Learning the Shift: Transcript–Protein Discordance as a Predictable Signal of Anoikis Resistance During EMT

When an epithelial cancer cell detaches from its surroundings, it is supposed to die — a self-destruct program called anoikis kills cells that lose contact with the extracellular matrix, and metastatic cells survive only by evading it. Almost everything we know about this evasion comes from measuring messenger RNA, because RNA sequencing is cheap and standardized. But mRNA levels match protein levels only 40–60% of the time (A high-stringency blueprint of the human proteome), and the mismatch is not random noise: during detachment and the epithelial–mesenchymal transition (EMT — the program by which immobile, glued-together epithelial cells become migratory mesenchymal ones), key survival proteins are kept alive by blocking their degradation rather than by making more mRNA. The clearest example is Angiopoietin-2, which stabilizes the Snail protein against proteasomal destruction and thereby drives both EMT and anoikis resistance, with no corresponding rise in Snail transcripts (Epithelial-mesenchymal transition and anoikis resistance induced by Ang2 in salivary adenoid cystic carcinoma). RNA sequencing is structurally blind to this mechanism.

Meanwhile, the machine-learning community has built increasingly sophisticated models that predict single-cell protein abundance from single-cell RNA — totalVI, a Bayesian model that jointly denoises RNA and antibody-tagged protein counts (A Joint Model of RNA Expression and Surface Protein Abundance in Single Cells); scTEL, a Transformer that predicts surface proteins and classifies cell types (A joint analysis of single cell transcriptomics and proteomics using transformer); PIKE-R2P, a graph neural network over protein-interaction networks (PIKE-R2P: Protein-protein interaction network-based knowledge embedding with graph neural network for single-cell RNA to protein prediction); and scChord, a generative flow-matching model — a method that learns to transport points from an RNA distribution into a protein distribution — that avoids the over-smoothing of deterministic regression (scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation). Every one of these is trained and benchmarked exclusively on immune cells, and every one assumes the RNA-to-protein map is stationary. This proposal tests and exploits a different hypothesis: under matrix detachment, the map itself shifts, and the shift is predictable.

Core Claim

During matrix detachment and EMT, the discrepancy between a protein's abundance and what its mRNA would predict is not irreducible noise but a structured, learnable function of the cell's dynamic state — time since detachment and position on the epithelial–mesenchymal spectrum. We will build the first single-cell multi-omic time-course of detachment and EMT in epithelial cancer cells, show that current RNA-to-protein translators fail specifically on post-translationally stabilized survival proteins, and build a state-conditional translator that recovers them. The model has one mechanistic commitment, stated as a decomposition of the log protein level $y_p$ of protein $p$ in state $s$:

$$\log y_p = f_\theta\!\left(x_p, s\right) + \gamma\, h_p\, g(s)$$

where $f_\theta$ is a generative RNA-to-protein map (flow matching, following scChord) whose input now includes the state vector $s$ = (time since detachment, hybrid E/M score), $h_p$ is the annotated half-life of protein $p$, $g(s)$ is a shared context factor interpretable as proteasome load, and $\gamma$ is a single shared coefficient. The term $\gamma\,h_p\,g(s)$ is the discordance term: long-lived proteins accumulate out of proportion to their transcripts exactly when degradation is suppressed. If this term is unnecessary — if conditioning and the half-life prior add nothing over retraining the same architecture unconditionally — we will report that and the field will have learned that in-domain retraining suffices.

Research Program

The program has three parts. First, generate the missing data: a paired single-cell RNA and protein time-course of detachment and EMT. Epithelial cancer lines (SACC-83 and SACC-LM from the Ang2 study, an NSCLC line, and an ovarian line, chosen so the results speak to the clinical cohorts below) are sampled at 0, 2, 6, 24, and 48 hours of forced suspension on poly-HEMA-coated plates, half of them with TGF-β-induced EMT, half without, so that detachment and EMT effects can be separated. Protein measurement uses CITE-seq-style antibody tags extended to intracellular epitopes, so the panel covers the proteins that actually carry the mechanism: Snail, ZEB1, E-cadherin, N-cadherin, Vimentin, ITGB4, and anti-apoptotic markers — the categories that pure surface-marker CITE-seq and the candidate Western blots of the Ang2 paper each capture only partially. Hybrid E/M states — which carry the highest metastatic fitness and anoikis resistance (Intermediate cell states in epithelial-to-mesenchymal transition; Guidelines and definitions for research on epithelial-mesenchymal transition) — are explicitly preserved by measuring single cells rather than bulk.

Second, quantify the failure. We run every existing translator twice: zero-shot as published (immune-trained) and retrained in-domain on our data, and measure per-protein Pearson correlation, RMSE, and Correlation Matrix Distance (the metric on which scChord degrades as panels grow) (scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation). The pre-registered prediction is a specific failure signature: models lose accuracy on exactly the proteins whose abundance changes strongly while their transcripts stay flat — the Ang2/Snail pattern, which the deubiquitination literature says should recur (ATXN3 stabilizes ZEB1 in glioblastoma (ATXN3 deubiquitinates ZEB1 and facilitates epithelial–mesenchymal transition in glioblastoma); HECTD1 controls SNAIL degradation (HECTD1 regulates the expression of SNAIL: Implications for epithelial‑mesenchymal transition); radiation stabilizes CDC6 protein to drive EMT (Radiation-promoted CDC6 protein stability contributes to radioresistance by regulating senescence and epithelial to mesenchymal transition)).

Third, build and validate the conditional model. Validation is mechanistic, not just metric-driven: for the top predicted discordant proteins, we run cycloheximide chase and MG132 proteasome-inhibition assays — the exact protocol the Ang2 study used to prove Snail stabilization — and test whether the model's predicted discordance correlates with measured changes in protein half-life. We then ask a clinical question: does the learned discordance correction change how the five-gene anoikis risk score (HMGA1, PLK1, ETV4, PHLDA2, ITGB4) stratifies NSCLC patients, using the GSE126044 immunotherapy cohort where 87.5% of high-risk patients failed to achieve remission (Anoikis-related gene signature is associated with immune infiltration and predicts the prognosis of non-small cell lung cancer)? If ITGB4, a surface protein, is post-translationally stabilized during detachment while HMGA1 tracks its transcripts, the risk score's meaning changes per gene — a correction only this kind of model can supply.

Baselines

The naive reference point is using RNA as a direct proxy for protein, which is what clinical anoikis risk scores implicitly do; the ceiling on this approach is the 40–60% mRNA–protein correlation reported in the human proteome blueprint (A high-stringency blueprint of the human proteome), and in ovarian cancer this proxy failure had clinical consequences: transcriptomic classification found no actionable subtypes while paired proteomics found four, with TYMP and CD40/CD40L emerging as targets invisible at the RNA level (Integrated proteomics and scRNA-seq analyses of ovarian cancer reveal molecular subtype-associated cell landscapes and immunotherapy targets). The strongest existing competitor is scChord, which reports the lowest RMSE and highest Pearson correlation against totalVI, sciPENN, and scTranslator on the CBMC (GSE100866) and PBMC (GSE164378) benchmarks, though no exact RMSE or PCC values are reported on Lacuna for it; its published throughput — 100,000 cells trained in approximately 27 minutes with 229 MB of GPU memory — gives us a concrete compute-parity target. We will run it and its published rivals (totalVI, scTEL, PIKE-R2P) both zero-shot and retrained in-domain on identical training cells, identical protein panels, and an identical tuning budget, with per-protein PCC and RMSE on the discordant protein subset as the primary comparison; beating the baseline means materially higher per-protein correlation on stabilized proteins while matching it elsewhere, and correctly recovering bimodal protein distributions where unconditional models over-smooth. As a secondary classification baseline, scTEL reports 86.6% cell-type classification accuracy on PBMC versus 77.0% for sciPENN and 76.1% for Seurat (A joint analysis of single cell transcriptomics and proteomics using transformer), a number we will reproduce as a sanity check on our pipeline before trusting our new numbers. The key ablation is our own architecture with the state conditioning and the half-life term removed — i.e., an unconditional flow-matching translator retrained on the same detachment data. That ablation, not the immune-trained zero-shot runs, is the honest test of whether the mechanism we claim actually matters.

Why not just…?

Why not just run RNA-seq and pathway enrichment, as the anoikis-signature literature does? Because the mechanism at the heart of this proposal is definitionally invisible to it. The Ang2 study proved that anoikis resistance and EMT can be driven entirely by blocked protein degradation with unchanged transcripts (Epithelial-mesenchymal transition and anoikis resistance induced by Ang2 in salivary adenoid cystic carcinoma), and single-cell proteomics in esophageal cancer showed the functional state of tumors after therapy diverged from what transcriptomics suggested (High-dimensional single-cell proteomics analysis of esophageal squamous cell carcinoma reveals dynamic alterations of the tumor immune microenvironment after neoadjuvant therapy). RNA-seq will remain the cheap workhorse; the point of this work is to tell the field which RNA-based conclusions are unsafe proxies and when. That said, if our conditional model fails to beat in-domain baselines, the honest conclusion will be that RNA-seq plus retrained translation is enough for this regime — and our dataset will still be the first to show that.

Why not just retrain the existing models on your new detachment data? This is the real threat, and it is precisely our central ablation. If in-domain retraining of scChord closes the gap on stabilized proteins, the conditional machinery is unnecessary and the proposal collapses to a valuable benchmark paper. We expect it will not close the gap, for a reason internal to the baselines' own logic: a stationary map forced to fit multiple time points must average across states, and the same over-smoothing argument scChord uses against deterministic regression applies one level up — at the state level, mixing 2-hour and 48-hour cells produces a map that fits neither. The decisive experiment is a held-out time point: train on 0, 2, 6, and 48 hours, predict 24 hours. A state-conditional model can interpolate; a stationary one has no input through which to do so.

Why not just use a global protein-interaction network with random-walk propagation instead of contextual modeling? Because context-free networks measurably underperform: PINNACLE beat random-walk target nomination in 100% of cell types, GAT in 44.2%, and BIONIC in 11.5% at APR@10, and contextualized representations plus 3D structure gave a larger PD-1/PD-L1 binding-discrimination gap (0.011 versus 0.003) than context-free ones (Contextual AI models for single-cell protein biology) — and that model was built on healthy tissues and explicitly excludes perturbation-specific networks. Network priors are also known to carry literature bias toward well-studied proteins, which is why the Steiner-tree work couples them to epigenomic evidence to surface hidden hubs like p300 and β-catenin (Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling). We borrow the network as a prior inside $f_\theta$, not as the answer, and a pure random-walk baseline on our data will quantify the difference directly.

Proposed Experiments and Decisive Checks

The experimental sequence is: generate the time-course dataset (three lines, five time points, ±TGF-β, intracellular-inclusive CITE-seq panel); reproduce the published benchmark behavior of all baselines on CBMC GSE100866 (8,617 cells, 13 proteins) and PBMC GSE164378 (>161,000 cells, 228 proteins) to verify our pipeline; run zero-shot and in-domain baselines on the detachment data and define the discordant protein set; fit the conditional model and the unconditional ablation under matched compute; validate mechanistically with cycloheximide chase and MG132 on the top predicted stabilized proteins; and apply the discordance correction to the NSCLC five-gene anoikis signature against GSE126044 immunotherapy response. Three checks are decisive. If per-protein correlation on stabilized proteins does not improve over the in-domain unconditional ablation, the mechanistic claim is wrong. If predicted discordance fails to correlate with chase-measured half-life changes, the model is fitting artifacts rather than degradation biology. And if the discordance-corrected risk score stratifies GSE126044 no better than the raw transcript signature, the clinical payoff does not materialize — but the benchmark and the failure map would still stand, and would be released with the data as the first multi-omic detachment resource usable by anyone building the next generation of RNA-to-protein models.

References

  1. A high-stringency blueprint of the human proteome
  2. Epithelial-mesenchymal transition and anoikis resistance induced by Ang2 in salivary adenoid cystic carcinoma
  3. A Joint Model of RNA Expression and Surface Protein Abundance in Single Cells
  4. A joint analysis of single cell transcriptomics and proteomics using transformer
  5. PIKE-R2P: Protein-protein interaction network-based knowledge embedding with graph neural network for single-cell RNA to protein prediction
  6. scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
  7. Intermediate cell states in epithelial-to-mesenchymal transition
  8. Guidelines and definitions for research on epithelial-mesenchymal transition
  9. ATXN3 deubiquitinates ZEB1 and facilitates epithelial–mesenchymal transition in glioblastoma
  10. HECTD1 regulates the expression of SNAIL: Implications for epithelial‑mesenchymal transition
  11. Radiation-promoted CDC6 protein stability contributes to radioresistance by regulating senescence and epithelial to mesenchymal transition
  12. Anoikis-related gene signature is associated with immune infiltration and predicts the prognosis of non-small cell lung cancer
  13. Integrated proteomics and scRNA-seq analyses of ovarian cancer reveal molecular subtype-associated cell landscapes and immunotherapy targets
  14. High-dimensional single-cell proteomics analysis of esophageal squamous cell carcinoma reveals dynamic alterations of the tumor immune microenvironment after neoadjuvant therapy
  15. Contextual AI models for single-cell protein biology
  16. Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling