Chimeric antigen receptor (CAR) T cells are immune cells reprogrammed with a synthetic receptor that directs them to attack a patient's tumor. A decisive limit on their clinical benefit is persistence: whether the infused cells keep surviving, killing, and resisting burnout inside the hostile tumor environment. That durability is governed in large part by how the receptor's internal signaling segments are chained together — and the possible chains number over 100,000, of which a wet lab can realistically build and test under 10% (Modeling CAR Response at the Single-Cell Level Using Conditional Optimal Transport, Enhanced Prediction of CAR T-Cell Cytotoxicity with Quantum-Kernel Methods). Neural architecture search (NAS) faced an almost identical wall a decade ago — millions of candidate network designs, each absurdly expensive to train — and answered it with cheap proxies that score a design without training it, plus budgeted rules for deciding when an expensive evaluation is worth paying for (EZNAS: Evolving Zero-Cost Proxies For Neural Architecture Scoring, Rethinking Performance Estimation in Neural Architecture Search).
This proposal carries that machinery across domains. The plan is to build a search engine over CAR signaling architectures in which the "expensive full training" rung is a wet-lab assay or a heavy single-cell generative model, a free "zero-cost" rung scores receptors from sequence alone, and a cheap binary classifier sits in between; a policy decides, under a fixed budget, which candidates earn promotion to a more expensive rung. Both sides gain. Cell therapy gets a principled way to pick the few constructs worth building from a space nobody can screen. NAS methodology gets a testbed where its most stubborn open problem — proxies that can filter bad designs but cannot tell good from best — decides whether a patient's cells persist.
Core Claim
A fidelity ladder with an explicit budget-allocation policy can identify persistence-enhancing CAR signaling designs with roughly an order of magnitude fewer expensive evaluations than proxy-only scoring, oracle-only ranking, or random selection, while blocking designs that carry toxicity risk. The same algorithm, run on a benchmark where every candidate's true performance is known, directly attacks the documented "top 10%" resolution ceiling of zero-cost proxies (EZNAS: Evolving Zero-Cost Proxies For Neural Architecture Scoring). The biological payoff targets exactly the clinical failure mode that matters: the field already knows that memory-like rather than exhausted infusion products predict complete responses (An open CAR-T single-cell atlas to enable in-depth characterization and rational engineering of CAR-T products), and that armoring constructs with survival cytokines extends lifespan (XCL1-secreting CEA CAR-T cells enhance endogenous CD8+ T cell responses to tumor neoantigens to confer a long-term antitumor immunity, 323 TALEN®-based gene edited iPSC-derived NK (iNK) cells demonstrate enhanced antitumor activity); what is missing is a way to search the receptor-architecture space systematically instead of one construct at a time.
Research Program
The first workstream formalizes the receptor space as a discrete combinatorial object mirroring a NAS cell: a fixed number of structural positions, each filled by one of 13 signaling motifs (CD28, 4-1BB, CD40, LAT, IRAK1, CD3ζ, and so on) or a null segment, following the experimental design of the 246-construct dataset (Enhanced Prediction of CAR T-Cell Cytotoxicity with Quantum-Kernel Methods). The null segment plays the role of the skip-connection in NAS, and this is not a loose analogy: on NAS-Bench-201, differentiable relaxation methods collapsed onto parameter-free skip-connections and underperformed random selection (NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search), and a relaxed receptor search would predictably collapse onto truncated, non-signaling receptors. We therefore commit to discrete search throughout and treat this documented failure as a design constraint, not a research question.
The second workstream builds the fidelity rungs. Rung 0 is a zero-cost score computed from the receptor's amino acid sequence and motif grammar alone — analogous to the uninitialized network statistics that EZNAS evolves into scoring formulas, and motivated by the older finding that a network's wiring, even with random weights, already encodes much of its capacity (On Random Weights and Unsupervised Feature Learning). Rung 1 is a budgeted classifier that predicts high-versus-low tumor killing from motif composition, with the classical support vector machine from the quantum-kernel study (F1 = 0.73) as the reference implementation (Enhanced Prediction of CAR T-Cell Cytotoxicity with Quantum-Kernel Methods). Rung 2 is the high-fidelity oracle: the conditional optimal-transport model of Modeling CAR Response at the Single-Cell Level Using Conditional Optimal Transport, which maps a receptor sequence to a predicted distribution over an 82-gene panel defining cytotoxicity, memory, and exhaustion states across repeated antigen stimulation, and which reached an in-distribution $R^2 > 0.95$ on its 31-construct library. Rung 3 is the actual repeated-antigen-stimulation assay with single-cell sequencing at days 0, 6, and 12 — the unit of cost that makes this a genuinely budgeted problem, since each construct consumes weeks of wet-lab effort. We deliberately do not re-architect the transport model itself; unlike the oracle, its architecture is not the bottleneck.
The third workstream is the budget-allocation policy, which is the proposal's central technical contribution. We formalize the choice as a sequential assignment problem:
$$\max_{\pi}\;\mathbb{E}\big[\mathrm{recall}@k\big]\quad\text{subject to}\quad \sum_{x} c_{\pi(x)} \le B,$$
where the policy $\pi$ assigns each candidate receptor $x$ to the most expensive rung it reaches, $c_j$ is that rung's cost (sequence scoring is free, classifier inference is cheap, a transport-model rollout is moderate, one assay is weeks), $B$ is the total budget, and recall@$k$ is the fraction of the truly best $k$ persistence designs that end up in the final selected set. The form follows Budgeted Performance Estimation, which recast training-budget selection as an optimization problem and cut AmoebaNet's 3,150 GPU-day search to 0.33 GPU days (Rethinking Performance Estimation in Neural Architecture Search); here the slashed resource is assays rather than GPU time. The policy exploits a structural parallel that no one has used: stimulation timepoints are the biological analog of training logs. Just as early-epoch covariates tighten comparisons between model runs (Can Training Logs Make Model Comparisons More Precise?), a candidate's predicted day-0-to-day-6 trajectory carries information about its day-12 endpoint, so partial-stimulation predictions can serve as a cheap stopping signal before the full horizon is simulated or measured.
The final workstream adds the safety constraint that pure accuracy-driven search would ignore. Using the atlas's finding that hybrid cell states co-expressing T cell and monocyte markers predict severe neurotoxicity with an odds ratio of 6.067 (An open CAR-T single-cell atlas to enable in-depth characterization and rational engineering of CAR-T products), the policy penalizes any candidate whose predicted state distribution drifts toward that signature. Selection is then a two-objective problem — maximize persistence, bound toxicity — for which we adapt the pairwise-comparison multi-objective evolutionary machinery already developed for accuracy-versus-latency trade-offs in NAS (A Pairwise Comparison Relation-Assisted Multiobjective Evolutionary Neural Architecture Search Method With Multipopulation Mechanism).
Baselines
The do-nothing reference point is what the field does now: pick constructs to build essentially at random or by hand, at the sub-10% screening rate the wet lab can sustain (Enhanced Prediction of CAR T-Cell Cytotoxicity with Quantum-Kernel Methods). Concretely, we simulate uniform random selection of the assay batch and report the expected recall@$k$, which is vanishingly small against a 100,000-variant space and serves as the floor any method must clear to justify its existence.
The strongest existing competitor is conCAROT-guided selection: the conditional optimal-transport model of Modeling CAR Response at the Single-Cell Level Using Conditional Optimal Transport used as a standalone ranker and active-learning engine, with the next assay batch chosen by predicted day-12 persistence score plus prediction uncertainty, and no cheaper rungs. This is the strongest baseline because it is the only model validated to predict distributional single-cell responses to receptor perturbations in this space. We run it with the same total budget, the same assay slots, the same tuning effort on the transport model, and the same held-out constructs, so any difference is attributable to the ladder rather than to unequal compute. Beating it means higher recall@$k$ of the true top-$k$ persistence designs at equal assay budget, and higher Spearman rank correlation against day-12 outcomes on held-out constructs. A second serious baseline is random search with a well-tuned discrete evolutionary loop, because on NAS-Bench-201 random search and regularized evolution beat differentiable methods and are far from trivial to improve upon (NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search); we run it at matched evaluation budget on every benchmark.
The key ablation isolates the ladder itself: identical total budget, identical rungs, but all spend concentrated at a single fidelity — proxy-only search (rung 0/1 throughout) and oracle-only search (rung 2 throughout). If single-rung allocation matches the adaptive policy, the proposal's central claim is wrong, and the honest conclusion would be that the field needs only a better oracle, not a search engine around it. We also ablate the toxicity penalty by running the full policy unconstrained, to measure how many top-persistence designs it would have surfaced that carry the IAC signature.
Why not just…?
Why not just screen the whole library with pooled assays? The combinatorial space exceeds 100,000 variants and is explicitly called impossible to test exhaustively, with laboratories limited to under 10% of it (Modeling CAR Response at the Single-Cell Level Using Conditional Optimal Transport, Enhanced Prediction of CAR T-Cell Cytotoxicity with Quantum-Kernel Methods); pooled bulk screens additionally average over the cell-state diversity that determines persistence, which the transport-model paper shows is exactly the information summary statistics destroy.
Why not just run the expensive model on all 100,000 candidates, since inference is cheap? This is the sharpest objection and we take it head-on. Enumerating conCAROT forward passes is affordable, but the model was validated in-distribution ($R^2 > 0.95$ on its 31-construct library); nothing in the file establishes that its rankings remain reliable far outside that library, and the budgeted-estimation literature documents precisely this failure — one-shot estimators track local, early rankings while losing global correlation at full training (Rethinking Performance Estimation in Neural Architecture Search). The proposal's leave-one-construct-out experiment measures the oracle's rank stability directly. If the oracle proves uniformly reliable across the full space, the computational contribution shrinks to the active experimental loop — and the experiments are designed so that we would report that outcome rather than argue around it.
Why not just train a standard regressor on summary statistics? The classical support vector machine plateaus at F1 = 0.73 in the sparse-regime positions of the 246-construct dataset (Enhanced Prediction of CAR T-Cell Cytotoxicity with Quantum-Kernel Methods), and binary labels cannot represent the memory-versus-exhaustion mixture that defines persistence (Modeling CAR Response at the Single-Cell Level Using Conditional Optimal Transport). The proposal uses such a classifier — but only as a cheap middle rung that is allowed to be coarse, not as the final arbiter.
Why not just use DARTS or a super-network relaxation? Because the failure mode is already documented on both sides: relaxation collapses onto parameter-free skip-connections and underperforms random search on NAS-Bench-201 (NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search, Differentiable Neural Network Architecture Search), and the analogous biological collapse is a truncated, non-signaling receptor. We commit to discrete search and benchmark it against relaxation as a reported negative control rather than a candidate.
Why not just random search, since it beats many NAS methods? Random search is a real threat and is included as a first-class baseline on every benchmark. The counter-evidence is that unbudgeted evaluation is where the cost lives — AmoebaNet needed 3,150 GPU days before budgeting cut it to 0.33 (Rethinking Performance Estimation in Neural Architecture Search) — and in this domain the unit of waste is an assay, not a GPU hour. If random search matches the ladder at matched budget on both NAS-Bench-201 and the simulated assay loop, the proposal fails, and that comparison is a stated decisive check.
Proposed Experiments
The first experiment validates the allocation algorithm where ground truth is exhaustive. On NAS-Bench-201 — 15,625 architectures with every training result pre-computed, best CIFAR-10 accuracy 94.37% (NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search) — we run the ladder against EZNAS-A (Kendall $\tau = 0.65$ on CIFAR-10, versus 0.57 for NASWOT and Synflow), against random search, and against regularized evolution at matched evaluation budgets (EZNAS: Evolving Zero-Cost Proxies For Neural Architecture Scoring). The headline metric is recall@$k$ restricted to the top 10% of architectures, because that is precisely where EZNAS concedes proxies stop resolving differences; global Kendall $\tau$ is reported as context. Passing means matching or beating $\tau = 0.65$ globally while measurably improving top-decile recall.
The second experiment is the core biological test on the 31-construct repeated-antigen-stimulation library. Holding out constructs in a leave-$k$-out scheme, the ladder (trained only on the retained constructs and never shown the held-out sequences at high fidelity) must rank the held-out receptors, and its ranking is scored by Spearman correlation against the measured day-12 composite of the 82-gene persistence panel, against conCAROT-only and proxy-only baselines under identical compute. The same experiment doubles as the oracle-reliability check: if conCAROT's held-out rank correlation degrades with distance from the training library, that quantifies exactly the off-distribution gap the budget policy is built to manage.
The third experiment runs the full simulated assay loop on the 246-construct cytotoxicity dataset, treating each ground-truth label as an assay with a real cost. With budgets of 32 and 64 queries, we measure recall@$k$ of true high-cytotoxicity constructs under the ladder versus random selection, proxy-only, and oracle-guided active learning, with the classical support vector machine (F1 = 0.73) and the projected quantum kernel (median F1 = 0.75, peak 0.81) reported as calibration for what each rung can and cannot resolve (Enhanced Prediction of CAR T-Cell Cytotoxicity with Quantum-Kernel Methods). The toxicity-constrained variant of the policy is evaluated by whether its selected designs avoid predicted enrichment of the IAC signature drawn from the atlas (An open CAR-T single-cell atlas to enable in-depth characterization and rational engineering of CAR-T products). The decisive checks, in order: the ladder must beat random search and single-rung ablations at matched budget on both NAS-Bench-201 and the assay simulation, and the leave-one-construct-out result must show either reliable off-distribution oracle rankings (in which case the contribution narrows, honestly, to the experimental-budget loop) or a measurable reliability gap that the policy demonstrably exploits.
References
- Modeling CAR Response at the Single-Cell Level Using Conditional Optimal Transport
- Enhanced Prediction of CAR T-Cell Cytotoxicity with Quantum-Kernel Methods
- EZNAS: Evolving Zero-Cost Proxies For Neural Architecture Scoring
- Rethinking Performance Estimation in Neural Architecture Search
- An open CAR-T single-cell atlas to enable in-depth characterization and rational engineering of CAR-T products
- XCL1-secreting CEA CAR-T cells enhance endogenous CD8+ T cell responses to tumor neoantigens to confer a long-term antitumor immunity
- 323 TALEN®-based gene edited iPSC-derived NK (iNK) cells demonstrate enhanced antitumor activity
- NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search
- On Random Weights and Unsupervised Feature Learning
- Can Training Logs Make Model Comparisons More Precise?
- A Pairwise Comparison Relation-Assisted Multiobjective Evolutionary Neural Architecture Search Method With Multipopulation Mechanism
- Differentiable Neural Network Architecture Search