New AI instrument forecasts therapies to restore wellness in afflicted cells

In an initiative that could transform medication discovery, investigators at Harvard Medical School have developed an artificial intelligence model capable of pinpointing treatments that restore normal states in cells affected by disease.

Unlike conventional methods that usually assess one protein target or medication at a time in search of an effective remedy, the recent model, named PDGrapher and available free of charge, concentrates on multiple drivers of ailments and identifies the genes most likely to revert diseased cells back to a healthy function.

This tool also determines the most effective single or combined targets for therapies that rectify the disease mechanism. The research, published Tuesday in Nature Biomedical Engineering, received partial funding from federal sources.

By focusing on the targets most likely to reverse ailments, this novel methodology could accelerate drug discovery and development, unveiling treatments for conditions that have long been elusive to traditional techniques, the scientists emphasized.

“Conventional drug discovery resembles sampling countless prepared meals to find one that happens to taste ideal,” remarked the study’s senior author Marinka Zitnik, associate professor of biomedical informatics at the Blavatnik Institute at HMS. “PDGrapher operates like a master chef who comprehends the desired dish and precisely how to combine ingredients to achieve the preferred flavor.”

The innovative approach could expedite drug discovery and development and unlock therapies for conditions that have long eluded traditional strategies.

The traditional drug-discovery method — which emphasizes activating or suppressing a single protein — has succeeded with treatments like kinase inhibitors, medications that obstruct certain proteins utilized by cancer cells for growth and proliferation. However, Zitnik highlighted that this discovery model may falter when diseases arise from the interactions of multiple signaling pathways and genes. For instance, numerous groundbreaking medications developed in recent years — consider immune checkpoint inhibitors and CAR T-cell therapies — function by targeting processes in diseased cells.

The approach facilitated by PDGrapher, according to Zitnik, takes a broader perspective to identify compounds that can truly reverse detrimental signs of disease in cells, even if researchers do not yet fully understand which molecules those compounds may influence.

How PDGrapher operates: Mapping intricate connections and impacts

PDGrapher is a type of artificial intelligence tool known as a graph neural network. This instrument examines not only individual data points but also the relationships that exist between these points and the consequences they impose on each other. 

In the realm of biology and drug discovery, this methodology is utilized to chart the connections between various genes, proteins, and signaling pathways within cells, predicting the most appropriate combination of treatments that could correct the underlying cell dysfunction to restore healthy cellular behavior. Instead of meticulously testing compounds from extensive drug databases, the new model prioritizes drug combinations that are most likely to reverse disease.

PDGrapher identifies components of the cell that might be contributing to the disease. Subsequently, it simulates the effects of turning off or reducing these cellular components. The AI model then provides an assessment of whether a diseased cell would improve if specific targets were “addressed.”

“Rather than testing every conceivable recipe, PDGrapher poses the question: ‘Which combination of ingredients will transform this bland or excessively salty dish into a perfectly balanced meal?’” Zitnik noted.

Benefits of the new model

The researchers trained the tool using a dataset of diseased cells before and after treatment, allowing it to determine which genes to focus on to shift cells from a diseased state to a healthy one.

Next, they evaluated it on 19 datasets covering 11 types of cancer, employing both genetic and drug-based methodologies, challenging the tool to predict various treatment options for cell samples it had not previously encountered and for cancer types it had not seen.

The instrument successfully identified drug targets already recognized as effective but that were purposefully excluded during training to ensure the model did not merely memorize the correct answers. It also singled out additional candidates supported by emerging data. The model highlighted KDR (VEGFR2) as a potential target for non-small cell lung cancer, aligning with clinical findings. Additionally, it identified TOP2A — an enzyme currently targeted by approved chemotherapies — as a treatment target in certain tumors, bolstering evidence from recent preclinical studies that TOP2A inhibition may be effective in curbing metastasis in non-small cell lung cancer.

The model exhibited improved accuracy and efficiency compared with other similar tools. In previously unexamined datasets, it ranked the correct therapeutic targets up to 35 percent higher than alternative models and delivered results up to 25 times faster than comparable AI applications.

What this AI advancement signifies for the future of medicine

The new methodology may enhance the process of designing novel medications, the researchers indicated. This is because rather than attempting to predict how every possible alteration would impact a cell and subsequently seeking a useful medication, PDGrapher directly searches for specific targets that can reverse a disease characteristic. This expedites the testing of hypotheses and allows researchers to focus on fewer promising targets.

“Our overarching objective is to establish a clear roadmap of potential avenues to reverse disease at the cellular level.”

Marinka Zitnik, Blavatnik Institute

This tool could be particularly beneficial for complex ailments driven by multiple pathways, such as cancer, where tumors can outmaneuver drugs that target only one entity. Since PDGrapher identifies multiple targets involved in a disease, it could help circumvent this issue.

Furthermore, the researchers noted that after thorough testing to validate the model, it might one day be utilized to analyze a patient’s cellular profile and assist in designing tailored treatment combinations.

Finally, as PDGrapher identifies causal biological factors of diseases, it could aid scientists in understanding why certain drug combinations are effective — providing new biological insights that could propel biomedical discovery further.

The team is presently utilizing this model to address neurological conditions such as Parkinson’s and Alzheimer’s, examining how cells behave in disease and identifying genes that could assist in restoring their health. The researchers are also partnering with colleagues at the Center for XDP at Massachusetts General Hospital to discover new drug targets and map which genes or pairs of genes could be influenced by treatments for X-linked Dystonia-Parkinsonism, an uncommon inherited neurodegenerative disorder.

“Our overarching objective is to establish a clear roadmap of potential avenues to reverse disease at the cellular level,” Zitnik stated.

---

The research was partially financed through federal grants from the National Institutes of Health, National Science Foundation CAREER Program, the U.S. Department of Defense, and ARPA-H Biomedical Data Fabric program, as well as awards from the Chan Zuckerberg Initiative, the Gates Foundation, Amazon Faculty Research, Google Research Scholar Program, AstraZeneca Research, Roche Alliance with Distinguished Scientists, Sanofi iDEA-iTECH, Pfizer Research, John and Virginia Kaneb Fellowship at HMS, Biswas Computational Biology Initiative in collaboration with the Milken Institute, HMS Dean’s Innovation Awards for the Use of Artificial Intelligence, Harvard Data Science Initiative, and the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University. Partial backing was also provided by the Summer Institute in Biomedical Informatics at HMS and the ERC-Consolidator Grant.