You open a newly published paper. The results look promising — and then you reach Figure 3: a tangle of coloured dots joined by lines of different thickness, arrows pointing every which way. The caption reads, “Knowledge graph of gene–phenotype–disease associations.” You read the rest of the paper without difficulty. This one figure stops you cold for a few minutes, especially if you’re new to the lab or early in your career.
That graph you couldn’t read can give you all the information scattered across a domain and its neighbouring fields — but it asks for a different kind of literacy. It’s the same literacy you once needed before a cladogram, a metabolic pathway, a volcano plot, a heatmap, or a dot plot made sense at a glance. Knowledge-graph figures now show up routinely in computational biology, ecology, drug discovery, and biodiversity science, and that frequency will only grow as more fields adopt graph-structured databases. Once you understand its grammar, even the most complex-looking network becomes far more approachable.
This article won’t teach you how to build one. It will teach you how to read one — so that when a knowledge-graph figure appears, you know what it is actually claiming, what it is quietly hiding, and which questions to ask before you trust it.
What Is a Knowledge Graph?
At its simplest, a graph has only two ingredients:
- Things
- Relationships between those things
In graph terminology, the things are called nodes and the relationships are called edges.
You’ve read graphs your whole career under other names. A phylogenetic tree is a graph: taxa are nodes, branches are edges. A metabolic pathway diagram is a graph: metabolites and enzymes are nodes, biochemical reactions are edges. A food web is a graph: species are nodes, feeding relationships are edges.
What makes a knowledge graph different is that it adds structurally richer meaning to each connection. Every node belongs to a specific category — perhaps a gene, disease, protein, drug, species, habitat, or phenotype. And every edge is labelled to describe the exact relationship between two entities. Instead of merely showing that two things are connected, a knowledge graph tells you what each thing is and what kind of relationship joins them — “treats,” “inhibits,” “observed in.” For example:
- Aspirin treats headache.
- TP53 interacts with MDM2.
- A plant species occurs in a protected forest.
- A mutation causes a disease.
The underlying information isn’t new — but its structure, and what you can query out of it, is transformed. (If you want the bigger picture of why this matters for research, see what a scientific knowledge graph is.)
Anatomy of a Knowledge Graph
Here are the core components you need to recognise:
- Node — any entity the graph tracks: a gene, a drug, a species, a phenotype. Think of it as a taxon in a phylogenetic tree.
- Edge — a relationship between two nodes. Think of it as an arrow in a metabolic pathway.
- Directed edge — a relationship with a specific direction, where A → B is not the same claim as B → A. Think of a kinase phosphorylating a substrate: the reverse statement is false.
- Edge label (or edge type) — the name of the relationship, not merely its presence. Labelling an arrow “phosphorylates” rather than leaving it as an unlabelled line is the difference between a claim and a doodle.
- Hub node — a node with an unusually large number of connections, analogous to a keystone species in a food web or a master-regulator gene in a pathway.
- Subgraph — a selected portion of a larger graph, analogous to a single clade pulled out of a full phylogenetic tree.
Every knowledge-graph figure you’ll ever encounter is built from these six pieces, arranged differently.
Reading Edges: Direction, Type, and Provenance
Once you can name the parts, the real interpretive work is in the edges.
Directed versus undirected is the single most consequential distinction on the page. An arrowhead means the relationship runs one way. In a signalling network, A → B → C reads as a cascade; C → B → A reads as feedback. Same nodes, same edges, opposite biology — the only difference is which end the arrowhead sits on. Check for it before you interpret anything else.
Edge type carries the meaning, not the line itself. The most common misreading of a knowledge graph is treating every edge as equivalent because every edge is drawn the same way. “Associated with” and “causes” can occupy visually identical lines yet represent entirely different classes of claim. In PrimeKG, for instance, a drug–disease edge can mean an approved indication, an off-label use, or a contraindication — three very different clinical statements rendered as three indistinguishable lines.
Edge provenance — where the edge came from — is one of the layers most figures don’t show you at all. An edge might be curated from primary literature, predicted by a machine-learning model trained on the rest of the graph, or inferred from statistical co-occurrence in text. Link-prediction methods rank pairs of nodes by how closely they resemble known true relationships in a learned mathematical space; if a figure or paper mentions a “predicted interaction” or a “link-prediction score,” treat it as a candidate worth checking, not a finding.
These are three different epistemic statuses that look identical once drawn. If the methods section doesn’t specify provenance, the safest default is to treat every edge as a claim of unknown reliability rather than an established fact.
What Visualisation Choices Actually Encode
Visual attributes in a knowledge-graph figure are conventions, not universal facts. They vary by paper, by software, and by author — so always check the legend, and never assume a convention holds outside the figure in front of you.
| Visual element | What it usually means | What to watch out for |
|---|---|---|
| Node size | Degree (connection count) or an importance score | May be arbitrary — check the legend before assuming meaning |
| Node colour | Node type or ontology category | Palettes are often colour-blind-unfriendly; read the legend |
| Edge thickness | Relationship strength, confidence, or evidence count | Reflects data volume, not necessarily reliability |
| Edge colour | Relationship type or polarity (inhibition vs. activation) | Conventions differ between papers — never assume on your own |
| Arrow direction | The direction the relationship runs, source to target | Bidirectional arrows can mean symmetry, or simply that direction was unknown |
| Node position | Output of a layout algorithm, usually not meaningful | Proximity does not imply biological or conceptual closeness |
| Cluster / group | A densely connected region, often a functional module | Frequently an artefact of the layout, not discovered structure |
The layout-algorithm problem deserves special attention, because it’s the most persistent source of misreading. Force-directed layouts — Fruchterman–Reingold, ForceAtlas2, and their relatives — place connected nodes near one another purely to minimise visual clutter. Two species sitting adjacent in a GBIF occurrence graph are not thereby shown to be phylogenetically related; their proximity is a rendering decision, not a biological finding.
Equally important is what the figure has already removed before you ever see it. Every visualisation is a filtered subset of something larger. A diagram showing fifty nodes may be drawn from a graph containing 250,000 nodes and four million edges, as in PrimeKG. Entities are typically filtered out because they fall below some confidence threshold — but their absence from the picture says nothing about their absence from the underlying biology.
Where to Be Careful: Four Interpretive Checks
Completeness is implied but never guaranteed. A node with many connections looks important, but connectivity is a function of what the graph happened to include. The same drug can show forty connections in one knowledge graph and five in another, depending entirely on which source databases were incorporated during construction.
Correlation gets sold as connection. Many knowledge graphs draw edges from text mining — genes and diseases that are simply co-mentioned across abstracts. That is co-occurrence, not a curated biological relationship, and a pair of entities can accumulate an edge for no reason beyond appearing together frequently in unrelated contexts.
Absence of an edge is not absence of a relationship. A missing connection may mean the relationship hasn’t been studied, hasn’t been published, hasn’t yet been incorporated into the database, or simply falls outside the graph’s declared scope. A knowledge graph encodes what is known and curated — not what is true.
Hub bias distorts visual importance. Highly connected nodes draw the eye, but connectivity reflects curatorial and database choices as much as biology. TP53 carries thousands of connections in biomedical knowledge graphs because it’s the most heavily studied gene in cancer biology — not necessarily because it’s the most relevant node for your specific question. The node that matters most to you may be a quiet, low-degree one sitting at the edge of the figure.
Eight Questions to Ask Any Knowledge Graph
These are the basic questions to ask yourself whenever you study a knowledge graph:
| Question | Why it matters |
|---|---|
| 1. What do the nodes represent? | Identify whether the nodes are genes, proteins, drugs, diseases, species, or multiple entity types. Knowing the entities is the first step to reading the graph correctly. |
| 2. What do the edges mean? | Not all connections represent the same relationship. Check what an edge actually represents — treats, causes, interacts with, associated with, or something else. |
| 3. Are the edges directed? | Arrowheads indicate direction. In biological pathways and regulatory networks, direction can completely change the interpretation. |
| 4. What do node size, colour, and edge thickness represent? | Visual properties often encode node type, degree, confidence, or interaction strength. Never assume their meaning — always read the legend. |
| 5. What has been filtered out? | Most published figures show only a small subgraph of a much larger network. Missing nodes or edges may simply have been excluded for clarity or confidence thresholds. |
| 6. Where did the relationships come from? | Determine whether edges are curated from literature, experimentally validated, computationally predicted, or inferred from text mining. The source sets how much confidence to place in them. |
| 7. Does the position of the nodes mean anything? | Usually not. Most layouts are generated by algorithms that improve readability rather than represent biological or conceptual distance. |
| 8. How large is the complete graph? | A published figure may represent a tiny fraction of the full knowledge graph. Knowing the overall scale gives context for interpreting the visualisation. |
The Graph Hasn’t Changed — Your Ability Has
Return to Figure 3. The same tangle of coloured dots and weighted lines sits on the page. But now you check the legend for node types before assuming what you’re looking at. You look for arrowheads before assigning direction. You read the edge labels instead of treating every line as equivalent. You ask what was filtered out, and you remember that the layout is telling you nothing about biological distance.
The figure hasn’t changed. Your ability to read it has.
Knowledge graphs aren’t a new visual language invented to intimidate scientists without a computational background. They’re a specific, learnable grammar — nodes, edges, direction, type, provenance — and you’ve just learned the alphabet. The next time a graph figure appears in a paper you’re reviewing, try reading your own field’s data as a graph with Axy and see how much of it you can now interpret on sight. If you’d like to help build the shared map of science rather than just read it, apply to join below.
