Drugs and biomarkers found in mice often fail in people. One quiet reason is rarely measured directly: the link between a gene's messenger RNA (the transcript, which is cheap to measure) and its protein (which does the biological work) may itself differ between species. In human tissue, transcript levels track protein abundance only 40% to 60% of the time (A high-stringency blueprint of the human proteome). If a gene's transcript reports its protein faithfully in mouse cells but poorly in human cells, then mouse RNA findings about that gene cannot be carried over to human protein biology, however well the sequences match. We propose to measure this link gene by gene in both species using mutual information, meaning the number of bits a transcript measurement tells you about the protein level. We then ask whether any species difference is real or produced by a handful of genes, and whether it predicts which proteins a mouse-trained model fails to predict in human cells.
The data already exist. Paired single-cell RNA and surface-protein measurements (CITE-seq) are available for mouse spleen and lymph node, with 111- and 208-protein panels (totalVI), and for human blood, with 228 proteins in GSE164378 and 13 in GSE100866. Many of the antibody targets are orthologous immune markers. The deliverable is a per-protein "coupling atlas" with calibrated uncertainty, a formal robustness test of the cross-species comparison, and a test of whether coupling divergence explains where cross-species prediction breaks.
Core Claim
For each shared surface protein $g$ in species $s$, define coupling information as
$$I_s(g) = I(T_g; P_g \mid C),$$
the mutual information between the gene's own transcript count $T_g$ and its measured protein level $P_g$, computed within matched cell types $C$ and averaged over them. Conditioning on cell type matters. Much of the raw correlation between a transcript and its protein simply reflects that both are high in one cell type and low in another. That shared cell-type signal transfers across species almost by construction. The within-type coupling is the part that reflects post-transcriptional regulation.
We claim three things. First, $I_s(g)$ can be estimated with calibrated uncertainty from existing CITE-seq data. Second, the cross-species relation between $I_{\text{mouse}}(g)$ and $I_{\text{human}}(g)$ departs from what within-species dataset-to-dataset variation would produce. Third, the size of the departure, $|\Delta I(g)|$, predicts per-protein failure when an RNA-to-protein model trained on mouse is applied to human cells. The third claim ties an information-theoretic measurement to a practical outcome. The atlas stands even if the first two claims turn out negative, because a finding that coupling is conserved is also useful for translation.
Research Program
The work rests on four existing lines. RNA-to-protein generative models motivate the question and serve as the transfer testbed. totalVI models RNA and protein jointly and separates protein foreground from antibody background. scTranslator is a pretrained Transformer that handles unaligned gene sets. scChord combines a zero-inflated count decoder with flow matching to reproduce bimodal proteins such as CD4. All three were built and evaluated within one species, and none reports how much information is actually shared between transcript and protein. SATURN aligns cells across species using protein-language-model embeddings of gene sequences, and it shows orthologs drifting functionally (human Myoc). Its inputs are RNA and static sequence only, so it cannot see divergence that shows up only at the protein level. That gap is exactly what we measure.
The information measure comes from neuroscience's direct method: total entropy minus noise entropy on discretized responses (Information through a Spiking Neuron). The sampling-bias caution also comes from there: estimates made from sparse bins are biased upward. We reuse the logic but not the data model. Transcripts are discretized as raw UMI counts in quantile bins, and proteins as totalVI-denoised foreground values and, separately, as raw log counts. Bias is corrected by subsampling cells and extrapolating to infinite sample size, and a shuffle null is computed within each cell type. Estimates are cross-checked with a k-nearest-neighbor continuous estimator. We treat the split between RNA dropout and genuine discordance as a confound to control, not something to wave away. As a control, we compute $I$ between protein and a smoothed neighborhood transcript level, which reduces the effect of dropout.
The robustness test comes from Testing Most Influential Sets. That work derives exact influence of data subsets in least-squares regression and uses extreme value theory, the statistics of worst cases, to give $p$-values for whether the most influential subset is more influential than chance allows. Its stated limitation is that the theory covers linear least squares only. Our cross-species comparison is exactly such a regression, fitted across proteins:
$$I_{\text{human}}(g) = a + b\, I_{\text{mouse}}(g) + \varepsilon_g,$$
where $a$ and $b$ are intercept and slope and $\varepsilon_g$ is residual. A slope near 1 with small residuals means coupling is conserved. We ask whether the conclusion about $b$ is driven by a small set of proteins, for example a few bimodal activation markers. This keeps the method inside the regime where the theory holds and does not stretch it to neural networks. The authors report validity down to $N = 50$. If fewer than about 50 orthologous proteins survive panel matching, we will say so and report exact enumeration of small subsets as descriptive statistics, not calibrated $p$-values. A second, coarser analysis leaves out one cell type at a time to check whether divergence is carried by one population that exists in only one tissue.
Baselines
The naive reference point is "coupling transfers": assume $I_{\text{human}}(g) = I_{\text{mouse}}(g)$, or, for prediction, apply a per-gene linear transcript-to-protein map fitted on mouse to human cells. Any claim of divergence must beat this identity model by more than the noise floor.
The key reference for claim 2 is a within-species null. We compute the same regression between the two human datasets (GSE100866 against GSE164378, on their shared markers) and between the two mouse panels in SLN-all. Mouse against human departs from conservation only if its residual spread and slope deviation exceed the within-species dataset-to-dataset spread, estimated by bootstrapping over cells and donors. This null absorbs differences in platform, tissue, and antibody clone that would otherwise pass for species effects.
The strongest existing competitors for claim 3 are totalVI, scTranslator, and scChord, each trained on mouse SLN-all and evaluated on human GSE164378 for orthologous proteins, with SATURN macrogene embeddings used to align input genes where 1-to-1 orthology is missing. All models get the same cells, the same shared protein set, the same hyperparameter search budget tuned on a held-out mouse split only, and no human labels. Reported numbers exist only within species. scTranslator reports a 10.7% cosine-similarity gain over sciPENN and cTP-net and similarity above 0.83 under shuffled gene sets. scChord reports the lowest RMSE and highest Pearson correlation on GSE164378, but no numeric values appear in the file. SATURN reports 85.8% median label-transfer accuracy from zebrafish to frog. No reported result was found on Lacuna for mouse-to-human RNA-to-protein transfer, so we will establish that number ourselves.
For claim 3, the proposal "beats" the baselines if $|\Delta I(g)|$ ranks per-protein transfer error better than two decoder-derived predictors: each model's own mouse validation error, and each model's predictive uncertainty. We score this with Spearman correlation across proteins, using a bootstrap confidence interval that excludes the competitor's value. The ablation that isolates the key ingredient replaces conditional coupling $I(T;P\mid C)$ with unconditional $I(T;P)$. If the unconditional version predicts transfer failure equally well, then within-cell-type regulation is not what matters, and the claim shrinks accordingly.
Why not just…?
Why not just use per-protein Spearman correlation instead of mutual information? This is a real threat. For monotone relationships, rank correlation and mutual information tend to order proteins the same way. Mutual information should add value for proteins with bimodal or threshold-like relationships to their transcript, such as the CD4 and CD45RA shapes that scChord shows regression smooths away, and when a transcript is informative only about whether a protein is present, not how much. We will compute both and report their rank agreement across proteins. If agreement is very high and Spearman predicts transfer failure equally well, we will state that the cheaper statistic suffices and present mutual information as the calibrated upper-level measure, not a necessity.
Why not just train totalVI on mouse, test on human, and read off per-protein error? That gives a transfer error but not a cause. A protein can fail to transfer because its coupling diverged, because the antibody clone differs, because the cell type is absent from spleen or blood, or because the model is miscalibrated. The decoder cannot separate these, and Pica et al. warn that decoder accuracy can reflect features the biology does not use. Our design runs this baseline anyway and asks whether a model-free, per-gene quantity explains its errors beyond what the model's own diagnostics explain. If it does not, the decoder is enough.
Why not just drop proteins one at a time and eyeball the slope, or use Cook's distance? Ordinary leave-one-out misses joint influence, and fixed cutoffs ignore that we are searching for the worst subset. Testing Most Influential Sets shows this concretely. On the Law School data, a 77-point set that flipped a coefficient's sign was within natural variation, while a 17-point set was excessively influential ($p = 0.019$). We will report Cook's distance side by side. The formal test earns its place only if it changes which proteins are flagged.
Why not just rely on SATURN's sequence-based macrogenes to catch divergent orthologs? SATURN detects drift from RNA co-expression and sequence. It is blind to divergence at the protein-abundance level by design. We test directly whether proteins with large $|\Delta I|$ are the ones SATURN places in different macrogenes across species. Substantial overlap would mean sequence plus RNA already captures most of the signal. Little overlap would mean protein-level coupling is an independent axis.
Proposed Experiments
The first experiment builds the coupling atlas. We estimate $I_s(g)$ and $I_s(g)/H(P_g\mid C)$, the fraction of protein uncertainty the transcript removes, for all orthologous surface proteins across matched T, B, NK, and myeloid populations in SLN-all, GSE164378, and GSE100866. Cell types are matched by SATURN co-embedding plus canonical markers. Estimator validity is checked by subsampling convergence curves and shuffle nulls, and by recovering known coupling in data simulated from totalVI's generative model.
The decisive checks follow. The cross-species slope and residual spread must fall outside the within-species bootstrap envelope, or we conclude coupling is conserved at this resolution. The most-influential-set test must show whether any divergence rests on a small protein set. The leave-one-cell-type-out analysis must show whether divergence is carried by one population. Finally, $|\Delta I(g)|$ must rank per-protein mouse-to-human transfer error for totalVI, scTranslator, and scChord better than those models' own diagnostics, and the conditional version must beat the unconditional ablation.
The main confound is tissue: mouse lymphoid organs versus human blood. We will limit its reach with the within-species null and cell-type conditioning, and we will state plainly that a clean species effect ultimately requires tissue-matched mouse and human CITE-seq data.
References
- A high-stringency blueprint of the human proteome
- Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells
- A pre-trained large generative model for translating single-cell transcriptome to proteome
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- Towards Universal Cell Embeddings: Integrating Single-cell RNA-seq Datasets across Species with SATURN
- Information through a Spiking Neuron
- Testing Most Influential Sets
- Quantifying how much sensory information in a neural code is relevant for behavior