Discovering novel targets and drug repurposing - the DNA-repair driver hiding in BRCA-normal gastric cancer

The starting point was a biological knowledge-based deep-learning network - a PNET - built from Reactome, taking multi-modal input: methylation, transcriptomics, mutations, and proteomics. Its first advantage is structural: every node maps to a real biological entity or function (a gene, a protein, a pathway), so the model is interpretable by design rather than a black box you rationalise after the fact. It was trained to distinguish high-HRD from low-HRD gastric cases, then queried for the features that best separate them.

An explainable network is still hard to read - thousands of nodes and edges make it almost impossible to see what matters by eye. So the module layered a large language model over the network. The LLM reads the whole graph, identifies the key genes and pathways upregulated in HR-deficient patients, surfaces candidates a human would struggle to spot in the mesh, and runs its own literature search to connect them to existing drugs. The network supplies interpretable structure; the LLM supplies interpretation at scale.

For each node whose role was uncertain, the agent scored it across more than 300 connected tools and databases (UniProt, AlphaFold, DepMap, ChEMBL and others): druggability, cellular location, essentiality in healthy versus diseased cells, and any competing compounds already in development. The scored list is generated dynamically and used to sort nodes into targets, biomarkers, and surface receptors.

The network pointed at EZH2. Asked what best distinguishes high- from low-HRD cases, it flagged EZH2 dysfunction as the likely driver of the repair defect - a gene with no obvious link to BRCA, and one not prominent in the prior gastric-cancer literature.

The headline is how easily this would have been missed. EZH2's association with HRD status is real (Figure 1), but on its own it would not survive multiple-hypothesis correction: test thousands of genes one at a time and apply Bonferroni, and EZH2 drops below the line with the noise. It is invisible to transcriptomics alone - the signal only holds up once the biology-informed model integrates methylation, mutations, and proteomics alongside expression. Routine analysis would have buried a true vulnerability; the network surfaced it because it reasons over multi-modal data and pathway structure, not one gene at a time in isolation.

The finding is mechanistically plausible. EZH2 is a histone methyltransferase already implicated in regulating homologous recombination, so EZH2 dysfunction disrupting HR is biologically coherent rather than a statistical artefact.

Then the payoff. Across gastric-cancer cell lines, HRD score correlates significantly with sensitivity to tazemetostat - an FDA-approved EZH2 inhibitor originally developed for sarcoma. The network did not just name a gene; it pointed to a repurposable drug and a synthetic-lethal strategy for a defined cancer subtype: the more HR-deficient the tumour, the more it leans on EZH2, the more an EZH2 inhibitor should hurt it. This work was carried out and validated in collaboration with the Danish Cancer Institute.

Explainability tells you how a model reached an output; interpretability tells you which nodes matter and what each one is. A network whose nodes map to biology gives you the first; layering an LLM over it, plus scoring each node's role, gives you the second - and that is what turns a graph no one can read into a decision and a drug to try.

The same approach generalises to any network. The agentic layer scores any node in any causal network for druggability, location, essentiality, and competition, sorting nodes into targets, biomarkers, and surface receptors. It can also overlay established prior knowledge onto a causal network to test whether known interactions are recovered by the model - and the whole enrichment runs from a single prompt, the agent selecting its own tools and databases, in hours rather than weeks. The team submits the analysis and returns to validate the result rather than doing the integration by hand.

This is the exploratory first step, not the last word. Because every Hydra analysis ships with its executable, associated codebase, Pharos can attach it to a client's additional databases, internal documents, and proprietary data - turning a one-off exploration into a production workflow that drives the next experiments. The end state is a human-in-the-loop, self-improving loop: agentic solutions automate the experimentation a data scientist would otherwise run by hand - robustness across cohorts and batches, the effect of mutations on a target's predictive power - and encode that know-how so it compounds. Some discovery work cannot be written as fixed rules; that is exactly where learning beats programming.