A graph neural network propagates evidence across the knowledge graph to rank novel drug targets
Before any graph convolution runs, every candidate gene must exist as a node with a well-defined starting representation. Modern target-discovery pipelines build a heterogeneous biomedical knowledge graph — genes, proteins, diseases, pathways and drugs as nodes, curated relations as edges — and seed each node with a baseline multi-omic feature vector drawn from public and proprietary evidence sources.
Target prioritization pipelines are typically built on top of the Open Targets Platform, which aggregates genetic association, somatic mutation, known drug, affected pathway, RNA expression, animal-model and literature-mining evidence into a single gene–disease association graph spanning roughly 19,000 protein-coding genes and thousands of disease terms.
On top of this backbone, additional relation types are layered in: protein–protein interaction edges from STRING and BioGRID, pathway co-membership edges from Reactome and KEGG, co-expression edges derived from GTEx and single-cell atlases, and structural-similarity edges between protein domains. The result is a heterogeneous graph with roughly a dozen distinct edge types, each carrying its own semantics and reliability weight.
This graph is the substrate the GNN will operate on. Unlike a flat feature table used by classical machine-learning target scorers (random forests, gradient-boosted trees), the graph explicitly preserves relational structure — which genes interact, which pathways they share, which diseases they co-occur with — so that a node with sparse direct evidence can still borrow signal from well-characterized neighbors during message passing.
Each node's initial 128-dimensional feature vector concatenates several independent evidence channels: tissue-specific expression z-scores across ~50 GTEx tissues, genome-wide association study (GWAS) association strength and fine-mapping posterior probability, CRISPR essentiality scores from DepMap cell-line screens, pretrained protein-language-model embeddings (ESM-2 / ProtBERT) capturing sequence and structural context, and one-hot indicators for existing drug-target status.
All continuous channels are z-normalized per feature across the node population before being handed to the network, and missing channels — common for poorly characterized genes — are zero-imputed with an explicit missingness flag rather than silently dropped, so the model can learn to discount unreliable inputs.
Roughly 30% of the protein-coding genome — the so-called "ignorome" — has fewer than five dedicated PubMed publications. For these genes, direct features are sparse by construction; the graph's edge structure, not the node's own data, becomes the primary carrier of predictive signal once message passing begins.
A recurring challenge in target-graph construction is the "cold-start" node: a gene with a well-connected position in the interactome but almost no direct experimental characterization. Feature initialization treats these nodes symmetrically with well-studied genes — assigning the same 128-dimensional slot, populated as far as data allows — rather than excluding them, since the whole point of a graph-based model is to let structure compensate for missing direct evidence.
Final preprocessing steps include row-wise L2 normalization of the concatenated vector and a sanity pass that checks for degenerate nodes (isolated components with zero edges), which are flagged and either merged into the nearest pathway-defined community or removed from the training graph entirely to avoid unstable gradients during the first GNN layer.
With node features and graph structure in place, the network performs the operation that gives Graph Neural Networks their name: at each layer, every node aggregates transformed feature vectors from its immediate neighbors, forming a "message" that updates its own representation. Stacking several such layers lets evidence travel several hops across the knowledge graph.
The canonical Graph Convolutional Network (GCN, Kipf & Welling 2017) updates each node's representation by averaging transformed neighbor features through a symmetrically normalized adjacency matrix: h′ = σ(D⁻¹/² A D⁻¹/² h W). This is elegant and fast on moderate graphs but assumes the full graph fits in memory for a spectral-style normalization.
GraphSAGE (Hamilton et al. 2017) instead samples a fixed-size neighborhood per node and aggregates it with a learnable function — mean, max-pool, or an LSTM over a random neighbor ordering — before concatenating with the node's own previous-layer representation. Because it operates on sampled subgraphs, GraphSAGE scales to graphs with tens of millions of edges and, critically, generalizes inductively to nodes never seen during training — essential when new candidate genes are added to the knowledge graph after the model has already been trained.
Both architectures are implemented and benchmarked using PyTorch Geometric (PyG), the standard library for graph deep learning, which provides optimized sparse message-passing kernels and mini-batch neighbor loaders.
Stacking too many GNN layers causes "over-smoothing": after roughly 6–8 rounds of neighbor averaging, distinct node embeddings converge toward the same graph-wide average and lose discriminative power. Practical target-prioritization models rarely exceed 4–6 layers for exactly this reason.
A full-graph forward pass over a knowledge graph with millions of PPI, pathway and expression edges does not fit in GPU memory using naive full-batch GCN training. GraphSAGE-style neighbor sampling addresses this directly: for each mini-batch of target nodes, a fixed number of neighbors is sampled at each hop (for example, 25 at hop one and 10 at hop two), bounding both memory use and the exponential neighborhood blow-up that would otherwise occur in dense hub regions of the graph.
PyTorch Geometric's NeighborLoader implements this sampling scheme efficiently, streaming subgraphs to the GPU in mini-batches during training. This is what allows a target-prioritization GNN to be trained on the full ~19,000-node, multi-million-edge Open Targets-derived graph on a single GPU in a few hours rather than requiring a distributed graph-processing cluster.
Because the knowledge graph is heterogeneous — a PPI edge means something structurally different from a shared-pathway edge or a co-expression edge — plain GCN/GraphSAGE aggregation, which treats all neighbors identically, discards useful information. Two extensions address this:
Graph Attention Networks (GAT) learn a per-edge attention coefficient so that more informative neighbors contribute proportionally more to the aggregated message, rather than every neighbor being weighted equally or by fixed degree normalization.
Relational GCNs (R-GCN) instead learn a separate weight matrix per edge type (or a low-rank basis decomposition shared across types, to control parameter count), so a genetic-association edge and a co-expression edge propagate messages through distinct transformations. For multi-relational biomedical graphs this typically improves held-out ranking performance over relation-agnostic GCN baselines.
| Product | Indication | Trial Design | Key Result |
|---|---|---|---|
| GCN | Symmetric-normalized mean aggregation over full adjacency | Full-batch, best under ~10⁵ nodes | Simple, strong transductive baseline |
| GraphSAGE | Sample-and-aggregate (mean / pool / LSTM) over fixed-size neighbor samples | Mini-batch, scales to 10⁷+ edges | Inductive — generalizes to unseen nodes |
| GAT | Attention-weighted neighbor sum with learned edge coefficients | Moderate — attention adds compute | Down-weights uninformative neighbors |
| R-GCN | Relation-specific weight matrix per edge type (basis-decomposed) | Moderate–high with basis sharing | Best for multi-relational KGs like Open Targets |
After L rounds of message passing, each node holds a stack of L intermediate representations — one per layer — that summarize progressively larger neighborhoods. The aggregation step combines these into a single fixed-length embedding per gene: a compressed vector that jointly encodes the node's own features and its multi-hop graph context.
Naively using only the final GNN layer's output as the node embedding discards useful information: early layers capture tight, local neighborhood structure (direct interactors), while later layers capture broader, more diffuse context — and, per the over-smoothing issue described earlier, later layers are also progressively noisier.
Jumping-Knowledge Networks (Xu et al. 2018) address this by letting the final representation adaptively combine outputs from every layer rather than only the last one — typically via concatenation followed by a linear projection, or via a learned per-node attention over layers. For target prioritization, concatenation across 4 layers followed by a projection down to 256 dimensions is a common configuration: it preserves both fine-grained local evidence (a gene's immediate PPI partners) and broader pathway-level context (its position within a disease module) in a single fixed-size vector suitable for downstream scoring.
A useful sanity check on a trained GNN is to project the 256-dimensional embeddings into two dimensions with UMAP and inspect whether the resulting layout recovers known biology without being told to. In practice, genes belonging to the same curated pathway (Reactome) or the same disease module (shared GWAS loci, shared PPI neighborhood) cluster together, forming on the order of several dozen visually distinct groups — evidence that the embedding space has organized itself around real biological structure rather than an arbitrary transformation of the input features.
This clustering also has a direct practical use: cosine similarity between embeddings supports "target-target retrieval" — given one validated target, retrieve the nearest neighbors in embedding space as mechanistically related candidates, independent of the downstream scoring head.
In internal benchmarks, genes within the same Reactome pathway show a median embedding cosine similarity roughly 3× higher than randomly paired genes — a strong signal that message passing successfully encodes pathway co-membership into the learned representation, not just raw feature similarity.
Because embedding generation is disease-agnostic — it depends only on graph structure and baseline node features, not on any single disease's labels — the same 256-dimensional vectors can be reused as a shared representation across many downstream scoring tasks. A single embedding pass over the full knowledge graph can feed prediction heads for dozens of diseases simultaneously, each trained independently on its own positive/negative target labels, without recomputing the expensive graph-convolution step per disease.
This multi-task reuse is one of the main practical efficiency arguments for the GNN approach over disease-specific feature-engineering pipelines: the costly part of the computation (message passing over millions of edges) is amortized once, while the cheap part (a small MLP head) is retrained per indication.
A dense embedding is not yet a decision. A lightweight supervised prediction head — typically a two-layer MLP with a sigmoid output — maps each node's 256-dimensional embedding to a single scalar score, trained to separate known approved or clinical-stage drug targets from randomly sampled genes, calibrated so the output can be read as a priority probability.
The prediction head sits on top of the frozen (or fine-tuned) GNN encoder and is deliberately kept small: a 256→64→1 multilayer perceptron with ReLU activations, dropout of 0.3 between layers to control overfitting on a relatively small positive-label set, and a final sigmoid to squash the output into a [0,1] priority score. Training minimizes binary cross-entropy between predicted score and label, using positive examples drawn from genes with an approved drug or a drug in active clinical trials (per Open Targets' known_drug evidence channel) and negative examples drawn from genes with no drug-target evidence and low prior literature association, downsampled to balance the classes.
Because the encoder was trained once and reused across tasks (see Stage 3), the scoring head itself is cheap to retrain per disease area — typically converging within a few dozen epochs on a single GPU.
Hard positive/negative labels are scarce relative to the ~19,000 candidate genes, so training also incorporates weak supervision: Open Targets' continuous overall association scores (themselves a weighted combination of genetic, somatic, expression and literature evidence) are used as soft regression targets in an auxiliary loss term alongside the primary binary classification objective. This lets the model learn from graded confidence rather than only hard labeled examples, improving score calibration for the large middle tier of genes that are neither clearly validated targets nor clearly irrelevant.
Distant supervision from clinical trial outcome data — including trials that failed for efficacy or safety reasons — is incorporated as an additional negative signal, since a gene that was tested as a target and failed is a more informative negative than an untested gene.
A raw sigmoid output from an MLP head is not automatically a well-calibrated probability — the model can be systematically over- or under-confident. Post-hoc Platt scaling (fitting a single-parameter logistic regression on held-out validation predictions) corrects this, so that a reported score of 0.8 corresponds, empirically, to roughly an 80% chance of the gene resembling a true positive target under the training distribution.
For interpretability, gradient-based attribution methods such as integrated gradients, or graph-specific tools like GNNExplainer, can be applied to a high-scoring node to identify which edges and which upstream neighbors contributed most to its final score — turning an otherwise opaque prediction into a traceable evidence subgraph that a reviewing biologist can inspect and sanity-check before committing bench resources.
Held-out AUROC of 0.87 substantially outperforms a genetics-only baseline (AUROC ≈ 0.74) that ignores graph structure entirely — direct evidence that neighbor-propagated context, not just a gene's own features, materially improves target ranking quality.
The final output of the pipeline is deceptively simple: a sorted list. Every candidate gene's calibrated score, together with its supporting evidence subgraph, is compiled into a ranked shortlist that hands bench biologists a triaged, explainable starting point for CRISPR screens, biochemical assays, and other functional follow-up — rather than 19,000 undifferentiated genes.
Genes are sorted by calibrated priority score, and a shortlist is cut at a threshold set by available experimental capacity rather than an arbitrary statistical cutoff — commonly the top 15–30 candidates for a given disease area. Ties near the cutoff are broken using secondary signals: network centrality (betweenness or eigenvector centrality within the disease-relevant subgraph), tractability annotations (does the protein have a known druggable pocket, is it a member of a druggable protein family such as kinases or GPCRs), and diversity constraints that avoid over-representing a single pathway or protein family in the final list.
The shortlist is delivered together with each gene's top contributing evidence paths — the specific neighbors and edge types that most influenced its score — so the ranking is not a black-box output but an auditable trail back to genetic, expression, and interaction evidence.
A ranked list is a hypothesis generator, not a final answer. Shortlisted genes typically proceed through a funnel of increasingly expensive validation: pooled CRISPR knockout or knockdown screens in disease-relevant cell lines first, followed by arrayed validation of hits, followed by more resource-intensive in vivo models for the strongest surviving candidates.
Critically, validation outcomes — positive and negative — are fed back into the training set as new labels for the next model iteration, closing an active-learning loop: the GNN's own ranking decisions determine which genes get tested, and those test results directly improve the next generation of the model.
In published retrospective benchmarks, GNN-prioritized shortlists validate in downstream functional screens at roughly 2–3× the hit-rate of score-matched candidates chosen by simple genetics-only ranking — the graph-propagated context measurably improves precision at the top of the list, where it matters most.
The most valuable outputs of a GNN-based ranking are frequently not the genes with the strongest direct genetic or literature evidence — those are usually already well known and already being pursued — but genes with modest direct evidence that sit in structurally important positions within the disease-relevant subgraph: a hub gene several hops from the nearest GWAS locus, connected through a well-supported pathway rather than a well-supported single study.
Because message passing explicitly propagates evidence across such paths, the model can assign these structurally-supported-but-under-studied genes a competitive score, surfacing candidates that a purely feature-based or literature-frequency-based ranking would systematically overlook. This is the core value proposition of applying graph neural networks to target discovery: not replacing domain expertise, but scaling exactly the kind of multi-hop reasoning a human curator would do manually, across a graph far too large to review by hand.