A research team at St. Jude Children’s Research Hospital has leveraged AI to use a computational approach to design more effective bi-specific chimeric antigen receptor (CAR) T cells. The method, detailed in the journal Molecular Therapy, promises to overcome challenges that plague the effectiveness of bi-specific CAR Ts including poor CAR expression on the surface of CAR T cells and suboptimal cancer-killing properties.

“We have developed and validated a computational tool that can significantly accelerate the design of tandem CAR constructs with improved surface expression and anti-tumor function,” said senior author Giedre Krenciute, PhD, a member of the department of bone marrow transplantation and cellular therapy (BMTCT) at St. Jude.

The development of bi-specific antibody treatments is an approach aimed at improving cancer immunotherapies, where many current treatment forms only target one tumor-specific antigen. This approach has yielded some success, but addressing a single tumor-specific protein antigen is often not enough to effectively kill all the tumor cells.

“Chimeric antigen receptor (CAR) T cell therapy is a highly effective treatment for multiple malignancies. However, one limitation is tumor antigen-heterogeneity and downregulation, which allows tumor cells to evade conventional, monospecific CAR T cells,” the researchers wrote.

Targeting two or more cell surface antigens to address multiple drivers of cancer growth, bi-specific CAR T cells represent a logical next step in tackling tumor heterogeneity. But designing effective bi-specifics has been a challenge.

To address this, the St. Jude team developed a computational approach that allows them to screen a large number of theoretical bi-specific CAR designs and rank them to identify a small handful with the most promise for optimization and validation.

In this study, the team initially designed a bi-specific tandem CAR targeting the IL13Rα2 and B7-H3 antigens, which are often present in pediatric brain tumors. Unfortunately, their original construct failed to express on the T cell surface. But this failure led to a breakthrough.

“We describe how our originally designed constructs failed to express due to a ‘trouble region’ within the VH (variable heavy) of the IL13Rα2  scFv (single-chain variable fragment),” the researchers wrote. Through systematic testing of 24 tandem constructs and super-resolution microscopy, the researchers found that misfolding and intracellular accumulation was the likely causes of this design’s expression failure.

The team then turned to the protein design software AbLIFT to computationally modify the structure of the CAR’s variable regions, which addressed this misfolding to create greater cell surface expression. These optimized tandem CARs were then validated in vitro and in vivo, outperforming single-target CAR Ts.

“Our most compelling result is that we completely cleared tumors in four out of five mice with the CAR T cells that had the computationally optimized tandem construct,” said co-first author Michaela Meehl. “By contrast, all heterogeneous tumors treated with single-targeted CAR T cells grew back.”

Meehl is a graduate student at the University of Tennessee Health Science Center and was recently awarded a $130,000 grant from the NCI to fund three years of research to better elucidate CAR T function.

To evaluate generalizability, the researchers applied their computational approach to redesign several other tandem CARs. In each case, the optimized versions showed enhanced tumor-killing ability. “Our study also highlights the necessity of computational methods to guide the design of synthetic proteins, and that these methods can increase CAR T cell efficacy,” the researchers wrote.

To enable rapid bi-specific therapy development, the scientists trained an AI algorithm on known CAR structures, which integrated features such as folding stability, aggregation tendency, and functional motifs. Evaluations of these factors where then combined into a single “fitness score” to predict which designs would likely express and function well.

“We designed this computational tool to be broadly applicable to many different CARs,” said co-first author Kalyan Immadisetty, PhD, a member of the BMTCT at St. Jude. “It can screen roughly 1,000 constructs in a matter of days, greatly speeding up a process that would take many years if researchers created each one in the lab.”

The optimized CARs showed high efficacy even in tumors where not all cells expressed the targeted antigens. “In our heterogeneous in vivo tumor model in which 5% of tumor cells did not express either target antigen, we found that the tandem (C#24) CAR T cells still fully cleared the tumor in four of five mice,” the researchers wrote. The researchers noted that this may indicate that that bi-specific CARs induce bystander killing of antigen-negative cells when most of the tumor is still targetable.

The researchers said that the need for more computational tools and AI-driven design optimization will become more necessary that as treatments progress toward more complex synthetic proteins, such as bi-specific CARs, TRuCs, and synthetic cytokine receptors.

“Optimization of this tandem IL13Rα2–B7-H3 CAR was time-consuming, laborious, and expensive, highlighting the growing need for better methods to design new synthetic proteins,” they wrote.