AI-assisted approach identifies IRS4 as a promising drug target in multiple solid tumors

Leveraging AI capabilities combined with genetic data led St. Jude Children’s Research Hospital investigators to predict new cancer drug targets that carry a limited risk of side effects. For their study, published in Science Advances, St. Jude Children’s Research Hospital scientists used a mix of genetic cancer dependency data, artificial intelligence (AI), and naturally occurring mutations to prioritize safer cancer drug targets.

They focused their efforts on targets most likely to be effective while limiting unwanted toxicity, identifying IRS4 as a potential dependency across multiple tumor types. The work provides a proof of principle for evaluating potential toxicity early in the search for novel therapeutics.

Of the potential cancer therapeutics that enter phase 1 clinical trials, an estimated 85%–97% do not become Food and Drug Administration (FDA)-approved, partly due to toxicity in normal tissues. Such side effects can have long-term consequences, which are of particular concern for treating pediatric cancers, where therapy-related late effects can cause chronic health problems even decades later. Despite this challenge, toxicity is often evaluated late in the drug development process, a paradigm that the St. Jude group sought to address.

“Cancer drugs have a higher failure rate in clinical trials than any other major disease,” said corresponding author Samuel Brady, Ph.D., St. Jude Department of Pharmacy & Pharmaceutical Sciences. “Historically, we’ve usually prioritized efficacy when searching for drug targets, while toxicity has often been considered later. Our work shows that we may be able to predict toxicity at the outset, allowing us to focus only on the promising lower-toxicity targets, such as IRS4.”

Finding a low-toxicity cross-cancer dependency

To identify targets with both strong anticancer effects and low toxicity, a property called a high therapeutic index, Brady and colleagues developed a multistep screening approach. They began by looking at genes in the Dependency Map portal, a database of genes that cancer cells need for survival. They then filtered for targets similar to FDA-approved targeted cancer therapies, narrowing the list to 346 potential targets. Then, they used AI to search the scientific literature for people whose natural loss of these genes causes either no or small, manageable disruptions to their normal lives.

“When treating cancer, we often have to target proteins that are also present in normal cells, and that’s where a lot of toxicity risk arises,” Brady said. “So, we used an approach to identify genes that humans or mice can survive without, reducing the likelihood that targeting them will cause serious side effects.”

This process reduced the list to 25 candidates, including several known targets, giving confidence that the approach identified good targets, along with a set of previously unexplored possibilities. Among the remaining candidates, IRS4 stood out as uniquely promising because it had a potential binding pocket for a drug to target. However, the region of the protein with the pocket was not necessary for its procancer effects. This information may guide downstream efforts to target IRS4 therapeutically, such as through protein degradation.

The researchers found that IRS4 is rarely expressed in normal adult tissues, and previous research indicates that individuals lacking the gene are mostly healthy, with manageable thyroid-related effects. When the researchers removed IRS4 or degraded the IRS4 protein, cancer cells that expressed it were unable to grow.

“You can think of IRS4 as an on-off switch,” Brady said. “When cancer cells have it, they need it. That makes it an attractive target and offers a potential biomarker to identify which tumors are most likely to respond.”

The study demonstrates that IRS4 could serve as a therapeutic target in multiple cancer types, including pediatric malignant rhabdoid tumors, osteosarcomas, and some brain tumors, as well as adult cancers such as breast, lung, uterine, and stomach cancers. More broadly, the findings highlight the potential of incorporating toxicity prediction into the earliest stages of drug discovery.

“We can already cure a large percentage of childhood cancers,” Brady said, “but many survivors face lifelong health challenges caused by treatment. Our goal is not only to cure cancer, but to ensure that patients can live full, healthy lives afterward. By prioritizing targets with a higher therapeutic index from the start, we can develop therapies that are both effective and safer.”