weak independent lineage-level enrichment. Module composition varied with environment beyond phylogeny, including significant human-associated enrichment of tail (log2(OR)=2.21), head morphogenesis (1.98), and anti-defense (1.70), but no individual TerL lineage was significantly environment-enriched in 0/500 tests after FDR correction. ₁₈ These are not interchangeable evidence streams: module exchange and accessory functions may track environments even when complete prophage lineages do not. The result is consistent with the phage-defense-syndromes-and-arms-race interpretation, but the proposed human-associated arms-race mechanism remains a hypothesis because the analysis was annotation-based and observational. ₁₈

The eukaryotic-contamination study contradicts treating pooled environmental associations as transportable evidence: environment achieved random-CV R² = 0.35 but GroupKFold out-of-study R² = −0.30 and AUC = 0.56; within NEON, five-fold R² was +0.17 ± 0.06. ₂₂ The contamination study likewise found 8 confirmatory variants non-significant after FDR despite relaxed positive associations, and fraction-aware tests were non-significant (p = 0.767 and 0.898 relaxed; p = 0.780 and p = 0.793 strict). ₁₉

The lignin study adds a tension between strong bacterial and weak fungal reproducibility: 16S gave R²=0.979, p=0.001, whereas ITS gave R²=0.582, p=0.001; PERMDISP was significant for both (p=0.0004 and p=0.0001), and the ITS history effect was not detectable (R²=0.142, p=0.090). ₂₁ The Agora similarly shows rank-resolution tension: Cyanobacteriia × PSII was supported at class rank but stable at genus, family, and order; Mycobacteriaceae × mycolic acid had family-level producer d = +0.31 and consumer d = −0.19 but leaf_consistency 0.15. ₁₂

The FW300 data contradict equating production with catabolic capacity: tryptophan increased in WoM, had 231 significant Fitness Browser genes, and complete GapMind biosynthesis, while 0/50 P. fluorescens BacDive strains utilized it as carbon. ₁₄ The Harvard Forest data likewise contradict treating transcript shifts as flux: pmoA/pmoB and aceA/icl or aceB/glcB signals were nominal but none survived FDR across 14K KOs, and counts were contig-annotation counts rather than TPM-quantified expression. ₂₅

The Harvard Forest DNA/RNA comparison also sits in tension with the closest published precedent at the same site. Roy Chowdhury et al. 2021 reported that treatment had a larger effect on KEGG transcripts than on CAZymes, carbohydrate-active enzymes, with ~68% of differentially expressed CAZyme transcripts upregulated in heated soils, whereas this project found DNA and RNA responding comparably at 10-13% R²; the report notes the comparison is complicated by their different sampling timepoint, replication, and analytical pipeline, and that their CAZyme up-direction is consistent with its own chitinase up-result. ₂₅ The report further claims novelty for the paired DNA-pool versus RNA-pool warming-variance comparison at this site, including the sensitivity check indicating that the apparent DNA-versus-RNA imbalance is largely a horizon × incubation artifact, and for the local methanotrophy-up plus glyoxylate-up RNA signature, while acknowledging that both signatures have published precedents elsewhere. ₂₅

The IBD CCA result supports taxonomy–metabolite covariation but contradicts interpreting r = 0.964 as a complete causal mechanism because pathway abundance was absent, polyamines lacked an m/z bridge, and cross-cohort batch effects prevented metabolomic clustering. ₂₆ Its five-phage 94.7% coverage applies to tested PhageFoundry strains rather than UC Davis isolates and lacks in-vivo validation. ₂₆ The discoveries digest refines how pathway and metabolite layers should be compared. In Crohn’s disease, the 06_polyamine_urea pathway was CD-down at pathway level (OR=0.42), whereas polyamines as a metabolite class were CD-up (OR=14.6), the largest theme-level metabolite effect in the project. ₂₇ The digest treats both results as correct, because a metabolite pool reflects the difference between production and consumption rates plus dietary and host inputs. Pool and flux measurements are not interchangeable, and both belong in an analysis when available. ₂₇

The NMDC analysis adds a related evidence-boundary tension: its 11/13 negative pathway–metabolite directions and significant leucine and arginine associations support a weak community-scale BQH signal, but GapMind measures genomic potential rather than expression, all 33 Freshwater samples lacked paired metabolomics, and pH, temperature, and total organic carbon were unavailable as usable controls. ₁₆ Thus the result refines rather than overturns the distinction between covariation and mechanism: metatranscriptomics, broader paired metabolomics, and abiotic covariates are required to test whether the association reflects active synthesis and release rather than taxonomic or habitat confounding. ₁₆

The pitfalls evidence adds a methodological tension to the atlas’s schema-level bridge claim: the atlas reports 536 bridges and 51 of 66 projects spanning multiple tenants, whereas the pitfalls document shows that access, namespace migration, identifier collisions, synonymy, and value-space mismatch can make a syntactically valid join biologically wrong or incomplete. _10,11 This is not resolved by either count: live catalog discovery, genus and accession checks, NCBI-taxid-backed synonym reconciliation, and explicit bridge validation are required before interpreting cross-modal overlap. ₁₁

Open Directions

• Test metal fitness using concentration-relative-to-MIC normalization, phylogenetic independent contrasts, and genes important for metals but not other stresses; determine whether the 87.4% core enrichment persists after separating general stress from specialized resistance. ₆
• Profile the 149 novel metal candidates with PaperBLAST, InterPro, and structural prediction, then validate conserved candidates experimentally across metals and organisms. ₆
• Replace simple metal gene-presence scores with regulatory, expression, and module-activity models, using the 600 responsive module records and the UC1 structural-fitness cohort. ₆
• Execute UC2–UC5 on the live cluster, validate the UC1 cohort with per-residue pLDDT or structural features, and quantify value-space overlap for genome_id, ncbi_taxon_id, sample_id, and feature_id. ₁₀
• Add live catalog/schema/type discovery, explicit identifier audits, and value-overlap checks to every cross-tenant integration; test whether dotted Iceberg and fallback underscore namespaces return equivalent records before analysis. ₁₁
• Join the 866 genes with FBA flux predictions and TnSeq calls to test whether pathway enrichment distinguishes the 227 discordant from 639 concordant genes; add measured trace aromatic compounds and test the 9-of-11 aromatic-degradation discordant set. _1,3
• Use cross-strain proteomics and engineered ΔaroF, ΔaroG, and dgoA genotypes to test pathway-localized effects, and validate aciad3137 and aciad2176 with interaction or co-purification experiments. _1,9
• Test whether the 625 condition-specific genes replicate across additional carbon sources and technical replicates, and map FW300 fitness-important genes to GapMind biosynthetic, catabolic, regulatory, and housekeeping steps. _2,14
• Use the 19 FW300 WoM–Fitness Browser metabolite matches to test gene fitness against explicitly produced metabolites, separating E emergence from I amplification and adding direct consumption measurements; determine whether production-associated fitness effects persist across media and strains. ₁₅
• Build the WoM pathway-to-substrate/product lookup table required for GapMind matching, curate the 107 formula-only ModelSEED candidate sets, and test whether E/(E+I) metabolic novelty associates with pangenome openness or accessory genes. ₁₅
• Experimentally test the respiratory wiring model with an NDH-2 deletion, NADH/NAD⁺ measurements on quinate, glucose, acetate, lactate, and urea, direct flux measurements, and KO-based K03885/K00330–K00343 searches across the 27K-species comparison. Characterize ACIAD3522 separately because its respiratory annotation remains uncertain. Reanalyze the Stuani et al. (2014) quinate-versus-succinate proteomics for respiratory-chain proteins to test whether Complex I and NDH-2 are condition-specifically regulated, since the available abundances are standard-condition only. ₅
• Match BacDive GCA accessions directly to pangenome genome_id values, expand MIC and growth-inhibition extraction, and test whether genome-derived metal scores predict measured tolerance. ₂₉
• Experimentally validate the 44 high-confidence annotation-gap assignments, reconstruct the 38 false-negative cases with alternative gapfilling and curated media, and characterize the 50 EC-less dark reactions. ₈
• Repeat Caulobacter outer-membrane proteomics with biological replicates and targeted LptD, LptE, lptC2, Pal, Tol-Pal, and CCNA_01217 assays; perform iron-limitation, ferric-supplementation, and hemin RB-TnSeq experiments. ₂₄
• Reanalyze Harvard Forest with factorial treatment × horizon × incubation models, restore measured abiotic metadata, add direct organic DNA if available, and validate pmoA/pmoB, aceA/icl, and aceB/glcB with quantitative transcripts, proteomics, enzyme activity, and carbon-flux measurements. ₂₅
• Repeat the IBD CCA with pathway abundance and batch-aware metabolomics, test whether r = 0.964 survives held-out cohorts, and distinguish abundance, metabolite concentration, and pathway-flux contributions. ₂₆
• Reanalyze the IBD ecotype result with held-out-feature clustering, leave-one-species-out refitting, and within-substudy contrasts; determine whether the three independently supported candidates replicate. ₁₁
• Correlate NMDC proteomics and metabolomics layers with dark-gene carrier abundance and community pathway potential. First check the 346K and 3.1M observation counts against the live catalog, because the atlas attributes its counts to agent memory. _17,10
• Repeat the ADP1 proteomics–essentiality comparison with ADP1-only proteomics instead of 7-strain averages to test whether the 0.743 AUC persists. ₃
• Pair IBD polyamine pathway abundance with flux measurements to separate the production, consumption, and dietary/host inputs behind the OR=14.6 metabolite-class signal and the OR=0.42 pathway-level signal. ₂₇
• Reanalyze eukaryotic contamination with GroupKFold by study and within-study contrasts, adding extraction-kit, filtration, host-depletion, library-preparation, and measured-depth metadata. ₂₂
• Repeat the lignin experiment with n>=5 per group, intermediate passage points, DADA2 ASVs, matched technical metadata, UNITE-based ITS profiling, and functional measurements of beta-ketoadipate and protocatechuate pathways. ₂₁
• Test whether the Round-1 memory index of ~0.50 persists under controlled inoculum, passage number, and carbon-history designs, and compare community-history effects with organism-level fitness measurements. ₂₁
• Replicate the NMDC community BQH analysis across additional studies and paired Freshwater metabolomics, add pH, temperature, and total organic carbon to partial-correlation or mixed-effects models, and pair metatranscriptomics with metabolomics to test whether leucine and arginine associations reflect expressed pathway activity. ₁₆
• Improve KEGG compound annotation and replace string-based metabolite matching with identifier-based crosswalks, then test cysteine, histidine, lysine, glutamine, and proline pathways and resolve the 61 Unknown-ecosystem samples using NMDC ENVO annotations. ₁₆
• Reanalyze prophage ecology with dedicated prophage callers, genome-size-matched phylogenetic models, and independently validated TerL lineages to determine whether module-level environmental associations persist after separating intact prophages from domesticated remnants. ₁₈
• Test the prophage module–environment associations in NMDC samples with direct prophage detection, sample-level rather than genus-proxy burden estimates, and measured pH, temperature, depth, and nitrogen; determine whether the packaging–pH association is causal or habitat-confounded. ₁₈
• Compare structural, tail, and anti-defense modules with host defense inventories and experimentally measured phage susceptibility in human-associated and freshwater isolates to test the proposed arms-race interpretation. ₁₈
• Extend the Web of Microbes exometabolomic panel beyond the 6 strains at which multivariate GBDT returned AUC=0.500, and add direct consumption measurements, to test whether the 940 univariate KO-metabolite associations (454 production, 486 consumption) hold with more strains. ₂₈
• Sample Harvard Forest RNA across multiple dates and horizons to test whether the comparable 10-13% DNA and RNA treatment R² conceals an earlier transient, and replace contig-annotation KO counts with directly quantified transcript expression; separately, add the excluded nmdc_arkin quantitative metabolomics, natural organic matter, and proteomics layers, which supply chemical and protein measurements rather than RNA quantification. ₂₅
• Replace the CCA pilot with a sparsity-gated MOFA+ decomposition on more than the n=106 subjects, and refit without the 30-component per-modality PCA reduction suspected of over-aligning the modalities, to test whether CC1’s Cliff +0.498 separation persists while CC2-CC4 remain non-discriminating; test the CC1 polyamine and bile-acid loadings against paired flux or substrate-product measurements. ₂₆