EmergentBiome Molecular Atlas

Executive summary

The atlas is a frozen, source-audited map of microbial molecular evidence. It registers 7,965 lane-scoped MAG/proteome records from a rumen reference set, freshwater wetlands and mangrove sediments. These are source records rather than globally dereplicated genomes or independent field samples. 7,710 carry an ESM-2 protein embedding, a gLM2 genomic-context embedding and a functional payload. The remaining 255 are documented source gaps retained as explicit rows.

Among the data-complete records, 5,209 carry pipeline-normalized screening events within their source lanes. The best MCycDB and SCycDB hits remain separate from accepted KOfam genes and METABOLIC events. Another 2,501 Old Woman Creek records carry the source's DRAM annotations; a subset also has processed expression detection. Tool, database and counting differences still require harmonization across lanes and annotation routes. The number eligible for a harmonized cross-lane methane mechanism score is 0.

Ecosphere Blue develops EmergentBiome as a molecular evidence and inference platform. MethaNet is its first wetland methane application. Today it supports source-linked evidence review and measurement priorities for a defined project question. A scoped paid assessment is a commercial hypothesis whose demand, price and repeat use remain to be tested. Measured methane flux, calibrated site risk, final A–E tiers and carbon-credit decisions require exact sample links, abundance, environmental covariates, uncertainty and paired field outcomes. This atlas release has no exact molecular-to-flux pairs.

The landing page's three curated evidence cases are a separately scoped September 26 projection of the ontology. They sit outside this August 10 atlas denominator.

Verified release snapshot

Each figure below is rendered from the machine-checked release ledger; the build stops if any count disagrees with it. Payload availability, shared-schema normalization, cross-lane mechanism comparability and field validation are separate rungs, so their counts differ by design.

Payload release state: ready for scoped diligence. Cross-lane mechanism comparability and field validation remain pending. Search indexing: off (noindex).

The EmergentBiome evidence graph

Each genome record sits at the center of an evidence model. Embeddings, functional annotations, genomic context, quality checks and validation requirements attach to it as separate, typed relationships. A neighbor in embedding space is therefore never mistaken for a measured function, and a missing measurement stays visible as a gap.

The queryable implementation covers a 662-record proof of concept. A formal ontology slice, added on 26 September 2026, holds selected evidence for 145 atlas records and is described with the evidence cases on the landing page. Neither yet spans the full atlas.

Evidence model for one genome record. Solid arrows: evidence present today, feeding an evidence card. Dashed amber arrows: the validation path still required, from exact sample links to abundance and environment, then to field or process measurement. The diagram shows structure, not causation.

Evidence integrity and current scope

Before release, the warehouse was checked against its source tables, per-genome outputs, embedding protocols, taxonomy fields and the Old Woman Creek expression tables. The table lists the findings that shape how the atlas can be used.

Evidence matures through payload availability, signal within a protocol, harmonized comparisons across source lanes and annotation routes, and paired sample-level ecological validation. Detailed audit tables stay in the internal warehouse; this report shows the resulting evidence states and the validation agenda.

The tri-view evidence contract

A tri-view record carries an ESM-2 embedding, a gLM2 embedding and a functional annotation. Its evidence state records whether those payloads support a common quantitative interpretation. That state travels with the record into the freeze manifest, the candidate cards and the release checks.

Static version of this chart

gLM2 runs under two protocols and is compared only within a protocol: 5,209 records use one native and one shuffled gene-order window, and 2,508 Old Woman Creek records use ten of each. ESM-2 uses one 650-million-parameter model with a cap of 6,000 proteins per genome; the 121 capped records are flagged.

Source provenance and environmental readiness

Where each source comes from, how precisely its genomes can be placed in a sample today, and the next link needed. Together these gaps set a concrete agenda for abundance mapping, environmental context and field validation.

The proof-of-concept crosswalk covers a 662-proteome cohort; the atlas reference core holds 625 of those genomes (518 rumen, 107 wetland) alongside the registered mangrove and Old Woman Creek records. Pending, source-gap, mixed-resolution and unlinked rows stay visible as explicit states.

Three molecular views in one evidence graph

Each view answers a different question, has its own comparison set and carries its own validation gaps. A mechanism claim needs convergent evidence from the right view under a compatible protocol; nearness on a map is only a starting point.

For every record the report shows what evidence exists, its protocol, what each functional count measures and the claim the record supports. A cross-lane methane mechanism score requires a validated common feature table, locked tool and database versions, and source-aware controls.

ESM-2 geometry with measured limitations

The ESM-2 representation places 7,710 records in a 1,280-dimensional space. The September 29 geometry release recomputed the retained 662-record pilot with final-layer 33 pooling to match the later source lanes. All 662 passed input-identity, protein-count, finite-vector and reproduction checks. In a separate representation control, 23 selected same-DNA records ranked first among 2,501 Old Woman Creek targets. These checks establish configuration consistency for navigation; ecological transfer and flux prediction still need independent tests.

Raw cosine similarity in this space is strongly anisotropic: two random records have a mean cosine of 0.9930 (median 0.9942), and the median similarity to the global centroid is 0.9970. Cross-habitat neighbor edges therefore sit in a saturated range, with a median raw cosine of 0.999471. The atlas uses this geometry to navigate neighborhoods and keeps functional and validation evidence separate.

A stricter test counts mutual neighbors: pairs in which each record is among the other's 35 closest across the full atlas. Raw space holds 15,626 mutual mangrove–wetland pairs, 111 rumen–wetland pairs and 367 rumen–mangrove pairs. After each dimension is standardized, 13,873 mangrove–wetland pairs remain and 486 mutual pairs involving a rumen genome remain.

A separate, one-way comparison asks which member of the 625-genome reference core (518 rumen, 107 wetland) is closest to each record. Outside the core, 1,403 of 2,501 wetland and 2,823 of 4,584 mangrove records point to a rumen genome, as do 10 of the 26 selected candidates outside the core. The median closest-match similarity outside the core is 0.997; the random-pair atlas median is 0.994. These values describe representation geometry, not a calibrated measure of biological equivalence. The core is mostly rumen, so a rumen match nominates a record for review; it does not establish shared biology, transfer between sources or methane flux.

Taxonomy explains part of the mangrove–wetland continuity. Among mutual pairs with usable phylum labels, 60.4% match exactly and 79.5% match after conservative synonym normalization. GTDB release metadata exists only for the reference core, so source and taxonomy-release effects are confounded. Harmonized taxonomy and phylogeny-aware null models are needed before neighborhood enrichment can be read as functional convergence.

UMAP is the default map view because it keeps local neighborhoods readable. Diffusion, t-SNE and PCA are offered for comparison; source structure and projection distortion must be assessed in the reconciled release. No projection is evidence on its own, and link membership is always computed in the full representation.

References

1. Lin et al. 2023, Science: the ESM-2 protein language model
2. Scientific Reports 2025: medium-sized protein language models in transfer learning
3. Coifman and Lafon 2006: diffusion maps
4. McInnes, Healy and Melville 2018: UMAP
5. van der Maaten and Hinton 2008: t-SNE
6. Communications Biology 2022: evaluating dimension-reduction methods
7. Old Woman Creek wetland microbiome study, mSystems

Molecular niche-space map

Every record with an ESM-2 embedding appears on the map, colored by source lane. Positions come from a two-dimensional projection; links come from the full 1,280-dimensional representation, so switching projections moves points but never changes which links exist.

Gold lines join the 26 selected candidates outside the reference core to their closest reference-core genome. Teal lines are mutual neighbor pairs and gray lines one-way neighbor pairs; the map draws the 2,200 most similar of the 59,496 directed cross-habitat neighbor edges. All of them join mangrove and wetland records, and 99.8% are mutual. Gold rings mark the 36 case-study candidates: select one to read its evidence card, or hover over any point.

Static version of this map

Candidate evidence cards

The candidate layer asks which evidence exists for each review hypothesis, and which comparisons that evidence can support. Reference-core cards (P) are reselected from the corrected cross-habitat neighbor fraction, QC and stable record identity. Mangrove cards (M) are ranked by embedding geometry and QC. Old Woman Creek cards (O) carry the source's annotations and, where present, processed expression detection.

The matrix and the wheel show availability and eligibility rather than strength. They separate ESM-2, gLM2, functional payload, cross-lane mechanism readiness, expression detection, QC, taxonomy and sample context. Mechanism strength, activity and any causal link to flux need direct evidence.

Static version of the matrix

Functional metric harmonization

A methane-marker density requires the same accepted gene definitions and denominators across sources. Raw annotation-hit rows do not meet that test because one gene can produce several hits, and the number of hits per protein differs by lane and tool. The pipeline-normalized lanes record accepted KOfam genes and METABOLIC events as explicit events and keep the best MCycDB and SCycDB hits separate. MCycDB hits await a validated family map before entering a methane score. Old Woman Creek retains source annotations under a separate contract. Version, database and feature definitions remain to be reconciled across all lanes.

The raw-row columns are diagnostics that show why raw hits cannot be compared. A combined index built on them is quarantined and plays no part in any ranking here. Its methane component tracked the index closely (Pearson r = 0.730), and all of its top 500 records were mangrove, a pattern consistent with counting differences between sources rather than biology. The event tables support screening and audit within their lanes. A quantitative cross-lane ranking requires a separate validation gate.

Old Woman Creek adds expression evidence and a field-validation lane

The Old Woman Creek wetland lane (MUCC v1) adds processed metatranscriptome detection across 133 source sample columns. Expression support exists for 1,948 MAGs: 1,358 have at least one processed methane-associated expressed-gene row and 1,868 have sulfur-associated rows. These are detection signals from deposited processed tables; comparing activity levels needs expression normalization first.

The warehouse also stages 275 chamber-flux rows (188 valid in the source), 5,280 porewater rows (1,563 valid) and 29,280 half-hourly, gap-filled tower-flux rows. It includes 694 exploratory FlashWeave associations, 126 of which pass the current stability filter, and 3 descriptive WGCNA modules (excluding the unassigned grey module).

The source study includes paired geochemical context, including porewater methane concentration. In this atlas release, 0 of 133 sequencing sample contexts have an authoritative exact sample-depth-environment-flux join in the integrated warehouse, and 133 remain blocked for our ecological validation. Concentration is distinct from a paired chamber-flux outcome. Until an exact flux join is established, the staged flux records provide site and time context and cannot be attributed to a specific MAG, expression signature or network edge.

Sample-linkage readiness

This chart groups mangrove records at the most precise environmental context available today. Futian records are grouped by site and month, each context spanning several depth samples with chemistry metadata. MSM records are grouped by the source's sample groups and BioSample sets. No record is yet assigned to a single physical sample.

These groups show a field team where abundance mapping, metadata reconciliation and validation measurements would pay off first. Sample-level methane-risk estimates become possible only after per-genome abundance, exact sample assignment, environmental conditions and flux or process validation are in place.

From molecular evidence to environmental readiness

The atlas advances toward sample-level decisions when validated genome features roll up to physical samples, sites and monitoring periods. Today the reference core and both mangrove sources carry pipeline-normalized screening events, while Old Woman Creek carries source annotations and expression detection. Cross-lane quantitative mechanism comparability remains pending. A defensible sample-level score needs four gated layers.

Field work is how the atlas learns. Sampling across habitats, restoration stages, salinity gradients, depths and seasons would widen the map, expose source-specific blind spots and test candidate signatures under blue-carbon conditions, provided each sample arrives with clean provenance, abundance, environmental measurements and a validation target.

The next operational output is a sample-readiness layer. Once samples are mapped, each would be labeled scoreable, monitor more, needs metadata, needs abundance, needs environmental covariates or needs flux validation. Current records do not support calibrated sample-risk scores; the labels would guide sampling while the evidence grows.

How the atlas enters a flux-model test

The proposed first field study would follow 36 plots across two microsites and two seasons, producing 144 repeated sample-events at one site. Plots are the replicate unit; seasonal revisits test transfer at the same points. Site access, funding, permits, sampling and paired flux outcomes are still required. The design tests whether molecular evidence adds predictive value beyond measured environmental conditions.

The ontology fixes the meaning of genes, pathways, samples and measurements. The evidence graph records feature lineage, source, QC, identity confidence and uncertainty before a model uses them, then returns model outputs to reviewable evidence cards. Grouped plot folds and a prespecified dry-season holdout would test this single-site design. Independent sites are needed to test geographic transfer. No predictive gain or calibrated risk is reported from the current atlas.

What comes next

The atlas rests on a source-audited warehouse of 7,965 registered genome records across four source lanes. It already supports payload auditing, neighborhood exploration, protocol-aware candidate review, expression-detection queries, metadata-gap priorities and validation-study design.

Its evidence states are explicit. 5,209 records carry pipeline-normalized screening events, 2,501 carry Old Woman Creek source annotations, and 0 pass the cross-lane mechanism-comparability gate. Keeping those states separate protects decisions from pipeline artifacts.

The next build should deliver one validated mechanism-feature table across sources, harmonized taxonomy with phylogeny-aware null models, reconciled gLM2 protocols, exact sample and abundance mappings, and paired field or process outcomes with uncertainty. Those steps would permit a quantitative cross-lane mechanism test and a calibrated sample-risk model.