Credit: FlashMovie/Getty Images
Traditional drug discovery has often followed the “one drug, one target” paradigm—developing compounds against individual proteins thought to drive disease. This strategy produced notable successes such as kinase inhibitors and HER2-targeted antibodies, but it can fall short for complex diseases involving multiple pathways. Many transformative therapies, including checkpoint inhibitors and CAR T cells, emerged by targeting broader cellular processes rather than single proteins.
A causally inspired graph neural network
Reporting in Nature Biomedical Engineering, researchers from Harvard Medical School and collaborators have now developed PDGrapher, an artificial intelligence platform that uses graph neural networks to identify therapeutic interventions capable of reversing disease phenotypes. Instead of predicting how a given drug changes cells, PDGrapher tackles the inverse problem: determining which set of genes or pathways must be targeted to shift a diseased cell back toward health.
As senior author Marinka Zitnik, PhD, associate professor of biomedical informatics in the Blavatnik Institute at HMS explained, “Traditional drug discovery resembles tasting hundreds of prepared dishes to find one that happens to taste perfect. PDGrapher works like a master chef who understands what they want the dish to be and exactly how to combine ingredients to achieve the desired flavor.”
By embedding diseased and treated gene-expression states into networks, the model learns latent representations of these states and predicts optimal interventions. PDGrapher incorporates protein–protein interaction and gene regulatory network data to approximate causal relationships, enabling it to highlight combinations of targets likely to be effective.
Performance across cancers and datasets
The team evaluated PDGrapher across 38 datasets, including both genetic (CRISPR knockouts) and chemical (drug treatments) interventions spanning 11 cancer types. According to the researchers, in held-out samples—even from cancers the model had not seen before—PDGrapher consistently outperformed competing AI methods, ranking true therapeutic targets up to 13% more accurately and training up to 25 times faster.
Importantly, PDGrapher recovered targets of approved drugs excluded from training. For example, it identified KDR (VEGFR2) as a key target in non-small cell lung cancer, aligning with clinical evidence. It also prioritized TOP2A, an enzyme targeted by existing chemotherapies, as a candidate in lung adenocarcinoma, supporting emerging evidence that TOP2A inhibition may suppress metastasis.
Insights into mechanisms and drug repurposing
Beyond predicting targets, PDGrapher can help clarify mechanisms of action. In proof-of-principle analyses, the model traced how vorinostat (an HDAC inhibitor) and sorafenib (a multikinase inhibitor) shift gene networks toward healthier states, underscoring the potential for mechanism discovery and drug repurposing.
Because the model ranks thousands of genes based on their contribution to disease reversal, it can also reveal alternative targets within the same pathway. This flexibility may enable design of combination therapies that address redundancy and resistance mechanisms in cancer biology.
Toward precision medicine applications
By highlighting the drivers most likely to reverse disease, PDGrapher could streamline drug development and reduce the need for broad, inefficient screening. In the future, the model may even be applied to patient-specific data to recommend individualized treatment strategies.
The team is now extending PDGrapher to neurodegenerative diseases, including Parkinson’s, Alzheimer’s, and the rare condition X-linked dystonia-parkinsonism. As Zitnik summarized: “Our ultimate goal is to create a clear roadmap of possible ways to reverse disease at the cellular level.”