A knowledge graph drug discovery approach represents genes, proteins, diseases, and pathways as a connected network rather than a spreadsheet, letting algorithms traverse relationships a human reviewer would never spot manually. This form of biomedical knowledge graph is only as good as the biology encoded in its edges and the graph neural network (GNN) that walks them, and the resulting target hypotheses still need experimental confirmation. The appeal for computational biologists lies in scale, but the discipline lies in knowing where the network signal ends and noise begins.
Key takeaways
- A biomedical knowledge graph encodes genes, proteins, diseases, drugs, and pathways as nodes connected by edges representing curated or literature-derived relationships.
- Public resources such as Open Targets, PrimeKG, Hetionet, and STRING provide the underlying data that computational biologists use to build and query disease networks.
- Graph neural networks (GNNs) learn patterns across these networks to rank candidate genes and proteins, a core technique in network medicine drug targets discovery.
- Signal-to-noise remains the central limitation: incomplete curation, literature bias, and structurally plausible but biologically irrelevant paths can inflate false positive rates.
- Every network-derived hypothesis still requires orthogonal experimental validation before it can inform a target identification and validation program.
What a biomedical knowledge graph is
A biomedical knowledge graph is a structured network in which nodes represent biological entities, genes, proteins, diseases, drugs, phenotypes, and pathways, and edges represent relationships between them, such as "gene associated with disease" or "drug targets protein." Each edge is typically drawn from a specific evidence source, a curated database, a clinical dataset, or text mined from the literature, and carries some indication of confidence or provenance. This structure lets a disease network AI system query relationships several steps removed from a starting point, for example tracing a gene through a pathway, a protein complex, and a comorbid condition to reach a candidate target that no single dataset would reveal on its own.
The value of a biomedical knowledge graph comes from integration rather than any single data type. PrimeKG, a widely cited precision medicine knowledge graph, merges 20 high-quality resources into a single network of more than 129,000 nodes and roughly 4 million relationships spanning 10 biological scales, from protein perturbations to approved drug indications, as described in its founding paper in Scientific Data. That scale is what allows biological network analysis to surface indirect connections, such as a drug's off-label mechanism overlapping with an underexplored disease pathway, that would be invisible in any one source database.
Because the graph is a simplification of biology, the choices made when it is built (which relationships to include, how confidence is scored, how contradictory evidence is resolved) shape everything downstream. A knowledge graph is not a ground truth about disease mechanism; it is a queryable hypothesis space built from the evidence available at the time of construction.
Key public knowledge graph resources
Several open, well documented resources anchor most academic and industry work in this space, and each emphasizes a different slice of biology. Open Targets, a public-private partnership coordinated through the European Bioinformatics Institute, aggregates evidence from 22 data sources, including genetic associations, gene expression, animal models, and known drugs, into a single scored network connecting targets and diseases, as detailed in its platform description. It is built specifically to support systematic target prioritization rather than open-ended exploration.
Other resources take a broader approach to network medicine drug targets. Hetionet integrates data from 29 public sources into a heterogeneous network of more than 47,000 nodes and 2.25 million relationships spanning genes, compounds, diseases, and anatomical structures, and was built explicitly to support drug repurposing predictions, as described in its original network study, published in eLife. PrimeKG, covered above, adds depth on rare disease coverage and clinical annotations. Underlying many of these graphs sits protein interaction data such as that curated by STRING, which scores protein-protein associations across sequenced genomes using a combination of experimental, curated, and computational evidence.
Computational biologists typically combine these resources rather than choosing one. A protein interaction backbone from STRING might anchor mechanistic plausibility, while Open Targets contributes genetic and clinical evidence, and a broader graph like PrimeKG or Hetionet fills in disease and drug relationships that neither source captures alone. The practical skill is not just building the merged network, but understanding which edges in it are well supported and which are structurally convenient placeholders.
| Resource | Primary network focus | Typical use case |
|---|---|---|
| Open Targets | Target-disease evidence scoring across genetics, expression, and known drugs | Systematic target prioritization |
| PrimeKG | Multimodal, multi-scale disease relationships | Rare disease and clinical annotation queries |
| Hetionet | Heterogeneous drug-disease-gene network | Drug repurposing hypothesis generation |
| STRING | Protein-protein association confidence | Mechanistic plausibility and pathway context |
Graph neural networks for target prediction
GNNs learn numerical representations of nodes and edges by aggregating information from a node's neighbors across multiple hops of the network, which lets them capture patterns that simple network statistics miss. Applied to a biomedical knowledge graph, a GNN can be trained to predict missing or future edges, most usefully a plausible but previously unrecorded link between a gene and a disease, or between a compound and a target. This edge prediction task underlies most contemporary approaches to network medicine drug targets, and it has been applied across drug-target interaction prediction, drug repurposing, and mechanism discovery, as summarized in a broad review of graph-based methods in drug discovery and development.
The practical workflow generally follows a few recurring steps:
- Assemble and harmonize a knowledge graph from curated databases, interaction networks, and literature-derived relationships, resolving identifier mismatches across sources.
- Train a GNN or related embedding model on the graph to learn vector representations for genes, proteins, diseases, and compounds that preserve network structure.
- Score candidate edges, such as untested gene-disease or drug-target pairs, using the learned embeddings, ranking them by predicted relevance.
- Filter high-scoring candidates against orthogonal evidence, tractability, and safety data before nominating any for experimental follow-up.
- Feed experimental results back into the graph and retrain, closing the loop between computational prediction and wet lab confirmation.
These models excel at pattern recognition across scales that a human reviewer cannot hold in mind simultaneously, but they inherit every gap and bias present in the underlying graph. A GNN trained on a network with sparse coverage of a rare disease will simply have less signal to work with, regardless of how sophisticated the architecture is.
From network signal to target hypothesis
A high network score is a hypothesis about biological relevance, not evidence of it, and treating the two as equivalent is the most common misstep in applying these methods. In practice, computational biologists use network centrality, predicted edge confidence, and cross-source concurrence as a triage layer, narrowing thousands of candidate genes to a shortlist that a bench team can realistically pursue.
This mirrors, and extends, current field practice in target identification more broadly, where genetic association data and multi-omics integration have already become standard triage tools; network-based scoring adds a layer that captures indirect, multi-hop relationships those methods handle less naturally. Because a network can flag a gene as connected to a disease for reasons ranging from strong genetic evidence to a single low-confidence text-mined sentence, prioritization frameworks generally combine graph-derived scores with independent lines of evidence, such as protein interaction density from STRING or clinical association data from Open Targets, described in the genetics-focused review of the platform's evidence model.
The output of this stage is rarely a single "answer." It is typically a ranked, annotated shortlist, useful raw material for a target identification and validation program, but it still requires the functional and mechanistic work that turns a network hypothesis into a validated target.
Limitations and validation challenges
Signal-to-noise is the defining limitation of biomedical knowledge graphs, and it stems from two compounding problems: incomplete curation and literature-derived noise. Knowledge graphs are necessarily incomplete because no database captures every true biological relationship, and they simultaneously contain edges of uneven reliability, since literature-mined relationships depend on natural language processing that can misidentify entities or infer associations that do not hold mechanistically. A graph can therefore generate large numbers of structurally plausible but biologically irrelevant paths between a gene and a disease.
The most common failure modes recur across projects regardless of which resource anchors the graph:
- Literature bias, where frequently studied genes accumulate disproportionately more edges regardless of true disease relevance.
- Incomplete curation, where rare diseases and understudied pathways are systematically underrepresented.
- Structurally plausible false positives, where multi-hop paths connect a gene and a disease without any mechanistically meaningful relationship.
- Version drift, where public resources update on different schedules, creating inconsistencies across a merged, multi-source graph.
Independent experimental validation is the field's answer to this problem, and recent work illustrates why it matters. A validated evidence pipeline, published in Nature Communications, tested knowledge graph-derived drug predictions for Fragile X syndrome against preclinical transcriptional data, finding strong correlation between the automatically extracted paths and the experimentally measured gene and pathway changes. That result underscores that network-derived evidence gains real value only once tested against orthogonal, empirical data rather than graph structure alone. Without that step, a knowledge graph can just as easily reinforce existing literature biases, since well-studied genes accumulate more edges regardless of true biological importance, as it can reveal genuinely novel biology.
The practical implication for teams building or buying into these systems is to treat graph output as a filter that reduces search space, not a discovery in itself. Reproducibility across independent graphs and datasets, and consistency with orthogonal genetic or functional evidence, are the closest available proxies for confidence until a candidate reaches the bench.
Knowledge graph drug discovery is reshaping target strategy
A knowledge graph drug discovery strategy works best when it is treated as one evidence stream among several rather than an oracle. Combining network-derived candidate rankings with genetic association strength, tractability assessments, and expression data gives a more defensible shortlist than any single method alone, and it is consistent with how target identification and validation programs already triage candidates before committing bench resources.
The technology's trajectory points toward tighter integration between graph-based candidate generation and the experimental pipelines that confirm or reject those candidates, an arc covered in broader terms in the ongoing look at AI in drug discovery across the full pipeline. As curation improves and validation feedback loops mature, the gap between an interesting network signal and a defensible target hypothesis should narrow, though it will not close on its own.
This article was produced in accordance with Drug Discovery News' Editorial Policies.
