Session 1

Introduction to Biological Networks and Knowledge Graphs

General

This session is about representing biomedical knowledge as a graph, and about the specific ways that representation misleads us if we are not careful.

We want to understand what a biological network is, what a knowledge graph is, and — the part people usually skip — how the two differ. A co-expression network is computed from measurements: correlate every gene against every other gene and keep what clears a threshold. A knowledge graph is read out of recorded facts: somebody looked at the evidence and wrote down that this gene is implicated in that disease. They look identical when drawn and they fail in opposite ways, so telling them apart is the first thing worth learning.

So we build both, over the same genes, and spend the rest of the session on what each one can and cannot say — starting from clinical ICD-10 codes, recovering information the graph appears not to have, and finally connecting it to the molecular data used in Session 2.

The graph is deliberately small — 881 nodes — so that every result can be checked by eye. Everything here scales to graphs of millions of nodes without changing.

The data are:

File	Description
/data/session-1-data/kg_nodes.csv	Node table — id, type (gene / disease / icd10), name, extra
/data/session-1-data/kg_edges.csv	Edge table — source, target, type, weight, evidence
/data/session-1-data/kg_evidence.csv	The same gene–disease edges split by kind of evidence
/data/session-1-data/icd10_map.csv	MONDO disease id → ICD-10 code and label
/data/session-1-data/coexpr_expression.csv.gz	500 patients × 737 genes of TCGA-BRCA expression

Built from Open Targets (CC0 1.0) and the MONDO Disease Ontology (CC BY 4.0) by data-prep/build_kg_data.py. That script downloads ~1.1 GB and cuts it down to a breast-cancer-focused subset small enough to read by eye: 881 nodes (760 genes, 90 diseases, 31 ICD-10 codes) and 1,768 edges. We have already downloaded and processed this data and it is available in the file locations indicated in the table above.

We want to assess the breast cancer disease environment by placing it in the context of 24 reference diseases - including other cancers, autoimmune and inflammatory conditions, neurological, and cardiometabolic disease - we want to look at the relationships between those disease.

Genes are identified by Ensembl gene ID accessions, the same identifier space as the TCGA-BRCA transcriptomics matrix in Session 2, so the two datasets are compatible. The expression file contains a matrix of expression values filtered to the genes in our graph. This allows us to build a computed and a curated network over exactly the same genes.

Part 1 — Networks and Knowledge Graphs: Key Concepts

1.1 Graphs

A graph is a set of nodes and a set of edges between them. Degree, paths and components all follow from those two sets. This part establishes the vocabulary, using a diagram and a four-node toy example, and shows the one operation that makes graphs worth the trouble: composing edges into a path produces an answer nobody stored.

1.2 Correlation versus Curation

Two networks can look very similar, but encode fundamentally different information.

Our gene expression network is a computed network its edges are calculated from numerical data according to our experimental design; in this case we calculate the correlation amongst genes pairwise and then retain only those relations that pass a threshold value that we set. The edges are a statistical outcome from our data as a result of our method. It is entirely data driven and may reveal relationships that are novel.

A curated knowledge graph reads its edges out of a database of asserted facts (often with qualifying or quantifying features). It is prior knowledge, not a discovery set on its own.

Both types of network can fail to accurately capture the properties we may be interested, but the reason for that failure is fundamentally different; neither is fixable by better code. In a computed network a spurious edge is noise generated by the algorithmic process used to generate it and decisions made in the experimental design. In a knowledge graph a missing edge means no-one has recorded that association yet, either because there is no evidence that it should exist or because it hasn’t been formally encoded in a knowledge source. A spurious edge would arise from an incorrect assertion in a knowledge source - so provenance and evidence associated with these data is crucial to assess when building such graphs.

1.3 What makes it a knowledge graph

Typed nodes, typed edges, and provenance. Ours has three node types (gene, disease, icd10) and three edge types (associated_with, is_a, maps_to), with an association score and evidence count on every gene–disease edge — a separate table splitting every gene–disease edge by the kind of evidence behind it. That is the “specialised” in a knowledge graph is a specialised class of network.

1.4 Building one of each

We then build both networks over the same 737 genes and compare them.

The same gene can occupy a completely different environment in the two networks. BRCA1 co-expresses with the proliferation programme — tubulin-adjacent mitotic genes invovled in cell division and PALB2 and FANCD2, which are genuine repair partners. The knowledge graph may capture the former, but is unlikely to capture the latter. Meanwhile the curated graph contains well-known associations and relations between genes, diseases, and ICD codes for hereditary breast–ovarian diseases; many of these are long established and present in a whole host of databases, ontologies, and other sources. TP53 is at the centre of a 27-disease hub in the curated graph, but is essentially invisible in the co-expression network (strongest |r| = 0.18), because its role is primarily effected through modulation of protein stability rather than differential transcriptional regulation.

The computed network has a flexibility in construction that the curated one does not. Sweeping across the correlation threshold from 0.3 to 0.8 allows us to modify the way that the same data is represented from a graph of 39,593 edges to one of 200. Nothing in the data implicitly says which (if any) is correct. The knowledge graph has exactly 1,768 edges defined simply from what we include and the sources we choose; we can of course vary that, but it is a fundamentally different consideration.

Part 2 — Building a Knowledge Graph with NetworkX

2.1 Construction

A knowledge graph is a loop over a node table and an edge table. We write it out explicitly before switching to the helper, so that nothing is magic.

One subtlety gets its own treatment: is_a is genuinely directional, but we work undirected because degree and components behave better that way. The direction is preserved as an edge attribute — without it, climbing the ontology in Part 3 walks downwards half the time and returns plausible wrong answers rather than errors.

2.2 Shape

We read the graph’s summary statistics and ask what each one is really telling us. The density is low, the clustering coefficient is near zero (the graph is close to bipartite), and the hubs are all diseases.

That last one is a construction artefact: the build kept the top 30 genes per disease, so diseases were always going to be the hubs. The habit this part is trying to build is asking, every time, whether a graph property came from the data or from how the data was cut.

2.3 Annotation sparsity

The most important idea in the session. A missing edge has three possible causes — the relationship does not exist, nobody has studied it, or it was recorded against a different term — and the graph cannot tell us which.

We meet a concrete case: basal-like breast carcinoma, one of the most studied breast cancer subtypes, has zero gene associations here, because the evidence was filed against the near-synonymous term triple-negative breast carcinoma.

Part 3 — Practical: Querying the Knowledge Graph

3.1 Starting from ICD-10

ICD-10 is the vocabulary hospitals actually use. Only 19 of our 90 diseases carry a code — and the gap is not spread evenly: 3 of 63 breast subtypes against 16 of 27 other diseases.

The reason is not a bad data source. ICD-10 subdivides breast cancer anatomically — by quadrant — so there is no code for “triple-negative” or “luminal A”, and there cannot be one. Morphology lives in a separate classification entirely. This is a granularity mismatch between the clinical and molecular vocabularies, and no alternative database fixes it.

Climbing the is_a hierarchy recovers a code for 82 of 90 — with the caveat that an inherited code describes the ancestor, not the subtype we started from.

3.2 Which diseases share genes?

Projecting the bipartite gene–disease graph onto diseases alone turns “how similar are these two diseases?” into a structural question. Breast and ovarian cancer share BRCA1, BRCA2 and BRIP1 — hereditary breast-ovarian cancer syndrome, recovered from graph structure without being told the two diseases were related.

But they also “share” tubulins. TUBB, TUBA1B and the topoisomerases show up in exactly the same list, not because of shared biology but because both cancers are treated with taxanes and anthracyclines. The overall association score cannot distinguish the two — but the evidence types can. Filter to genetic and somatic evidence and every tubulin disappears, leaving the real hereditary genes. The advantage of carrying provenance is that we can ask what kind of claim is this, not just how strong is it.

The diseases group themselves. Community detection on that causal graph then recovers the clinical taxonomy — a cancer community, an autoimmune/inflammatory community, a metabolic one — without ever being told those categories exist. Structure nobody encoded turns out to be recoverable from structure somebody did.

3.3 Connecting to the omics data

Finally, the bridge to Session 2. The TCGA matrix uses versioned Ensembl IDs (ENSG00000012048.23) and Open Targets does not, so a naive join matches nothing at all — not fewer things, nothing. An empty join is the safe failure; a partial one is what reaches publication.

All five PAM50 subtypes predicted in Session 2 are disease nodes in this graph, so a subtype prediction becomes an entry point rather than a label.

3.4 Both networks at once

The closing section puts the co-expression network from Part 1 back alongside the curated graph, and asks the question the session title actually poses: can molecular data strengthen the evidence for a gene–disease relationship?

Take the 16 genes with causal evidence for breast cancer and ask what co-expresses with them. Three things come back.

Corroboration. Ten of 120 causal gene pairs co-express — and the strongest are BRCA2–BRIP1, BARD1–BRIP1, BARD1–BRCA2: the homologous-recombination complex, recovered twice over. The graph knows because families were sequenced; the matrix knows because they are transcribed together. Two independent routes, one biology.

A confound we had already removed. The seven strongest partners carrying any curated breast-cancer edge are TOP2A, TYMS, TUBA1B, TUBA1C, TUBB, TOP1 and CDK6 — precisely the chemotherapy targets Section 3.2 just filtered out. They come back because they are proliferation genes, not because the filtering was wrong. Two sources agreeing is only evidence when their errors are independent, and here they are not.

And 184 candidates nothing can sort. Genes that co-express with the causal set and carry no curated breast-cancer edge at all. BLM, EXO1, FANCD2, RAD51, MSH6 — real genome-stability genes worth following up — sit interleaved with BUB1B, HMMR, KNL1 and RRM2, which are proliferation artefacts. No column in the table separates them; that judgement needs biology.

None of the gaps or mismatches along the way are bugs. They are what working with real biomedical knowledge is like, and recognising them and what they might mean is the skill.

Where this leads

Session 3 builds LLM agents — tools, MCP and skills — and Session 4 turns them on graphs like this one, queried from multi-omics profiles. Everything in this session — the coverage gaps, the inherited codes, the technically-correct-but-misleading edges, and the candidate list nothing in the data can sort — is what those agents have to get right, and what we need to check them against.