Synthetic synNotch-gated CAR T cell circuits enable next-generation immunotherapy by programming T cells to discriminate between cancer cells and healthy tissue. In collaboration with Prof. Hernández-López's lab at Stanford University, SDLabs optimizes the multi-dimensional circuit design space to maximize tumor cytotoxicity while maintaining safety and T cell durability.

The optimization landscape presents a three-way trade-off between killing power, patient safety, and therapeutic durability. Tighter receptor binding and stronger expression boost tumor killing but also increase off-target toxicity and accelerate T cell exhaustion. The choice of costimulatory signaling domain fundamentally shapes whether the engineered T cell prioritizes speed (CD28) or endurance (4-1BB). The synNotch gating receptor acts as a tunable antigen-density filter, but its affinity must be carefully calibrated — too tight and it triggers on healthy tissue, too loose and it misses tumors. See the full landscape description below.

CAR-T therapy has transformed blood cancer treatment, but solid tumors remain a challenge. The target antigens on solid tumors (such as HER2) are also present on healthy cells — just at lower levels. Conventional CAR-T cells cannot distinguish high-density (tumor) from low-density (healthy) antigen expression, causing dangerous off-target toxicity. A two-step synNotch-CAR circuit addresses this by gating CAR expression on antigen density, but the circuit's performance depends on at least six interacting design variables spanning receptor affinity, expression level, hinge geometry, and signaling architecture. Exhaustive screening of this space is impractical with cell-based assays that take days per condition.

The Chimera hierarchy navigates the three-way trade-off automatically. The optimizer first satisfies tumor cytotoxicity (the top-priority objective), then within the tolerance band it improves healthy cell sparing, and finally maximizes T cell persistence. This mirrors the clinical priority: a therapy must kill the tumor effectively, then be safe, then last long enough to prevent relapse. The Gaussian Process model learns the non-linear interactions between receptor affinities, expression levels, and signaling domains — discovering, for example, that a moderate-affinity CAR paired with 4-1BB costimulation achieves nearly as much killing as an aggressive CD28 design while dramatically improving persistence.

In 20 experiments (5 iterations of 4), the optimizer reaches >83% tumor cytotoxicity, >50% healthy cell sparing, and >35 days persistence. The Chimera hierarchy drives a clear three-phase behavior: early iterations maximize cytotoxicity by exploring high-affinity CAR and strong promoter combinations. Once cytotoxicity stabilizes within the 10% tolerance, the model shifts to improving sparing — learning that moderate synNotch affinity (~200 nM) and lower promoter strength dramatically reduce off-target activation. In the final phase, the optimizer discovers that switching from CD28 to CD28+4-1BB costimulation preserves most of the killing power while extending persistence by 50%. The progress chart shows cytotoxicity plateauing by iteration 2, sparing improving through iteration 4, and persistence climbing as the optimizer exploits the costimulatory domain insight.

Objective convergence — Cytotoxicity stabilizes first, then Sparing and Persistence improve as Chimera shifts focus:

Parallel coordinates — Visualizing the three-way trade-off across all evaluated circuit designs:

The core challenge is a three-way trade-off between killing power, patient safety, and therapeutic durability. No single parameter setting maximizes all three objectives — the optimizer must find the best compromise.

CAR receptor affinity is the primary lever for tumor killing, but it comes at a direct cost to safety. Tighter-binding CARs (lower Kd) activate more readily on tumor cells, but they also begin to recognize the low levels of antigen present on healthy tissue. The optimizer discovers that a moderate affinity around 20–30 nM provides strong killing while preserving the selectivity window. Very high-affinity CARs (Kd < 5 nM) are a deceptive local optimum — they score well on cytotoxicity but fail the safety objective.

The synNotch receptor acts as a tunable antigen-density gate. Its affinity determines the threshold at which the circuit "switches on." A synNotch with moderate affinity (~200 nM Kd) creates a sharp sigmoidal response: silent on healthy cells (low antigen density) but fully active on tumors (high density). Too tight a synNotch (Kd < 100 nM) fires on healthy cells too, collapsing the safety margin. Too loose (Kd > 400 nM) and the circuit fails to activate even on tumors. The sweet spot around 150–250 nM is where the antigen-density discrimination is strongest.

Promoter strength creates a direct tension between potency and durability. Stronger promoters drive higher CAR expression upon activation, boosting tumor killing. But they also increase the risk of "tonic signaling" — low-level CAR expression even without antigen engagement — which chronically stimulates the T cell and accelerates exhaustion. The optimizer learns that moderate promoter strength (0.3–0.5 a.u.) provides sufficient killing while avoiding the exhaustion trap.

The costimulatory domain fundamentally shapes the speed-vs-endurance trade-off. CD28 delivers rapid, powerful activation and the highest immediate cytotoxicity, but it drives T cells toward terminal differentiation and exhaustion within 2–3 weeks. 4-1BB promotes memory T cell formation and can sustain functional T cells for over 40 days, but at the cost of slower, weaker initial killing. The dual CD28+4-1BB design emerges as the Pareto-optimal choice — capturing most of CD28's killing power while retaining much of 4-1BB's persistence advantage.

Hinge length has a moderate effect on immune synapse formation. An intermediate hinge (~45 amino acids) provides optimal spacing for the CAR to engage the target antigen. Shorter hinges restrict access to membrane-proximal epitopes, while longer hinges can reduce signal transduction efficiency. The effect is less dramatic than affinity or costimulation choices, but it fine-tunes performance at the margins.