Testing whether calorie restriction breaks post-translational inflammation switches in pre-malignant liver

When cells are chronically inflamed, the signals that keep them alive and push them toward cancer often do not show up in their RNA. Instead, they are written in chemical modifications of proteins — for example, small ubiquitin-like tags called SUMO that attach to the NF-κB protein p65 and shield it from the cellular garbage-disposal system, keeping pro-survival signaling on without any change in p65's mRNA (Liu et al.). Meanwhile, the best evidence that diet can prevent cancer — calorie restriction (CR) reducing liver tumors in a carcinogen mouse model — rests almost entirely on RNA sequencing, which showed that restriction "restores" the carcinogen-disturbed transcriptome toward normal (Duan et al.). This creates an unresolved tension: mRNA levels predict protein levels only 40–60% of the time across the human proteome (Adhikari et al.), and the very nodes that sustain inflammatory oncogenesis (p65, FRA1, and their upstream regulators) are controlled almost purely by post-translational modification (Liu et al.; Wang et al.; Clarke et al.).

The core claim of this proposal is that calorie restriction's protection against inflammation-driven liver cancer is executed post-translationally — by restoring ubiquitin-mediated degradation of stabilized p65 and by suppressing modification-dependent transcription-factor activation — and that transcriptomic "restoration" alone cannot detect or predict this effect. We will test this directly in the DEN carcinogenesis model with graded CR, using proteome-wide and modification-specific mass spectrometry rather than inference, and we will quantify exactly how far RNA-to-protein prediction models trained on nutrient-replete data degrade under caloric deprivation. In short: one direction (discordance) supplies the methodological lens, a second (inflammatory oncogenesis) supplies the candidate mechanism, and the third (calorie restriction) supplies the perturbation regime where all three meet.

Core Claim

Three linked hypotheses. First, in chronically inflamed pre-malignant liver, NF-κB activity is sustained at the protein level by SUMO2/3 shielding of p65, as Liu et al. demonstrated in HBV-associated tissue (SUMO2/3 and p65 protein correlated at $R = 0.800$, $p = 0.006$ in their 13-patient cohort), so a transcriptomic assay of the same tissue would report a clean, falsely reassuring picture. Second, graded caloric deficit — not protein restriction, which fails to reproduce CR's systemic anti-inflammatory effects such as lowered circulating TNF-α (Mitchell et al.) — collapses these post-translational switches: p65 loses SUMO shielding, reacquires ubiquitination, and re-enters proteasomal clearance, while the HDAC6-dependent deacetylation of FRA1 at Lys-116 that IL-6 uses to activate stemness programs (Wang et al.) is attenuated. Third, the conditional distribution $P(\text{protein} \mid \text{RNA})$ is not static: because CR reprograms translation machinery and autophagic clearance (Derous et al.), models calibrated on fed cells will systematically mispredict the proteome of restricted cells, and the size of that shift is itself a measurable, publishable quantity.

This is deliberately different from the existing two-way intersections. Duan et al. connected CR to liver cancer but only through bulk RNA-seq and candidate immunohistochemistry. Liu et al. and Wang et al. connected inflammatory oncogenesis to transcript-protein discordance but never asked whether nutrition modulates the PTM switches they found. The steatotic-liver multiomics study (Multiomics analyses decipher intricate changes in the cellular and metabolic landscape of steatotic livers upon dietary restriction and sleeve gastrectomy) brought single-cell resolution to dietary remodeling but stopped short of neoplastic progression and measured no proteome. Nobody has measured whether CR's chemoprevention is visible — or invisible — at the protein-modification layer where inflammatory oncogenesis actually operates.

Research Program

The program has three linked aims built around one animal experiment. We use the DEN-induced hepatocarcinogenesis model in C57BL/6 mice exactly as Duan et al. did, with four arms — control, graded CR, DEN, DEN+CR — plus a fifth DEN+protein-restriction arm to rule out the simpler dietary confound, since Mitchell and colleagues showed that the somatic-energy reallocation and anti-inflammatory effects are driven by caloric deficit specifically, not protein intake (Mitchell et al.). Sampling occurs at two stages: the pre-malignant inflammatory window and established tumors. The two-stage design matters because Liu et al. found that SUMO2/3 overexpression stabilizes p65 in inflamed tissue yet inhibits proliferation in fully malignant HCC lines — a stage-dependent paradox we must resolve rather than ignore.

Aim one is proteomic ground truth. For each arm and stage we collect bulk RNA-seq, TMT-labeled whole-proteome mass spectrometry, and two modification-enriched layers: SUMO2/3 immunoprecipitation followed by p65 detection and ubiquitin-remnant profiling to test directly whether CR restores p65's ubiquitination and proteasomal clearance, and acetylomics to quantify FRA1 Lys-116 deacetylation along the lines of Wang et al.. The decisive contrast is within the DEN+CR arm: if p65 protein abundance and SUMO occupancy fall while p65 mRNA is unchanged or falls only modestly, the mechanism of protection is post-translational, and RNA-only readouts of chemoprevention are quantitatively misleading. We will report the per-gene mRNA-protein correlation in each arm; Duan et al. reported no global correlation coefficient for their model, so no reported result exists on Lacuna for the DEN model's baseline transcript-protein correlation, and our first contribution is to establish it. As a ceiling reference, bulk pan-cancer data give $r \approx 0.5$ across genes and $r \approx 0.3$ across samples (Deangeli et al.).

Aim two is single-cell and spatial resolution of the inflammatory microenvironment, because Han et al. showed in esophageal cancer that transcriptomics misses functional immune rewiring (such as CCR4+CCR6+ immunosuppressive macrophages and CD16+ NK activation) that mass cytometry captures, and the steatosis study showed CR and bariatric surgery have opposite effects on myeloid infiltration and hepatocyte zonation within the same tissue (Multiomics analyses decipher intricate changes in the cellular and metabolic landscape of steatotic livers upon dietary restriction and sleeve gastrectomy). We run CyTOF with a panel designed from Han's findings plus intracellular p-p65, SUMO2/3, and FRA1, on the same livers as aim one, so that discordance between immune-cell abundance (visible in RNA) and immune-cell activation state (visible only in protein) can be mapped cell by cell.

Aim three is computational: quantifying the shift in $P(\text{protein} \mid \text{RNA})$ and using network integration to find hidden nodes. We calibrate a Proteomizer-style RNA-to-protein regressor on the fed arms and evaluate on the restricted arms, reporting Pearson correlation and error per arm; Deangeli et al. showed such models already fail across unseen tissues (their multilayer-perceptron pipeline lifted bulk Tx-Px correlation from $r = 0.27$ to $r = 0.68$ on matched data, but gains did not generalize across contexts), so our question is whether energetic state is a distinct and larger axis of failure. We also apply the Prize-Collecting Steiner Tree approach of Huang et al. — which integrates transcriptomic and phosphoproteomic prizes over a protein-interaction network to surface "Steiner nodes" like CREBBP and β-catenin that show no expression shift but carry the signaling — to the DEN+CR contrast, testing whether the protective CR program runs through unmeasured, non-differentially-expressed hubs that a transcript-factor enrichment analysis, which is what identified SP1 in Duan et al., structurally cannot find.

Baselines

The "do nothing" reference point is the analysis exactly as performed by Duan et al.: bulk RNA-seq of the four DEN-model arms plus candidate immunohistochemistry of PCNA and cleaved caspase-3. It is the strongest competitor for the claim "CR protects against DEN tumorigenesis" because it is the incumbent evidence in this very model; we run it on our own mice with the same sequencing depth and the same differential-expression and transcription-factor-enrichment pipeline, so any improvement is attributable to the added layers, not to cohort or protocol differences. Beating it means showing a concrete case where the RNA-only analysis gives the wrong sign or magnitude of protection — for example, NF-κB pathway genes restored transcriptionally while p65 protein and its SUMO occupancy persist — and demonstrating that protein-layer features predict tumor burden across arms with lower error than transcript features alone, evaluated by held-out prediction on the graded-CR dose response.

The strongest existing mechanistic method is candidate-protein immunoassay of the p65-SUMO axis following Liu et al., who used co-immunoprecipitation and ubiquitination assays in HBV-HCC tissue to establish the stabilization mechanism (SUMO2/3 and p65 protein correlation $R = 0.800$, $p = 0.006$). We run these same assays on our tissue as the targeted gold standard; our contribution beyond it is scale and context — proteome-wide and modification-enriched MS in a dietary-intervention in vivo setting, which their small in-vitro-plus-13-patient design could not address. Beating it means finding effects invisible to the candidate panel, which is exactly what the Steiner-node analysis tests.

On the computational side, the strongest existing RNA-to-protein predictor benchmarked on Lacuna material is the Proteomizer pipeline of Deangeli et al., which reaches $r = 0.68$ Tx-Px Pearson correlation on matched TCGA/CPTAC bulk data from a baseline of $r = 0.27$ for naive approaches (their own reported numbers; the natural baseline correlation across the pan-cancer data is $r \approx 0.5$ per gene). We retrain it on our fed arms with the same architecture and tuning budget and apply it to restricted arms; the fair comparison is identical inputs, identical held-out genes, and identical compute. It wins if its restricted-arm predictions match measured proteomes as well as fed-arm predictions do; our claim of a fed-restricted distribution shift is supported only if its accuracy drops by a margin we report honestly. The key ablation isolates the discordance ingredient: PCST network integration run with transcript-only prizes versus integrated transcript-plus-protein prizes, on the same network. If the integrated version surfaces nodes that predict CR protection and the transcript-only version does not, the discordance layer is doing work; if both perform equally, the RNA layer suffices and our central claim about hidden post-translational execution is weakened.

Why not just…?

Why not just sequence the RNA, since Duan et al. already showed transcriptomic restoration predicts protection? Because restoration of mRNA is not restoration of function. Across the human proteome, mRNA explains only 40–60% of protein-abundance variance (Adhikari et al.), and the specific node that maintains inflammatory NF-κB signaling in liver — SUMO-shielded p65 — was shown to operate with no change in p65 mRNA at all (Liu et al.), while central NF-κB regulators like TRRAP and CTDSPL2 likewise show no mRNA change during activation (Clarke et al.). Our experiment settles this empirically: if DEN+CR livers show restored NF-κB-related transcripts but persistent p65 protein and SUMO occupancy, the RNA-only conclusion is demonstrably incomplete in this exact model.

Why not just use an off-the-shelf model like scTranslator, scTEL, or totalVI to predict the proteome from the transcriptome and skip the mass spectrometry? This is the most serious alternative, and we do not argue around it — we test it. The evidence for concern is that Proteomizer-style models already lose their gains on unseen tissues and protocols (Deangeli et al.), and flow-based models like scChord degrade on correlation-structure accuracy as protein dimensionality scales (Zhang et al.), because these models implicitly assume $P(\text{protein} \mid \text{RNA})$ is stationary while CR rewrites translation and autophagic clearance (Derous et al.). But if the models transfer to restricted tissue within noise, the cheap route wins for future studies and we will say so plainly; that outcome would itself be a useful negative result, and our aim-three design reports it either way. What the models cannot do even if they transfer is measure SUMO occupancy, ubiquitin remnants, or FRA1 acetylation — the mechanistic heart of the proposal — so the PTM layer requires direct measurement regardless.

Why not just do Western blots for p65, SUMO, and FRA1? Targeted assays are necessary but not sufficient. The candidate-marker approach is biased toward known nodes; unbiased network integration in glioblastoma showed the critical coordinators (CREBBP, β-catenin, p300) were "Steiner nodes" with no expression change that candidate screens skip (Huang et al.), and bulk assays homogenize the spatially opposite effects CR has across tissue compartments, such as disrupting periportal hepatocyte zonation while shifting myeloid density (Multiomics analyses decipher intricate changes in the cellular and metabolic landscape of steatotic livers upon dietary restriction and sleeve gastrectomy). We keep the targeted assays as the mechanistic gold standard and add the unbiased layer precisely to catch what they miss.

Why not just use protein restriction in vitro instead of graded calorie restriction in mice? Because it is the wrong intervention. Controlled titration studies showed that the systemic adaptations and the anti-inflammatory, insulin-sensitizing effects are driven by the caloric deficit itself, with pure protein restriction failing to lower TNF-α or improve glucose homeostasis (Mitchell et al.; Mitchell et al.). Since our hypothesis concerns TNF-α-linked SUMO2/3-p65 stabilization, protein restriction is an invalid proxy; we include it only as an active control expected to fail, which makes the causal attribution to energy deficit rather than amino-acid composition.

Decisive checks

Three outcomes decide the proposal. First, in DEN+CR versus DEN livers, p65 SUMO occupancy and protein abundance fall while p65 mRNA does not track the fall — the post-translational-execution hypothesis — or they do not, in which case transcriptomic restoration is a faithful surrogate and we report that. Second, in the stage comparison, the SUMO2/3 effect flips sign between pre-malignant and tumor tissue, resolving the paradox Liu et al. observed in vitro. Third, the Proteomizer-style model retrained on fed arms loses a quantified amount of accuracy on restricted arms, establishing energetic state as a distinct generalization axis for RNA-to-protein models — or it does not, and the field learns that the cheap inference route survives metabolic perturbation. As a clinical-translation anchor, we test whether protein-layer (but not transcript-layer) features stratify human HCC prognosis, mirroring how Han et al. used chemokine features to stratify TCGA-ESCC survival ($p = 0.000073$ with their ridge-penalized Cox model).

References

  1. Small ubiquitin-related modifier 2/3 interacts with p65 and stabilizes it in the cytoplasm in HBV-associated hepatocellular carcinoma
  2. Dietary restriction protects against diethylnitrosamine-induced hepatocellular tumorigenesis by restoring the disturbed gene expression profile
  3. A high-stringency blueprint of the human proteome
  4. The inflammatory cytokine IL-6 induces FRA1 deacetylation promoting colorectal cancer stem-like properties
  5. Signalome-wide mapping of the NFκB pathway in T-cells reveals novel targets for immunotherapy
  6. The effects of graded levels of calorie restriction: II. Impact of short term calorie and protein restriction on circulating hormone levels, glucose homeostasis and oxidative stress in male C57BL/6 mice
  7. The effects of graded levels of calorie restriction: XI. Evaluation of the main hypotheses underpinning the life extension effects of CR using the hepatic transcriptome
  8. Multiomics analyses decipher intricate changes in the cellular and metabolic landscape of steatotic livers upon dietary restriction and sleeve gastrectomy
  9. The effects of graded levels of calorie restriction: I. impact of short term calorie and protein restriction on body composition in the C57BL/6 mouse
  10. Proteomizer: Leveraging the Transcriptome-Proteome Mismatch to Infer Novel Gene Regulatory Relations
  11. High-dimensional single-cell proteomics analysis of esophageal squamous cell carcinoma reveals dynamic alterations of the tumor immune microenvironment after neoadjuvant therapy
  12. Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling
  13. scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation