Core Concepts

Grounding

General-purpose LLMs answer biology questions from statistical memory — plausible, often wrong, never checkable. Grounding means anchoring an agent's reasoning to a curated knowledge graph: the model retrieves real, vetted associations (microbe → disease, metabolite → pathway, drug → target) and cites the nodes and edges it used.

Graphomics is the layer that constructs and serves that grounding, rather than consuming someone else's. The knowledge graph is built by MapForge with a human review gate, so what you ground on has been vetted — not scraped or generated.

The knowledge graph

MicroMap is the first knowledge graph built with MapForge — a microbiome graph spanning taxa, diseases, metabolites/compounds, drugs, pathways, proteins, and papers, connected by typed relationships (e.g. ASSOCIATED_WITH_DISEASE, PRODUCES, TARGETS). It is served through a REST API and exposed to agents through MCP tools.

The graph is part of a federated composite (graphomics) that can route queries across discipline-specific constituents (microbiome, genomics, transcriptomics, metabolomics, proteomics) as they are populated. See the Schema reference.

Validated execution

Workbench turns analysis into reproducible pipelines assembled from validated nodes — 55 of them, covering data I/O, transforms, statistics, bioinformatics (16S, shotgun, diversity, differential expression), ecological modeling, and machine learning. A pipeline is a directed graph of nodes; it runs on an orchestration engine and produces versioned artifacts. You can also publish your own nodes (raw Python, a notebook, or a model script) into your organization's catalog.

Provenance and write-back

Every meaningful action produces a Decision, and every finding produces an Assertion — a confidence-scored, organization-scoped claim (subject → predicate → object) linked to the analysis and evidence that produced it. The chain is:

Experiment → Analysis → Assertion → (subject / object entities)
                     ↘ Decision (the action that produced it)

This is what "the loop closes" means: a pipeline result can be written back into the graph as an Assertion carrying the workflow id, statistics, and evidence references, then queried later via lineage. Assertions are versioned — a newer one can supersede an older one — so the graph carries a defensible, queryable history of how each claim came to be.

Organizations and isolation

Access is organization-scoped. Your credentials resolve to an organization_id, and reads, writes, and assertions are filtered to your organization. Workbench additionally enforces role-based access control and per-organization storage isolation. See Authentication and Workbench roles.