Patient-Matched Leukocytes as a Confounder Reference for Scarce Circulating Tumor Cells

A liquid biopsy draw yields perhaps one to ten tumor cells that have shed into the bloodstream. These are called circulating tumor cells, or CTCs. Methods that read their gene expression to infer which organ the cancer came from or how aggressive it is have no way to tell a tumor program from an imprint of the patient's other illnesses. Chronic inflammation, kidney failure, diabetes and medications all change the transcriptional state of circulating cells. With only three cells, nothing about the CTCs themselves can separate "this tumor is stressed" from "this patient is inflamed."

The same blood draw, however, also captures thousands of the patient's own white blood cells. Today's pipelines treat them only as contamination to filter out. Those leukocytes record the same systemic host state. We propose to use them as a per-patient reference for the confounder. We first estimate a host-state signature from the abundant leukocytes. We then remove from each CTC only the part of its expression that this host state predicts, before tracing the lesion or classifying the cancer.

Core Claim

The obstacle shared by both research directions is a mismatch between where the confounder lives and where the data are scarce. Comorbidity confounding is a property of the patient, not of the tumor cell. Every circulating cell in that patient is therefore exposed to it, and the leukocytes are exposed in large numbers. Our claim is twofold. First, a host-state representation estimated from a patient's own leukocytes predicts a measurable share of the non-tumor variation in that patient's CTCs. Second, conditioning on it makes tumor-of-origin and cancer-type inference robust to host-state shifts, in the 1–5-cell regime where current methods break down. If the first part is false, meaning host effects in leukocytes do not carry over to epithelial tumor cells, the proposal fails. We therefore test that link directly before building anything on top of it.

Relation to Prior Work

Two lines of work motivate the gap. CTC-Tracer transfers organ-of-origin labels from primary-tumor single-cell atlases onto CTCs. It uses unsupervised domain adaptation, meaning it aligns the feature distributions of two datasets without labels in the target. Its matching step is cycle-consistent, pairing clusters only when each is the other's nearest match across the two domains. It reports roughly 95% average transductive tracing accuracy, but it degrades below about 30 target cells. It models the shift purely as the tissue-to-blood transition and treats co-captured white blood cells only as something to discard. Hassani et al. classify six cancer types from CTC mRNA in ctcRbase with F1 0.9638. That figure falls to 0.7776 on an independent prostate cohort (GSE153514), a drop consistent with unmodeled patient-level shift, although the paper does not attribute it.

Three lines of work supply methods but constrain the design. f-scLVM separates confounding factors from biological ones with a linear factor model. Regressing out a dense confounder in 49,300 retinal cells revealed hidden subpopulations. It needs many cells per estimate, which a CTC sample cannot supply but a patient's leukocyte compartment can. Wasserstein feature reweighting suppresses dementia-driven speech features to detect depression, reaching F1 81.0%. It depends on a balanced two-by-two factorial cohort and 32 dense features, a luxury oncology does not offer. CCVAE splits a variational autoencoder's latent space into label-specific and background parts. It reaches 84% versus 69% for a conditional VAE on EEG. However, it needs supervised labels for each factor, and its authors report instability above 32 latent dimensions and incomplete disentanglement when labels are correlated. We take CCVAE's partition idea. We take the defining signal for the nuisance block from the reference cells rather than from comorbidity labels we do not have.

MARS and Tabula provide cross-tissue embeddings. Neither models patient context, and Tabula's authors name metadata integration as open work. We use them as frozen encoders, not as competitors on the host-state question.

Research Program

The model has three parts that share a gene-expression decoder. A host encoder pools a patient's leukocytes into a vector $h_p$ for patient $p$. Cell-type identity is first removed from those leukocytes, so that $h_p$ captures shared systemic programs such as interferon response, glucocorticoid exposure and hypoxic stress, rather than shifts in leukocyte composition. A tumor encoder maps each CTC to $z$. The decoder models counts with a zero-inflated negative binomial, a count distribution with extra zeros to handle the dropouts typical of single-cell sequencing:

$$x_{g} \sim \mathrm{ZINB}\big(\mu_g = \exp(a_g(z) + b_g^{\top} h_p),\ \theta_g,\ \pi_g\big)$$

Here $x_g$ is the count of gene $g$ in one cell, and $a_g(z)$ is the cell-intrinsic tumor term. The vector $b_g$ is the gene's response to host state, $\theta_g$ is its dispersion, and $\pi_g$ is its dropout probability. The host-response weights $b_g$ are tied across cell types up to a learned per-lineage scale. This tying is the one assumption that lets thousands of leukocytes inform how three CTCs should be corrected. An adversarial penalty discourages $z$ from predicting $h_p$. Downstream tracing (with CTC-Tracer's alignment applied to $z$) and classification then operate on $z$ alone.

We will build this in three stages. Stage one tests the transfer assumption on the Tabula Muris Senis atlas, where an organism-level variable, age, acts on immune and epithelial cells of the same animal. Stage two builds semi-synthetic host-shift benchmarks. We inject host programs into CTC datasets: programs estimated from PBMC (peripheral blood mononuclear cell) datasets, and the f-scLVM-style dense factors. The injection uses the same tied-response model plus a held-out nonlinear variant, so the method is not graded only on its own generative assumption. Stage three evaluates on real CTC data, down-sampled to the clinical 1–5-cell regime.

Baselines

The naive reference is CTC-Tracer applied unchanged, together with the ctcRbase DNNraw classifier applied unchanged, to the same down-sampled, host-shifted inputs. The reported numbers on clean data are about 95% average transductive tracing accuracy for CTC-Tracer, and F1 0.9638 on ctcRbase and 0.7776 on GSE153514 for DNNraw. No result was found on Lacuna for any CTC method under host-state shift or at 1–5 cells. We will rerun both baselines ourselves on identical splits, so that every number in the comparison comes from one code path.

CTC-Tracer is the strongest existing method. It is the only one that already adapts across a biological shift on CTCs. Beating it means a statistically significant gain in tracing accuracy at injected host-shift strengths that are clinically plausible, with no loss on clean data.

The second baseline is per-patient leukocyte mean-centering. For each gene we subtract the patient's leukocyte deviation from the cohort mean, then run CTC-Tracer. This is the cheapest version of our idea and the most important ablation, because it uses the same reference cells without the tied ZINB response model.

The third baseline is f-scLVM fitted on the pooled leukocytes. Its dense factors are projected onto CTCs by linear regression. This isolates whether nonlinearity matters.

The final ablation removes the tie between leukocyte and CTC response weights. If performance does not drop, the reference cells are not doing the work we claim.

All methods get the same input genes, the same pretrained encoder initialization where applicable, and an equal hyperparameter search budget on a validation split that is disjoint by patient.

Why not just…?

Why not just regress out recorded comorbidity covariates? No CTC resource in the file, ctcRbase included, carries comorbidity annotations. Diagnosis codes are also coarse proxies for the molecular host state that actually reaches cells, such as a treated versus untreated autoimmune flare. Where metadata exist, we will add a covariate-regression arm to compare against. We expect leukocyte-derived $h_p$ to match or exceed it, since $h_p$ measures the state rather than its label.

Why not just per-patient centering on the leukocytes? This is the real threat, and it is the second baseline. It differs from our method in assuming every gene shifts equally in leukocytes and tumor cells. If host programs are largely lineage-general, centering may capture most of the gain. In that case the paper's contribution reduces to the observation that co-captured leukocytes should be kept, which is still useful, and we would report it as such. The tied, lineage-scaled model is justified only if it beats centering on the Tabula Muris Senis transfer test and on the nonlinear injection benchmark.

Why not just fine-tune a single-cell foundation model on CTCs? Tabula reaches 0.9810 on hPancreas annotation. However, it sees one cell at a time with no patient context, so it cannot tell a host-induced state from a tumor state. We use it as the encoder backbone in both our model and the baseline. Any gain then comes from the reference conditioning, not from representation quality.

Why not just drop known inflammation and stress genes? Gene exclusion is a crude form of feature masking, close in spirit to Young et al. but chosen by hand. It also discards real tumor biology, since stress and immune-evasion programs are part of metastatic behavior. We include a masked-gene arm using standard pathway sets. Our method should beat it specifically on tumor-intrinsic stress signal that is preserved in the semi-synthetic benchmark.

Proposed Experiments and Decisive Checks

The first decisive check runs on Tabula Muris Senis. We estimate $h$ from each animal's immune cells and ask whether it predicts age-associated drift in that animal's epithelial cells, measured as held-out log-likelihood gain over a model without $h$. A null result here would end the project's central premise before any CTC work.

The second check uses the CTC-Tracer benchmark suite and ctcRbase. We down-sample to 1, 3, 5 and 30 CTCs per sample, inject host programs at graded strengths, and plot tracing accuracy and macro-F1 against shift strength for every arm. The prediction is that the unconditioned baselines degrade with shift and with fewer cells, while our method's curve stays flatter, most visibly below 30 cells.

The third check is external validation on GSE153514, where the published 0.7776 F1 serves as the anchor. Leukocytes from those samples can be used where they are present. If they are absent, the check reduces to evaluating whether the learned $z$ alone transfers better.

Finally, we test for masking in the sense of Bast, where a comorbid condition made an abnormal signal look normal. We construct injections in which the host program pushes CTCs toward leukocyte-like expression. We measure CTC-versus-leukocyte detection recall, where CTC-Tracer reports about 99% on clean data, to see whether conditioning on $h_p$ recovers cells that the baselines misclassify as white blood cells.