Who Loses the Rhythm When Nights Warm? Cell-Type-Resolved Diurnal Deconvolution of Leaves Under Asymmetric Warming

Over much of the land surface, nights are warming faster than days. Chamber experiments on winter wheat show that this matters for yield: shifting the same warming toward the night cost more grain and more water-use efficiency than warming day and night equally. The standard explanation is that warm nights speed up leaf respiration and burn carbohydrates. That explanation comes entirely from whole-leaf or whole-pot measurements, which average over very different cell types: mesophyll that photosynthesizes, guard cells that open and close the leaf's pores, and phloem companion cells that load sugar for export. The same averaging problem has misled biologists before. In mouse skin, bulk measurements made immune-cell daily clocks look weak, and single-cell data showed the cells were strongly rhythmic but out of step with each other.

We propose to measure where in the leaf the daily transcriptional rhythm changes when night warming narrows the day–night temperature range. The core tool is a deconvolution method that splits a bulk RNA sample (ground-up tissue) into cell-type contributions. Unlike existing methods, it fits the whole time series at once and requires each cell type's contribution to follow a smooth 24-hour curve with its own phase and amplitude. A small single-nucleus sequencing reference anchors the method, and held-out single-nucleus samples test it. Dense, inexpensive bulk time courses then carry it across warming treatments. The aim is to separate three explanations for a weakened bulk rhythm under night warming that bulk analysis cannot tell apart: the mix of cell types shifted, cell types drifted out of phase with one another, or rhythms within a cell type weakened or scattered.

Core Claim

A weaker daily rhythm in a bulk leaf transcriptome under asymmetric warming is ambiguous. We claim it can be broken into three parts. The first is a change in cell-type mixture. The second is a phase shift between cell types, which bulk time courses can identify once the method shares rhythmic structure across time points. The third is a loss of amplitude or synchrony within a cell type, which bulk data cannot identify and which therefore needs a small, targeted single-nucleus measurement.

The deliverable is a method and an experiment. The method is a time-structured deconvolution that recovers phases specific to each cell type from bulk diurnal series. The experiment is a wheat chamber study that reproduces the symmetric and asymmetric warming contrast of Qiao et al. (2014) and reports which part of the leaf carries the night-warming signature.

Why this gap, and what prior work does and does not do

The climate side motivates the question. Qiao et al. (2014) grew winter wheat in enclosed-top chambers. Asymmetric warming (+2.5 °C day, +3.5 °C night) reduced grain yield by 15.6% and water-use efficiency by 21.8%. Symmetric warming (+3 °C day and night) reduced them by 10.3% and 15.6%. Cheng et al. (2024) report that warmer nights can raise next-day photosynthetic capacity through nighttime respiratory mobilization, at a carbon cost. Lobell (2007) links changes in day–night temperature range to national cereal yields, but notes that this range moves together with radiation and cloud cover in field data. That collinearity is why we use chambers, where the day and night temperatures can be set independently.

All three studies explain the effect through bulk physiology. None asks which cells carry it.

The single-cell side supplies both the warning and the tools. tauFisher predicts the time of day from one transcriptome sample and showed that an apparently weak bulk rhythm can come from cells that are out of phase. It predicts a sample's time. It does not recover phases for each cell type from bulk mixtures, and it has never been run under temperature manipulation or in plants. We reuse it to check the internal time of each bulk sample.

Count Bridges model integer RNA counts with a birth–death process, which is a natural fit for count data. They split bulk samples into latent single-cell counts whose total matches the observed bulk. The method treats every sample separately and has no time structure.

POPPY aligns bulk tumour profiles to single-cell representations using the Cell Ontology, a mammalian hierarchy of cell types. That dependence makes it a poor starting point for plants, so we do not build on it.

The breast-cancer deconvolution benchmark (Performance of tumour microenvironment deconvolution methods) constrains our design. It found that when one cell lineage dominates a sample, related minority types are misassigned to it. A leaf, dominated by mesophyll, is exactly this case. The same benchmark found BayesPrism among the most robust tools under this imbalance, which makes it our main competitor.

Research Program

The model treats each gene $g$ in a bulk sample taken at time $t$ under treatment $c$ as a mixture over cell types $k$:

$$ b_{gc}(t) = \sum_k p_{kc}\, s_{gk}\,\big(1 + a_{gkc}\cos(\omega t - \phi_{gkc})\big) + \varepsilon, $$

where the terms are defined as follows. $p_{kc}$ is the share of cell type $k$ under treatment $c$. $s_{gk}$ is the baseline expression signature taken from the single-nucleus reference. $a_{gkc}$ and $\phi_{gkc}$ are the amplitude and phase of gene $g$ in cell type $k$. $\omega = 2\pi/24\,\mathrm{h}^{-1}$ is the daily frequency. $\varepsilon$ is count noise; a negative binomial likelihood is used in practice. Phases are pooled hierarchically across genes within a cell type, so a rare type such as guard cells borrows strength from the whole series instead of being estimated one time point at a time. That pooling is the step meant to escape the dominant-lineage misassignment problem.

The model has a deliberate limit. Suppose cells within a type have phases scattered with spread $\sigma$, approximated as a normal distribution wrapped around the 24-hour cycle. Their average is a cosine damped by a factor of $e^{-\sigma^2/2}$, the expected value of $\cos$ under such phase noise. Bulk data therefore see only the product $a\,e^{-\sigma^2/2}$, so loss of synchrony within a type cannot be told apart from loss of amplitude. We resolve that one term directly. Nuclei are extracted from flash-frozen flag leaves at two time points per treatment, and within-type scatter is measured in the single-nucleus data, with tauFisher-style time scores computed per nucleus.

The experiment reproduces the CON, ETs, and ETa treatments of Qiao et al. (2014) and adds a night-only warming arm to strengthen the contrast. Bulk flag-leaf RNA is sampled every 3 hours over 48 hours with four biological replicates. Single-nucleus sequencing covers end-of-day and end-of-night under all treatments. The searched index lists no published plant leaf single-cell atlas, so we build that reference ourselves. Nighttime leaf respiration is measured alongside, so that cell-type phase changes can be related to carbon loss.

Baselines

The naive reference is ordinary rhythm detection on bulk data, using harmonic regression of the JTK/Lomb–Scargle type, with no deconvolution. It reports only the overall change in bulk amplitude and phase. The proposal must show that its decomposition explains this bulk change, not merely restates it.

The strongest existing competitor is a two-stage pipeline: run BayesPrism on each time point separately, then fit cosine curves to its cell-type expression estimates. We choose it because the breast-cancer benchmark ( Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures) found it among the most stable tools when one lineage dominates. That paper's summary gives qualitative rankings only. No reported result was found on Lacuna for any deconvolution method on plant tissue or for recovering phases specific to each cell type.

A second competitor is Count Bridges ( Count Bridges enable Modeling and Deconvolving Transcriptomic Data) run per sample, followed by the same cosine fitting. It is restricted to about 2,000 rhythmic genes because of its compute cost. Its reported advantages over CIBERSORTx and MuSiC have no numerical values in the file.

All methods use the same reference signatures, the same gene set, and the same hyperparameter search budget, tuned on simulated data only. The key ablation is our own model with the shared time structure removed, fitting each time point independently. This isolates whether pooling across time points is what helps.

Beating the competitors means lower error in recovered phase, measured as hours on the circle, for guard cells and companion cells in held-out single-nucleus samples, with no loss on mesophyll. For time-of-day accuracy, we read tauFisher's reported bulk result as a rough target: accuracy 0.875 with an RMSE of 2.669 h (tauFisher). That figure is for predicting a sample's time, not a cell type's phase, so it is only an approximate reference. Cell-type phase recovery that is coarser than about 2.7 h is not useful for this question.

Why not just…?

Why not just run single-nucleus sequencing on every sample? This is a real threat, and we take it seriously. The research file argues that plant single-cell work suffers from artifacts of enzymatic dissociation. That argument applies to protoplast preparation, not to nuclei isolated from flash-frozen tissue, so it does not rule out single-nucleus sequencing. The real argument is scale. A design with four treatments, 16 time points, and four replicates is 256 libraries, and adding field seasons multiplies that. Nuclear RNA also lags cytoplasmic RNA, which matters for transcripts that change quickly over the day. We therefore do single-nucleus sequencing where it is irreplaceable, on within-type scatter and as validation. If held-out validation shows bulk-derived phases are no better than noise, the honest conclusion is that cheaper single-nucleus sequencing is the right route, and we would report that.

Why not just run BayesPrism or CIBERSORTx-style deconvolution at each time point? This is the two-stage baseline above. Its weakness is that each time point independently faces mesophyll dominance, which the breast-cancer benchmark shows misassigns minority lineages. Our ablation tests directly whether shared time structure fixes this. If the two-stage pipeline matches the full model, the contribution reduces to the experiment and the three-part decomposition, and we would say so.

Why not just measure respiration and gas exchange? Gas exchange gives the carbon flux we ultimately care about, and we measure it. But it reports one number per leaf and cannot say which cells, or which phase changes, produce it. That is the limitation that leaves the explanations of Qiao et al. (2014) and Cheng et al. (2024) inferential.

Proposed Experiments and Decisive Checks

The first check is simulation. Bulk series are built by mixing the single-nucleus reference with known phase shifts and known within-type scatter, across mesophyll shares from 60% to 95%, following the purity-gradient design of Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures. The method must recover phase shifts between cell types. It must also correctly report that a planted within-type scatter shows up only as reduced effective amplitude, and is not misattributed to a phase shift.

The second check is held-out biology. Phases recovered from bulk data at end-of-day and end-of-night are compared with single-nucleus pseudobulk from separate plants.

The third check is the scientific test itself. Under ETa relative to ETs, we ask whether the drop in bulk amplitude is carried by one cell type, whether it reflects a phase shift or a within-type change, and whether that component tracks the measured rise in nighttime respiration across replicates. One possible outcome is that the effect concentrates in companion cells or guard cells instead of mesophyll. Another is that the leaf responds uniformly. Either outcome replaces the current guess, based only on whole-leaf averages, with a measurement.