Every measurement of how fast a protein is destroyed in a tissue today answers the wrong question at the wrong resolution. The standard tool — the cycloheximide (CHX) chase assay, in which a drug freezes the ribosomes so that translation stops and the remaining protein can only decay — requires grinding up millions of cells at each time point, so the single half-life it reports is an average across tumor cells, stromal cells, and immune cells that may differ wildly in their degradation machinery. The one method that does measure protein lifetimes in intact tissue with spatial detail, DELTA, requires a genetically engineered tag knocked into the target protein's gene in a mouse line, which confines it to one protein at a time and to animal models, never human biopsies. Meanwhile, spatial transcriptomics — sequencing RNA while keeping each measurement's physical coordinates in the tissue section — gives us exquisitely detailed maps of which cell types sit where, but only of RNA, and RNA degradation is empirically almost uncorrelated with protein degradation ($r = 0.147$ across 68 paired prostate samples in Comparative analysis of mRNA and protein degradation in prostate tissues indicates high stability of proteins).
This proposal combines the three: run the CHX chase on intact, freshly cut tissue sections rather than dissociated cells, quantify protein decay per spatial location by multiplexed immunofluorescence registered to a spatial transcriptomics section from the same block, and fit a kinetic deconvolution model that unmasks the bulk-averaged decay curve into cell-type-specific half-lives. The result would be the first maps of endogenous protein turnover across native tissue microenvironments — including human clinical specimens — and a direct test of whether spatial niches (hypoxic tumor cores, fibroblast barriers) systematically speed up or slow down protein destruction in the cells that inhabit them.
Core Claim
The central claim is that cell-type-specific protein degradation rates are identifiable from aggregated, spatially resolved chase measurements, provided the mixture composition is pinned down independently. Consider one spatial location $i$ (an imaging field or capture spot) followed over chase time $t$ after CHX treatment. If cell type $k$ destroys the target protein at rate $\lambda_k$, the measured signal is a sum of exponentials:
$$S_i(t) \;=\; \sum_{k} w_{ik}\, a_k\, e^{-\lambda_k t} \;+\; \epsilon_i(t),$$
where $w_{ik}$ is the proportion of cell type $k$ at location $i$, $a_k$ is its initial protein abundance, and $\epsilon$ is measurement noise. Fitted naively, this is badly ill-posed: a mixture of exponentials with free weights, amplitudes, and rates is close to unidentifiable from four or five time points. The key move is to not fit the weights at all. A co-registered spatial transcriptomics section, deconvolved with an established tool, supplies $w_{ik}$ from static RNA data; the kinetic model then only estimates the two small parameter sets $\{a_k\}$ and $\{\lambda_k\}$ per protein, shared across all locations. Spatial transcriptomics, in other words, is used as the prior that makes the chase assay's heterogeneity masking — the same averaging problem documented for bulk sequencing in Genome 2011 — solvable rather than merely observed.
This differs from the closest existing work in a specific way. The integrative FMR1–FTO study in gastric cancer (Integrative single-cell, spatial, and bulk transcriptomics reveal an FMR1–FTO axis linked to the immune-excluded phenotype in gastric cancer) is the only paper in the file that uses both CHX chase and spatial transcriptomics on the same biological question, and it leaves the gap wide open: its chase assays ran in dissociated 2D cell culture, its Visium analysis (GSE251950) was purely static RNA, and nothing connected the two. Our proposal measures the decay kinetics in situ, in the same spatial geometry the transcriptomics describes. DELTA (DELTA: a method for brain-wide measurement of synaptic protein turnover reveals localized plasticity during learning) proves spatially resolved turnover measurement is biologically informative — it found environmental enrichment shortens PSD-95 lifetime brain-wide by roughly 19% — but it needs HaloTag knock-in mice per protein, whereas we target endogenous, untagged proteins in human tissue. And the deconvolution literature, from stereoscope's probabilistic cell-type mapping (Single-cell and spatial transcriptomics enables probabilistic inference of cell type topography) through the count-process models of Count Bridges enable Modeling and Deconvolving Transcriptomic Data, has only ever unmixed static counts, never time-series decay curves; as the Count Bridges analysis itself makes clear, linear mixture models have no dynamic term to estimate per-type decay rates.
Research Program
The program has two tightly coupled components, an assay and an estimator, and neither works alone.
The assay is an ex vivo spatial chase. Fresh surgical specimens (or fresh-frozen slices from rapid autopsy programs) are cut into serial sections: one goes to a spatial transcriptomics platform for standard deconvolution, and adjacent sections go into short-term slice culture, receive CHX, and are fixed at chase intervals (0, 2, 4, 8 hours). Protein abundance per spatial location is quantified by cyclic immunofluorescence against a panel of targets — on the order of ten to twenty proteins per experiment, chosen where both antibodies and single-cell reference profiles exist — and registered to the transcriptomic coordinates by shared histology. Serial-section registration across a 55-micron Visium spot's worth of tissue depth introduces some coordinate blur, which we treat as part of the noise model and measure directly with co-stained landmarks. A genuine technical risk, stated plainly: CHX penetration and slice viability over an eight-hour chase in 300–500 micron or thinner sections are unproven at this scale, and multi-hour global translation arrest is itself a stressor that can secondarily alter degradation machinery. The first phase of the work is therefore a feasibility campaign — measuring penetration uniformity via puromycin-lacking translation readout, and restricting chase windows to the interval over which a no-drug control slice keeps a stable proteome and transcriptome. If the usable window is only four hours, the model accommodates it; short-lived proteins are exactly the ones where cell-type differences should be largest.
The estimator is the anchored kinetic deconvolution written above. The mixture weights come from deconvolving the co-registered transcriptomic section with a top-performing tool; the systematic benchmarking study of 18 tools across 50 datasets (A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics) establishes Cell2location and Tangram as strongest for large spot counts and CARD for smaller cohorts, and we follow its guidance per dataset size. Because the weights are fixed by this external prior, the per-protein fit reduces to a constrained nonlinear regression with a handful of parameters, regularized with a weak log-normal prior on half-lives (most mammalian proteins live hours to days). A second model layer, optional but scientifically central, allows $\lambda_k$ to vary smoothly with spatial niche labels — the hypoxic and immunosuppressive hubs that Starfysh (Starfysh integrates spatial transcriptomic and histologic data to reveal heterogeneous tumor–immune hubs) identifies from histology plus transcripts — so the output is not just per-cell-type half-lives but a test of whether the same cell type degrades the same protein faster or slower in different microenvironments. Morphological priors from adjacent H&E, of the kind reviewed in Deep Learning in Spatially Resolved Transcriptomics: A Comprehensive Technical View, provide an independent check on the registered weights where histology is visibly compartmentalized.
The scientific payoff is concrete. The FMR1–FTO axis is the pilot system: the gastric cancer study showed by in vitro CHX chase that the RNA-binding protein FMR1 shields the demethylase FTO from proteasomal destruction, and that FTO-high tumor nests sit behind fibroblast barriers, but it could not ask whether FTO is more stable inside those nests, where FMR1-expressing tumor cells actually reside. Our method asks exactly that question in the tissue. Targets like the NRF2 pathway, whose stability is set by the deubiquitinating enzyme USP35 (The deubiquitinating enzyme USP35 regulates the stability of NRF2 protein), extend this to stress-response programs where stromal and malignant cells are expected to differ sharply.
Baselines
The do-nothing reference point is the conventional bulk CHX chase: dissociated cells or homogenized tissue from the same specimens, Western blot or mass spectrometry at the same time points, one half-life per protein. This is the field's standard practice and the correct null model; it is also what the prostate comparative study effectively quantified the limits of, since its bulk macrodissected punches completely masked cellular heterogeneity. Our method beats it only if the recovered per-type half-lives genuinely differ from the bulk value in a way that replication confirms, and if a model comparison shows the single-exponential bulk fit is rejected on spot-level data. No reported result found on Lacuna for spot-level CHX chase in intact tissue, because no such measurement exists; the bulk baseline numbers will be generated in-house on the same specimens.
The strongest existing competitor for the estimation half of the problem is static spatial deconvolution applied per time point — treating each chase time point's protein image as a static map and unmixing it with a tool such as CARD or Cell2location, then reading half-lives off the per-type trajectories post hoc. This is the most plausible way to get our answer without building anything new, so we will run it on identical data with an equal tuning budget (same reference atlas, same spot counts, same compute). The performance reference comes from the Starfysh paper, which reports deconvolution accuracy of RMSE ≈ 0.09 on its primary breast cancer spatial cohort (Starfysh integrates spatial transcriptomic and histologic data to reveal heterogeneous tumor–immune hubs); our anchored method should recover cell-type weights at least that well while additionally producing half-life estimates, which the per-time-point static approach should struggle with because early time points carry almost no kinetic information and per-time-point unmixing errors propagate multiplicatively into rate estimates. A second strong competitor is Count Bridges (Count Bridges enable Modeling and Deconvolving Transcriptomic Data), whose discrete count-process machinery could in principle be extended to time series; we will test its static model on our synthetic time-series benchmark to quantify what the missing dynamic term costs. No reported number for protein-turnover recovery exists for any of these, since none has been applied to chase data.
The essential ablation isolates the anchoring ingredient: the same kinetic model fitted with freely estimated weights versus ST-anchored weights versus randomly permuted weights. If free-weight fitting matches anchored fitting on our synthetic benchmarks, the central design claim — that spatial transcriptomics is what makes the kinetic unmixing identifiable — is wrong, and the anchoring should be dropped. Conversely, if permuted weights destroy half-life recovery while anchored weights preserve it, the ingredient is vindicated.
Why not just…?
Why not just infer protein stability from spatial transcript abundance or degradation-machinery gene expression? Because the empirical decoupling is stark: across 68 paired prostate samples, mRNA degradation and protein degradation correlated at $r = 0.147$ (Comparative analysis of mRNA and protein degradation in prostate tissues), and the mechanistic case studies in the file — FMR1 shielding FTO, USP35 stabilizing NRF2, USP3 stabilizing EPHA2 (USP3 promotes osteosarcoma progression via deubiquitinating EPHA2 and activating the PI3K/AKT signaling pathway) — all act through direct physical binding that changes half-life without changing target transcript levels. A ligase-expression proxy would be blind to exactly the regulatory events this proposal is built to detect. That said, this alternative is a real threat for coarse questions, so our experiments include it head-on: on the same specimens, we will test whether an E3/DUB expression score predicts the measured per-type half-lives, and report the correlation either way.
Why not just use DELTA? DELTA already delivers spatially resolved, single-synapse protein lifetime measurement, and its effect sizes — a 19% brain-wide PSD-95 lifetime shortening, 40–50% GluA2 turnover acceleration restricted to hippocampal CA1 — show the biological resolution such measurements achieve. But it requires a HaloTag knock-in line per protein, which means one protein per experiment, mouse models only, and no path to human clinical specimens like the breast and gastric cancer cohorts already profiled by spatial transcriptomics. Our approach trades DELTA's temporal and subcellular precision for endogenous targets, human tissue, and multiplexing, and we will be explicit that the two are complementary rather than competing: on proteins where both are possible, DELTA-style measurements in mouse slices would serve as an orthogonal validation of our recovered half-lives.
Why not just FACS-sort the cell types and run standard CHX chases on each? This would give clean per-type half-lives with today's tools, and it is the alternative most likely to be proposed in review. It falls short for the specific reason that motivates the spatial program: dissociation destroys the microenvironment, and degradation rates are set by precisely the signaling contacts and local conditions — fibroblast barriers, hypoxic cores — that dissociation removes. Sorted cells measure half-lives of cells that no longer inhabit their niche. Rather than argue, we will run the experiment: for one target, compare half-lives from FACS-sorted populations against in situ recovered values for the same nominal cell type. If they agree everywhere, spatial niche modulation is a smaller effect than hypothesized and the in situ machinery adds less than claimed; if they diverge specifically in niche-defined compartments, the divergence itself is the result.
Proposed Experiments and Decisive Checks
The first experiment is a fully controlled synthetic ground truth: mix cultured cell lines with independently measured in vitro half-lives (classic bulk CHX Western blot) in known proportions, pellet and section them as pseudo-tissue, run the spatial chase, and test whether the anchored model recovers the input half-lives and proportions. Recovery error on $\lambda_k$ below roughly 30% across mixing ratios spanning 5–90% per type, with the weight recovery at or better than the Starfysh reference of RMSE ≈ 0.09, is the pre-registered success criterion; the free-weight and permuted-weight ablations run on the same data.
The second experiment applies the full pipeline to human breast cancer specimens modeled on the HER2-positive cohort of Spatial deconvolution of HER2-positive breast cancer delineates tumor-associated cell type interactions (36 sections, 8 patients), with a target panel including HER2, FTO, and pathway readouts where the deconvolution reference (stereoscope resolved 8 major cell types there) is solid. The decisive scientific test is whether per-type half-lives differ by more than replication noise — for instance, whether FTO is measurably longer-lived in tumor-nest cells than in stromal cells, as the in vitro FMR1 mechanism predicts but no one has shown in tissue. The third experiment is the FACS-versus-in-situ comparison above, and the fourth is the niche-modulation test: fitting half-lives as functions of Starfysh-style hub labels and asking whether the same cell type's degradation rate shifts across niches by an effect size comparable to the 19–50% shifts DELTA demonstrated for plasticity, which would establish that microenvironment, not just cell identity, sets protein stability.
If the feasibility campaign shows that intact slices cannot sustain a usable chase window, or if synthetic recovery fails at achievable time-point counts, the fallback is a reduced claim: two time points plus the anchored weights still yield a per-type relative-stability ranking, which the transcript-proxy baseline cannot provide given the $r = 0.147$ decoupling. The negative results would themselves be worth reporting, since no one has yet measured whether translation arrest is even chemically coherent in intact human tissue sections.
References
- Comparative analysis of mRNA and protein degradation in prostate tissues indicates high stability of proteins
- As we come to the end of 2011, several members of the Genome Biology Editorial Board give their views on the state of play in genomics
- Integrative single-cell, spatial, and bulk transcriptomics reveal an FMR1–FTO axis linked to the immune-excluded phenotype in gastric cancer
- DELTA: a method for brain-wide measurement of synaptic protein turnover reveals localized plasticity during learning
- Single-cell and spatial transcriptomics enables probabilistic inference of cell type topography
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics
- Starfysh integrates spatial transcriptomic and histologic data to reveal heterogeneous tumor–immune hubs
- Deep Learning in Spatially Resolved Transcriptomics: A Comprehensive Technical View
- The deubiquitinating enzyme USP35 regulates the stability of NRF2 protein
- USP3 promotes osteosarcoma progression via deubiquitinating EPHA2 and activating the PI3K/AKT signaling pathway
- Spatial deconvolution of HER2-positive breast cancer delineates tumor-associated cell type interactions