Skip to content

Graphs, Biology, and Knowledge Graphs

…and what a KG does for biot

Friday intro talk


Agenda

  1. What is a graph? (5 min)
  2. Why biology is secretly all graphs (5 min)
  3. Knowledge graphs: adding meaning to edges (10 min)
  4. A KG for biot: what we built and what it enables (10 min)

Part 1 — What is a graph?


The simplest useful data structure

A graph is just things and relationships:

  • Nodes (vertices): the things — a protein, a paper, a person
  • Edges: the relationships — binds, cites, knows
  (KRAS) ──inhibited_by──> (sotorasib)
     │
  mutated_in
     │
     ▼
  (lung adenocarcinoma)

That's it. Everything else is detail.


Why not just use tables?

Tables answer: "what are the properties of X?" Graphs answer: "what is X connected to, and through what?"

Question SQL Graph
Properties of protein X easy easy
Direct interactions of X one JOIN one hop
Everything within 3 hops of X 3 nested JOINs, painful traverse(X, depth=3)
Shortest path X → Y don't built-in

The deeper the relationship chain, the more a graph wins.


A few concepts you'll hear today

  • Directed vs undirected — A inhibits B has a direction
  • Labeled/typed — nodes and edges have types (Protein, Disease; treats, causes)
  • Path — a chain of hops: gene → protein → pathway → disease
  • Neighborhood — everything N hops from a node
  • Community — a cluster of nodes more connected to each other than to the rest
  • Hub — a node with many edges (IL-6, TNF-α… the usual suspects)

Part 2 — Graphs in biology


Biology is natively graph-shaped

Biological system Nodes Edges
Protein–protein interaction (STRING) proteins binds / complexes
Metabolic pathways (KEGG, Reactome) metabolites, enzymes reactions
Gene regulation TFs, genes activates / represses
Disease genetics variants, genes, diseases associated_with
Microbiome taxa, metabolites produces / consumes
Drug action compounds, targets inhibits / agonizes

None of these fit cleanly in a table. All of them are literally drawn as graphs in every textbook.


The payoff: inference by connection

Classic examples of graph reasoning in biology:

  • Guilt by association — unknown protein interacts with 5 DNA-repair proteins → probably DNA repair
  • Drug repurposing — drug D targets protein P; P sits in pathway of disease X → candidate for X
  • Pathway gaps — two metabolites, no known enzyme between them → there's probably an enzyme to find

The common pattern: missing edges are hypotheses.


Part 3 — Knowledge graphs


A knowledge graph = graph + meaning

A KG is a graph where every edge is a typed, sourced claim:

(Mucin-1) ─[upregulated_by, source: Smith 2021, confidence: 0.9]→ (E. coli challenge)

Three things distinguish a KG from "just a graph":

  1. Ontology — controlled types (Protein, Disease, Chemical…)
  2. Provenance — every edge cites where it came from
  3. Entity resolution — "IL-6", "interleukin-6", "IL6" → one node

Where do the edges come from?

Curated databases: UniProt, MONDO, PubChem, Reactome — high precision, slow, generic.

Literature extraction (what we do): LLM reads papers → extracts (entity) ─[relation]─> (entity) triples → resolved and merged into the graph.

One paper ≈ dozens of structured claims. One corpus ≈ a queryable map of a research field.


Entity resolution — the hard part

Same concept, many names. If you don't merge them, your graph is confetti.

Our approach (battle-tested on real corpora):

  • Normalize names → norm:<name> canonical keys (primary connector)
  • Public ontologies (UniProt, MONDO) as supplement — only ~42% coverage on ag-bio text
  • Name-agreement guard (fuzzy match ≥ 0.85) — a wrong merge is worse than a missed one

Real catch: resolver wanted to merge Mucin-1 into pleckstrin. Guard rejected it.


Communities: structure for free

Run Leiden clustering on the graph → densely connected clusters emerge.

In practice these clusters are topics: a mastitis-immunity cluster, a gut-microbiome cluster — nobody labeled them, the edge structure did.

Used for: summarization ("what's in this graph?"), search routing, and spotting which topics two programs share.


Part 4 — A KG for biot


What biot has today

  • Research programs, each producing papers, experiments, data
  • Literature search with embeddings (pgvector)
  • biot kg CLI — thin read client over the graph

The problem KGs solve for biot: knowledge is siloed per program. Program A's findings never meet Program B's — except inside someone's head.


What we built (real numbers)

Per-program KGs from literature, then merged into one connected graph:

Programs merged 5
Nodes 16,304
Edges 52,857
Entities shared across ≥2 programs 5,171
Entities in all 5 programs 23

Those 5,171 shared entities are the bridges — where one program's knowledge flows into another's.


Gap inference: the killer demo

Question: what does the combined graph know that one program alone doesn't?

IL-6, viewed from one program alone: 2 neighbors. IL-6 in the connected 5-program graph: 20 neighbors.

→ 18 relationships relevant to your program that your program's own literature never surfaced. Each one is a lead: a mechanism, a biomarker, an intervention someone else already studied.

1-v-1 becomes 5-v-5.


How it plugs into biot

papers ──> graffold-ingest ──> Parquet (source of truth)
                                   │
              ┌────────────────────┴──────────────┐
              ▼                                   ▼
        pgGraph (Postgres)                 Spanner Graph (GCP)
        self-hosted serving               managed serving
              ▲                                   ▲
              └──────── graffold-api / biot kg ───┘

embeddings ──> pgvector (biot literature search, joins on entity id)
  • Graph DBs are swappable projections — one env flip between them
  • Vector search stays in pgvector; graph carries entity IDs for joins

What this enables for biot users

  • "What connects X and Y?" — path queries across all programs
  • "What do we already know about Z?" — one query instead of five literature reviews
  • Cross-program leads — gap inference as a standing report, not a lucky hallway conversation
  • Grounded AI answers — agent cites graph edges with paper provenance, not vibes

Takeaways

  1. Graphs = things + relationships; the model fits whenever connections carry the knowledge
  2. Biology is graph-shaped — pathways, interactions, regulation
  3. A KG adds types, provenance, and entity resolution — that's what makes it trustworthy
  4. For biot: merged program KGs already surface 10× more context around shared entities
  5. Missing edges are hypotheses — the graph is a lead-generation machine

Questions?


Appendix: try it

biot kg search "IL-6"
biot kg neighbors <entity-id>
biot kg communities

Docs: wiki.graffold.com — benchmarks, architecture, SDK reference.