Unmixing Kinase Activity: Time-Resolved Phosphoproteomic Deconvolution with miRNA-Perturbation References

When a researcher measures phosphorylation — the phosphate tags that switch proteins on and off within seconds to minutes — in a tumor biopsy, the signal is a weighted average over hundreds of thousands of cells. A rise in phosphorylated JNK could mean the rare, aggressive stem-like cells that drive drug resistance turned a pathway on, or it could mean surrounding non-malignant stroma did. Existing computational unmixing tools answer exactly this question for messenger RNA counts, but nothing answers it for phosphorylation. The gap matters because microRNA-based therapeutics, a major class of experimental cancer drugs, change protein phosphorylation without changing mRNA levels — Nie et al. showed that a miR-148b mimic suppresses the MAP3K9 protein and downstream JNK phosphorylation in renal cancer cells while the MAP3K9 mRNA stays flat. In that regime, RNA-based unmixing tells you who is in the room but says nothing about what any of them are doing.

This proposal builds a method that unmixes bulk phosphoproteomic time courses into per-subpopulation kinase-activity trajectories, and validates it where the ground truth is purchasable: cell-line mixtures with a rare, FACS-sortable cancer stem cell (CSC) fraction, perturbed with synthetic microRNA mimics and profiled over time with a microscale phosphoproteomics workflow. The core trick is that a perturbation that hits subpopulations differently creates temporal fingerprints that make an otherwise ill-posed mixture problem identifiable.

Core Claim

Bulk phosphoproteomic measurements can be decomposed into cell-state-specific signaling trajectories, provided the reference signatures are time-indexed and perturbation-induced rather than static. Formally, for phosphopeptide $i$ at time $t$ in a mixture,

$$y_i(t) \;=\; \sum_{k} \pi_k(t)\, a_{ik}(t) \;+\; \varepsilon_i(t),$$

where $\pi_k(t)$ is the fraction of subpopulation $k$ and $a_{ik}(t)$ is the per-cell abundance of phosphopeptide $i$ in that subpopulation. With static references, $\pi_k$ and $a_{ik}$ are not separately identifiable — this is precisely why current deconvolution methods freeze $a_{ik}$. I claim that when subpopulations respond to a perturbation with distinct kinetics (a mimic that represses a kinase cascade in CSCs but not in bulk-differentiated cells), the temporal diversity of $a_{ik}(t)$ breaks the degeneracy, and that matched RNA-seq at sparse time points anchors $\pi_k(t)$ from the composition side while the phospho data carries the activity side. The output is an attribution: which subpopulation drove which phosphorylation change, and when.

No existing method attempts this. Count Bridges deconvolves bulk data into single-cell count profiles, but its machinery is built on Poisson birth-death dynamics over integer RNA counts; phosphorylation stoichiometry is a continuous ratio in $[0,1]$ and phosphopeptide intensities are continuous, so the model assumptions fail by construction. POPPY maps bulk transcriptomes onto single-cell atlases with ontology-aware contrastive learning, but is observational, static, and purely transcriptomic. Nie et al. connect miRNA mimics to phosphorylation, but via a single static Western blot on one homogeneous cell line. The proposal sits deliberately in the hole between all three.

Research Program

Three components. First, a statistical model — call it KinaseMix — for continuous-intensity mixture deconvolution with time-varying per-type signatures. It will use a regularized linear likelihood on log phosphopeptide intensities (Gaussian or heavy-tailed, matching the 8–10% coefficient of variation reported for the measurement workflow in SOP-Phos), smoothness priors on $a_{ik}(t)$, optional beta-likelihood handling for stoichiometry ratios on sites where phospho and total forms are both quantified, and an RNA-derived constraint on $\pi_k(t)$. Deliberately, the model will be a sparse linear dynamical system rather than a foundation model: PerturbCausal showed that virtual cell models (GEARS, CPA, Geneformer, STATE) inflate predicted perturbation effect sizes roughly 23-fold and lose to plain ElasticNet on causal-recovery benchmarks, which is a direct warning against heavyweight generative machinery for exactly this kind of perturbation-response modeling. Second, a reference-signature protocol: FACS-sort the rare CD24⁺CD44⁺ESA⁺ CSC subpopulation (0.1–0.9% of the total population, per the PANC-1 assay of Lu et al.) and the complement fraction, deliver a miR-200a mimic (established to force mesenchymal-to-epithelial transition and cut invasion to 0.33-fold and migration to 0.22-fold in these cells), and profile both fractions separately at 0, 1, 4, 12, 24, 48, and 72 hours with the SOP-Phos microscale workflow, which quantifies 6,548 ± 21 phosphopeptides from as few as 2,500 cells with <1% missing values (SOP-Phos). DIA spectra will be processed with in-silico libraries from DeepPhospho, whose predicted fragment intensities reach a median Pearson correlation of 0.968 on RPE1 DDA data and raise phosphopeptide identification by up to 25.9% over experimental libraries, avoiding the need to build project-specific DDA libraries. Third, a benchmark of controlled mixtures: recombine sorted fractions at known proportions spanning 0.5%–50% CSC content, with and without mimic perturbation, and with perturbation applied to only one fraction, yielding ground truth for both $\pi_k$ and $a_{ik}(t)$.

Two design choices deserve justification. Using mimics rather than, say, a kinase inhibitor, is not decoration: mimic-induced protein repression without mRNA change (Nie et al.) is the regime where RNA-only unmixing is blind by construction, which is exactly the regime the benchmark must contain. And the rare-CSC setting is not arbitrary: Palmer et al. showed population averages fail to resolve rare pluripotent subpopulations, and Lin et al. showed endocrine resistance emerges from selection of such cells (SOX2 up 74.8-fold in resistant sublines), so attribution of signaling to the rare fraction is the clinically meaningful question.

Proposed Experiments

Experiment 1 (identifiability and accuracy): on the known-proportion mixtures, evaluate fraction recovery (RMSE on $\pi_k$) and trajectory recovery (correlation between estimated and measured $a_{ik}(t)$ for held-out phosphopeptides not used in fitting). The key manipulation is a "constant composition, changing activity" series where $\pi_k$ is fixed but only the CSC fraction is perturbed — any method that infers activity from composition alone must fail here, and the size of that failure is the headline number. Experiment 2 (detection limit): dilute the CSC fraction across 5%, 1%, and 0.5% and measure the lowest fraction at which KinaseMix correctly attributes a mimic-induced pJNK-like change to the CSC compartment. The practical floor is set by the reference data, not the mixture: the sorted CSC fraction must yield ≥2,500 cells for SOP-Phos, so with a 0.1–0.9% CSC frequency this demands scaling up the culture, which we budget for. Experiment 3 (blind attribution): prepare mixtures where the perturbation target is hidden from the analyst; the method must call which fraction was hit and at what lag, scored as AUROC over subpopulation–timepoint pairs. As a downstream sanity check, resolved per-subclonal kinase states can be tested as drug-response predictors on TCGA bulk cohorts, where the relevant reported comparator is SubNetDL's median AUROC of 0.74 across ten cancer–drug pairs and 0.77 in external immunotherapy cohorts (SubNetDL) — though this is context, not the primary claim. For phosphorylation deconvolution itself, no reported result found on Lacuna for any prior method on this task; the benchmark we build is the first.

Baselines

The naive reference point is doing nothing: read the bulk phospho time course at face value and attribute the observed phosphorylation change to the whole sample uniformly. This is what nearly all miRNA-mimic studies in the file actually do, including Nie et al. and the ovarian cancer mimic study of miR-101-3p. It is the strongest competitor whenever the true per-type activities happen to move together, and Experiment 1's divergent-activity series is designed precisely to quantify when it breaks. The strongest existing method is a two-stage transfer: run POPPY on matched bulk RNA-seq to estimate $\pi_k(t)$ — POPPY reports up to a 714% increase in patient composition correlation and a 73% silhouette-score gain on its melanoma immunotherapy cohorts, making it the best-documented composition estimator available — then apply those fractions to the phospho mixture equation with static reference signatures, i.e., KinaseMix minus time-varying activity and minus perturbation-induced references. Both stages and both methods get the same input data, the same compute, and the same hyperparameter search budget (a fixed grid over regularization and smoothness settings), and KinaseMix is beaten only if it fails to exceed this baseline on trajectory recovery in the divergent-activity series. Two ablations isolate the proposal's ingredients: (a) KinaseMix with perturbation-time references replaced by resting-state signatures, which tests whether temporal diversity rather than model structure does the work; and (b) KinaseMix given true $\pi_k(t)$ (from the mixing protocol) instead of RNA-estimated fractions, which separates errors in activity estimation from errors in composition estimation.

Why not just…?

Why not just sort the cells and measure each fraction directly? For cell lines, this works and is exactly our ground-truth protocol; the honest answer is that the proposal's value lies where sorting cannot go. Clinical biopsies are too small and heterogeneous for reliable multi-way sorting, sorting itself stresses cells while phosphatases strip phosphate groups within seconds of lysis — a documented hazard in SOP-Phos — and rare fractions below roughly 2,500 cells fall under the microscale workflow's input floor. The decisive experiment is applying KinaseMix to an unsorted, unperturbed-then-perturbed bulk culture and checking its attribution against a small sorted validation aliquot; if sorted measurement and deconvolution agree in that setting, the unmixing earns its place as the scalable version of sorting.

Why not just deconvolve the RNA and be done? This is the real threat, because RNA unmixing is mature and POPPY-style composition estimates are demonstrably good. But composition is not activity. In the mimic regime the mRNA does not move while the phosphoproteome does (Nie et al.), so an RNA-derived $\pi_k$ combined with static references attributs activity changes to whoever is present, not whoever acted. There is also a documented failure mode on the composition side itself: the breast cancer benchmark of deconvolution methods showed that high tumor purity causes algorithms to misclassify normal cell types as malignant, so even the fractions can be wrong exactly where the question matters most. Experiment 1's constant-composition series settles this head-on: RNA-only transfer is blind there by construction, and if it nonetheless matches KinaseMix on the realistic mixed series, the simpler route wins and I would report that.

Why not just use phosphoflow cytometry for single-cell phosphorylation? Phosphoflow — measuring phospho-to-total ratios per cell by flow cytometry, as used to monitor STAT phosphorylation in the tofacitinib lupus trial — gives true single-cell resolution and I will use it as an orthogonal validation readout on a handful of anchor proteins. Its limit is multiplexing: a flow panel covers a few phospho-epitopes per run, against tens of thousands of phosphopeptides from SOP-Phos-DIA, and it depends on antibody availability for each site. The two are complements, and the proposal uses each for what it is uniquely good at: phosphoflow to check KinaseMix's per-protein attributions, and mass spectrometry for the system-wide measurement no panel can reach.

Decisive Checks

The proposal stands or falls on three numbers from the mixing benchmark: trajectory-recovery correlation for held-out phosphopeptides in the divergent-activity series (KinaseMix must beat RNA-transfer and the do-nothing baseline, with confidence intervals from replicate mixtures); the lowest CSC fraction at which blind attribution succeeds (target: 1%, against a naive-visibility floor near the several-percent level implied by the breast cancer deconvolution benchmark failure to resolve rare lineages at high purity); and the gap between resting-reference and perturbation-reference ablations, which must be substantial for the central claim about temporal, perturbation-induced signatures to survive. If the resting-reference ablation performs as well, the honest conclusion is that static phospho references suffice and the temporal machinery is unnecessary — a useful negative result, but a different paper.