Cancer and Immunotherapy · global
Redesigning CAR-T Recognition Proteins: AI Designs Show Better Tumor Control in Mouse BCMA Experiments
Immune cells’ ability to fight cancer effectively depends on how they recognize their targets and when they activate. A new study used AI to design CAR-T binding proteins, improving tumor control in mice and identifying clues to premature activation from unsuccessful designs; efficacy and safety in humans have yet to be verified.
Enabling immune cells to recognize cancer is only the starting point for treatment. If the recognition machinery activates before encountering a tumor, the cells may become exhausted prematurely, weakening their subsequent attack. A research team at Memorial Sloan Kettering Cancer Center (MSK) in the United States is trying to use artificial intelligence to redesign the recognition proteins in CAR-T cells, while identifying which molecular features allow them to act at the appropriate time.
The study was published in *Nature Biomedical Engineering* on September 9, 2026. According to MSK, in mouse tumor experiments targeting BCMA, CAR-T cells equipped with AI-designed binding proteins performed significantly better at controlling tumor growth than comparator designs using binding proteins from existing FDA-approved therapies. BCMA is an important therapeutic target in multiple myeloma; these results represent proof of concept in animal models and do not yet support the conclusion that the new designs are more effective in human patients.
CAR-T therapy genetically modifies T cells to carry chimeric antigen receptors that recognize targets; the binding protein within each receptor serves as the recognition component that makes contact with cancer cells. The team designed new proteins from scratch, incorporated them into receptors, and tested their function in vitro and in animal models. A research briefing in the same journal noted that BindCraft generated the batch of binding protein designs with the highest success rate in the study, while ProteinMPNN’s sequence design approach inspired the team’s development of the CARPNN workflow.
The experiments revealed that a protein’s ability to bind its target is insufficient to predict the performance of the CAR-T cell as a whole. The researchers found that when binding proteins carried excessive positive charge, the cells were more likely to activate prematurely in the absence of cancer cells. Amino acid composition and binding location also affected function, extending the design question from “Can it bind the target?” to “How does the cell respond after binding?”
Results for other targets further demonstrated the limits of this approach. According to reports from MSK and Medical Xpress, some CD22 designs triggered activation upon contact with healthy cells. After modifying the binding proteins with CARPNN, the team corrected this problem observed in the experiment; however, AI failed to produce effective CD19 binding proteins. Examining successes alongside failures allowed the researchers to turn experimental results into a basis for the next round of design.
The value of this work lies in connecting protein design with validation of cellular function, rather than selecting candidate molecules solely on the basis of computational predictions. Before advancing to clinical use, the new proteins’ recognition specificity, sustained activity, and safety in humans still need to be established. Tumor control in mice and the correction of abnormal activation in a single design are insufficient to demonstrate that the entire platform can already prevent damage to healthy tissue.
MSK also disclosed that a provisional patent application has been filed based on the research findings, and that study leader Caleb Lareau serves as a consultant to Cartography Biosciences. For now, these findings offer a research direction for improving the recognition component of CAR-T cells; whether they can become a treatment available to patients remains a question for subsequent clinical development.