Overview Data & Annotations Search & Explore Batch Annotation AI Agent Release notes

Knowledge Graph

A biomedical network of genes, diseases, drugs, variants, and the relationships that connect them, assembled from published databases.

Engineering notes, not a contract

This page describes the current shape of the graph. Entity types, relationship types, and exact counts change with each data release. For stable contracts, use the documented graph API and follow the release notes.

What it is

A single connected structure where every node is a real biological entity and every edge is a claim from a published source. The point of the graph is that you can take a question that normally needs three databases, a spreadsheet, and a careful read of five papers, and answer it with one query.

Everything in the graph is read only. The graph is rebuilt from source on a release schedule.

What lives in it

The exact set of entity and relationship types depends on the current data release. The shape you can count on is a graph covering these domains.

Genes and proteins

Canonical gene and protein entities, their aliases, their domains, and their expression patterns.

Diseases and phenotypes

Ontology-backed diseases and phenotypes with hierarchy, so a query for a broad disease can include more specific forms.

Variants and regions

Common and disease-associated variants, plus regulatory regions that link to genes.

Drugs and targets

Drugs, their targets, known indications, and pharmacogenomic associations with variants.

Pathways and functions

Biological pathways, molecular functions, and cellular components, with membership links to genes.

Studies and evidence

GWAS studies, credible sets, and other provenance that keeps an edge traceable to where it came from.

Edges carry evidence

A relationship in this graph is not a flat boolean. Each edge carries as much of the following as its source provides.

Where it came from. The source database and the version of that database.
How strong it is. A score, a p value, an effect size, or a curated confidence label, whichever the source publishes.
What kind of claim it is. The edge type tells you whether the link is an association, a causal claim, a co-membership, or a simple containment.

Because every edge is scored and traceable, two callers can walk the same path and still disagree on whether to trust it, and both are right to do so. The graph gives you the evidence.

How queries work

There are three kinds of queries you can run against the graph.

Lookup

Give it an entity and get back a profile: the entity, its immediate neighbours, and the kinds of links that exist. This is what the web portal uses to render entity pages.

Traversal

Start from one or more seeds and walk outward, optionally filtering by edge type, direction, or score. Useful when a question looks like 'from this gene, find everything connected by a specific kind of evidence'.

Analysis

Ask a question about shape rather than about a specific entity. Enrichment, similarity, shortest path, or overlap between two sets. These queries use the graph as a substrate for statistics.

The AI agent uses all three kinds of queries under the hood. When you ask a research question in natural language, it picks the shape of query that fits.

What it powers

The same graph sits behind several surfaces of the platform.

Entity pages. Every gene, disease, drug, and variant page is a rendered slice of the graph or some part of it.
Search. Search uses the graph to return matches and their context.
AI agent. The agent runs scored traversals, so it reasons over structured evidence.
Variant lists. When you run enrichment on a variant list, the graph is what tells you which genes, diseases, and drugs are connected to the variants in your list.

What it is

Everything in the graph is read only. The graph is rebuilt from source on a release schedule.

Edges carry evidence

A relationship in this graph is not a flat boolean. Each edge carries as much of the following as its source provides.

Where it came from. The source database and the version of that database.

How strong it is. A score, a p value, an effect size, or a curated confidence label, whichever the source publishes.

What kind of claim it is. The edge type tells you whether the link is an association, a causal claim, a co-membership, or a simple containment.

Because every edge is scored and traceable, two callers can walk the same path and still disagree on whether to trust it, and both are right to do so. The graph gives you the evidence.

What it powers

The same graph sits behind several surfaces of the platform.

Entity pages. Every gene, disease, drug, and variant page is a rendered slice of the graph or some part of it.

Search. Search uses the graph to return matches and their context.

AI agent. The agent runs scored traversals, so it reasons over structured evidence.

Variant lists. When you run enrichment on a variant list, the graph is what tells you which genes, diseases, and drugs are connected to the variants in your list.