SBU Researcher Uses AI to Design Safer Personalized CRISPR Gene Therapies

Stony Brook University researcher Eric Josephs is combining massive biochemical screening with transparent machine learning to prevent off-target DNA cuts and tailor genome editing to individual patients.

Eric Josephs

Stony Brook, NY, Nov 20, 2026 - Eric Josephs, Empire Innovation Associate Professor of Biomedical Engineering at Stony Brook University, has been studying how artificial intelligence can be used to treat genetic diseases more effectively. Backed by a seed grant from Stony Brook’s AI Innovation Institute (AI3), his research recently appeared in Trends in Biotechnology, a prominent review journal. It focuses on engineering CRISPR-based gene therapies for safer genome editing.

Medicine has experienced a revolutionary shift in the treatment of genetic diseases since the FDA first approved therapeutics using CRISPR ribonucleoproteins (RNPs). Short for Clustered Regularly Interspaced Short Palindromic Repeats, CRISPR was originally discovered as a natural immune defense system in bacteria and archaea.

It uses a guide (RNA) molecule and a cutting enzyme (Cas9) to recognize the right target site in a cell’s DNA sequence, match the region with a given genetic sequence as well as its variants, and remove and replace these sequences to restore normal cell function. A kind of semantic “find-and-replace” operation to fix genetic mistakes. Today, researchers are using CRISPR to cut the section of a cell’s DNA that’s causing the disease by altering it in a highly specific way without killing the cell. 

Josephs said, “There has been a strong movement by the U.S. government and industry to broaden the use of this technology to treat many different chronic diseases, starting with recently FDA-approved treatments for sickle cell disease and a rare liver disorder in newborn babies.” But moving these tools from a controlled laboratory environment into diverse human populations introduces serious risks.

Human DNA is billions of nucleotides long, and while everyone carries harmless natural variations across populations, a single person’s genome also contains many healthy sequences that look nearly identical to a disease-causing target. Because bacteria evolved CRISPR to fight rapidly mutating viruses, its components can tolerate slight mismatches and get tricked by these similar-looking sequences—accidentally cutting healthy DNA and introducing new mutations that can trigger serious side effects, including cancer. In their article ‘On-target and off-target activities of CRISPR therapeutics across scales,’ Josephs and his collaborators from Stony Brook, Clemson University, and the University of North Carolina at Greensboro discussed the challenges with existing solutions.

Josephs explained, “To reduce off-target errors, researchers have engineered variants of the Cas9 enzyme that loosen its grip on DNA. So if there is a significant enough mismatch between the genetic sequence about to be cut and the guide RNA molecule, the molecule has enough time to fall off before the enzyme makes the cut.” This causes two issues: weaker binding often dulls the therapy's potency at the actual disease target, and because the solution is not tailored to a specific DNA sequence, a single-letter mismatch at an off-target site can still sometimes trigger an accidental cut.

Addressing a pressing need for highly personalized CRISPR gene therapies, Josephs’ team generated from 65,000 up to 2 million CRISPR combinations by introducing simultaneous changes to both the Cas9 protein and the guide RNA, and then tested these variants against 110 intended DNA targets. “The goal was to identify which variations behave the way we want them to, so that we can continue to investigate them further as safer candidates in the development of these therapeutics.”

These enzyme-guide combinations generated roughly seven million experimental measurements—a dataset too vast for human inspection alone. Josephs partnered with AI Consultant Deboparna Banerjee through Stony Brook’s AI Innovation Institute to decipher the results, training interpretable machine learning models (models whose decision-making steps researchers can trace) to understand exactly how specific variations in the Cas9 protein interact with individual letters along the guide RNA. 

“Rather than forcing researchers to run massive, multi-million-variant laboratory screens from scratch for every new disorder,” Josephs said, “These AI models will help establish the underlying biophysical rules of how these enzyme-guide mutations work together.” While the team has completed the initial sequencing runs, the machine learning analysis is still underway, with researchers actively training and refining the models to extract those predictive rules from the dataset.  

These human-understandable and scientifically insightful rules, rather than black-box predictions, are critical to converting the technology from a fox into a hedgehog—from a versatile laboratory research tool ("foxes" that know many things) into specialized human therapies ("hedgehogs" that do one big thing safely). 

Josephs has received a small grant from the New York Center for Biotechnology to continue to develop AI tools for these goals. Looking ahead, he points to a clinical pipeline where predictive design rules work hand in hand with patient-specific genome sequencing, allowing researchers to check a patient's DNA for hidden off-target risks and computationally tailor a custom CRISPR therapy before treatment begins.

To extend the project, Josephs is collaborating with AI3 Affiliate Faculty Chenyu You, assistant professor in the Department of Applied Mathematics & Statistics, and Dongyan Tan, associate professor in the Department of Pharmacological Sciences at the Renaissance School of Medicine. By establishing transparent design principles, the Stony Brook researchers are laying the groundwork for gene therapies that can be safely adapted to individual patients and rare genetic conditions alike.

News Author

Ankita Nagpal