When doctors give a local treatment first — radiation, freezing, heat — and an immune cell therapy second, the timing matters enormously. The local treatment kills tumor cells and spills their antigens, briefly putting the tumor into a sensitized state in which infused T cells can work; given too early or too late, the same cells fail or become exhausted. Yet the receptors those T cells carry are engineered in the opposite regime: in flat plastic dishes and static binding assays, against pristine targets, with no regard for the volatile microenvironment the priming treatment creates. This note proposes to close that loop. The claim is that the best mutation for a therapeutic T-cell receptor depends on the temporal shape of antigen release produced by the priming therapy, and that neither the mutation nor the schedule can be optimized alone without leaving substantial efficacy on the table.
Two threads in the literature set up the gap. On the engineering side, Overcoming T cell tolerance to tumor self-antigens through catch-bond engineering used site-saturation mutagenesis of T-cell receptor binding loops to produce variants whose bond with the tumor antigen strengthens under mechanical pulling (a "catch bond"), extending bond lifetime from roughly 0.2 seconds for the wild type to about 5 seconds, clearing prostate tumors in mice, and resisting exhaustion in single-cell profiles. But the variant was selected in an isolated biophysical chamber and tested in a fixed xenograft protocol with no priming therapy and no timing optimization. On the scheduling side, Case report: Cryoablation as a novel bridging strategy prior to CAR-T cell therapy for B cell malignancies with bulky disease shows ablation priming works in patients but leaves infusion timing unstandardized, and Spatially optimized 125I brachytherapy: a novel immunomodulatory approach based on spatially fractionated radiation therapy principles shows a low-dose-rate brachytherapy source can release antigen continuously over weeks as a deliberate "trickle." Nobody has asked which receptor kinetics suit which release profile — and the two literatures point in opposite directions, since a catch bond needs a defined mechanical threshold and antigen density that a slow trickle may never supply, while an acute burst may overdrive it.
Core Claim
Site-directed mutagenesis of a therapeutic receptor should be conditioned on the dynamically remodeled, post-priming tumor state rather than on a static assay, and the priming schedule should be chosen jointly with the mutation. We propose to build and test a co-design loop with three pieces: a set of primed-state gene sets derived from time-course data after local therapy, used through functional enrichment analysis to represent the microenvironment as a moving target rather than a fixed one; a mutation-conditioned state model that predicts how residue-level receptor variants rewire downstream T-cell gene networks, extending existing perturbation predictors from small molecules to engineered protein sequences; and a closed-loop experimental validation pairing engineered receptors with deliberately matched and mismatched priming kinetics in mice, with single-cell profiling of the infused cells to check whether the model's predicted states occur in vivo. The in vitro–in vivo translation gap enters as the central constraint: the whole point is that a variant that looks best in a dish is not necessarily the variant that performs best inside a primed living tumor.
Research Program
The first component builds the representation of the primed state. Standard enrichment collections — Gene Ontology, KEGG, the MSigDB compendium — are derived from baseline tissue profiles and are blind to transient post-ablation states. The problem is documented: Importance of collection in gene set enrichment analysis of drug response in cancer cell lines showed that enrichment results swing dramatically with the choice of reference collection, that a data-driven co-expression collection (HGSK, built by clustering 311 cancer cell lines) captured the top enriched pathway for a drug in 40–60% of cases across 727 cell lines and 138 drugs from the Cancer Genome Project, and that collection differences reached $p = 1.24 \times 10^{-33}$ by rank-sum test. We will do the same thing for time instead of drugs: cluster post-radiation and post-ablation time-course profiles (the acute-phase cascades mapped by Peng and Li and the danger-signal kinetics with finite half-lives measured by Gameiro et al.) into gene sets keyed to a time-since-priming index, so that any single-cell or bulk profile of a treated tumor can be scored for "how primed, and in which direction" rather than only for baseline identity.
The second component is the mutation-conditioned model. DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction is the closest precedent: it links drug substructure fragments to 13 functional pathways and transfers from 225,781 cell-line samples in DepMap to 555 clinical samples in The Cancer Genome Atlas, reaching an AUC of 0.957 on clinical Sorafenib response (versus 0.811 for the next best baseline), and — critically for us — its own ablation shows that removing the pathway-level readout collapsed average AUC from 0.856 to 0.662, evidence that pathway-level representation, not raw gene lists, is what carries signal across the culture-to-patient gap. But DeepSADR takes chemical SMILES as input and models static, single-agent response; it cannot represent a residue substitution or a schedule. Cell-Level Virtual Screening supplies the other half of the mechanism: it predicts how a perturbation rewires gene-gene dependency networks, not just expression snapshots, using an amortized surrogate loss that decomposes network estimation into pairwise regressions, achieving a Hits@1 of 0.1250 on its DDR-Bench retrieval task versus 0.0179 random and 0.0357 for molecular fingerprints — but it is conditioned on small molecules and broad knockdowns, never residue-level protein variants, and was validated only on LINCS L1000 cell lines. Our model replaces the chemical input with receptor sequence embeddings of the kind validated in TCR-EML: Explainable Model Layers for TCR-pMHC Prediction, whose contact-region explanations achieved a Binding Region Hit Rate above 0.71 on the TCR-XAI structural benchmark of 274 crystal complexes and an ROC-AUC of 0.999 with ProteinBERT on top-100 unseen peptides; the training follows the proven two-stage recipe of pre-training on large in vitro screens (in the spirit of the ~10 million-variant emibetuzumab library of Co-optimization of therapeutic antibody affinity and specificity using machine learning models that generalize to novel mutational space, where models reached 93% bin-classification accuracy and Spearman correlations of 0.87 for affinity and 0.67 for specificity, and generalized to 15 CDR sites outside the training library) and fine-tuning on scarce in vivo single-cell signatures, of the kind in An open CAR-T single-cell atlas — 414,000 cells across 102 patients, with known exhaustion, memory, and toxicity-associated programs such as the ICANS-associated cells at odds ratio 6.067. The model's output is the joint objective: predicted tumor-infiltrating state of a variant, indexed by the primed-state gene sets from component one, as a function of priming modality and interval.
The third component closes the loop experimentally, described in the experiments section below.
Baselines
The "do nothing" reference point is the current clinical default: an unmodified receptor and concurrent or ad-hoc sequencing. In mice this means wild-type TCR156 T cells infused on a fixed schedule with no priming therapy; the catch-bond paper reports that the wild type fails to eliminate PC3-PAP-A2 tumors in the NSG xenograft model, so the reference point for survival is essentially no tumor control. Any variant–schedule pair that does not beat untreated wild-type controls on tumor growth inhibition and T-cell persistence is a failed design regardless of its in vitro numbers.
The strongest existing competitor is the catch-bond engineering pipeline itself, run exactly as published in Overcoming T cell tolerance to tumor self-antigens through catch-bond engineering: select variants (S32Ma, S30E32Qa) by bond lifetime under a Biomembrane Force Probe against a fixed antigen-coated surface — reported peak lifetimes around 5 seconds versus roughly 0.2 seconds for wild type, with resting 3D affinity essentially unchanged — then infuse on the static xenograft protocol that cleared tumors. This is the right comparator because it is the only published workflow that already combines mutagenesis, in vivo efficacy, and single-cell enrichment, and it achieves complete tumor clearance in its setting. We will run it with the same receptor panel, the same NSG PC3-PAP-A2 model, the same cell dose and tumor line, and the same tuning budget of assay runs per variant, the only difference being that its schedule is fixed and its selection ignores priming. Beating it means showing a variant–schedule pairing that outperforms the best fixed-schedule variant on tumor growth inhibition and on the fraction of infused cells avoiding the exhausted transcriptional program, specifically under priming conditions where its static selection misleads it.
The key ablation isolates the dynamic-enrichment ingredient: the identical co-design model with the primed-state gene sets replaced by off-the-shelf static collections (MSigDB C2/C5 or KEGG), holding the architecture, data, and compute fixed. DeepSADR's own ablation — average AUC falling from 0.856 to 0.662 when the pathway readout was removed — sets the expectation that representation choice, not architecture, carries much of the transfer; if our dynamic gene sets do not beat static ones at predicting in vivo T-cell states in held-out conditions, the central claim is wrong and we would report that. A secondary ablation drops the mutation conditioning entirely (one shared model per receptor backbone), testing whether residue-level input matters at all or whether schedule alone explains the variance.
Why not just…?
Why not just give everything at once? Concurrent administration is the logistically simple alternative, and the material argues against it on mechanism: Sun et al. warn that concurrent checkpoint delivery amid continuous radiation-driven antigen release accelerates terminal T-cell exhaustion, and Development of a 177Lu-labeled EphA2-targeting cyclic peptide combined with an HPK1 inhibitor found that targeted radiation alone suppresses T-cell activation through p-AKT and p-ERK signaling — its 67.2% tumor growth inhibition came only from adding an HPK1 inhibitor to rescue T-cell function, i.e., from correcting a sequencing-induced deficit. The vascular-normalization window documented by Yu et al. makes the same point for delivery: cells and drugs must arrive inside a transient perfusion window or they fail to reach the tumor core. Concurrent delivery is therefore a real, runnable arm in our mouse study rather than a dismissed strawman; we will include it and expect it to lose, but the experiment settles it.
Why not just affinity-mature the receptor as hard as possible? This is the industry-standard alternative, and the file contains direct evidence it hits a wall: the emibetuzumab study shows that pushing affinity through contact-site mutations drives a Pareto trade-off with non-specific binding that only a learned co-optimization broke (its EM2 variant gained 1.28-fold affinity while cutting non-specific binding by 70%). The catch-bond paper makes a sharper point for T-cell receptors specifically: its best variants left static resting affinity essentially unchanged, so a pure affinity-maturation screen would never have selected them — the gain shows up only under mechanical load at the cell-cell interface. A maximally affine TCR is also a tonic-signaling TCR, prone to exhaustion under sustained antigen, which is precisely what the priming kinetics control. We will still run an affinity-maximized variant as a comparator arm, because if it matches the co-designed variant in vivo, the dynamic-conditioning claim collapses; we predict it will show earlier exhaustion markers in the single-cell profiles.
Why not just use DeepSADR or CellVS-Net as they are? These are the strongest transfer models in the file, and we reuse their techniques rather than dismiss them — but DeepSADR's input space is chemical SMILES at a fixed baseline state (no residue-level variants, no schedules), and CellVS-Net is conditioned on small molecules and whole-gene knockdowns with only cell-line validation. Neither can answer a question of the form "does mutation S30E32Q at interval day-3 after cryoablation beat S32Ma at day-0?" Our contribution is precisely to re-parameterize their machinery — pathway-level readout, network-state prediction, two-stage transfer — around receptor sequence and priming schedule. If a reviewer believes a modest wrapper would suffice, the ablation above answers it empirically: the architecture is held fixed and only the representation and conditioning change.
Why not just run a factorial mouse experiment — a few mutants crossed with a few schedules? This is the most serious alternative, because for a handful of variants it would indeed work. The honest answer is combinatorial: variant space after site-saturation at a few hotspots is thousands to millions deep, and schedules multiply it by modality, dose, and interval, so the factorial is only feasible after the model has nominated a shortlist. Our decisive check is therefore deliberately factorial in miniature: three receptor variants (wild type, catch-bond, affinity-matured) crossed with two priming profiles (acute cryoablation burst versus continuous low-dose-rate trickle in the style of the 125I framework) and three infusion intervals, with the model's ranking predicted in advance. If the factorial outcome disagrees with the model's ranking on more than a minority of cells, the model earns no claim to search the larger space, and we say so.
Proposed Experiments
The computational evaluation comes first, on held-out data. Primed-state gene sets are scored by whether they retrieve the correct time-since-priming window from held-out post-treatment profiles, benchmarked against static collections using the same rank-based comparison logic that separated HGSK from MSigDB in the CGP screen. The mutation-conditioned network model is evaluated on retrieval tasks in the style of DDR-Bench and DTR-Bench (built from LINCS L1000 perturbations), where the reported CellVS-Net Hits@1 of 0.1250 against 0.0179 random and 0.0357 fingerprint baselines gives the scale of what counts as a meaningful margin; our claim to test is that residue-level conditioning improves retrieval of the matched downstream T-cell state relative to a sequence-blind version of the same model at equal parameter count.
The wet-lab core tests the co-design hypothesis where it matters, in vivo. In the NSG PC3-PAP-A2 prostate model used by the catch-bond study, we cross the receptor arms (wild-type TCR156; the published catch-bond variants; an affinity-matured comparator; and one to two model-nominated variants) with priming arms: acute cryoablation delivering a rapid antigen burst per the two freeze–thaw protocol of Zhang et al., and a continuous low-dose-rate trickle modeled on the 125I spatially-fractionated design of Sun et al., each at three infusion intervals. Before infusion, each variant is characterized by Biomembrane Force Probe across a range of antigen densities and forces, testing the mechanistic prediction that the burst profile favors force-strengthened variants while the trickle under-stimulates them. Primary endpoints are tumor growth inhibition (with the 67.2% TGI of the 177Lu/HPK1 combination in immunocompetent melanoma as an orienting magnitude, not a matched comparison, since our model is immunodeficient) and, secondarily, the fraction of tumor-infiltrating cells in exhaustion programs, measured by single-cell profiling analyzed with the scVI-plus-Milo differential-abundance workflow of the CAR-T atlas. The decisive readout is an interaction effect: a significant crossover in which variant wins under which priming profile, which no fixed-schedule or fixed-variant pipeline could produce by construction. A final check on the translation-gap ingredient compares each variant's in vitro ranking (BFP lifetime, static affinity, network prediction from baseline profiles) against its in vivo ranking: our claim predicts the two rankings will decorrelate specifically under priming, and that the primed-state enrichment score is what restores the correspondence — the same alignment-of-domains effect DeepSADR demonstrated when its pathway readout carried cell-line pretraining into patient response.
References
- Overcoming T cell tolerance to tumor self-antigens through catch-bond engineering
- Case report: Cryoablation as a novel bridging strategy prior to CAR-T cell therapy for B cell malignancies with bulky disease
- Spatially optimized 125I brachytherapy: a novel immunomodulatory approach based on spatially fractionated radiation therapy principles
- Importance of collection in gene set enrichment analysis of drug response in cancer cell lines
- Acute-phase plasma proteomics of rabbit lung VX2 tumors treated by image-guided microwave ablation
- Combination Therapy with Local Radiofrequency Ablation and Systemic Vaccine Enhances Antitumor Immunity and Mediates Local and Distal Tumor Regression
- DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
- Cell-Level Virtual Screening
- TCR-EML: Explainable Model Layers for TCR-pMHC Prediction
- Co-optimization of therapeutic antibody affinity and specificity using machine learning models that generalize to novel mutational space
- An open CAR-T single-cell atlas to enable in-depth characterization and rational engineering of CAR-T products
- Development of a 177Lu-labeled EphA2-targeting cyclic peptide combined with an HPK1 inhibitor for synergistic anti-tumor effects
- Vascular normalization: reshaping the tumor microenvironment and augmenting antitumor immunity for ovarian cancer