Clonotype-Anchored Batch Harmonization for Tracking Abscopal Immunity

When surgeons or radiation oncologists destroy one tumor with focused radiation or heat, patients occasionally see distant, untreated metastases regress too — the abscopal effect. The mechanism is an in-situ vaccination: local immunogenic cell death releases tumor antigens, dendritic cells carry them to draining lymph nodes, and newly primed CD8+ T cells circulate out to attack lesions the physician never touched. Combinations of local ablation with anti-PD-1 antibodies make this reproducible in mice and meaningful in the clinic — adding radiotherapy to checkpoint inhibitors roughly doubles the response rate outside the irradiated field in metastatic lung cancer (Efficacy and safety of immune checkpoint inhibitors with or without radiotherapy in metastatic non-small cell lung cancer: A systematic review and meta-analysis), and stereotactic radiation plus anti-PD-1 eradicated distant tumors in 50% (6 of 12) of mice in a bilateral liver cancer model (Stereotactic body radiation therapy (SBRT) increases anti-PD-1 antitumor activity by enhancing the tumor immune microenvironment in mice with metastatic hepatocellular carcinoma).

The bottleneck is measurement, not mechanism. The abscopal response is a trajectory: naive T cells in the draining lymph node, effector expansion in blood, infiltration and either exhaustion or killing in distant lesions. The only datasets that capture it — paired single-cell RNA and T-cell-receptor sequencing across tumor, lymph node, and serial blood draws, such as the 21-patient neoadjuvant HCC cohort of Wang et al. (576 Comprehensive single cell sequencing analysis of paired tissue and blood samples from HCC patients treated with neoadjuvant anti-PD-1 therapy reveals treatment-induced clonal T cell dynamics) — are also maximally contaminated by technical batch effects: each tissue is dissociated, prepared, and sequenced under different conditions on different days. Naively integrating such data is a trap in both directions. Treat anatomical compartment or sampling day as a "batch" to remove, and standard harmonization methods erase the very trajectory you are trying to measure; refuse to integrate, and technical noise drowns the rare neoantigen-reactive clones that mediate the effect.

Core Claim

The two nuisance structures in abscopal data must be separated, and current methods cannot separate them. Technical batch — sequencing run, dissociation date, platform — is noise to remove. Anatomical compartment and time — lymph node versus blood versus distant tumor, day 0 versus day 7 — is the signal, a continuous biological gradient. I propose an integration method, clonotype-anchored harmonization, that uses the shared T-cell receptor sequence as a natural pairing signal to remove technical noise while provably protecting the rare expanding clones and the continuous activation-exhaustion gradient that define abscopal immunity.

The anchor works like the "cross-well sampling" trick that made lightweight imaging models beat billion-parameter baselines: Huang et al. trained a masked Siamese network by pairing cells that received the same perturbation but sat in different wells, so that only plate-level noise differed between the pair (Efficient Cell Painting Image Representation Learning via Cross-Well Aligned Masked Siamese Network). In the abscopal setting the analogue is exact: two cells with the identical TCR clonotype, sampled from the same compartment and time point but processed in different technical batches, differ only by technical noise. Pulling those pairs together removes batch effects without touching compartment or time structure, because both are held fixed within the pair. This is the pairing signal that no imaging or transcriptomic integration method has exploited.

Research Program

The method is a conditional autoencoder $f_\theta$ that embeds each cell $x_i$ into a latent code $z_i$, conditioned on the biological condition $s_i$ = (compartment, day, treatment arm) so that condition structure is preserved by construction rather than removed. Let $b_i$ denote the technical batch, $c(i)$ the TCR clonotype, and $\mu_c$ the centroid of clone $c$ in latent space. Training minimizes

$$\mathcal{L} = \mathcal{L}_{\text{rec}} \;+\; \lambda_a \!\!\sum_{\substack{c(i)=c(j),\; s(i)=s(j)\\ b(i)\neq b(j)}} \!\!\lVert z_i - z_j\rVert^2 \;+\; \lambda_r \sum_c w_c\,\mathcal{D}(\mu_c) \;+\; \lambda_m \mathcal{L}_{\text{corr}}$$

The four terms do distinct jobs. $\mathcal{L}_{\text{rec}}$ is standard reconstruction. The anchor term $\lambda_a$ pulls same-clone, same-condition, cross-batch pairs together — this is the cross-well analogue and the proposal's key ingredient. The rare-clone protection term weights each clone centroid by $w_c$, its measured expansion fold-change between time points, and penalizes collapse of expanding clones ($\mathcal{D}$ measures how distinguishable $\mu_c$ remains from neighboring clone centroids); this targets the documented failure mode in which post-hoc and adversarial methods cannot recover a signal the upstream encoder already discarded — Nguyen et al. state this limitation explicitly for their scBatchProx adapters (Federated-inspired Single-cell Batch Integration in Latent Space). Finally, $\mathcal{L}_{\text{corr}}$ is a correlation-preserving objective in the spirit of the Corr-MSE loss of Yi et al., which was needed precisely because batch-mixing metrics reward methods that collapse intra-cell-type variation (Benchmarking deep learning methods for biologically conserved single-cell integration).

Two design choices deserve justification against prior work. Unlike BERMUDA, which aligns only cluster pairs whose MetaNeighbor similarity exceeds a threshold (BERMUDA: a novel deep transfer learning method for single-cell RNA sequencing batch correction reveals hidden high-resolution cellular subtypes), we never gate alignment on cross-batch cluster similarity, because in unilateral ablation experiments the treatment-induced populations (phagocytosing Batf3-lineage dendritic cells, debris-associated myeloid states, per Khalil et al. (In situ vaccination with defined factors overcomes T cell exhaustion in distant tumors)) have no counterpart cluster in untreated tissue, and a threshold either leaves their batch noise uncorrected or falsely merges them with resting cells. Unlike Adversarial Information Factorization, which strips batch information with a discriminator on the latent code (A novel batch-effect correction method for scRNA-seq data based on Adversarial Information Factorization), we never adversarially remove compartment or time information — those are condition inputs, not adversaries.

Proposed Experiments

The primary testbed is re-analysis of the Wang et al. HCC neoadjuvant cohort (576 Comprehensive single cell sequencing...): 21 patients, matched primary tumor, adjacent normal tissue, draining lymph node, and peripheral blood at baseline, day 7, resection, and adjuvant phases, with paired TCR sequencing — the only dataset in this literature that contains both the abscopal-style trajectory and the clonotype labels the anchor requires. The pre-specified endpoints are: first, whether integrated analysis detects the day-7 circulating clonal expansion that predicted pathologic response in the original study, in more of the non-responder patients (an abscopal signal that naive per-compartment analysis misses), and second, whether the exhaustion continuum — naive to PD-1-high Eomes-high terminal exhaustion, the state Khal et al. showed is selectively eliminated in responding lesions — survives integration as a continuous gradient rather than being collapsed into a few discrete clusters. Quantitatively, we score batch removal and biological conservation with the scIB-E and Corr-MSE family of metrics from Yi et al., plus a new clone-recovery metric: the fraction of responder-associated clonotypes whose latent trajectories connect all four compartments without being absorbed into bystander clusters.

The second experiment is prospective: a bilateral H22 murine hepatocarcinoma cohort following the protocol of Xu et al. (Stereotactic body radiation therapy (SBRT)...), with 8 Gy in 3 fractions to the primary leg tumor plus anti-PD-1, and multi-compartment scRNA+TCR sequencing at baseline, day 7, and endpoint. The endpoint is biological, not just methodological: mice in which the harmonized analysis reveals a clone traffic pattern matching the day-7 expansion signature should be the ones whose distant tumors regress — Xu et al. report a 50% (6 of 12) distant cure rate under this regimen, giving a natural stratification target.

The third experiment is translational cross-cohort projection. Khalil et al. projected their murine CMP-treatment bulk signature onto human melanoma RNA-seq and found that signature-matching patients had a progression-free survival hazard ratio of 0.29 (In situ vaccination with defined factors...) — but bulk deconvolution cannot tell clonal expansion from altered cell-type ratios. We will redo this projection at single-cell resolution: harmonize the murine compartments and the human cohort with clonotype-anchored integration (cell-type matched, since human cells carry no murine TCR anchor; the anchor is dropped and only the condition-conditional structure and correlation losses remain), then test whether the single-cell CMP signature stratifies patients more sharply than the bulk hazard ratio of 0.29. No reported result found on Lacuna for a single-cell-resolution cross-species abscopal signature projection; this would be the first.

Baselines

The "do nothing" reference point is per-compartment analysis without any integration — exactly what Wang et al. performed, where each tissue is processed through standard pipelines separately and clones are matched across compartments by raw TCR sequence identity. It is the reference because it produced the field's headline finding (day-7 peripheral expansion predicting response), so any integration method must recover at least that and add something. We will reproduce their clone-matching result on their public cohort and ask whether harmonization finds additional cross-compartment structure.

The strongest existing competitor is the top-performing integration method from the Yi et al. benchmark of 16 deep-learning methods on immune atlases — Domain Class Triplet Loss combined with the Corr-MSE loss — run with technical batch as the batch label and cell type as the class label. This is the strongest because Yi et al. showed it retains continuous developmental trajectories (validated on a 584,944-cell lung atlas) while eliminating center-specific bias, and because its Corr-MSE term enhanced intra-cell-type biological conservation by 6.61% on average across the benchmarked methods; no competing method in the file reports a better conserved-integration number on immune data. It will be given identical input data, the same encoder capacity, and the same tuning budget (identical hyperparameter search grid over embedding dimension, learning rate, and loss weights), so any difference is attributable to the clonotype anchor rather than compute.

A second baseline is post-hoc latent correction: scBatchProx FiLM adapters with proximal regularization applied on top of a standard PCA embedding, per Nguyen et al., who report a 3–8% relative gain in overall embedding quality across baselines with under 5 seconds of runtime. It is included because its speed makes it the pragmatic reviewer's favorite, and because its documented upstream-dependence limitation is precisely what the anchor is designed to fix — if the anchor method beats scBatchProx only when upstream PCA has already destroyed rare-clone structure, the comparison isolates the mechanism.

The key ablation removes the anchor: the identical conditional autoencoder with the same reconstruction, correlation, and rare-clone-protection terms, but with $\lambda_a = 0$, so same-clone cross-batch pairs are no longer pulled together. The difference between this and the full method on clone-recovery and exhaustion-gradient metrics is the cleanest measure of whether the TCR pairing itself — not the encoder capacity, not the auxiliary losses — carries the value.

Beating the baselines means, concretely: higher clone-recovery fraction on the Wang et al. cohort (more responder-associated clonotypes with intact four-compartment trajectories), a preserved exhaustion continuum scored by the Corr-MSE-style within-type correlation rather than discrete-cluster collapse, and in the mouse cohort, a day-7 signature that stratifies eventual distant-tumor regressors from non-regressors better than per-compartment analysis does.

Why not just…?

Why not just run a standard integration tool with compartment as a covariate? This is the most serious alternative, and scVI-style conditioning on known covariates is close in spirit to the condition-conditional encoder proposed here. The evidence against its sufficiency is Yi et al.'s overcorrection finding: when integration is optimized or scored for batch mixing, the result is a "gradient decline" in intra-cell-type information — exactly the naive-to-effector-to-exhaustion gradient that is the object of study. The covariate approach also does nothing for the rare-clone problem, since a clone below 0.1% frequency exerts no gradient on the aggregate loss. The decisive experiment is the anchor ablation above: if covariate conditioning alone matches clone recovery, the anchor is unnecessary and the proposal collapses to a benchmarking study. I consider this the most likely way the idea could fail, and the ablation is designed to detect exactly that.

Why not just track TCR clones by sequence identity, with no representation learning at all? Wang et al. already did this and found the day-7 expansion signal — no new method required. The honest answer is that this suffices for clone identity tracking but not for the phenotypic questions that follow: whether a tracked clone is exhausted or effector in the distant lesion, how its transcriptional state shifts across compartments, and how treatment-induced myeloid and dendritic states — which carry no TCR at all — compare across tissues. Those comparisons require a shared latent space, and the TCR-matching-only approach provides none. Still, the per-compartment clone-matching baseline is retained precisely because if it answers every pre-specified endpoint as well as the anchored method, the machinery is not justified.

Why not just apply scBatchProx post-hoc? It is fast (under 5 seconds) and needs no raw-data access, but Nguyen et al. state the limitation themselves: a post-hoc method cannot recover a biological signal the upstream embedding discarded. Ablation-driven neoantigen-reactive clones are rare enough that standard PCA or VAE encoders plausibly collapse them into bystander clusters before any correction runs. The head-to-head on the Wang et al. cohort, where the true expanding clones are known, directly tests whether post-hoc correction recovers them; if it does, the upstream-encoder portion of this proposal is over-engineered, though the anchor could then be grafted onto scBatchProx itself as a cheaper variant.

Why not just use bulk RNA-seq across the tissues and skip single-cell batch headaches entirely? Bulk was the translational path taken by Khalil et al. (hazard ratio 0.29 from bulk signature matching), and it avoids the whole batch-correction problem. But two findings in the file show it loses the mechanism: Wang et al. established that response prediction came from specific rare clonotypes shared across compartments, which bulk averaging cannot resolve, and Khalil et al. showed the effect depends on the selective absence of exhausted PD-1-high T cells within tumors while sparing lymph nodes and spleen — a spatial distinction bulk cannot make. The single-cell CMP-signature re-projection experiment is the direct test of whether this resolution buys clinically meaningful stratification over the bulk 0.29 hazard ratio.

Why not just do ablation alone in the mouse arm, without combination therapy or multi-tissue sequencing? Because Xu et al. report that SBRT monotherapy achieved a 0% distant cure rate versus 50% (6 of 12) with anti-PD-1 added, and Khalil et al. found local CD40/TLR4 stimulation gave only partial distant control without checkpoint blockade. The mouse arm is powered on the combination regimen precisely because that is the only setting where an abscopal signal exists to be tracked.

Decisive Checks

Three results would decide the proposal's fate. First, on the Wang et al. cohort, the full method must beat both the top Yi et al. integrator and the anchor-free ablation on clone-recovery and exhaustion-gradient metrics — if the ablation ties, the anchor is dead and the covariate-conditional design alone is the finding. Second, post-hoc scBatchProx must demonstrably lose the rarest known expanding clones that the anchored encoder retains, converting Nguyen et al.'s stated limitation into a measured failure. Third, in the mouse cohort, the day-7 harmonized signature must stratify eventual distant regressors from non-regressors; if the per-compartment clone-matching baseline stratifies equally well, the field has learned that the simpler analysis suffices for this biology, and the method's value narrows to the no-TCR myeloid compartments. Each outcome is informative; none is a rescue.