Separating cell loss from gene switching in multi-organ stress time courses

When a plant is heat-shocked or a tissue is starved, bulk RNA sequencing (grinding a whole organ and measuring the average expression of every gene) often shows that whole programs, such as chloroplast genes, go down. Two very different biological stories produce that same signal. In the first, the cells that express those genes die or shrink in number. In the second, surviving cells actively switch the genes off. Current tools resolve this by assumption. Deconvolution methods, which estimate cell-type proportions from a bulk sample using a single-cell reference, treat every change as a change in proportions. Time-course co-expression analysis treats every change as gene regulation inside a fixed population.

We propose a way to tell the two stories apart from cheap bulk data, using an experimental design rather than a better reference. The design profiles several organs of the same organism over time. Our working assumption is that a given cell type, for example vascular parenchyma or epidermis, runs a largely shared intrinsic stress program wherever it sits, while its abundance differs from organ to organ. If so, the same unknown per-cell-type expression trajectories appear in several bulk mixtures with different weights. That is the extra leverage that makes "how many cells" and "what each cell does" separately estimable. Testing that assumption is itself a scientific result. It asks how much of what the field calls tissue-specific stress response is really differences in cell composition.

Core Claim

A joint model fitted across organs and time points can recover two things that single-organ or single-time-point analyses provably cannot. The first is the time-varying cell-type proportions within each organ. The second is the time-varying expression of each cell type. The model uses three ingredients: smooth trajectories in time, a count-based likelihood, and a shared per-cell-type stress program with a penalized organ-specific deviation. The size of that fitted deviation answers the motivating biological question directly. Two outcomes are possible. If deviations are small and validated against sparse single-nucleus measurements, organ-specific bulk stress signatures are largely compositional. If deviations are large, organs genuinely reprogram the same cell type differently.

Where the field stands

The tissue-specific stress literature motivates the problem but works at the resolution of whole organs. In Populus euphratica, Tissue-Specific Transcriptome Analysis Reveals Multiple Responses to Salt Stress in Populus euphratica Seedlings profiled leaf, phloem, xylem and root at three salt doses. It reports distinct programs per organ: ion transporters in leaves, redox enzymes in roots. Because each organ is a different cell mixture, those differences could be compositional. Physio-biochemical and dynamic transcriptome comparison of heat tolerance in wild and cultivated tomato accessions followed whole leaves over eight time points of 42°C heat and recovery. It found chloroplast processes collapsing in the sensitive M82 line, with membrane leakage 1.61 times higher than in wild LA1589 after 24 hours. Whole-leaf profiling leaves open whether that collapse is mesophyll death or repression. Both studies supply data and a concrete question, but no method.

Deconvolution supplies the method family, and its known failures constrain the design. Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures benchmarked nine tools on 38,000 pseudo-bulk mixtures. It found that related lineages are absorbed into the dominant one at high purity, with BayesPrism and DWLS performing best overall. That warning matters for leaves, which are dominated by mesophyll. Two other methods relax the fixed-reference assumption, but only within a single sample. Deep autoencoder for interpretable tissue-adaptive deconvolution and cell-type-specific gene analysis (TAPE) adapts its signature matrix to a target tissue in an unsupervised second stage. Cross-modal transfer learning for mapping bulk transcriptomes at cellular level (POPPY) aligns bulk samples to single-cell foundation-model embeddings. Neither couples samples across time or organs. Neither has been tested under the large-scale rewiring of a stress shock; the tomato study counts over 10,000 differentially expressed genes per accession.

Count Bridges enable Modeling and Deconvolving Transcriptomic Data argues that methods built for continuous values handle integer sequencing counts poorly. We adopt that lesson through a count likelihood, but not its generative bridge machinery. Finally, trajectory models such as Learning Explicit Single-Cell Dynamics Using ODE Representations (Cell-MNN) and STODE require single-cell or spatial data at every time point. We borrow only their premise that cell states move smoothly through time, and apply it to bulk mixtures.

Research Program

The model treats each bulk count $y_{otg}$ (organ $o$, time $t$, gene $g$) as negative-binomial. Its mean is the sequencing depth times $\sum_k w_{otk}\,\mu_{otkg}$. Here $w_{otk}$ is the proportion of cell type $k$, and $\mu_{otkg}$ is that cell type's expression. Cell-type expression is written as

$$\log \mu_{otkg} = \log s_{kg} + b_{okg} + \sum_{r=1}^{R} c_{kr}(t)\,U_{rg} + \epsilon_{otkg}.$$

In this expression, $s_{kg}$ is the unstressed reference profile from a baseline single-cell atlas. $b_{okg}$ is a static organ offset fitted at time zero. $U_{rg}$ are $R$ shared stress programs, which are gene loadings learned from data. $c_{kr}(t)$ are smooth, cell-type-specific activation curves shared across organs. $\epsilon_{otkg}$ is an organ-specific intrinsic deviation with a sparsity penalty $\lambda$. Proportions $w_{ot\cdot}$ follow smooth trajectories on the simplex, separately for each organ.

The identifiability argument is simple counting. With $O$ organs that share the curves $c_{k\cdot}(t)$ but mix cell types differently, each time point gives $O$ independent mixtures of the same $K \times R$ unknown activations. Composition and intrinsic change separate when the organ proportion matrix has rank at least $K$, or when the low-rank program assumption makes $R$ small. We will state this condition formally and check it empirically on simulated designs.

The penalty $\lambda$ is the honest weak point. A large $\lambda$ forces the "shared program" answer and would bias the biological conclusion. We therefore do not tune $\lambda$ on bulk fit. Instead we choose it by error on held-out single-nucleus anchor time points, and we report the full curve of conclusions as $\lambda$ varies. Fitting is maximum a posteriori with spline-parameterized curves on a GPU. We also run a cheap Laplace or bootstrap step to obtain uncertainty on the composition-versus-intrinsic split.

Baselines

The first naive reference has two poles, and both encode "do nothing." The all-intrinsic pole holds proportions fixed at baseline and attributes every bulk change to gene regulation. This mirrors the co-expression analysis of the tomato study. The all-compositional pole is CIBERSORTx-style static deconvolution with the unstressed reference at each time point, which attributes every change to proportions. Any useful method must beat both on the attribution metric defined below.

The strongest existing competitor is BayesPrism run independently per time point. It was a top performer in the breast cancer deconvolution benchmark. Unlike linear tools, it also returns a posterior estimate of sample-specific cell-type expression, so it can in principle capture intrinsic change on its own. That benchmark reports BayesPrism's ranking, not a numeric error, and we found no reported result on Lacuna for BayesPrism, or any deconvolution method, on longitudinal stress series. Our own runs will therefore define its number.

TAPE with its adaptive stage, run per time point, is the second strong competitor because it explicitly updates the signature matrix. TAPE is reported to beat CIBERSORTx, Scaden and MuSiC on mean absolute error, but the file gives no figure. All methods receive the same reference atlas, the same bulk counts and the same marker gene set. Each gets an equal hyperparameter search budget, tuned on the same simulated validation split. We count a method as beaten if our model has lower error on per-cell-type log-fold-change recovery and on attribution, with non-overlapping bootstrap intervals across simulation seeds.

Three ablations isolate the key ingredient. The first removes cross-organ sharing by fitting each organ separately with temporal smoothing kept. If this ablation matches the full model, the multi-organ idea is not doing the work. The second removes temporal smoothness. The third replaces the count likelihood with Gaussian loss on log-normalized data.

Why not just…?

Why not just run BayesPrism at each time point and smooth the outputs afterwards? This is the real threat. BayesPrism already estimates per-sample cell-type expression, and post hoc smoothing removes much of the time-point jitter. Its estimate of intrinsic change in a single sample is still weakly identified: one mixture cannot separate a cell type's abundance from its expression level without a prior. Under the purity effect documented in the breast cancer benchmark, that prior favors absorbing minority signal into the dominant lineage. The simulation with known ground truth settles this. If smoothed BayesPrism matches our attribution error in the regime where proportions and expression change simultaneously, we will report that and abandon the method claim. The biological question about compositional versus intrinsic organ signatures would remain.

Why not just use TAPE's adaptive stage? TAPE adapts one signature matrix to one tissue context. Applied per time point, it has no coupling that stops it from explaining a proportion drop as a signature change, or the reverse. The single-organ, smoothed ablation of our model is close to "TAPE plus time." Comparing it to the full model measures exactly what cross-organ sharing adds.

Why not just collect single-nucleus data at every time point and organ? That would answer the question directly, and we use it as ground truth where we can afford it. Eight time points times three organs times two accessions, with biological replicates, is far beyond the budget of a typical stress lab. The open problems in the file note that dense temporal single-cell replication is the bottleneck. Our design spends single-nucleus sequencing on two anchor time points and bulk sequencing on the rest. We include a baseline that simply interpolates between anchors, so the value of the bulk series is measured rather than assumed.

Proposed Experiments

The first experiment uses controlled simulation with known answers. We build pseudo-bulk time series from the longitudinal single-cell datasets in the Cell-MNN suite: embryoid body, about 16,000 cells over 25 days; CITE-seq, about 31,000 cells over 7 days; and Multiome, about 33,000 cells over 7 days. In these datasets both composition and per-cell-type expression genuinely change over time. We create three to four synthetic "organs" by resampling cells with organ-specific, time-varying proportions. These include a high-purity organ dominated by one type, imitating the purity trap. In one condition intrinsic change is shared across organs by construction. In graded violation conditions, we inject organ-specific expression shifts into a chosen fraction of cell types by swapping in cells from shifted time points.

Metrics are proportion mean absolute error and correlation between estimated and true per-cell-type log-fold changes. The central metric is attribution error: the absolute difference between the estimated and true fraction of each gene's bulk change that is due to composition. The decisive check is how attribution error degrades as sharing is violated. A method that only works when the assumption holds exactly is not useful.

The second experiment is a real multi-tissue analysis. We fit the model to the Populus euphratica data (SRP116293), using salt dose as the ordered axis in place of time; with three levels, this is a weak test of smoothness. The question is how many of the 2,849 leaf-specific and 679 root-specific differentially expressed genes the model reassigns to composition. A baseline poplar single-cell atlas is needed for the reference; if none can be obtained, this experiment becomes exploratory.

The third experiment is new data and the decisive biological test. We repeat the tomato heat-shock design of Shi et al. in M82 and LA1589 with the same eight time points. We sample leaf, stem and root by bulk RNA sequencing, and add single-nucleus profiling of all three organs at 0 hours and at two held-out time points. The model is fitted without the held-out nuclei. It is then scored on predicting their cell-type proportions and per-cell-type expression, against BayesPrism, TAPE and anchor interpolation.

An independent physiological check comes from the leakage and chlorophyll measurements. If the model attributes M82's chloroplast decline mainly to mesophyll loss, the inferred mesophyll fraction should track membrane leakage across time, and it should not do so in LA1589. If it attributes the decline to repression, it should not. Either outcome is informative. What would sink the proposal is the model failing to beat smoothed per-time-point BayesPrism on the held-out nuclei.