Researchers at Harvard Medical School have developed an artificial intelligence model that could transform the way medicines are discovered. Known as PDGrapher, the system moves beyond the traditional paradigm of drug development, which for decades has centered on identifying and inhibiting single proteins thought to drive disease.

While this single-target approach has yielded major advances, such as kinase inhibitors that disrupt cancer cell proliferation, it often fails in conditions shaped by complex networks of genes and signaling pathways. To overcome this limitation, PDGrapher tackles what researchers call the inverse problem: instead of testing how drugs alter cells, it predicts which sets of molecular targets must be modulated to restore healthy cellular function.

Built as a graph neural network, PDGrapher maps relationships among genes, proteins, and pathways to reveal therapeutic strategies. Trained on cellular datasets collected before and after treatment, the model learned to distinguish healthy from diseased states and to forecast which combinations of targets could revert pathology.

Validation tests across 19 datasets and 11 cancer types showed PDGrapher could recover known therapeutic targets that had been withheld from training. It correctly flagged KDR (VEGFR2) as a driver in non-small cell lung cancer and identified TOP2A, already targeted by approved chemotherapies, as a candidate in certain tumors. These predictions aligned with clinical evidence, reinforcing the model’s ability to capture genuine biology.

In benchmarks, PDGrapher ranked true therapeutic targets up to 35% higher than competing AI models while delivering predictions 25 times faster. Its efficiency derives from directly forecasting disease-reversing targets, rather than conducting exhaustive virtual screenings of drug libraries. By suggesting combinations of targets, it may also help counter therapeutic resistance, a common challenge when tumors adapt to single-agent therapies.

Initial applications focus on oncology and neurodegenerative diseases—fields where single-target strategies have struggled. The team is using PDGrapher to explore therapies for Parkinson’s disease, aiming to restore dopamine-producing neurons, and for Alzheimer’s disease, by intervening in pathways linked to neuronal death and plaque formation. At Massachusetts General Hospital, collaborators are testing its potential for X-linked Dystonia-Parkinsonism, a rare movement disorder.

Importantly, PDGrapher is available as open-source software, allowing researchers worldwide to input their own cellular data and obtain target predictions. This democratization of advanced drug discovery tools could accelerate progress across diverse disease areas, including rare conditions that often receive limited attention.

Looking forward, the Harvard team envisions using PDGrapher to analyze patient-specific cellular profiles, paving the way for personalized treatment combinations. Realizing that vision will require rigorous clinical validation, but the model marks a significant step toward precision medicine—where therapies are designed not just for a disease, but for the molecular architecture of each individual’s illness.

Source.