Every RNA-to-protein prediction model in circulation shares one quiet assumption: the cell that made the mRNA is the same cell whose proteins you want to know. The outermost layer of skin breaks this assumption in the most extreme way biology offers. As epidermal cells finish their terminal differentiation, they destroy their own nucleus and organelles and become corneocytes — dead, keratin-packed bricks sealed into a lipid mortar that forms the body's main physical barrier. In that final layer, mRNA abundance is zero while the structural proteins that do the actual work (loricrin, involucrin, filaggrin, corneodesmosin) are at their lifetime maximum and persist for weeks. All existing models are blind exactly where the barrier is built.
This matters because the field has been reading epidermal transcriptomes as proxies for barrier function. Single-cell RNA sequencing of human epidermis resolves basal stem cell states through the granular layer, but halts there because anucleate corneocytes yield no libraries Single cell transcriptomics of human epidermis identifies basal stem cell transition states. Meanwhile, systematic proteome profiling shows mRNA levels track protein levels only 40–60% of the time A high-stringency blueprint of the human proteome, and barrier assembly is governed almost entirely by post-translational processing — proteolytic cleavage of profilaggrin, transglutaminase cross-linking, kallikrein-mediated desmosomal degradation — that no transcriptome can see Proteomics, pathway array and signaling network-based medicine in cancer. The proposal here is to build and validate the first RNA-to-protein translation model designed for this discontinuity: one that reads the transcriptome of the last living keratinocytes and predicts the cross-linked proteome of the dead layer they are about to become.
Core Claim
The claim is that transcript-protein discordance in terminal differentiation is not unmodeled noise but a structured, learnable physical process: the cornified proteome is a time-integrated, persistence-weighted function of the transcriptome along the differentiation trajectory. Formally, for protein $i$, we predict its abundance in the stratum corneum as
$$\hat{P}_i \;=\; \int_{t_{\mathrm{GRN}}}^{t_{\mathrm{SC}}} \alpha_i \, m_i(t)\, e^{-d_i\,(t_{\mathrm{SC}}-t)}\, dt,$$
where $m_i(t)$ is the mRNA trajectory of protein $i$ between the granular state ($t_{\mathrm{GRN}}$, the last measurable transcriptomic state) and cornification ($t_{\mathrm{SC}}$), $\alpha_i$ is a learned per-protein translation efficiency, and $d_i$ is a learned per-protein loss rate that captures how long a deposited protein survives. The integral form comes from the established biology of discordance: stable proteins persist long after their transcripts are gone, while short-lived signaling proteins track mRNA closely — the "shadow protein" dynamics documented in memory research, where protein and transcript lifetimes diverge by orders of magnitude Persistently Increased Expression of PKMzeta and Unbiased Gene Expression Profiles Identify Hippocampal Molecular Traces of a Long‐Term Active Place Avoidance Memory and “Shadow” Proteins. Cornified envelope proteins sit at the extreme of this spectrum, which is precisely why same-cell regression models fail there.
Three pieces of prior work motivate and constrain this design rather than duplicating it. Wang et al. supply the input side: a 17,553-cell atlas of human epidermis (GSE147482) resolving basal states BAS-I through BAS-IV, spinous, and granular communities, with RNA velocity already estimating the $m_i(t)$ trajectories — but measuring transcripts only Single cell transcriptomics of human epidermis identifies basal stem cell transition states. scChord supplies the machinery: a two-stage model pairing a zero-inflated negative binomial variational autoencoder (which absorbs the dropouts and over-dispersion of single-cell count data) with optimal-transport conditional flow matching, which learns straight-line transport from RNA to protein space and captures multimodal protein distributions that deterministic regressors collapse scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation. What no model supplies is the cross-death mapping: scChord, scTranslator, and scTEL are all trained and evaluated on paired same-cell data from immune or cancer contexts — cord blood, peripheral blood, MALT, influenza cohorts, TCGA/CPTAC bulk tissues scChord, A pre-trained large generative model for translating single-cell transcriptome to proteome, A joint analysis of single cell transcriptomics and proteomics using transformer — and none addresses tissue where the target proteome belongs to cells that no longer have RNA. HEIST, the spatial graph foundation model, makes the adjacent contrast: its Voronoi cell-proximity graphs assume contiguous nucleated cells and have never been applied to stratifying skin, where the barrier is an anucleate lipid-embedded layer HEIST: A Graph Foundation Model for Spatial Transcriptomics and Proteomics Data.
Research Program
The program has three components: a model, a benchmark, and a disease application.
The model, which we call Corneal-Translate, extends the scChord architecture with two new ingredients. First, the lag-and-persistence head: instead of transporting RNA embeddings of a cell into that same cell's protein space, it integrates along the RNA-velocity trajectory from a cell's current state to the terminal cornified state, weighting each protein by its learned survival rate $d_i$. Second, a pathway-informed prior on the output geometry: the cornified envelope is a covalently cross-linked assembly of loricrin, involucrin, envoplakin, periplakin, and filaggrin, so the model's protein space carries a protein-protein interaction prior over these complexes. This directly targets the documented failure mode of scChord — its correlation-matrix-distance (CMD, a metric for whether predicted proteins co-vary correctly) degrades as target panels scale into hundreds of proteins, because the flow model loses the global co-regulation structure scChord. The cross-linking prior is exactly the structure a free-form generative model lacks when the target is a macromolecular assembly rather than independent surface markers.
The benchmark, which we call CornSeq, must be built, because no reported result found on Lacuna for paired epidermal RNA-protein data — and for good reason: CITE-seq antibody panels cover 10–200 cell-surface markers on intact single-cell suspensions and cannot profile insoluble, cross-linked structural matrix proteins A joint analysis of single cell transcriptomics and proteomics using transformer. CornSeq pairs layered tape-strip sampling of the stratum corneum with LC-MS/MS quantification of cornified envelope proteins (loricrin, involucrin, filaggrin, corneodesmosin, TGM1/3, KLK5/7) from healthy and psoriatic donors, matched to scRNA-seq profiles of the viable layers from the same donors, processed in the GSE147482 pipeline. The imiquimod psoriasiform BALB/c model, in which epidermal thickness is measurable (~117 μm diseased versus ~65 μm treated) and treatment response can be tracked Microneedle-assisted percutaneous delivery of methotrexate-loaded nanoparticles enabling sustained anti-inflammatory effects in psoriasis therapy, provides the in vivo perturbation arm, with depth-resolved drug exposure measured in the style of dermal open flow microperfusion, where probes sit at 0.5–1.0 mm dermal depth and a 0.2 mm depth difference significantly alters measured concentrations Kinetics of Clobetasol-17-Propionate in Psoricial Lesional and Non-Lesional Skin Assessed by Dermal Open Flow Microperfusion with Time and Space Resolution.
The disease application addresses a genuine clinical ambiguity. In plaque psoriasis, two conflicting things happen at once: inflammatory transcription runs hot (NF-κB-driven IL-1β transcription, with secretion additionally gated by caspase-1 at the protein level Punicalagin Alleviates Psoriasis by Inhibiting NF-κB-Mediated IL-1β Transcription and Caspase-1-Regulated IL-1β Secretion), while the thickened stratum corneum physically sequesters topical drugs — lesional dermal exposure on day 1 of clobetasol treatment was 0.854 ng·h/ml versus 2.142 ng·h/ml in non-lesional skin, recovering to 2.768 ng·h/ml after 14 days as the plaque score fell from 6.3 to 2.4 Kinetics of Clobetasol-17-Propionate in Psoricial Lesional and Non-Lesional Skin Assessed by Dermal Open Flow Microperfusion with Time and Space Resolution. Nobody can currently tell whether early treatment failure reflects drug never reaching the target (the physical trap) or the target being engaged but transcriptionally or post-translationally unresponsive. Corneal-Translate separates these: if the granular-layer transcriptome of lesional skin predicts a healthy cornified proteome but the measured proteome is deficient, the failure is post-translational assembly; if the predicted proteome is itself deficient, the failure is upstream transcription. No existing study or model makes this distinction, because pharmacokinetic and omic measurements have never been coupled Kinetics of Clobetasol-17-Propionate.
Baselines
The do-nothing reference point is raw transcript abundance used directly as a protein proxy — the implicit assumption of most epidermal transcriptomic studies. It is the strongest competitor for the descriptive question because it costs nothing, and it has a documented performance ceiling: mRNA explains only 40–60% of protein variance across the human proteome A high-stringency blueprint of the human proteome. On CornSeq we will report its per-protein Pearson correlation; beating it means a materially higher correlation on the cornified panel, which we expect to be easy for the structural proteins and hard for the processing enzymes.
The strongest existing method is scChord, the flow-matching translation model with the best reported results on the standard benchmarks: lowest RMSE and highest Pearson correlation on both GSE100866 (cord blood, 13 proteins) and GSE164378 (peripheral blood, >161,000 cells, 228 proteins), outperforming totalVI, sciPENN, and scTranslator, with a 229 MB GPU footprint and roughly 27 minutes of training on 100,000 cells scChord. No numeric RMSE or PCC value is reported on Lacuna for scChord, so we will reproduce its numbers with its released code as our operating reference. The fair comparison runs scChord twice: once zero-shot on CornSeq (its immune-trained form, which isolates the domain gap) and once fine-tuned on CornSeq with the same GPU budget, cell counts, and tuning search as Corneal-Translate. Beating scChord means matching it on its home benchmark GSE164378 within reproduction noise on RMSE and PCC, then exceeding it on CornSeq on three metrics: per-protein PCC and RMSE, Wasserstein distance between predicted and measured protein distributions, and CMD on the full cornified panel — the metric on which scChord itself reports degradation at panel scale, and on which our cross-linking prior is aimed. We will also run scTranslator zero-shot, since its Gene Positional Encoding makes it align-free to unseen gene and protein panels and it reports a 10.7% cosine-similarity increase over sciPENN and cTP-net with unaligned similarity above 0.83 A pre-trained large generative model for translating single-cell transcriptome to proteome; it is the natural test of whether bulk-cancer pretraining transfers to a stratifying tissue.
The key ablation removes exactly the proposal's two new ingredients: the same ZINB-encoder plus flow-matching stack is trained to map granular-layer RNA directly to cornified protein without the trajectory integral and without the cross-linking prior. If the full model does not beat this ablation on CornSeq, the lag-and-persistence mechanism is doing no work and the proposal's core claim is wrong; we would report that honestly and fall back to the simpler finding that immune-trained models simply do not transfer to skin.
Why not just…?
Why not just tape-strip and run bulk RNA-seq plus LC-MS/MS? This is the most serious alternative, and we adopt half of it: layered tape-strip LC-MS/MS is precisely the ground-truth half of CornSeq. What the bulk approach cannot do is three things the model does. Bulk tape strips blend the distinct basal sub-states (BAS-I through BAS-IV occupy different micro-niches with different markers like ASS1 and GJB2) into one average, losing the lineage information that determines which trajectory a cell is on Single cell transcriptomics of human epidermis identifies basal stem cell transition states; tape stripping measures only surface concentrations and cannot resolve depth, the same limitation that motivated depth-resolved microperfusion over tape stripping in the pharmacokinetic literature Kinetics of Clobetasol-17-Propionate in Psoricial Lesional and Non-Lesional Skin Assessed by Dermal Open Flow Microperfusion with Time and Space Resolution; and a measurement is not a predictive model — it cannot interpolate to unseen donors, diseases, or computational perturbations, which is the point of a translation model and is what enables pseudo-knockout style inference A pre-trained large generative model for translating single-cell transcriptome to proteome. If a skeptical reviewer believes bulk measurement answers every question worth asking, the psoriasis experiment is the decisive case: bulk data cannot attribute treatment failure to sequestration versus assembly failure, because that requires coupling a per-cell-state transcriptome to the deposited proteome.
Why not just run CITE-seq (or spatial transcriptomics) directly on skin? CITE-seq measures what antibodies can reach: oligonucleotide-conjugated surface markers on dissociated, intact single cells, in panels of roughly 10 to a few hundred proteins A joint analysis of single cell transcriptomics and proteomics using transformer. The stratum corneum's functional proteins are the opposite case — insoluble, covalently cross-linked, embedded in a lipid matrix, and belonging to cells that cannot survive dissociation. Spatial transcriptomics fares no better for a simpler reason: corneocytes have destroyed their RNA, so the most functionally critical 10–20 μm of the tissue is invisible to any transcriptomic modality Single cell transcriptomics of human epidermis identifies basal stem cell transition states. We concede the partial version of this objection: a small CITE-seq run on the viable epidermal layers would validate the model's same-cell predictions where CITE-seq does work, and we include it as a sanity check rather than fight the objection.
Why not just use linear or elastic-net regression on the paired data? Deterministic regression fits the conditional mean of protein given RNA, which over-smooths and misses multimodal protein distributions entirely — the specific failure scChord documented against regression baselines, reconstructing bimodal CD4 and heavy-tailed CD45RA distributions where regressors collapse them scChord. Simple integration methods also fall short in head-to-head comparisons: Seurat's linear-style integration reached 76.1% cell-type accuracy versus 86.6% for the transformer-based scTEL A joint analysis of single cell transcriptomics and proteomics using transformer. Still, we will run ridge regression on CornSeq as a baseline, because if the cornified proteome turns out to be a nearly linear integral of the granular transcriptome, the elaborate machinery is unjustified — and that finding would itself be worth reporting.
Proposed Experiments and Decisive Checks
Experiment one establishes the benchmark and the headline result. Train Corneal-Translate on the GSE147482 trajectories plus CornSeq training donors; evaluate held-out donors on per-protein PCC, RMSE, distributional Wasserstein distance, and CMD across the cornified panel, against the transcript-proxy baseline, ridge regression, zero-shot and fine-tuned scChord, and zero-shot scTranslator. The decisive comparison is the fine-tuned scChord ablation: the claim is vindicated only if the lag-and-persistence integral plus cross-linking prior wins specifically on structural assembly proteins and on CMD, not merely on average error. Sanity check on GSE164378 confirms no regression on the standard immune benchmark relative to the reproduced scChord numbers.
Experiment two tests the psoriasis attribution question. Apply the trained model to lesional versus non-lesional granular-layer transcriptomes from psoriatic donors and compare predicted versus tape-strip-measured cornified proteomes. The pre-registered outcomes: if predicted proteomes are healthy but measured ones are deficient, post-translational assembly failure is implicated; if predicted proteomes are deficient, transcriptional failure is. The imiquimod arm, with epidermal thickness tracked from ~117 μm toward ~65 μm under microneedle-delivered methotrexate Microneedle-assisted percutaneous delivery of methotrexate-loaded nanoparticles enabling sustained anti-inflammatory effects in psoriasis therapy, tests whether the model's predicted proteome normalizes on the same timescale that plaque scores and drug exposure recover in the human microperfusion data Kinetics of Clobetasol-17-Propionate in Psoricial Lesional and Non-Lesional Skin Assessed by Dermal Open Flow Microperfusion with Time and Space Resolution.
Experiment three stresses the model's generalization, following the established pattern of cross-dataset transfer evaluation in this literature A joint analysis of single cell transcriptomics and proteomics using transformer: train on healthy donors, test on psoriatic; train on the mouse model, test on human. A model that only interpolates within one donor set is not a translation model. The single number that would make the program: per-protein PCC on held-out donors' cornified structural proteins substantially above the 0.4–0.6 ceiling that raw transcript abundance sets elsewhere in the proteome A high-stringency blueprint of the human proteome, sustained on the CMD metric where prior generative models degrade. If instead the same-cell ablation matches the full model, the honest conclusion is that the RNA-to-protein discontinuity of cornification is wider than any transcriptome-integrated model can bridge, and barrier assessment must remain proteomic — itself a result that would redirect how the field reads epidermal transcriptomes.
References
- Single cell transcriptomics of human epidermis identifies basal stem cell transition states
- A high-stringency blueprint of the human proteome
- Proteomics, pathway array and signaling network-based medicine in cancer
- Persistently Increased Expression of PKMzeta and Unbiased Gene Expression Profiles Identify Hippocampal Molecular Traces of a Long‐Term Active Place Avoidance Memory and “Shadow” Proteins
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- A pre-trained large generative model for translating single-cell transcriptome to proteome
- A joint analysis of single cell transcriptomics and proteomics using transformer
- HEIST: A Graph Foundation Model for Spatial Transcriptomics and Proteomics Data
- Microneedle-assisted percutaneous delivery of methotrexate-loaded nanoparticles enabling sustained anti-inflammatory effects in psoriasis therapy
- Kinetics of Clobetasol-17-Propionate in Psoriatic Lesional and Non-Lesional Skin Assessed by Dermal Open Flow Microperfusion with Time and Space Resolution
- Punicalagin Alleviates Psoriasis by Inhibiting NF-κB-Mediated IL-1β Transcription and Caspase-1-Regulated IL-1β Secretion