Restricted-Spline Transfer Curves for RNA-to-Protein Prediction in Single Cells

When a cell makes more messenger RNA for a surface protein, the protein on its surface does not simply rise in proportion. It may stay near background until transcription crosses a threshold, or it may level off once membrane trafficking saturates. Current models that jointly analyze single-cell RNA and protein, such as variational autoencoders (generative models that compress each cell into a low-dimensional code and reconstruct its measurements from it), pass that code through an unconstrained neural network. Such a network can fit the curve where cells are plentiful. Nobody can read the curve off it, and nothing controls what it does for cell states at the extremes of expression, which are exactly where predictions for rare or unseen cell types are made. We propose to replace that network, for each measured protein, with an explicit dose-response curve over its own gene's denoised RNA level. The curve is a restricted cubic spline: a smooth piecewise cubic that is forced to be a straight line beyond its outermost breakpoints. Epidemiologists use this tool to find thresholds and plateaus in exposure-risk data.

The claim is narrow and testable. If RNA-to-protein curves are fitted inside the probabilistic model, so that both RNA dropout and the protein's antibody background are accounted for, and if their tails are forced to be linear, then two things should follow. Protein prediction for held-out cell types should improve over neural decoders, with the gain concentrated at extreme RNA levels. And each protein's threshold or saturation point should come out as a number with an uncertainty attached. If a simpler post-hoc curve fit on an existing model's output does as well, that finding is itself the result. The experiments are designed to tell these outcomes apart.

Core Claim

Restricted cubic splines appear across the clinical literature as a way to locate tipping points: a 25 g/day sugar threshold for obesity risk in Sugar Is the Key Cause of Overweight/Obesity in Sugar-Sweetened Beverages (SSB), an L-shaped protective plateau in An L-shaped link between the composite dietary antioxidant index and human papillomavirus infection in women, and a U-shaped curve in Association between night sleep duration and depression among Chinese elderly. These papers motivate the curve shapes we expect. They also supply the tool's key property: linear tails keep estimates stable where data are sparse. They do not deal with count noise, dropout, or tens of thousands of features.

On the multi-omic side, the existing models split along an interpretability line. totalVI jointly models paired RNA and protein counts from a shared 20-dimensional code. It uses a negative binomial likelihood for RNA and a two-component negative binomial mixture that separates true protein signal from ambient antibody background. Its decoders, however, are multilayer perceptrons. scTEL translates RNA to protein with a Transformer-LSTM and reports the best protein error on CITE-seq PBMC, but it offers no explicit curve. At the other end, scETM and Mowgli are interpretable because their decoders are linear, so they cannot express saturation or thresholds. The closest spline work is ZISS. It fits zero-inflated smoothing splines of gene expression along a one-dimensional pseudotime, but it uses RNA only, has no tail restriction, and makes Gaussian-mixture assumptions. clonealign shows that an explicit dosage link (copy number to expression) can anchor integration across modalities, but its dosage function is linear or capped.

None of these models estimates a per-protein, tail-restricted RNA-to-protein curve inside a count likelihood that handles background noise. That is the gap this proposal fills.

Model

We keep totalVI's encoder, its RNA likelihood, and its protein background mixture. We change only how the protein's foreground mean is produced. For protein $j$ in cell $n$, let $m_{nj}$ be the model's denoised log expression of the cognate gene, decoded from the cell's code $z_n$. The foreground mean is

$$ \log \mu_{nj} = s_j(m_{nj}) + w_j^{\top} u_n + b_{j,\text{batch}(n)}, $$

where the terms are as follows. $s_j$ is a restricted cubic spline with $K$ knots placed at quantiles of $m_{\cdot j}$, so it is linear below the first knot and above the last. $u_n$ is a small (2 to 4 dimensional) projection of $z_n$ that captures translation and trafficking differences unrelated to the cognate mRNA. $w_j$ is a weight vector, and $b$ is a per-batch intercept.

Because $s_j$ takes the denoised RNA level as input rather than raw counts, it is fitted against an estimate of the true RNA level. The aim is to avoid the downward flattening that Direction 1 attributes to fitting splines on zero-inflated counts. The low-rank offset is kept deliberately small, so that the spline does the explaining, and its share of explained variance is reported.

We choose $K$ for each protein from 3 to 7 by held-out likelihood. This resolves the knot-selection concern that ZISS raises: knot positions follow the data's quantiles, and model selection sets their number. From each fitted $s_j$ we extract saturation points (where the slope drops below a fraction of its peak), thresholds (where the curvature peaks), and the slope in each tail. Posterior samples of the spline coefficients give uncertainty for all three quantities.

Research Program

The work has three parts. The first is a semi-synthetic study. We take real CITE-seq RNA, simulate protein counts from known curves (linear, threshold, saturating, and non-monotonic), and add totalVI-style background. This tests whether the model recovers the curve shape and the location of the inflection point, and whether recovery degrades gracefully as sparsity increases.

The second part, the main test, is prediction on real data under distribution shift. We hold out entire cell types, or batches from another tissue, during training and then predict their proteins from RNA alone.

The third part is a descriptive atlas of RNA-to-protein curves across immune proteins, checked for consistency across datasets. A curve that is real should have a similar saturation point in PBMC, MALT, and H1N1-infected PBMC after batch intercepts are accounted for. Ablations isolate the two ingredients that carry the claim: fitting the curve inside the likelihood, and restricting the tails.

Baselines

The naive reference is a per-protein linear model on log-normalized cognate RNA. It is the dosage assumption clonealign makes, and it shows how much any curvature buys. A second naive point predicts each protein's training-set cell-type mean using labels from an RNA classifier. In noisy settings, the Open Problems benchmark (Defining and benchmarking open problems in single-cell analysis) found that simple models often beat deep ones.

The strongest existing methods are totalVI and scTEL. scTEL reports the "lowest RMSE and highest Pearson correlation across all tested protein markers compared to Seurat, totalVI, and sciPENN" on CITE-seq PBMC. However, no reported result was found on Lacuna for the numeric protein RMSE or Pearson value. The only number given is cell-type classification accuracy, 86.6% for scTEL against 77.0% for sciPENN, and that is not the metric under test here. We therefore rerun both baselines.

All methods get the same training cells, held-out splits, and preprocessing. Each method gets the same hyperparameter search budget: 30 configurations chosen by the same validation likelihood or MSE. totalVI's latent dimension is matched to ours, and scTEL uses its published architecture. Beating them means significantly lower per-protein RMSE on held-out cell types across at least three of the five CITE-seq datasets, with a paired test across proteins. On in-distribution cells, our model only needs to match them within noise.

There are three ablations, each of which removes one claimed ingredient. The first swaps the restricted spline for an ordinary cubic spline with the same knots, which isolates the tail constraint. The second swaps it for a linear term, which isolates curvature. The third fits restricted splines post hoc to totalVI's denoised RNA and protein outputs, which isolates in-model fitting.

Why not just…?

Why not just fit restricted splines post hoc to totalVI's denoised outputs? This is the most serious threat, because it gives readable curves at almost no cost. Its weakness is that totalVI's denoised protein values already come from an MLP. A post-hoc curve can only describe what that network does, including its behavior in the tails, and it cannot propagate the background-mixture uncertainty into the curve. The post-hoc ablation tests this directly. If it matches our held-out-cell-type error and recovers the semi-synthetic inflection points equally well, we will report the curve atlas as a post-hoc analysis tool and drop the claim that the decoder needs to change.

Why not just regularize totalVI's MLP more heavily, or use a monotone network? Weight decay and monotonicity constraints shrink overall flexibility, but they do not fix what the function does beyond the data. A monotone network also rules out the non-monotonic shapes that the clinical spline literature documents. We include totalVI tuned under the same search budget, with weight decay in the search space. The comparison that settles this question is error in the top and bottom 5% of cognate RNA levels on held-out cell types, because tail control is the property that separates the two.

Why not use knot-free penalized smoothing splines as ZISS does? ZISS shows that penalized splines beat zero-inflated additive models on sparse RNA, with a mean squared error of 0.169 against 0.368. It never compared against restricted splines with quantile knots, and penalized splines still extrapolate with cubic-like behavior unless constrained. We add a penalized-B-spline variant of $s_j$ with the same likelihood. If it ties on tail error, we will prefer it, since it removes knot selection.

Proposed Experiments

The data are the CITE-seq PBMC benchmark (161,764 cells), MALT (8,412), H1N1 PBMC (53,201), Monocytes (37,112), and SLN-all (about 33,000 cells, with 111 and 208 protein panels), all as used by totalVI and scTEL.

For distribution shift, we use leave-one-cell-type-out splits in PBMC and cross-tissue transfer from PBMC to MALT. For panel transfer, we use SLN-all's held-out proteins.

The metrics are per-protein RMSE and Pearson correlation. We also measure the coverage of 90% predictive intervals, since scTEL's quantile loss makes this comparison fair. Separately, we report error restricted to cells in the extreme 5% of cognate RNA levels, which is where the tail constraint should matter.

On the semi-synthetic data, the metrics are absolute error in recovered threshold and saturation location, and accuracy in classifying the curve shape.

There are two decisive checks. The first: does the restricted spline beat the unrestricted cubic spline and totalVI specifically in the tails of held-out cell types, while tying them in the middle of the RNA range? If not, the tail story is wrong. The second: do extracted saturation points agree across PBMC, H1N1, and MALT more closely than points extracted post hoc from totalVI? If not, the interpretability claim reduces to the post-hoc tool.