Bulk RNA sequencing of a tumor averages the messenger RNA of millions of cells. Computational "deconvolution" tries to undo that averaging by estimating what fraction of the sample came from each cell type. The estimates are matched against reference profiles taken from single-cell sequencing. Two problems keep these estimates from being trusted for the questions clinicians care most about, such as how many exhausted versus memory CD8 T cells a tumor contains. First, bulk and single-cell data come from different lab protocols and platforms, so their expression values differ systematically for technical reasons. This is a "batch effect," and it corrupts the match against the reference. Second, the standard way to check deconvolution is to build artificial bulk samples by summing single-cell profiles. That kind of test never contains the bulk-versus-single-cell technical gap, so it cannot measure how badly that gap hurts. We propose to use T-cell receptor sequences to address both problems. Each T cell carries a receptor gene assembled by random recombination, so a receptor sequence acts as a natural barcode for a family of cells descended from one ancestor (a "clone"). Unlike gene expression, the barcode is the same string of letters whatever machine read it.
The core idea follows. Suppose a patient has bulk RNA-seq plus single-cell RNA-seq with receptor sequencing from the same tumor. Every clone seen in both is an anchor. The single-cell data tell us which functional states that clone's cells occupy. The bulk receptor reads tell us how abundant the clone is in the bulk sample, and that count is not distorted by expression batch effects. Together these give constraints on the T-cell state proportions in the real bulk sample that do not depend on expression. We use the constraints in two ways. They supervise a small correction that maps single-cell reference profiles into bulk-platform space. They also provide the first real-bulk benchmark for fine T-cell state deconvolution. After training on a matched cohort, the learned correction is applied to bulk-only cohorts that have no single-cell data.
Core Claim
Clone-anchored constraints can learn the technical mapping between single-cell references and bulk tissue RNA-seq, and generic batch correction cannot do this as well. We claim that a platform correction learned this way will lower the error of fine CD8 T-cell state estimates on held-out real bulk samples, relative to the same deconvolution model with no correction or with a correction learned from expression alone. We also claim the gain comes from state-resolved clone information, not merely from knowing the total T-cell amount. The second claim is the one a skeptic will attack, and the experiments isolate it directly.
Relation to Prior Work
The motivating evidence comes from the deconvolution benchmark of Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures. It tested nine tools on 38,000 simulated breast cancer mixtures and found two things: failures at high tumor purity, and weak resolution of fine immune lineages. BayesPrism and DWLS were the most reliable on granular immune types. That benchmark is built from summed single-cell profiles, so it measures algorithmic error with no platform gap. Our proposal adds the missing half: a real-bulk evaluation where the truth comes from receptor barcodes.
Two recent methods supply modeling ideas without addressing receptors. Count Bridges deconvolves bulk counts into latent single-cell count profiles using integer-valued stochastic dynamics. POPPY aligns bulk transcriptomes to single-cell atlases through contrastive learning guided by the hierarchy of cell types. Both rely only on expression. POPPY's alignment can say "T cells" but cannot check any clone-level claim.
On the batch-correction side, two methods supply the correction machinery, and our ablations test both. BERMUDA aligns only similar cluster pairs across batches so that populations are not merged. scBatchProx learns lightweight per-batch affine adapters on embeddings, with a proximal penalty that keeps them near the original structure. Its ablations show that dropping this penalty lets the adapters destroy biological structure. We borrow the adapter-plus-proximal-penalty design. What changes is the supervision signal: our correction is supervised by clone anchors rather than by distribution matching, which is what lets it avoid merging exhausted and effector states that differ in abundance between batches.
The closest conceptual precedent is LCL. LCL treats cells that share an engineered lineage barcode as positive pairs, in order to separate heritable expression programs from transient noise. LCL uses synthetic barcodes in vitro, works only at single-cell resolution, and ignores batches. We use natural receptor barcodes, and we use them to bridge bulk and single-cell platforms.
One constraining paper is Exploiting antigen receptor information to quantify index switching. It used receptor uniqueness to show that multiplexed sequencing on patterned flow cells leaks sequences between libraries, with a median leakage of 3.9% (interquartile range 1.7–7.3%). For us this means a clone shared between two libraries is not automatically a true anchor, and anchors must be filtered with an explicit leakage model. The repertoire classifiers DeepRC and LaDM³IL work on receptor sequences alone and ignore cell state. They are not competitors here. We note them only to mark that we do not attempt to classify patients from their receptors.
Research Program
Write $y$ for a bulk expression vector, $x_k$ for the reference profile of T-cell state $k$ taken from single-cell data, and $g_\phi$ for a gene-wise affine map from single-cell to bulk platform. The map $g_\phi$ is regularized toward the identity, following scBatchProx. Deconvolution then fits state proportions $p$ by
$$y \approx \sum_k p_k\, g_\phi(x_k).$$
For anchor clone $c$, let $\pi_{ck}$ be the fraction of its single-cell members in state $k$, and let $f_c$ be its share of bulk TCR-beta reads. One confound must be handled before the reads can be used. Cells in different states carry different amounts of receptor mRNA, so a read fraction is not a cell fraction. We therefore introduce per-state receptor expression factors $r_k$, estimated from single-cell data and refined during fitting. The clone-implied cell mass in state $k$ is then
$$q_k \propto \sum_{c \in A} \frac{f_c\, \pi_{ck}}{\sum_j \pi_{cj} r_j},$$
where $A$ is the set of accepted anchors. Clones never seen in the single-cell data make the anchor set incomplete. For that reason we use $q$ as a one-sided constraint: the deconvolved proportion of each state within T cells should be at least the anchored mass $q_k$. We do not require the two to be equal.
Training learns $\phi$ and $r$ on a matched cohort by minimizing the expression fit plus a hinge penalty on violations of these bounds. The deconvolution backbone is DWLS-style weighted least squares, chosen for speed. We also test a BayesPrism-style likelihood.
Anchors pass a contamination filter. For each clone, we compare its read count in the target library with its counts in other libraries from the same sequencing run. Using a leakage prior centered on the reported 3.9% rate, we reject any clone whose presence could be explained by bleed-through.
At deployment, the learned $\phi$ is frozen and applied to bulk-only cohorts such as TCGA breast cancer. When those cohorts have bulk receptor reads, we optionally add a total-T-cell abundance constraint.
The required data are tumors with bulk RNA-seq together with single-cell RNA-seq plus receptor sequencing from the same specimen. The file does not document which public cohorts include paired receptor data. Assembling the matched cohort is therefore the first deliverable. Candidate sources are the breast atlases referenced in the deconvolution benchmark, and paired tissue and blood repertoire designs such as that of Wei Wang et al.. If fewer than roughly 20 matched specimens can be found, the real-bulk arm shrinks and the controlled arm described below carries the main result.
Baselines
The do-nothing reference is BayesPrism and DWLS run directly on bulk data with an uncorrected single-cell reference. These are the two tools the breast benchmark found best on granular immune lineages. They are the strongest existing methods for this exact question, and every other arm is measured against them. The file gives qualitative rankings but no numeric error. There is no reported result found on Lacuna for BayesPrism or DWLS error on fine CD8 T-cell states, so we will establish these numbers ourselves on identical data.
The strongest correction competitor is scBatchProx-style adapters, with the same architecture and proximal penalty, fitted by distribution matching between bulk and pseudobulk rather than by clone anchors. Its reported benefit is a 3–8% relative gain in embedding quality on single-cell integration benchmarks. That is a different task, so the figure is context rather than a target. We also run BERMUDA-style cluster-matched alignment of the reference to bulk, and ComBat followed by deconvolution.
The key ablation is an abundance-only constraint. It uses bulk receptor reads only to bound total T-cell fraction and discards the state composition $\pi_{ck}$ of each clone. It isolates whether the state-resolved clone information is what helps.
All arms share the same reference, gene set, backbone, patient splits, and hyperparameter search budget. Beating the field means lower root-mean-squared error and higher correlation for fine CD8 states on held-out patients, with confidence intervals from patient-level bootstrap that exclude the best baseline.
Why not just…?
Why not just trust simulated pseudobulk benchmarks and use BayesPrism? Pseudobulk made from the reference platform has no platform gap by construction. It therefore cannot reveal the error that matters on real bulk. The controlled experiment below measures that gap directly. If uncorrected BayesPrism loses little accuracy when the gap is introduced, the premise of the proposal is wrong and we will report that.
Why not just add bulk receptor abundance as a covariate? This is the real threat. Receptor reads give a clean total T-cell count, which could fix most of the purity-driven error by itself. The abundance-only ablation tests this exactly. If it matches the full method on fine states, we conclude that clone identity helps only through abundance, and the state-resolved claim fails.
Why not just correct the batch effect with ComBat or distribution-matching adapters? Global and distributional corrections assume shared composition across batches. Without anchors they tend to merge states whose proportions differ between batches, a failure documented by BERMUDA and by the ablations of scBatchProx. The scBatchProx-style arm with no clone supervision tests whether anchors add value beyond a well-regularized correction.
Proposed Experiments
The controlled experiment uses single-cell datasets with receptor data processed in two technical batches. We build the reference from batch A and pseudobulk from batch B. Proportions are then known exactly, the technical gap is real, and receptor strings are shared across batches. We sweep tumor purity from 5% to 95%, following the benchmark design, and the fraction of clones available as anchors from 0% to 100%. The sweep traces how error falls as the number of anchors grows.
The real-bulk experiment trains on matched specimens and tests on held-out patients. Truth for the held-out patients comes from their own clone-implied proportions, which are never used in training.
A contamination check injects synthetic cross-library leakage at the reported rates. It confirms that the filter rejects spurious anchors, and it shows how much error unfiltered anchors introduce.
The decisive result has three parts. The full method must beat uncorrected BayesPrism and DWLS on fine states, beat anchor-free adapters, and beat the abundance-only ablation on held-out real bulk. A secondary transfer check applies the frozen $\phi$ to TCGA breast cancer and tests whether estimates of exhausted-state fractions become less dependent on purity.
References
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- BERMUDA: a novel deep transfer learning method for single-cell RNA sequencing batch correction reveals hidden high-resolution cellular subtypes
- Federated-inspired Single-cell Batch Integration in Latent Space
- LCL: Contrastive Learning for Lineage Barcoded scRNA-seq Data
- Exploiting antigen receptor information to quantify index switching in single-cell transcriptome sequencing experiments
- Modern Hopfield Networks and Attention for Immune Repertoire Classification
- A Label Disambiguation-Based Multimodal Massive Multiple Instance Learning Approach for Immune Repertoire Classification
- 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