Building Biomedical Knowledge Graphs for Target Identification at Scale

Target identification is the first and arguably most consequential step in the drug discovery pipeline. Choose the wrong target, and billions of dollars and years of effort are wasted pursuing a molecule that will never reach patients. Despite its importance, target identification remains largely driven by literature review, expert intuition, and serendipity.

At Gimerny AI, we built GimernyDiscover to systematize this process. At its core is a biomedical knowledge graph containing 2.1 billion edges that connects proteins, genes, diseases, pathways, compounds, cell types, and phenotypes into a unified computational substrate for reasoning about biology.

Why Knowledge Graphs?

Biological knowledge is inherently relational. A gene encodes a protein. That protein participates in a signaling pathway. That pathway, when dysregulated, contributes to a disease phenotype. A compound modulates the protein's activity. These relationships form a graph, and reasoning about drug targets requires traversing this graph to find non-obvious connections.

Traditional bioinformatics approaches treat each data type in isolation. Genomics teams analyze GWAS hits. Proteomics teams study protein interactions. Literature teams review publications. Each group produces ranked lists of candidate targets, and a committee attempts to synthesize these disparate signals into a coherent prioritization. This process is slow, subjective, and prone to anchoring bias.

Knowledge graphs unify these data types into a single structure where computational methods can reason across boundaries. A gene with a strong GWAS signal for a disease becomes more interesting if the protein it encodes sits at the intersection of two dysregulated pathways and has a druggable binding pocket. These multi-hop inferences are natural in a graph but nearly impossible to make systematically from separate spreadsheets.

Data Integration Architecture

Our knowledge graph integrates data from over 40 sources spanning five categories.

Molecular interactions include protein-protein interactions from STRING, BioGRID, and IntAct; kinase-substrate relationships from PhosphoSitePlus; and metabolic reactions from KEGG and Reactome. Genetic associations encompass GWAS Catalog, ClinVar, and OMIM, with effect sizes, population frequencies, and confidence scores preserved as edge attributes. Expression data incorporates tissue-specific expression from GTEx, single-cell atlases from the Human Cell Atlas, and disease-specific differential expression from GEO. Chemical data spans ChEMBL bioactivity data (IC50, EC50, Ki values), DrugBank approved drug information, and PubChem compound properties. Literature-derived relationships are extracted from 35 million PubMed abstracts using our custom NLP pipeline, which identifies gene-disease, drug-target, and pathway-phenotype relationships with 89% precision.

The integration challenge is substantial. Entity resolution across databases requires mapping between different identifier systems (UniProt, Ensembl, HGNC, PDB). Confidence scoring normalizes evidence strength across heterogeneous sources: a crystallography-validated protein interaction carries different weight than a high-throughput yeast two-hybrid hit.

Heterogeneous Graph Neural Networks

Standard GNNs treat all nodes and edges as the same type. In a biomedical knowledge graph, this is woefully inadequate. A protein node has different features (sequence, structure, domains) than a disease node (prevalence, heritability, tissue specificity), and a "phosphorylates" edge has different semantics than an "is_associated_with" edge.

We use heterogeneous graph neural networks (HGNNs) that maintain type-specific transformation matrices for each node and edge type. Our architecture extends the Relational Graph Convolutional Network (R-GCN) framework with several innovations.

Type-specific message functions allow each edge type to have its own learned message function, capturing the distinct semantics of different biological relationships. Attention-weighted aggregation uses multi-head attention to weight messages from different relationship types, learning that, for example, direct physical interactions should be weighted more heavily than text-mined associations for druggability prediction. Subgraph sampling addresses the challenge of training on a 2.1B edge graph, which does not fit in GPU memory; we use a biased random walk sampling strategy that preserves local neighborhood structure while maintaining computational tractability.

From Embeddings to Target Predictions

The trained HGNN produces embeddings for every entity in the graph. For target identification, we formulate the problem as link prediction: given a disease node, which protein nodes are most likely to be connected by a "therapeutic_target_for" edge?

We score candidate targets using a combination of the learned embeddings and a separate druggability model. The embedding similarity captures biological plausibility (is this protein involved in disease-relevant biology?), while the druggability model evaluates practical considerations (does this protein have a suitable binding pocket? Is it selectively targetable? What is the competitive landscape?).

Our validation on 847 known target-disease pairs from the Open Targets platform shows a recall@50 of 73%, meaning that the true target appears in the top 50 predictions for 73% of diseases tested. More importantly, in prospective evaluations with three pharmaceutical partners, GimernyDiscover has identified 12 novel targets that passed experimental validation, with four now in active lead optimization campaigns.

Lessons Learned

Three key lessons have emerged from building and deploying this system at scale. First, data quality dominates model architecture. Spending engineering time on entity resolution, confidence calibration, and source weighting has yielded larger improvements than any architectural innovation. Second, temporal awareness is essential. Biology is a living knowledge base, and new papers publish daily. Our weekly update pipeline processes approximately 15,000 new papers per week, extracting and integrating new relationships. Targets that were unknown six months ago can become top predictions after a single high-impact publication. Third, explainability drives adoption. Pharmaceutical scientists will not trust a black-box prediction. Every target recommendation from GimernyDiscover comes with a subgraph explanation showing the specific paths through the knowledge graph that support the prediction, along with the underlying evidence sources.

Building a knowledge graph of this scale is a multi-year engineering effort, but the payoff is transformative: systematic, data-driven target identification that surfaces opportunities no individual expert could see.