AI Points to Drug Repurposing Opportunities for Diseases Without Cures
By Deborah Borfitz
October 31, 2024 | Harvard scientists have succeeded in using artificial intelligence (AI) to identify promising drug candidates for diseases for which there are no examples of successful treatment. The graph foundation model, known as TxGNN, opens untold opportunities to repurpose drugs for the thousands of diseases afflicting small populations and exacting a huge toll in terms of both economics and human suffering, according to Marinka Zitnik, Ph.D., assistant professor of biomedical informatics at Harvard Medical School.
“Existing drug-repurposing AI models are generally centered around one disease,” she says, mostly because they are often trained on data specific to that disease. That approach limits their ability to nominate potential uses across multiple conditions.
When dealing with a relatively small amount of available data, it can be extremely challenging to do the training with large language models using the “one disease, one model” methodology, explains Zitnik, who is also an associate faculty member at the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University. So, she and her team trained a single model across many diseases in a manner allowing the predictive power from diseases with a larger amount of molecular data to be transferred to diseases having much sparser associated information.
TxGNN is a “zero-shot” drug-repurposing model, meaning it predicts therapeutic candidates for diseases with limited or no treatment options—a list that includes all but 5% to 7% of the more than 7,000 rare diseases worldwide. The focus is on drugs already approved by the U.S Food and Drug Administration (FDA).
The underlying principle is that effective drugs directly target disease-perturbed networks or indirectly propagate therapeutic effects through disease-associated networks, Zitnik says. TxGNN itself has two modules, one predicting drug indications and contraindications and the other explaining the knowledge graphs that connect a queried drug to a queried disease.
In a recently published study in Nature Medicine (DOI: 10.1038/s41591-024-03233-x), TxGNN was used to rank nearly 8,000 drugs approved or in clinical trials, based on their treatment utility and side effects. The target list comprised more than 17,000 diseases, 92% lacking pre-existing indications and known molecular target interactions.
Many of the top-ranked drugs aligned well with off-label prescribing already going on, and the predictions were better than eight other drug-repurposing AI models in terms of accuracy and contraindications, she notes.
Clinical trials could soon be designed to test some of the identified drugs. Earlier this year, Every Cure announced it had secured $48.3 million from the Advanced Research Projects Agency for Health, a new federal agency advancing research that otherwise could not be readily accomplished. Every Cure subsequently announced that it was partnering with AI drug development company BioPhy to evaluate high-value drug-disease matches and drug-repurposing opportunities, focusing on the simulation and optimization of clinical trials and their endpoints.
AI for Good
The researchers have made the TxGNN tool available for free and are hopeful that clinician-scientists will use it in their search for new therapies, says Zitnik. The website provides a visual interface where users can enter their disease and drug names of interest to view the model’s predictions and the medical knowledge it relied on when making those calculations. For those with the interest and expertise in training the model on their own private data or improving upon the published version of TxGNN, the code is also freely available, she adds.
Since a preprint version of the paper has been on medRxiv.org for many months, the Harvard team is already partnered with the Chan Zuckerberg Initiative, which brings together dozens of patient-led organizations centered around different rare diseases. Several major pharmaceutical companies have also reached out for assistance in reproducing the code and model internally, where TxGNN was being retrained on both public and internal data.
From the perspective of the Harvard team, TxGNN revolves around the notion of AI for public good for diseases that are underserved, she says. “There are many, many thousands, perhaps millions of patients that are undiagnosed with [rare] diseases and, when they are diagnosed, they do not have effective treatments to manage [them]... and the problem is in large part is due to a lack of incentives for developing drugs for those types of diseases because the patient populations are relatively small.
“The primary motivation,” she continues, is to “provide some initial resource that could help [patient- organizations] screen drugs for these rare diseases.” Individual foundations have been working hard to build registries and biobanks to learn more about these poorly characterized conditions, which up to now have lacked any type of AI models pointing to drugs for potential clinical adoption.
Current foundation models are centered around diseases for which large, curated cohorts exist at National Institutes of Health and other institutions such as Protein Data Bank. Zitnik’s lab wanted to leverage cutting-edge advances in AI to look beyond these “low-hanging fruit” to areas that are currently underserved by medical research and care—while acknowledging that aspects of the predictive tool create commercial opportunities that others are free to explore.
Matching Process
A foundation model, unlike algorithms trained in a narrow manner (e.g., single phenotype or disease), are trained across many diseases where each one can be thought of as a task to be performed—in the case of TxGNN, a total of 7,000 tasks, says Zitnik. The goal was to create a high-performing model with general-purpose capabilities and the question was how to best accomplish that.
This led Zitnik and her team to the idea of combing through data in 1.2 million electronic medical records (EMRs) to come up with novel predictions that did not exist in the training data. That is, they had TxGNN randomly match drugs to diseases, focusing on the combinations where the drug was tied to a disease for which it was not indicated.
The top drug-disease picks were soundly confirmed by the significantly higher level of off-label prescribing of the same medicine or another in the same class for each condition relative to how often investigators saw a random drug being prescribed off-label for the disease. In a separate analysis, TxGNN also favorably ranked 10 drugs newly approved by the FDA. “That suggests to us that the novel predictions this model makes for a broad range of diseases can be promising drug candidates,” says Zitnik.
The study was conducted in collaboration with digital health experts at Mount Sinai who have phenotyped and curated EMRs at their institution, Zitnik says. They are the ones who built the EMR-linked dataset, across 478 diseases and 1,200 drugs, which was used for training TxGNN.
Predictions Explained
The so-called Explainer module of TxGNN speaks to its usability since “few want to follow black-box AI models,” Zitnik says. The users here are drug developers and clinical researchers whose decisions are, of necessity, limited by budgets and concerns of drug efficacy and patient safety—be that which drug screening lead or off-label prescription use pattern to follow up on.
It was therefore crucial that the model not only pair a specific drug candidate with a disease but also indicate the “pieces of existing medical knowledge” it most relied on when making that prediction, says Zitnik. These knowledge graphs include information about genes, the pathways genes are associated with, and the pathways drugs act on.
The explanations are designed to help drug developers identify the predictions that are more likely to succeed in downstream wet lab or clinical research applications, Zitnik says. They can also alert users if the model relied on data that is untrustworthy or took confusing “shortcuts” due to gaps in existing medical knowledge. It might thereby prompt the generation of missing data to improve the model for specific drugs or diseases, she adds.
Where the model aligns with current medical understanding, the predictions are inherently more trustworthy. In the latest study, predictions made by TxGNN about drugs for three rare conditions were assessed by a dozen human evaluators—five clinicians, five clinical researchers, and two pharmacists—who reported that the explainability feature enhanced their confidence level in the forecasts.
The next steps include focused case studies where researchers examine one disease and its predicted drug match. The work will be led by rare disease foundations with access to clinical data on the specific patient populations they represent, says Zitnik, noting that her team's quantitative benchmarking analyses did not rely on patient-level data.
Merging Models
As previously reported, the Zitnik lab has also been focused on the design of new drug molecules—again using foundation models enabling AI to self-learn. One of the larger, ongoing projects is development of the PINNACLE model for single-cell biology that supports therapeutic research by tailoring its outputs to the biological contexts in which it operates, leveraging contextual learning advances in natural language processing and transformers.
While TxGNN is trained on clinically relevant medical data, PINNACLE is trained on molecular proteomics data. The two models are “completely separate from each other,” at least for now, Zitnik says.
Moving forward, the plan is to create a single agent system that would “consider [both] predictions around drugs that can be derived from molecular transcriptional information about gene activation... and patient-level health record data,” says Zitnik. The objective is to learn if using predictions from the molecular information could be used to screen drug candidates from the clinical world to produce even better insights—meaning, “multidimensional endpoints” considering safety, efficacy, effort of manufacturing, and delivery of those drugs to patients.
The team will first need to work on improving the TxGNN model further by addressing its limitations. “One important consideration is the size of the data this model is trained on,” Zitnik says. Construction of the medical knowledge graphs and training of TxGNN across so many diseases required significant effort by the research team.
“Ideally, we’d want to retrain this model one time and in such a way that it would easily accommodate new data as it becomes available,” she continues. Recreating the dataset every month or year would be impractical and thus require the application of techniques for continual curation and learning and to enable predictions that also consider new FDA approvals.
Patient-level information will also need to be integrated with the medical knowledge graphs to provide personalized drug-repurposing predictions, says Zitnik. The current version of TxGNN is publicly available in part because it was not trained on patient-level data from EMRs—that was used only to evaluate model predictions. If both types of data were being leveraged, model performance might be further boosted with information about real-world treatment regimens and patient response in terms of potential adverse events and disease progression.
Additionally, Zitnik and her team hope in the future to be creating knowledge graphs with more comprehensive information on host–pathogen interactions. For infectious diseases, good predictions will require modeling not just human biology but the biology of external pathogens—including whether they are viral or bacterial, what they do to the body, and the molecules they bind to.