Separating Ligand Bias from Tissue Bias in Bulk Phosphoproteomics

When a drug targets a G protein-coupled receptor (GPCR) — the family behind roughly a third of all medicines — it does not simply switch the receptor on. The receptor is a flexible molecule, and different ligands stabilize different shapes, so a ligand can preferentially drive one intracellular pathway (say, G-protein signaling) while barely engaging another (say, β-arrestin recruitment). This is called biased agonism, and it is the reason drugs like the opioid oliceridine, whose effective doses were pinned down clinically at an ED50 of 18.45 µg/kg in Establishing the Median and 95% Effective Doses of Oliceridine..., promise analgesia with less respiratory depression. But nearly everything we know about bias comes from measuring signaling in one engineered, homogeneous cell line — while real drugs act on tissues containing many cell types whose effector machinery differs. When we measure a mixed tissue's response, the readout is an average, and we cannot tell how much of the apparent pathway preference belongs to the ligand and how much is an artifact of which cells happened to be present.

I propose to make that separation measurable. The plan is to challenge a heterogeneous tissue model with a panel of agonists spanning a bias spectrum, read out the response with multiplexed quantitative phosphoproteomics (mass spectrometry that counts, per protein site, how much phosphate is attached), and fit a factor model that explicitly splits the observed signaling into a ligand-intrinsic part shared across cell types and a cell-type-specific part. The result is a per-cell-type map of G-protein versus β-arrestin engagement that no bulk assay and no single-line assay can give, plus an estimate of how much the tissue's composition distorts the bias a clinician would infer from a bulk measurement.

Core Claim

Apparent signaling bias measured in a mixed tissue factorizes into two pieces: a ligand-intrinsic bias trajectory, and a cell-type-specific gain that reflects each lineage's effector stoichiometry. If we measure a reference "response tensor" — phosphoproteome responses of each isolated cell type to each ligand across early time points — then a low-rank factorization of that tensor, combined with cell-type abundances estimated from the same samples, can (i) recover per-cell-type pathway activation from a bulk measurement of a challenged tissue, and (ii) predict responses to held-out ligands it never saw. The factorization is the new ingredient: existing proteomic deconvolution only estimates static cell-type fractions, and existing single-cell signaling models work off small predefined antibody panels.

Research Program

The motivating evidence is already in the literature, split across three strands. First, bias is real and structural: Computationally designed GPCR quaternary structures bias signaling pathway activation engineered CXCR4 dimers that kept normal G-protein signaling but abolished β-arrestin recruitment, and confirmed the downstream consequence with ERK phosphorylation kinetics — but exclusively in recombinant HEK293T cells, deliberately leaving native heterogeneous tissue untested. Second, GPCR signaling is radically lineage-dependent: The function of GPCRs in different bone cells catalogs how parathyroid hormone 1 receptor (PTH1R) drives bone-forming cAMP pathways in osteoblasts yet distinct cascades in osteoclasts and chondrocytes, with receptor expression across five skeletal lineages resolved only at the transcript level. Third, the measurement and deconvolution machinery now exists but has never been pointed at this problem: Deciphering Cell Type Abundance in Proteomics Data Through Graph Neural Networks gives GraphDEC, the strongest method for estimating cell-type fractions from proteomics, but it deconvolves baseline composition, not perturbation responses or phosphorylation; Cell‐to‐cell and type‐to‐type heterogeneity of signaling networks: insights from the crowd proves that models accounting for cell-to-cell variability beat homogeneous models, but only through a 36-antibody mass-cytometry panel under kinase inhibitors, with no untargeted depth and no GPCR pharmacology.

The program has three parts.

The reference tensor. I will build a mouse bone co-culture model containing the four main skeletal lineages (mesenchymal stromal cells, osteoblast-lineage cells, osteoclast-lineage macrophages, chondrocytes), chosen because PTH1R's lineage-divergent signaling is the best-documented case of one receptor producing opposing phenotypes by cell type. Each sorted lineage is treated separately, ex vivo, with a ligand panel spanning the bias axis — a balanced agonist, G-protein-biased and β-arrestin-biased analogues, and vehicle — sampled at 1, 5, 15, and 60 minutes, the seconds-to-minutes regime where phosphorylation cascades run long before any transcription changes, as the single-cell phosphoproteomics device work in Device and Method for Parallel Measurement of Phosphoproteome and Proteome from Single Cells makes explicit. Quantification uses TMT isobaric tagging, whose 16–18 multiplexed channels accommodate the ligand-by-time grid for one cell type per run; the same chemistry has already mapped signaling-pathway dysregulation in Xiaoxuan Bai et al.. Pathway activation scores are computed from phosphosite motifs: protein kinase A substrate sites report the Gs-cAMP axis, protein kinase C sites the Gq axis, and ERK and related sites the arrestin/ERK arm that the CXCR4 dimer paper used as its downstream bias readout.

The factor model. Let $s$ index phosphosites, $c$ cell types, $\ell$ ligands, and $t$ time points. The reference tensor is modeled as

$$R_{s,c,\ell,t} \approx \sum_{k=1}^{K} a_{k,\ell,t}\; g_{k,c}\; v_{k,s}$$

where $a_{k,\ell,t}$ is the ligand-intrinsic bias trajectory of component $k$ (shared across cell types), $g_{k,c}$ is a cell-type gain (the system-bias factor), and $v_{k,s}$ loads component $k$ onto a pathway module of phosphosites. A bulk measurement of a challenged tissue is then $y_{s,\ell,t} = \sum_c p_c R_{s,c,\ell,t} + \varepsilon_{s,\ell,t}$, where $p_c$ are cell-type fractions obtained from the same samples' baseline proteome using GraphDEC. The model is fit by alternating least squares; a small $K$ (two to four) is the scientific hypothesis that bias lives on a low-dimensional axis. This continuous-intensity model deliberately avoids the discrete-count machinery of Count Bridges enable Modeling and Deconvolving Transcriptomic Data, which is built for integer RNA counts and does not transfer to mass-spectrometry ion intensities, and it sidesteps the RNA-to-protein mismatch that makes transcriptomic deconvolvers like POPPY in Cross-modal transfer learning for mapping bulk transcriptomes at cellular level inapplicable to phosphorylation. Batch disentanglement across TMT runs follows the invariant/spurious latent split demonstrated in Conditionally Invariant Representation Learning for Disentangling Cellular Heterogeneity.

Tissue application and ground truth. The intact co-culture, and separately dosed mouse bone in vivo, is challenged with the same ligand panel and measured in bulk — this is the input the method must explain. Ground truth comes from two directions. First, from challenged co-cultures we FACS-sort the lineages post-hoc and measure each type's phosphoproteome directly, capturing any paracrine crosstalk that the isolated-type tensor misses; the discrepancy between factor-model predictions and these direct measurements quantifies microenvironment effects, exactly the confound bulk assays cannot see. Second, Deep Visual Proteomics defines single-cell identity and heterogeneity supplies the spatial arm: its imaging, AI segmentation, and 200 nm-precision laser microdissection isolate phenotype-matched cell groups from intact challenged tissue, and although its authors state phosphoproteomics is not yet incorporated, its >5,000-protein, ~100-cell readout (quantitative reproducibility Pearson r = 0.96) independently verifies composition and total-proteome shifts where phosphorylation is out of reach.

Proposed Experiments

Evaluation rests on three tests. First, pseudo-bulk benchmarks: mix the reference tensor's per-type profiles in known proportions (the same pseudo-bulk strategy GraphDEC used) and score recovery of per-cell-type pathway activation by concordance correlation coefficient (CCC) and root-mean-square error, the metrics used in the GraphDEC and DREAM-challenge literatures. Second, held-out-ligand prediction: train the factorization on a subset of the ligand panel and predict per-type responses to a withheld biased ligand, mirroring the missing-conditions task in the DREAM challenge. Third, the tissue test on intact co-cultures and in vivo dosed bone, scored against the sorted-type direct measurements. As a secondary, generalization check, the same pipeline is applied to a μ-opioid receptor system with oliceridine in a mixed neuronal/glial culture, since the clinical doses are established (Establishing the Median and 95% Effective Doses...; the companion anesthesia trial Estimation of median effective doses... reports no exact ED50/ED95 number in its abstract summary, so only the first paper's values anchor dosing). A small mass-cytometry arm with a fixed panel corroborates the canonical axes cheaply. Success means per-type pathway scores within the noise floor of the sorted-type ground truth and a calibrated statement of how far bulk-inferred bias deviates from ligand-intrinsic bias in real tissue.

Baselines

The naive reference point is doing nothing: interpret the bulk phosphoproteome of the challenged tissue as if every cell responded identically. It is the de facto standard in pharmacological tissue assays, and it is beaten if any deconvolution predicts per-type responses materially better than the bulk average when scored against sorted-type ground truth. The strongest existing method is a two-stage pipeline combining GraphDEC for cell-type fractions with the reference tensor used as a fixed lookup table — the bulk response is explained by composition-weighting each type's measured response to that ligand, with no factorization and no generalization. GraphDEC is the right competitor for the composition stage because it is the strongest published proteomic deconvolver: it reports a CCC of 80.7% on seven synthetic single-cell proteomics benchmarks, a 15% improvement over the next best method (CellDART), and an average CCC of 93% on cross-species proteomics, outperforming Tangram by 26.4% Deciphering Cell Type Abundance in Proteomics Data Through Graph Neural Networks. It is weakest exactly where this proposal lives, since it estimates static fractions and never touches perturbation responses or PTMs; running it fairly means giving it the same reference data, the same composition inputs, and the same tuning budget, and beating it means better per-type response recovery on held-out ligands, where the lookup table has no entry and the factor model does. The strongest learned competitor is the DREAM-challenge winning approach from Cell‐to‐cell and type‐to‐type heterogeneity of signaling networks: insights from the crowd: models that use cell-type identity and basal signaling states to predict single-cell phosphorylation, which achieved an RMSE of 0.849 with median correlation 0.761 and median R² of 0.40 on the missing-marker task, and up to 48% improvement on missing conditions. We retrain that model family on our tensor (the original is CyTOF-based, so the comparison is architecture-to-architecture on identical data and compute) and beat it if we exceed its RMSE and correlation on per-type prediction while also recovering interpretable bias axes it does not produce. Finally, the key ablation removes the proposal's central ingredient: fit each cell type's response to each ligand independently, with no shared low-rank ligand factors. If the factorized model does not beat this independent-fit ablation on held-out ligands, the factorization — and the ligand-intrinsic-versus-tissue-bias separation it licenses — is not earning its keep, and I would report that.

Why not just…?

Why not just use single-cell RNA-seq or spatial transcriptomics? This is the cheapest alternative and it fails on two counts documented in the file. Transcript levels correlate weakly with protein abundances, which is precisely why GraphDEC was built for proteomics rather than reusing transcriptomic deconvolvers Deciphering Cell Type Abundance..., and GPCR functional selectivity plays out in phosphorylation within seconds to minutes, far ahead of any transcriptional change, as the trace-input phosphoproteomics work emphasizes Device and Method.... The bone GPCR atlas itself is transcript-only and explicitly leaves proteomic signaling states unmeasured The function of GPCRs in different bone cells. RNA data can supply composition priors, and we use it only for that.

Why not just use mass cytometry (CyTOF)? This is the most serious threat, and I will not argue it away: for the canonical Gs/Gq/arrestin axes, a phospho-specific CyTOF panel at true single-cell resolution would corroborate much of the per-type story, which is why a CyTOF arm is built into the experiments. Its documented limits are the fixed panel — 36 antibodies across more than 80 million single-cell measurements in the DREAM challenge, constrained by antibody availability and cost Cell‐to‐cell and type‐to‐type heterogeneity... — and the inability to discover unexpected phosphosites or pathway crosstalk, the kind of untargeted finding that deep proteomics has repeatedly delivered, such as the uncharacterized cell-cycle regulator C7orf50 found by Deep Visual Proteomics Deep Visual Proteomics.... The decisive comparison is run head-to-head: if the panel matches the untargeted deconvolution on the known axes, the proposal's value narrows to discovery, and I would say so.

Why not just quantify bias with BRET reporters in HEK293T cells? That is the field standard and it is exactly the practice whose blind spot this proposal targets: Computationally designed GPCR quaternary structures... quantified dimer-driven bias beautifully in homogeneous HEK293T and noted nothing about native tissue, and the same homogeneous-line limitation holds for the allosteric FFAR1 study, whose intracellular-loop bias switch was validated only in bulk IP1 assays A non-canonical mechanism of GPCR activation. A single transformed line cannot express the lineage-specific effector stoichiometries that the bone review shows produce opposing PTH1R phenotypes. The experiments settle it: if per-type gains $g_{k,c}$ come out uniform across lineages, the homogeneous-cell assumption was fine and the tissue-bias concern was overstated.

Why not just run CIBERSORTx, MuSiC, or plain least-squares regression on the reference profiles? Off-the-shelf deconvolvers are composition tools for static data; on proteomics they are already dominated by GraphDEC, with CellDART 15% and Tangram 26.4% worse on the benchmarks cited above. Plain non-negative least squares of the bulk response onto per-type reference vectors is a real ablation, not a straw man — it is precisely the independent-fit ablation described in the Baselines section — and the honest risk is that with a dense, fully-sampled ligand-by-type grid it may match the factor model. The factorization's claim to superiority is confined to the sparse regime: predicting held-out ligands and unmeasured cell-type/ligand combinations, which least squares cannot do by construction. That is the regime the held-out-ligand experiment is designed to test.

Decisive Checks

Three findings would make or break the proposal. If the low-rank factorization fails to beat the independent-fit ablation on held-out ligands, the shared-bias-axis hypothesis is wrong and the honest fallback is the lookup-table baseline. If the sorted-type ground truth from challenged co-cultures diverges sharply from isolated-type predictions, the divergence itself becomes the headline: it quantifies paracrine crosstalk, the microenvironmental effect that both bulk assays and single-line assays are blind to. And if bulk-inferred bias in the in vivo bone samples deviates substantially from the recovered ligand-intrinsic bias — because bone's cellular composition shifts with disease or age — then the practical payoff is concrete: a corrected, tissue-aware readout of what a biased agonist is actually doing to each cell type, which is the quantity that clinical dose-finding of the oliceridine kind has never had access to.