CD45 Isoform Epitopes as a Single-Cell Test of Splicing-Driven Protein Discordance

Methods that predict a cell's protein levels from its RNA assume that one gene yields one protein signal. That assumption breaks when a gene's RNA is spliced, meaning cut and rejoined in different ways, into several versions that encode different protein forms. Some of the most widely used single-cell protein assays already measure such forms without anyone treating them as splicing data. CITE-seq tags cells with antibodies and reads RNA and surface proteins from the same cell. Its antibodies against CD45RA and CD45RO recognize different spliced products of one gene, PTPRC.

We propose to use these antibody pairs as a protein-level readout of splicing at the scale of 100,000 cells. We build an RNA-to-protein model in which, for the epitope-carrying proteins, the predicted amount factors into two parts: how much the gene produces overall, and what fraction of its splice versions carry the epitope. The second part is predicted from the cell's expression of splicing regulators. We then use the same data to show how mixing cell types in bulk tissue can fake, or hide, a splicing switch.

The problem matters because the field treats three facts separately. mRNA and protein agree only 40–60% of the time (A high-stringency blueprint of the human proteome). Large isoform switches occur in genes whose total expression does not change (Detecting Splicing Variants in Idiopathic Pulmonary Fibrosis from Non-Differentially Expressed Genes). Bulk tissue averages over cell types. The IPF study is the clearest example of the gap: it reports a TOM1L1 isoform ratio inverting from roughly 77:23 to 27:73 in six bulk lung samples. Nothing in current tooling can say whether individual cells switched or the tissue's cell mix changed, or whether the switch reaches the protein. We want a setting where all three questions have measurable answers in the same cells.

Core Claim

For a surface protein recognized by an antibody that binds only some splice versions of its gene, part of the RNA-to-protein discordance in single-cell data is splicing-driven. That part varies within cell types. It is predicted by splicing-regulator expression and not by the gene's own mRNA count.

Gene-level translation models can reproduce the population distribution of such a protein without separating the two sources. They do so by encoding cell identity in a learned latent state, a compact vector summarizing the cell. Our claim is that they fail exactly where cell identity and splicing state come apart. Examples are transitional cells that carry both CD45RA and CD45RO, and donors or batches not seen in training. A model that explicitly factors "gene output × epitope-carrying isoform fraction", with the two antibodies sharing one gene, should do better in those regimes.

The claim is falsifiable. If the strongest gene-level model matches the factored model on within-cell-type residuals and on held-out batches, we conclude that cell-state latents already absorb splicing for this protein class. We would report that as the result.

Model

We build on the protein decoder of totalVI. totalVI is a variational autoencoder that maps paired mRNA counts and surface-protein counts into a shared 20-dimensional latent space. It models each protein as a mixture of true signal ("foreground") and antibody background (Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells). We keep its background component and change only the foreground mean for epitope-restricted proteins.

For cell $c$ and an antibody $j$ that binds the isoform subset $S_j$ of a gene with $K$ annotated isoforms, the foreground mean is

$$ \mu_j(c) = \beta_j \, s(c) \sum_{k \in S_j} \pi_k(c), \qquad \pi(c) \in \Delta^{K-1}. $$

Here $\beta_j$ is an antibody-specific scale. $s(c)$ is the gene's total protein output, predicted from the gene's mRNA count and the latent state. $\pi(c)$ is the cell's isoform-usage vector, the fraction of each splice version, which lies on the probability simplex $\Delta^{K-1}$ (non-negative entries summing to one). $\pi(c)$ is output by a small network whose input is restricted to the expression of annotated RNA-binding and splicing-factor genes. This follows the logic of CellSpliceNet, which predicts neuron-type splicing in C. elegans from sequence plus single-cell splicing-factor expression (Spearman ρ = 0.88). That work stops at the RNA level; we take its trans-regulator input and make protein the training signal.

Several antibodies on one gene share $s(c)$ and $\pi(c)$, and this coupling is what makes the epitope fractions identifiable. With 3′-tagged RNA we cannot see PTPRC exon junctions directly. We therefore only claim to recover the epitope-carrying sums $\sum_{k\in S_j}\pi_k$, not the full isoform vector.

Research Program

The primary data is GSE164378, the PBMC atlas of more than 161,000 cells with 228 surface proteins used by scChord. scChord explicitly reports CD45RA as a heavy-tailed protein. Our first task is an audit. We will list every panel antibody whose epitope is known to depend on exon content, check whether CD45RO and CD45RB are present, and map each antibody to its isoform subset $S_j$ using the transcript annotation.

We expect PTPRC to be the anchor case and the list of other cases to be short. The scope is narrow on purpose: this is a case where the protein measurement is itself a splicing assay, which almost nothing else in the file offers. PBMC5k/10k (A Joint Model of RNA Expression and Surface Protein Abundance in Single Cells) and GSE100866 serve as held-out batches, subject to the same audit.

The analysis has three parts.

The first is a decomposition study that needs no new model. For every panel protein, we split the variance of the residual between its own-gene mRNA and its protein into between-cell-type and within-cell-type parts, using the law of total variance. We then ask whether epitope-restricted proteins carry more within-type discordance than other proteins. We also ask whether that within-type part is explained by splicing-factor expression beyond what the gene's own mRNA explains.

The second part fits the factored model against the baselines below.

The third part is the masking experiment. We build pseudobulk samples by summing cells. In "composition-only" mixtures we change the proportions of naive and memory T cells while leaving within-type splicing unchanged. In "switch-only" mixtures we hold proportions fixed and reweight cells within each type toward high or low inferred RA fraction.

On each pseudobulk we apply the "Six Region" analysis of Deng et al. That analysis plots gene fold change against differential-splicing significance and calls genes in "Region 4" when splicing shifts but total expression does not. We then test whether a joint estimate of cell-type proportions and cell-type-specific epitope fractions can tell the two mixture scenarios apart. The count-based deconvolution of Count Bridges and the reference-embedding approach of POPPY both deconvolve gene-level counts only. We borrow the idea of treating per-cell profiles as latent variables under a sum constraint, but apply it to isoform-fraction readouts.

Finally, we run an orthogonal RNA-level check. Naive and memory T cells from a PBMC donor are sorted and sequenced with joint short- and long-read RNA sequencing, quantified with MPAQT. MPAQT is a Poisson model that combines short-read depth with full-length long reads and corrects platform biases (Accurate isoform quantification by joint short- and long-read RNA-sequencing). The model's inferred epitope fractions should match the long-read PTPRC isoform fractions in each sorted population. MPAQT supplies the method here. Its reported gains (Pearson r = 0.91 against RT-qPCR on MAQC, and exon-inclusion change r = 0.81 joint versus 0.33 short-read-only in hESC-derived neurons) are why we do not trust short reads alone for this check.

Baselines

The naive reference predicts each protein from its own gene's normalized mRNA with a per-cell-type linear fit, using annotated cell types. This is the "gene equals protein, adjusted for cell type" assumption. Any model must beat it on within-type residual correlation to show it captures more than cell identity.

The standard generative baseline is totalVI, run with default architecture on the identical train/test split, the same gene selection, and the same GPU budget.

The strongest competitor is scChord. It trains a zero-inflated negative binomial autoencoder, which absorbs dropout and overdispersion, and then learns an optimal-transport flow in the latent space to generate proteins. The paper reports that this recovers the bimodal CD4 and heavy-tailed CD45RA distributions and gives the lowest RMSE and highest Pearson correlation on GSE164378 against totalVI, sciPENN and scTranslator. However, no reported result was found on Lacuna for scChord's exact RMSE or Pearson correlation values. We will therefore rerun it ourselves with its released settings and the same tuning budget as our model: an equal number of hyperparameter trials, selected on the same validation cells.

Beating scChord means three things. First, higher within-cell-type Pearson correlation for the epitope-restricted proteins on held-out cells. Second, lower error on held-out batches (PBMC5k/10k). Third, a lower rate of physically inconsistent predictions, defined as cells where the predicted RA and RO signals imply an epitope-carrying fraction sum above one. Scores on proteins without splice-restricted epitopes should not change, since the model is identical there.

Two ablations isolate the key ingredient. In the first, the coupling is removed: each antibody gets its own unconstrained mean from the same inputs. In the second, the splicing-regulator input to $\pi(c)$ is replaced by the full latent state.

Why not just…?

Why not just use totalVI or scChord as they are, since their latent state already knows naive from memory T cells? This is the real threat. scChord already reproduces the CD45RA distribution. If cell identity fully determines CD45 isoform use, the factored model adds nothing to accuracy. The experiment that settles it is the within-type and held-out-batch comparison. If the gene-level models match ours there, we report that splicing discordance for this protein class is absorbed by cell state. The decomposition study and the masking analysis still stand as contributions in that case.

Why not just add splicing-factor genes as features to a per-protein ridge regression? This is cheap, and it could capture most of the within-type signal. The coupling-removed ablation is exactly this comparison with a nonlinear learner. If ridge with splicing-factor features matches the coupled model, the structured factorization is unnecessary, and we will say so. scChord's argument that deterministic regressors collapse multimodal proteins to an average applies to bimodal markers. Whether it applies to within-type residuals is untested, so we treat this baseline as live.

Why not just measure isoforms directly with single-cell long-read sequencing? That would give direct RNA isoform truth. But it would not give paired protein at this scale, and it would not tell us whether the switch reaches the protein. No reported result was found on Lacuna for single-cell long-read isoform quantification paired with surface protein. We use bulk long reads on sorted populations only as validation.

Why not just run linear deconvolution (for example CIBERSORTx) on bulk isoform ratios? Linear deconvolution returns cell-type proportions. It does not return cell-type-specific isoform fractions, so it cannot separate a composition shift from a within-type switch. Count Bridges also reports that it outperforms CIBERSORTx and MuSiC on gene-level deconvolution. In the pseudobulk experiment, we run linear deconvolution followed by proportion-weighted type means as a baseline for classifying scenarios, so this answer is tested rather than assumed.

Decisive Checks

The first check compares within-cell-type correlation between predicted and observed CD45RA and CD45RO on held-out GSE164378 cells, and on PBMC5k/10k after training on GSE164378. The factored model is compared against scChord, totalVI, the ridge-plus-splicing-factor model and both ablations. A gain concentrated in RA+RO double-positive transitional cells supports the claim.

The second check is the pseudobulk masking test. The aim is to show that composition-only mixtures produce spurious Region-4 calls under the Deng et al. analysis, and that the joint estimator classifies composition-only versus switch-only mixtures above the linear-deconvolution baseline.

The third check compares inferred epitope fractions in sorted naive and memory cells against MPAQT long-read PTPRC isoform fractions. A failure in any one of these three checks would identify which part of the claim does not hold.