Predicting Matrix Protein States from Transcripts, Across the Dish-to-Joint Gap

Cells sit inside a structural mesh called the extracellular matrix (ECM), whose health depends on a running balance between synthesis of structural proteins like collagen and aggrecan, and their enzymatic degradation by proteases such as MMP-13. When that balance tips — as in osteoarthritis — the tissue collapses, and every existing therapy for it has failed in clinical translation. Meanwhile, nearly all discovery work on this problem measures mRNA, because RNA sequencing is cheap and standardized. This is a known trap: mRNA levels predict actual protein abundance only 40% to 60% of the time (A high-stringency blueprint of the human proteome), and in cancer drug screening, models trained on proteins beat models trained on RNA head-to-head (median Pearson correlation 0.5467 versus 0.5311; Drug Response Modeling across Cancers: Proteomics vs. Transcriptomics).

Two further failures compound this. First, most ECM discovery is done in flat 2D chondrocyte cultures dosed with an inflammatory cytokine, and what rescues cells in a dish routinely fails in a loaded, sheared, immune-competent joint. Second, computational models that do translate RNA measurements into predicted protein abundances (scChord) have only ever been trained on suspended blood cells and a few dozen immune surface markers — never on insoluble structural matrix proteins, protease-inhibitor stoichiometry, or sugar-chain modifications that sit on the protein after translation and, as Characterization of aberrant glycosylation associated with osteoarthritis based on integrated glycomics methods showed, change on over 27% of measured matrix glycopeptides in osteoarthritic cartilage.

This proposal builds a single model that addresses all three failures at once: a generative RNA-to-protein translation model whose output layer targets the matrisome (the full inventory of matrix structural proteins, proteases, and their inhibitors), whose protein-protein correlation structure is guided by a physical interaction prior, and whose readout is explicitly domain-adapted from in vitro perturbation data to in vivo joint outcomes.

Core Claim

The central claim is that predicting in vivo ECM protein state from transcriptomics requires two ingredients that no current model has, and that adding them together is what makes the dish-to-joint prediction possible: (1) a network-guided generative translation head that predicts not just protein abundances but the correlation structure among matrix builders, proteases, and inhibitors — because ECM fate is decided by the stoichiometric balance between MMPs and TIMPs, not by any single protein's level (The effect of chronic periodontitis on serum levels of matrix metalloproteinase-2 (MMP-2), tissue inhibitor of metalloproteinase-1 (TIMP-1), interleukin-12 (IL-12) and granulocyte-macrophage colony-stimulating factor (GM-CSF) showed that inhibitor levels rise 16-fold in disease and still fail, so a univariate predictor is structurally blind); and (2) a permutation-invariant adaptive readout, of the kind DeepSADR proved necessary for crossing the cell-line-to-patient domain shift in oncology, retrained here to map in vitro cytokine-perturbed chondrocyte states onto in vivo cartilage degradation outcomes.

The concrete testable prediction: a model with both ingredients will predict in vivo matrix protein state and OARSI-scored cartilage damage from in vitro transcriptomic perturbation screens more accurately than (a) transcript abundance used directly as a proxy, (b) the strongest existing RNA-to-protein translation model retrained on the same data, and (c) the same architecture with either ingredient ablated.

Research Program

The work proceeds in three stages.

The first stage extends the RNA-to-protein translation machinery to matrix biology. The starting method is the two-stage generative design of scChord — a variational autoencoder that absorbs single-cell dropout noise, followed by a flow-matching map (a generator that learns to transport points from the transcript distribution to the protein distribution along straight trajectories). We keep that backbone but change three things. The output panel becomes the matrisome: collagens, aggrecan, fibronectin, MMPs, TIMPs, and the glycosyltransferases and trafficking factors (MGAT4C, MGAT5B, ALG3, ALG5, PLOD3) that post-translationally mature matrix proteins and were shown to shift in inflamed cartilage (Characterization of aberrant glycosylation associated with osteoarthritis based on integrated glycomics methods). The training signal becomes site-level proteoform abundance where it exists, not just total protein, because Regulation of post-Golgi LH3 trafficking is essential for collagen homeostasis demonstrated that collagen transcripts can be perfectly normal while the modifying enzyme is diverted to lysosomes and fibrils collapse — a failure mode invisible to any total-abundance predictor. And, critically, the flow objective gains a correlation-matching term: a penalty on the distance between the predicted protein-protein correlation matrix and the measured one, estimated over random protein subsets. This directly attacks the documented weakness that flow-based translation degrades on the Correlation Matrix Distance metric as panels scale toward 228 proteins (scChord), and it imports the insight from Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling that interaction-network priors recover functional regulators (their CREBBP, β-catenin, p300 "Steiner nodes") that raw single-modality screens miss entirely.

The second stage is the domain transfer. In vitro training data comes from cytokine- and drug-perturbed chondrocyte screens (IL-1β stimulation protocols as used in NUMB attenuates posttraumatic osteoarthritis by inhibiting BTRC and inactivating the NF-κB pathway, Blockade of the Notch Pathway by Inhibition of Fucosylation Ameliorates Chondrocyte Inflammation and Extracellular Matrix Degradation, and N6-methyladenosine (m6A) methyltransferase WTAP-mediated miR-92b-5p accelerates osteoarthritis progression, plus LINCS L1000-style perturbation profiles). The in vivo target comes from paired cartilage datasets: the GSE51588 osteoarthritic-cartilage expression cohort, the 393-glycopeptide LC-MS/MS cartilage cohort of Characterization of aberrant glycosylation associated with osteoarthritis based on integrated glycomics methods, and ACLT/DMM animal-model outcome data from the papers above, where interventions (pcDNA-NUMB, SGN-2FF, shWTAP) have known in vivo OARSI outcomes. Following the DeepSADR recipe, low-level translation weights are pretrained on the in vitro screens and frozen; only a permutation-invariant set-transformer readout and the outcome predictor are fine-tuned on the in vivo paired data. The readout must learn to reweight pathway-level interactions differently in the joint than in the dish — exactly the mechanism whose removal collapsed DeepSADR's patient-response AUC from 0.856 to 0.662.

The third stage turns the model into a screening instrument. Given an untested perturbation's transcriptomic signature in vitro, the model predicts the in vivo matrisome state — specifically the MMP:TIMP balance and structural-protein proteoform state — and ranks perturbations by predicted matrix preservation. The evaluation uses drug-disease retrieval logic modeled on DDR-Bench from Cell-Level Virtual Screening: can the model retrieve, from a held-out panel, the interventions that actually protected cartilage in the ACLT and DMM models, using only their in vitro signatures?

Proposed Experiments

Experiment 1 (translation fidelity on existing benchmarks) retrains our network-guided model and the strongest competitors on the paired CITE-seq benchmarks GSE100866 (~8,600 cord-blood cells, 13 proteins) and GSE164378 (>161,000 peripheral-blood cells, 228 proteins), the exact evaluation sets of scChord. Metrics are RMSE, Pearson correlation per protein, and Correlation Matrix Distance at both panel sizes. This establishes that the correlation-guiding term does not damage ordinary translation accuracy before any ECM claims are made. No exact RMSE or PCC value for scChord is reported in the retrieved material — the file states only that scChord achieved the lowest RMSE and highest PCC against totalVI, sciPENN, and scTranslator, so the comparison must be run head-to-head rather than against a quoted number.

Experiment 2 (matrisome translation, the real test) applies the same models to cartilage data: predicting matrix protein and site-level glycopeptide abundance from transcript profiles, trained on the healthy side of the glycomic cohort and evaluated on the osteoarthritic side, with GSE51588 providing an independent transcript-only check. Success means beating the same competitors on matrisome targets specifically, including FN1 site N528 and ACAN site N333, where health-to-disease glycan switching is documented.

Experiment 3 (dish-to-joint transfer) evaluates the full pipeline: frozen translation backbone, adaptive readout fine-tuned on paired in vivo data, predicting OARSI-ranked cartilage protection from in vitro signatures alone. The decisive metric is retrieval Hits@1 on held-out protective interventions, compared against molecular fingerprints and against raw expression snapshots as in Cell-Level Virtual Screening. A wet-lab confirmation arm is feasible: the model's top-ranked untested perturbation and its bottom-ranked one are administered in the ACLT rat protocol shared by the NUMB and SGN-2FF studies, with OARSI scoring and glycopeptide LC-MS/MS as outcomes.

Baselines

The naive reference point is using transcript abundance directly as a proxy for protein state — what most ECM literature effectively does. Its measured ceiling is known: RNA and protein agree only 40–60% of the time (A high-stringency blueprint of the human proteome), and on the ProCan-DepMapSanger benchmark an RNA-trained model reached a median Pearson correlation of 0.5311 against 0.5467 for the protein-trained model (Drug Response Modeling across Cancers: Proteomics vs. Transcriptomics). We will run this baseline as simple per-gene transcript-to-protein regression fit on the same training splits and tuning budget as our model.

The strongest existing competitor is scChord (scChord), the current best RNA-to-protein translation model, which outperformed totalVI, sciPENN, and scTranslator on RMSE and PCC while training 100,000 cells in roughly 27 minutes with 229 MB of GPU memory. We will retrain it from its published recipe on every dataset we use, with the same compute and hyperparameter-search budget as our model, and compare on identical held-out folds. totalVI (Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells) serves as the established generative-alternative baseline, run the same way. Beating scChord means matching or exceeding its RMSE and PCC on GSE100866/GSE164378 while substantially improving Correlation Matrix Distance at the 228-protein scale and on the matrisome panel — its two documented weak points.

Two ablations isolate the proposal's key ingredients. Removing the correlation-matching term (leaving plain flow matching) tests whether network guidance is what fixes high-panel correlation degradation; the expectation, grounded in scChord's documented CMD deterioration at 228 proteins, is a significant drop on the matrisome panel. Removing the adaptive readout in favor of standard pooling — the exact ablation that collapsed DeepSADR's AUC from 0.856 to 0.662 — tests whether domain-adaptive readout is what carries the dish-to-joint prediction; if pooled readout transfers equally well, the adaptive component is unnecessary and we would report that honestly.

Why not just…?

Why not just use RNA-seq expression snapshots of MMP13 and COL2A1, as the ECM field does? The evidence against this is direct in the retrieved material. Transcript and protein abundance agree only 40–60% of the time (A high-stringency blueprint of the human proteome); head-to-head modeling showed RNA is strictly inferior to protein as input (0.5311 vs. 0.5467; Drug Response Modeling across Cancers: Proteomics vs. Transcriptomics); and specifically in matrix biology, WTAP-mediated m6A marks degrade TIMP4 transcripts post-transcriptionally so that inhibitor protein falls far below what transcription suggests (N6-methyladenosine (m6A) methyltransferase WTAP-mediated miR-92b-5p accelerates osteoarthritis progression), while disease collapse proceeds even when TIMP-1 levels rise (The effect of chronic periodontitis on serum levels of matrix metalloproteinase-2 (MMP-2), tissue inhibitor of metalloproteinase-1 (TIMP-1), interleukin-12 (IL-12) and granulocyte-macrophage colony-stimulating factor (GM-CSF)). A transcript-only pipeline is blind to exactly the stoichiometric and post-translational failures that destroy matrix. Experiment 2 quantifies this directly.

Why not just fit a standard regression or use totalVI instead of a two-stage generative model? This is a fair challenge on cost grounds: regression is faster at inference and avoids multi-step ODE integration, a latency issue the scChord authors themselves flag. But the retrieved evidence shows conditional-mean regression over-smooths multimodal protein distributions (it collapses the bimodal CD4 and heavy-tailed CD45RA distributions into their averages), and scChord's probabilistic rectification strictly outperformed totalVI, sciPENN, and scTranslator on RMSE and PCC (scChord). Since matrix fate depends on rare-tail imbalance between proteases and inhibitors, losing the tails loses the phenotype. totalVI remains in the baseline suite precisely so this claim is tested rather than asserted; it is a real threat on speed, and if it matches accuracy on matrisome targets at a fraction of the cost, the generative component should be dropped.

Why not just align in vitro and in vivo expression profiles with standard batch-correction or pooling? The closest existing work tested this directly: replacing DeepSADR's adaptive set-transformer readout with standard pooling collapsed clinical response AUC from 0.856 to 0.662, and global-expression baselines (GANDALF, WISER) reached only 0.811 against DeepSADR's 0.957 on clinical Sorafenib (DeepSADR). The mechanism reason is that pathway-level interactions reweight between dish and organism, so pooled global alignment averages away the shift instead of adapting to it. We will include a batch-corrected pooled-readout variant as the ablation above, so the comparison is empirical, not rhetorical.

Why not just screen compounds by molecular fingerprints and target binding? On the drug-disease retrieval benchmark, molecular fingerprints achieved a Hits@1 of only 0.0357, barely above the 0.0179 random baseline, while cell-level network prediction reached 0.1250 — and predicted network rewiring even beat measured expression snapshots, because dependency networks filter noise (Cell-Level Virtual Screening). Matrix protection is an emergent tissue-level property, not a binding event, so this gap should be even larger here. We include fingerprint retrieval as a baseline in Experiment 3.

Decisive Checks

The proposal stands or falls on three outcomes. First, the correlation-guided translation must beat unguided flow matching on Correlation Matrix Distance at high panel sizes — if it does not, network guidance is dead weight and the simpler scChord backbone should carry the rest. Second, matrisome targets must show larger transcript-protein discordance than immune surface markers, and the model must capture it: the Yu et al. glycomic shifts (Characterization of aberrant glycosylation associated with osteoarthritis based on integrated glycomics methods) and the Banushi et al. trafficking failure (Regulation of post-Golgi LH3 trafficking is essential for collagen homeostasis) predict it will, since both are post-transcriptional mechanisms invisible to RNA. Third, the dish-to-joint retrieval must clearly separate true cartilage protectors from failures using in vitro signatures alone, and the ACLT confirmation arm must agree with the model's ranking. If all three hold, the result is a reusable instrument: any lab's cheap transcriptomic perturbation screen becomes a predictor of in vivo matrix fate — the piece that has been missing every time a dish-proven therapy reached a joint and failed.