Abstract Details

F3 - Cancer - Targeted Gene and Cell Therapy

330: Discovery of Novel CARs for Solid Tumors Using Senti REVEAL™, a Massively Parallel Technology Platform Comprising Pooled Screening, Machine Learning, and Lab Automation

Type: Oral Abstract Session

Presentation Details

Session Title: Targeted Gene and Cell Therapy I






Introduction. CAR-T and CAR-NK therapies have limited efficacy in solid tumors due to multiple factors including the immunosuppressive tumor microenvironment and the limited potency of the signaling domains of current approved CAR therapies. More potent CAR signaling domains may address this limitation by activating different pathways (or combinations) that increase robustness to the immunosuppressive tumor microenvironment. FDA-approved CAR therapies for liquid tumors tend to use a limited set of intracellular domains, highlighting a need for new signaling domains for solid tumors. Conventional approaches of screening and optimizing CAR designs are manual labor intensive, low throughput, and can suffer from operator and batch error.
Methods. We developed a massively parallel technology platform, REVEAL™ (Research Engine for Validation of Engineered Asset Libraries), comprising (1) pooled screens of CAR libraries of 10,000s to 100,000s of signaling domains, (2) machine learning models to predict performance, and (3) parallel clonal validation of 100s to 1,000s of candidates using a purpose-built automated liquid handler. Here we used CAR-NK cells, but we anticipate that this framework can be used for other types of effector cells such as CAR-T.
We designed libraries of CARs with a fixed extracellular domain targeting CEA, an antigen expressed on a range of solid tumors including colorectal cancer, and over 50,000 combinatorial signaling domains consisting of arrays of subdomains derived from native receptors. High-performing CARs in the libraries were discovered using a sort-seq approach: cells expressing CARs driving robust degranulation (CD107a+) were sorted, and CARs enriched in these cells were identified via NGS. We used these data to train machine learning models mapping CAR structure (i.e., the position and identity of constituent subdomains) to performance. In addition to reproducing experimental CAR performance, these models were used to predict CAR structures not present in the dataset due to cloning or coverage constraints. The highest performing library CARs and model-predicted CARs were then clonally validated in parallel using a liquid handler-based workflow that automated viral production, NK transduction, co-culture assay setup, and imaging-based readout of target killing.
Results. Our library screens and modeling indicate, and our clonal validations confirm, that signaling domains enriched for TRAF-binding motifs have high performance in the context of anti-CEA CAR-NK cells. Among highest performing CARs, however, TRAF domains are typically accompanied by other functional motifs, constituting a more diverse signaling array than would be found in a single native immune receptor. Successful CARs often comprised multiple canonical immune signaling types, i.e., signal 1 (CD3z), signal 2 (one or more costimulatory motifs) and signal 3 (cytokine signaling), demonstrating that a variety of signaling features can be successfully packaged into a single compact CAR. Additionally, these CARs contained sequences from a wide variety of sources outside of canonical CAR domains such as CD3z, CD28, 4-1BB, indicating the untapped potential for designing new CARs for solid tumors.
Conclusion. The REVEAL™ framework described here is capable of extensive optimization of CAR structures, and could potentially be applied to development of potent CAR cell therapies targeting solid tumors and hematologic malignancies beyond B cell cancers.

Tony Hua, Marcus Gainer, Brian Garrison, Kanya Rajangam, Timothy K. Lu, Nicholas W. Frankel

Senti Biosciences, South San Francisco, CA"