Intact-Tissue Bulk RNA-seq as the Artifact-Free Anchor for Cell-Type Drought Maps in Crops

To see which cells in a crop leaf or root actually change their metabolism under drought, you need single-cell RNA sequencing. Plant cells cannot be read one at a time until their walls are digested away with enzymes, a step called protoplasting. That digestion is itself a sudden osmotic and oxidative shock. It switches on many of the same genes drought does: stress hormone signaling, heat-shock proteins, and the machinery that copes with reactive oxygen species. So every plant single-cell atlas of drought carries a technical artifact that resembles the biology being studied, and it is present in every cell. The usual remedies are to delete a list of "protoplasting genes" or to discard the most stressed-looking cells. Both risk throwing away the drought response itself. Ordinary bulk RNA-seq of intact tissue has the opposite problem. It is free of the digestion artifact, but it averages over epidermis, mesophyll, and vascular cells, so it cannot say where a metabolic shift happens.

This note proposes to use each measurement to repair the other. Bulk RNA-seq of intact tissue is the only view of the plant that has never been digested. It can therefore serve as the "zero-digestion" reference against which a cell-type-specific artifact is estimated and subtracted from single-cell profiles of the same plants. We will build a count-level model that fits three things jointly: single-cell data collected at several digestion durations, paired bulk data from intact and digested tissue, and the drought time-course. The model returns cell-type-resolved drought effects with the artifact removed. We will apply it to a concrete open question in upland cotton. Is drought-induced proline synthesis and antioxidant rewiring a uniform low-level shift across the leaf, or a strong shift confined to specific cell layers?

Core Claim

The digestion artifact cannot be identified from plant single-cell data alone, because there are no undigested cells to compare against. It becomes identifiable once intact-tissue bulk is treated as the endpoint of a digestion-duration curve. The claim is falsifiable in two ways. If the artifact is removed correctly, the model should reconstruct held-out intact bulk samples better than a model fit to digested references. It should also recover planted cell-type-specific drought effects in semi-synthetic data that a gene-list deletion approach misses.

A second, biological claim follows. Bulk co-expression analysis in cotton links a proline-correlated module to drought tolerance and names WRKY22 as a hub (Mining Key Drought-Resistant Genes of Upland Cotton Based on RNA-Seq and WGCNA Analysis). We predict that, once the artifact is removed, much of that module resolves to a minority of cell types, and that cell-type-resolved expression predicts measured proline and lipid-peroxidation (malondialdehyde, MDA) levels better than bulk expression does.

Where This Sits Relative to Prior Work

Three lines of work supply parts of the method and also show what is missing.

The first is bulk deconvolution, which means estimating which cell types make up a bulk sample and what each contributes. SCDC ensembles several single-cell references, weighted by how well each reconstructs the bulk, to absorb protocol differences (SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References). POPPY aligns bulk and single-cell encoders through a cell ontology (Cross-modal transfer learning for mapping bulk transcriptomes at cellular level). Count Bridges unmixes bulk into single-cell profiles while keeping integer count geometry (Count Bridges enable Modeling and Deconvolving Transcriptomic Data). All three treat the single-cell reference as ground truth for each cell type's expression and treat bulk as the thing to be explained. We reverse the direction of trust for the artifact component: bulk becomes the reference that corrects the single-cell profiles. None of the three was evaluated on plants, and none faces an artifact shared by every reference cell. In our setting, SCDC's reference weighting cannot help, because every available reference is digested.

The second is artifact disentanglement. CRADLE-VAE splits each cell's latent state into basal biology, perturbation effect, and technical artifact. It learns the artifact by contrasting cells that pass and fail quality control, and it raises the QC pass rate of generated cells from 70–80% to 93–98% on human Perturb-seq data (Cradle-VAE: Enhancing Single-Cell Gene Perturbation Modeling with Counterfactual Reasoning-based Artifact Disentanglement). This paper motivates keeping stressed cells rather than discarding them. It also constrains our design: its counterfactual depends on having some artifact-free cells, and protoplasted plant data has none. That missing contrast is exactly what intact bulk and the digestion-duration series provide.

The third is plant single-cell foundation models. scPlantLLM, a transformer pretrained on plant single-cell data, reports zero-shot annotation and cross-batch integration (Foundation Model: A New Era for Plant Single-cell Genomics). It is useful here for labeling cell types. However, the general batch-integration methods it resembles, including BBKNN (BBKNN: fast batch alignment of single cell transcriptomes), BERMUDA (BERMUDA: a novel deep transfer learning method for single-cell RNA sequencing batch correction reveals hidden high-resolution cellular subtypes), scAdapt (scAdapt: Virtual adversarial domain adaptation network for single cell RNA-seq data classification across platforms and species), and latent adapters such as scBatchProx (Federated-inspired Single-cell Batch Integration in Latent Space), all remove differences between batches. A shift present in every batch is invisible to them. BERMUDA and scAdapt also align discrete cluster centroids, which risks collapsing the continuous drought progression we want to keep.

Research Program

The model works at the level of read counts. For gene $g$, cell type $k$, and condition $c$ (cultivar, watering, and drought day), let $m_{gkc}$ be the true log expression in undigested tissue. A single cell of type $k$ digested for duration $\tau$ has expected count

$$ \mathbb{E}[y_{g}] = s \cdot \exp\big(m_{gkc} + f_{gk}(\tau, c)\big), \qquad f_{gk}(0, c) = 0, $$

where $s$ is the cell's library size and $f_{gk}$ is the digestion artifact. The artifact is constrained to be monotone and saturating in $\tau$, and it may interact with drought stage, because already-stressed cells may respond to digestion differently. A bulk sample $j$ of intact tissue has expected count

$$ \mathbb{E}[b_{gj}] = L_j \sum_k p_{jk} \exp(m_{gkc}), $$

where $L_j$ is depth and $p_{jk}$ are cell-type proportions. Both likelihoods are negative binomial, and $m$ is shared between them.

Three sources of variation make the artifact identifiable. The digestion-duration series gives the shape of $f$ along $\tau$. Intact bulk pins the curve's value at $\tau = 0$, which cannot be measured at the single-cell level. Paired bulk of the same tissue before and after digestion gives $\sum_k p_{jk}$ times the artifact effect at the population level, and proportions that vary across drought days separate this into per-cell-type contributions. The artifact is given low rank across cell types: one shared response plus a cell-type-specific scaling and a sparse residual. This keeps the model estimable with modest sample counts.

The data plan builds on the public cotton drought time-course (NCBI accession PRJNA1240467). That study profiled cultivars 'Jin' (tolerant) and 'TM-1' (sensitive) at 0, 7, 10, 15, and 25 days, with proline, MDA, superoxide dismutase activity, and chlorophyll measured alongside (Mining Key Drought-Resistant Genes of Upland Cotton Based on RNA-Seq and WGCNA Analysis). We will reproduce the protocol and add three components at days 0, 10, and 15 for both cultivars. The first is leaf single-cell RNA-seq at three digestion durations. The second is paired intact and digested bulk from the same leaves. The third is manually separated epidermal peels and the remaining mesophyll-enriched tissue, profiled by bulk RNA-seq and qRT-PCR, as a coarse physical ground truth for where expression sits. Cell types will be labeled with scPlantLLM and marker genes. Cotton's allotetraploid genome will be handled by keeping homeologs as separate genes, so homeolog-specific artifact responses are not averaged away.

The metabolic analysis then asks a specific question for the proline-correlated module and for antioxidant and lipid-remodeling genes. For each, we will decompose the bulk drought change into three parts: a within-cell-type expression change, a change in cell-type proportions, and a part attributable to the artifact if a digested reference had been used naively. We then test whether cell-type-resolved scores predict per-plant proline and MDA better than bulk module scores.

Baselines

The naive reference point is bulk co-expression analysis as published for this dataset. It reports module–trait correlations of r = 0.81 for proline and r = 0.86 for MDA, and a 3.2-fold higher WRKY22 expression in 'Jin' at day 15 (Mining Key Drought-Resistant Genes of Upland Cotton Based on RNA-Seq and WGCNA Analysis). It sets the bar for trait prediction. We will refit it on our replicated data under the same cross-validation folds, holding out entire plants and the day-25 time point, and compare out-of-sample prediction of proline and MDA. Beating it means a significantly higher held-out correlation, judged by paired bootstrap over plants.

The practical community baseline is single-cell analysis after deleting a protoplasting-responsive gene set. We will define that set from our own paired intact-versus-digested bulk, so the baseline gets the same data we do, at a comparable false-discovery threshold. Drought effects are then estimated by standard per-cell-type pseudobulk differential expression.

The strongest existing deconvolution method on the file's benchmarks is SCDC. It reports a Pearson correlation of 0.991 with ground truth on a three-cell-line mixture (SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References). Count Bridges reports outperforming CIBERSORTx and MuSiC, but no reported number was found on Lacuna for its bulk deconvolution, and no reported result was found on Lacuna for any deconvolution method on plant tissue. We will run SCDC with all digestion-duration references as its multiple references, and Count Bridges with the pooled reference. Both get the same bulk samples and gene filtering, and each gets the same hyperparameter search budget as our model, measured in fits per fold. CRADLE-VAE will be run with cells grouped by digestion duration as its artifact contrast. This is the closest adaptation possible without intact cells.

The key ablation is our model with the bulk anchor removed, so that the $\tau = 0$ value comes only from extrapolating the digestion series. A second ablation drops the drought-by-artifact interaction. Beating the baselines means two things: lower held-out reconstruction error on intact drought bulk, and higher recovery of planted cell-type drought effects in the semi-synthetic benchmark below.

Why not just…?

Why not just delete the known protoplasting genes? This is the most serious competitor. Its weakness is overlap: the digestion response includes hormone, heat-shock, and oxidative genes that are also core drought genes, so deletion removes signal along with artifact. The semi-synthetic benchmark measures this directly by planting drought effects in artifact-responsive genes. Our model should recover them, while deletion recovers none of them by construction. If the planted effects in the real overlap set turn out to be negligible, meaning few drought genes are digestion-responsive, then deletion is adequate. We would report that as a negative result.

Why not just extrapolate the digestion-duration curve back to zero and skip the bulk? This is the no-anchor ablation. Early digestion responses are fast and likely saturate within the shortest feasible duration, so linear or monotone extrapolation will misjudge the intercept. The test is whether the no-anchor model's implied intact profile reconstructs held-out intact bulk as well as the full model does. If it does, the bulk anchor is unnecessary, and the method simplifies to a dose-response correction.

Why not just run SCDC or Count Bridges with the digested atlas as reference? Both treat the reference as correct for each cell type. A shared artifact therefore reappears in their estimated cell-type profiles, and wherever proportions shift with drought, part of it appears as a drought effect. SCDC's reference weighting cannot fix a bias that every reference shares. The comparison on reconstructing held-out intact bulk, and on the epidermal-peel ground truth, settles this.

Why not just use CRADLE-VAE? Its counterfactual needs some cells without the artifact. With only digested cells, the best available contrast is between durations, which captures differences in artifact strength but not its baseline level. We include this adaptation as a baseline rather than assume it fails.

Proposed Experiments

The semi-synthetic benchmark provides known answers. We take our day-0 single-cell data and estimate the artifact from paired bulk. We then plant drought effects of known size into chosen cell types, including genes inside the artifact-responsive set, and regenerate both single-cell counts at each duration and intact bulk counts from the model's generative process with resampled proportions. The metrics are correlation between estimated and planted cell-type log fold changes, and the false-positive rate for calling artifact-only genes drought-responsive. These are reported separately for genes inside and outside the overlap set.

On real data, the first check is held-out reconstruction. The model is fit without the day-15 intact bulk and must predict it at the count level, scored by negative binomial log-likelihood and by gene-wise correlation. The second is physical ground truth: deconvolved epidermal-versus-mesophyll drought changes for the proline and antioxidant genes should agree in sign and rank with peel-based qRT-PCR. The third is the trait test against the bulk co-expression baseline already described.

The decisive outcomes are as follows. If gene-list deletion matches our model on planted-effect recovery and trait prediction, the proposal's central ingredient is unnecessary. If the no-anchor ablation matches the full model on held-out intact bulk, the bulk data is unnecessary. If both comparisons favor the full model, and the proline module concentrates in a few cell types in 'Jin' but not 'TM-1', we will have the first artifact-corrected, cell-type-resolved account of drought metabolic reprogramming in a polyploid crop. We will also have a reusable recipe, intact bulk as the zero-digestion anchor, for any plant single-cell study of abiotic stress.