Optimizing CAR Costimulatory Domains Using Contrastive Learning and Optimal Transport on High-Throughput Screening Data

Chimeric Antigen Receptors (CARs) are engineered into immune cells to bestow specificity towards identified targets on cancer cells. CAR T cells have established themselves as therapies for B cell and plasma cell malignancies, where they induce prolonged remission in most patients, though fewer than half achieve durable control of their disease. CARs follow a modular design, with an extracellular portion conferring target-specificity, a transmembrane hinge, and two or more intracellular signaling domains. One of the signaling components - typically CD3z - initiates the signaling cascades crucial for T-cell activation, proliferation, attainment of cytotoxic function, and ultimately immune memory. The other signaling component, called the costimulatory domain, is crucial for ensuring proper and controlled immune responses. Absence of this domain results in T cell anergy, but the optimal costimulatory domains remain unknown.

Intracellular immune receptor signaling domains often comprise disordered regions lacking fixed tertiary protein structure. Their primary function is fulfilled by their primary structure consisting of one or more short (3 – 10 amino acid long) conserved sequences called Eukaryotic Linear Motifs. Adhering to this design principle, we systematically identified all transmembrane proteins within the Gene Ontology database associated with immune cell differentiation, proliferation, regulation or signaling. We specifically focused on proteins with known topologies in UniProt, extracting the primary sequences of the cytosolic portions. Through this process we curated a pool of over 1,200 distinct costimulatory domains, encompassing receptors not native to T cells, including cytokine/chemokine and inhibitory receptors.

We lentivirally transduced anti-CD20 CAR with this pool of costimulatory domains into CD8⁺ T cells from 2 healthy volunteers. We cultured the resulting CAR T in triplicate, alone, activated by CD2/CD28 beads, in the presence of Raji (target) cells or K562 (negative control) cells over a 14-day period. At the conclusion of the experiment, we immunophenotypically sorted the cells into 6 sets: distinguishing central memory, effector and naïve T cell subsets each characterized by low or high PD1 expression. We prepared targeted sequencing libraries from the sorted cells to identify the costimulatory domains.

Our screen yielded several promising novel ‘hits’ that we are evaluating through arrayed experiments. Notably, CD74 displayed enrichment in both the non-exhausted memory and effector subsets. Recognized as the Major Histocompatibility Complex (MHC) class II-associated invariant chain, CD74 is primarily known for its role in facilitating the assembly and trafficking of MHC class II molecules within cells. Beyond this canonical function, CD74 also harbors intracellular signaling capabilities that contribute to immune responses, engaging the ERK1/2, PI3K/AKT, and NF-κB pathways. By contrast, 4-1BB, which is used in most commercial CAR, activates a superset of these pathways, MAPK, PI3K/AKT and NF-κB. ERK1/2 is a subset of the MAPK pathway that primarily responds to growth factors and mitogens.

Large language models have found powerful application in biology. ESM-2 (Evolutionary Scale Modeling) is a foundation model developed by Meta AI, that maps primary sequences to a representation (‘embedding’) that captures evolutionary and structural information. To gain a predictive understanding of proportions of immunophenotypic subsets resulting from different costimulatory domains, we trained a contrastive learning model to learn a shared embedding between ESM-2 representations (a 5,120-dimensional space) and our experimentally determined proportions of immunophenotypic proportions under different co-culture conditions (a 24-dimensional space). We trained two dense neural networks with 2 hidden layers to project to a 20-dimensional shared embedding using 90% of our data, we achieved a moderate Fraction of Samples Closer than the True Match (FOSCTTM) of 0.45 in our validation sample. We further refined this shared embedding with fused Gromov-Wasserstein optimal transport to achieve a respectable FOSCTTM of 0.13.

Our model predicts proportions of immunophenotypic subsets for different costimulatory domains in the context of our anti-CD20 CAR. We plan arrayed testing of domains predicted to improve those in our previous pool.