In modern medicine, some of the most important signals are also the hardest to detect—tiny genetic variations that can shape a person’s risk for disease.

At NYU Langone Health, researchers are working to turn those signals into actionable insights while tackling a central challenge in biomedical AI: Models often perform well on specific datasets but fail to generalize to the real world.

To address this, NYU Langone recently completed its first DREAMx Challenge, a competition inviting teams at the institution to develop and rigorously test models that predict risk for breast cancer, coronary artery disease, and type 2 diabetes and using genomic data linked to electronic health records.

“AI and genomic science are transforming how we understand disease,” said Dafna Bar-Sagi, PhD, executive vice president and vice dean for science, chief scientific officer. “This challenge brings together expertise across disciplines to ensure these advances translate into meaningful insights for patient care.”

The effort was built on deciphEHR, a de-identified dataset combining genomic sequences with longitudinal clinical data from nearly 20,000 patients. Developed by the Center for Human Genetics and Genomics, led by Aravinda Chakravarti, PhD, the dataset represents one of the most deeply integrated genomic-clinical resources within a US health system. The challenge was led by Gustavo A. Stolovitzky, PhD, director of the Biomedical Data Science Hub and founder of the global DREAM Challenges.

“Patients know AI and genetics will shape the future of medicine,” said Dr. Stolovitzky. “Our job is to ensure these tools are accurate, fair, and trustworthy before they guide care.”

Competitors worked with polygenic risk scores, which estimate disease risk by aggregating the small effects of thousands of genetic variants. The goal is to identify patients who may benefit from earlier screening or prevention, potentially even years before symptoms appear.

But these models come with limitations. They can perform unevenly across populations and often fail to translate across health systems. DREAMx is designed to address those gaps through rigorous, real-world testing.

The 33 participating teams developed models using a training dataset and submitted predictions for validation data. Models were evaluated on two key criteria: whether genetics-based models could accurately predict high risk for disease, and whether the way teams used genetics data added value to predictions based on clinical data only.

To be deemed the best performer in each of the three disease areas, the models needed to be significantly statistically better than the other teams’ models as well as the previously published predictive models used as reference models.

One team, appropriately called GoGene, surpassed the reference models in identifying the signs of higher risk for both type 2 diabetes and coronary artery disease—by 40 percent and 10 percent, respectively. The team included Yiqiu Shen, PhD, assistant professor in the Department of Radiology at NYU Grossman School of Medicine, and his students Jiajian Ma, at the NYU Center for Data Science, and Wushuang Rui, who recently graduated from NYU with a master of science degree in biology.

For breast cancer, the KyDeciph team, consisting of Kyrillos Ibrahim, senior bioinformatics programmer from the Center for Human Genetics and Genomics, did 25 percent better than the reference model at recognizing the disease risk.

The winning teams, along with two teams who were finalists in the competition, plan to collaborate on a scientific paper to share what worked, what didn’t work, and why.

The competition was supported by NYU Langone’s high-performance computing infrastructure, including the Ultraviolet cluster and a secure cloud-based platform, Databricks, that allows researchers to analyze sensitive data without compromising privacy.

“Advancing AI in medicine requires more than powerful algorithms. It requires the infrastructure to test them rigorously and safely at scale,” said Nader Mherabi, executive vice president and vice dean, chief digital and information officer.

If the future of medicine is predictive, initiatives like DREAMx are testing whether that future can be both scientifically rigorous and clinically reliable.