Generative AI Redesigns CAR T Binders, Achieving Complete BCMA Tumor Control in Mice

CAR T-cell therapy genetically modifies a patient’s own immune cells to recognize and attack cancer. Although the approach has shown tremendous promise, particularly in blood cancers, its reach remains limited. According to a recent analysis, only about 4% of patients with advanced or metastatic cancer in the United States are currently eligible for CAR T-cell therapy, and closer to 3% are likely to benefit.

The laboratory of Dr. Caleb Lareau, a computational biologist at Memorial Sloan Kettering Cancer Center (MSK), aims to expand the impact of CAR T-cell therapy and other cancer treatments by using generative artificial intelligence to design better “binders.” These proteins recognize and attach to cancer cells, enabling the immune system to destroy them.

In a proof-of-concept study conducted in mouse models and published in Nature Biomedical Engineering, the MSK team found that its AI-designed binders significantly outperformed the binders used in existing FDA-approved CAR T-cell therapies targeting B-cell maturation antigen (BCMA).

The study also explained why current AI models struggled to design effective binders against CD19, a major target in blood cancers, and why some CAR T cells become activated prematurely in the absence of cancer cells. Such antigen-independent activation can exhaust CAR T cells before they encounter a tumor, thereby reducing therapeutic efficacy.

The ultimate goal is to replace the field’s traditional trial-and-error approach with a more detailed, evidence-based understanding of why some cancer-targeting proteins work while others fail.

According to Lareau, the same strategy could eventually enable the rapid development of individualized treatments for rare cancers. For example, when a patient has a rare blood cancer, sarcoma, or pediatric cancer for which no existing therapy is suitable, AI could potentially be used to develop a treatment capable of recognizing that patient’s cancer within weeks.

The study was led by three co-first authors: postdoctoral researcher Dr. Arthur Chow, graduate student Hoyin Chu, and research technician Ruofan “Betty” Li.

Using AI to Develop New Binding Proteins from Scratch

Several essential components are required to create an effective CAR T cell. The first is a target—a specific protein displayed on the surface of cancer cells that acts like a molecular flag. Ideally, this target should be found only on cancer cells. If it is also expressed on healthy cells, the engineered T cells may attack those cells as well.

The second component is a binder—a protein positioned on the surface of the engineered T cell that recognizes the target, attaches to it, and enables the T cell to destroy the target-bearing cell.

Figure 1. Summary of protein binder development using generative artificial intelligence. (Chow A, et al., 2026)

Current FDA-approved CAR T-cell therapies generally use binders known as single-chain variable fragments, or scFvs. These molecules are essentially repurposed fragments of existing antibodies produced by the immune system to recognize foreign substances.

AI-designed binders operate differently. Rather than being derived from existing antibodies, they are entirely new proteins designed from scratch. They can be engineered to attach to a specific region of a target protein. They are also smaller and more compact than scFvs, a characteristic that may help overcome some of the limitations of current CAR T-cell therapies.

Narrowing One Million Designs to the Best Candidates

Once the research team identifies a target protein on a particular type of cancer cell, it uses two existing AI tools to generate approximately one million potential binders and then progressively filters them to identify the most promising candidates.

The binders begin with similar three-dimensional structures inspired by naturally occurring proteins. However, each binder is encoded by a different DNA sequence and therefore has distinct physical and biochemical properties.

The designs produced by the first program are passed to a second AI model, which ranks them according to multiple criteria. Lareau compared the process to generating one million possible answers with one generative AI system and then asking a different AI model to evaluate and rank them.

This screening process reduces the initial pool to approximately 100–200 strong candidates. The number is small enough to be experimentally manageable while remaining large enough to preserve meaningful sequence and structural diversity.

The DNA sequences encoding these candidate binders are then sent to an external laboratory for synthesis, and the resulting materials can be returned in approximately one week.

Binders are proteins constructed from the 20 standard amino acids. Although candidate proteins may share highly similar overall structures, even subtle variations in their amino acid sequences can substantially affect their performance.

Generating candidate binders is therefore only half of the process. The team must also test their activity in laboratory cancer models. More importantly, the researchers aim not only to identify which binders succeed or fail but also to understand the molecular reasons for these differences.

CARPNN: An AI Model That Learns from Experimental Results

To achieve this goal, the researchers developed CARPNN, short for CAR Protein Neural Network. This customized AI tool learns from laboratory data and analyzes how subtle sequence differences among binders influence CAR T-cell performance.

The researchers found that small variations in binder properties could produce major differences in therapeutic activity. Important factors included electrical charge, amino acid composition, and the precise region of the target protein to which the binder attached.

This ability to explain why a particular binder succeeds or fails represents a level of mechanistic understanding that has previously been difficult to achieve in CAR research.

Rather than treating experimental screening as a simple process of selecting the best-performing candidate, CARPNN extracts general design principles from the results. These principles can then be applied to evaluate and improve future generations of binders.

Learning from Both Successful and Unsuccessful Designs

The researchers tested their AI-driven strategy against several targets already addressed by FDA-approved CAR T-cell therapies. These included BCMA, which is expressed on multiple myeloma cells, as well as CD19 and CD22, which are associated with other blood cancers.

The results obtained with BCMA were particularly striking. The MSK-developed binder controlled tumor growth in mice substantially better than the binder used in the FDA-approved therapy. The AI-designed binder achieved essentially complete tumor control, whereas the existing clinical product was unable to effectively restrict tumor growth in the experimental model.

However, the less successful experiments also yielded several important lessons.

First, the AI models failed to generate effective CD19 binders. This failure was informative because it revealed that current AI tools can struggle to move beyond previously established binding patterns and generate genuinely novel solutions.

Second, experiments involving CD22 showed that one of the team’s strongest binders was also prone to activating CAR T cells against healthy cells. Using CARPNN, the researchers were able to modify the binder and correct this problem.

Third, the team found that binders carrying excessive positive charge were more likely to activate CAR T cells prematurely, even when no cancer cells were present. This type of antigen-independent activation can cause T-cell exhaustion before the engineered cells reach and engage a tumor.

Fourth, the researchers observed that the generative AI models tended to favor alanine in their protein designs. A high alanine content, however, was associated with binding-related problems.

Together, these findings produced a new set of design rules involving electrical charge, amino acid composition, target-binding position, and other critical properties. CARPNN can learn from these rules and apply them when assessing future candidates.

In this way, the platform transforms conventional informed guesswork into a more systematic and evidence-based design process.

From Computer-Generated Proteins to Clinical Therapies

For Lareau, the significance of the study extends beyond the success or failure of individual binders against specific targets. Conventional development typically involves testing many candidates and selecting those that appear to perform best. In contrast, the team’s approach enables researchers to develop new CARs more systematically while generating broadly applicable rules that can guide subsequent rounds of therapeutic design.

The technology is not limited to CAR T-cell development. It could also be used to design binding proteins for other therapeutic applications, including precision drug conjugates and additional forms of targeted cancer treatment.

A major long-term objective of the laboratory is to use generative models to create new cancer drugs, advance them into clinical trials, and ultimately make them available to patients at MSK and elsewhere. The laboratory is currently approximately three years into this ten-year plan.

This work with generative AI represents a new research direction for the laboratory and reflects the type of timely, high-risk, high-reward science supported by MSK. For Lareau, pursuing AI-guided protein design was a major scientific commitment as an early-career faculty member. The striking results obtained with the BCMA-targeting binder provide encouraging evidence that the team is moving in the right direction.