Single-cell RNA sequencing (measuring the genes active in each individual cell) shows which cells respond to a drug. It is also expensive and it destroys the cell. Bulk RNA sequencing of a whole well is cheap, but it reports only the average of thousands of cells. That average cannot tell a small subpopulation changing strongly apart from every cell changing a little. Cell Painting, a microscopy assay that stains seven cell structures and images every cell, is cheap and resolves single cells, but it carries no gene-level information. This note proposes to let the microscope decide who is in the well and the bulk sequencer decide what those groups are expressing. Morphology assigns each imaged cell to a discrete state, such as a cell-cycle phase or a responder/non-responder shape class. The per-well fractions of those states are then treated as known mixing weights. Bulk RNA from a matched well is unmixed into state-specific expression by regression across many wells.
The central distinction from image-to-transcriptome generation is where the gene-level signal comes from. Here the transcriptomic signal is measured, not predicted from pixels. The image supplies only the partition of the well into cell states. Our method should therefore keep working in the regime where predicting transcriptomes from images fails: compounds that change genes without changing shape. In that regime it should at least recover the whole-well change correctly. The price is that the partition must be estimated consistently across plates and experiments. Technical batch effects therefore stop being a nuisance and become the main source of bias in the mixing weights.
Core Claim
Consider well $w$ treated with condition $c$. Let $b_w$ be its bulk expression vector, and let $h_{wk}$ be the fraction of its imaged cells assigned to morphological state $k$. We model
$$b_w = \sum_k h_{wk}\,(\mu_k + \delta_{k,c}) + \varepsilon_w,$$
where $\mu_k$ is the baseline expression of state $k$, $\delta_{k,c}$ is the condition-specific shift of that state, and $\varepsilon_w$ is noise. Unlike classical deconvolution, the mixing weights $h$ are observed from images rather than inferred. The unknowns are the state profiles.
Identifiability requires the fractions to vary across wells. Dose series, partial-response compounds, and deliberate mixtures supply that variation. We will also impose a low-rank penalty on the shifts $\delta_{k,c}$ across conditions, so they can be estimated when variation within a single condition is small. This penalty is an assumption, and the experiments below test whether it holds.
The claim is concrete. With batch-robust state assignment, bulk-plus-image data recovers state-specific expression shifts that bulk averaging hides, down to minority fractions of a few percent. It does so at a fraction of the sequencing cost of single-cell profiling. When the relevant states are invisible to the microscope, it falls back to the correct bulk answer instead of producing a confident wrong one.
Relation to Prior Work
The motivation comes from the heterogeneity-masking literature. Averaging hides the population actually driving a phenotype, for example rare stem-like cells behind tamoxifen resistance (Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance), or the question of whether a change is primary in one cell type or compensatory in another (RNA-seq co-expression network analysis reveals anxiolytic behavior of mice with Efnb2 knockout in parvalbumin+ neurons). Resolving Cellular Heterogeneity Masked by Bulk Sequencing frames this directly.
Three lines of deconvolution work unmix bulk data using single-cell sequencing references:
- SCDC ensembles several single-cell RNA references to reduce reference batch bias.
- Count Bridges samples integer single-cell counts constrained to sum to observed aggregates.
- POPPY maps bulk profiles into a single-cell foundation-model space with ontology-aware contrastive learning.
All three infer the mixing weights and inherit the reference's blind spots: a state absent from the atlas cannot be recovered. We take the opposite route. We measure the mixing weights optically and need no single-cell sequencing reference for the perturbed states.
The closest multimodal work constrains the design. PhenoSeq generates single-cell transcriptomic embeddings from Cell Painting images with a conditional diffusion model. It reports that morphologically silent perturbations such as Simvastatin in U2-OS cells cannot be predicted. That negative result is why we refuse to derive gene signal from pixels.
scGeneScope supplies the data and a second constraint. It provides treatment-matched imaging and single-cell RNA for 28 compounds, and it shows that models lose 5–7% balanced accuracy when moving to held-out experiments. So any image-derived quantity must be validated across experiments, not within them.
Morphology and gene expression profiling provide complementary information motivates the division of labor. Morphology is far more reproducible across replicates (57–83% versus 16–35% replicating for L1000 expression), which is what we need from mixing weights. It captures fewer independent dimensions, which is why genes must come from sequencing.
Predicting cell health phenotypes using image-based morphology profiling shows that Cell Painting predicts cell-cycle stage. This gives us one morphological partition with an independent transcriptomic ground truth.
On batch effects, Evaluating batch correction methods for image-based cell profiling finds that Harmony and Seurat-RPCA rank consistently among the top three on JUMP Cell Painting features, while linear methods fail across microscopes. We reuse these methods for the state-assignment step.
Set-DINO and CWA-MSN learn batch-invariant encoders by pairing images of the same perturbation across wells. Set-DINO does this over mean-pooled sets of 16 cells. Mean pooling is exactly the averaging we are trying to undo. We therefore keep per-cell state assignment and count cells, and we never average embeddings.
Research Program
The pipeline has three stages.
Encode and assign states. Single cells are encoded, either with CellProfiler features or a CWA-MSN-style encoder trained with cross-well pairs. Cells are then assigned to $K$ states in two ways. The first is supervised: cell-cycle phase, trained from nuclear features. The second is unsupervised: clusters fit on control and treated cells from training experiments only. Per-well fractions $h_{wk}$ are simply counts, with uncertainty from assignment probabilities propagated into the regression.
Estimate state profiles. We solve the model above with $\mu_k$ shared across conditions and a nuclear-norm penalty on the matrix of shifts $\delta_{k,c}$. The error model uses a negative-binomial likelihood on counts rather than squared error, respecting the count nature of sequencing that Count Bridges emphasizes.
Flag silent effects. We compute the part of each condition's bulk change that no reweighting of state fractions can explain. Large unexplained change combined with unchanged fractions is the signature of a morphologically silent effect. The output then reports it as a whole-well shift instead of attributing it to any state.
The analysis centers on identifiability and bias. We will derive how state-profile error scales with the spread of the fractions across wells and with assignment error. Ordinary regression theory predicts that errors in $h$ act like errors-in-variables and bias the estimated shifts toward zero. We will measure that bias empirically, both with no batch correction and after correction.
Baselines
Bulk averaging (naive reference). Compute condition-versus-DMSO log fold changes on pseudo-bulk, with no unmixing. Every state is assigned the same whole-well change. Any method that cannot beat this on state-specific shifts has learned nothing from the images. We run it on exactly the same pseudo-bulk.
Reference-based deconvolution. SCDC is the strongest classical competitor because it explicitly handles reference batch variation. It is run with scRNA references from DMSO wells only (the realistic cheap setting) and, separately, with references that include treated cells (an oracle that spends the sequencing we claim to save). Its reported figure is a Pearson correlation of 0.991 with true proportions on 6:3:1 cell-line mixtures (SCDC). That figure concerns proportions on a different dataset, not state-specific expression, so it bounds what proportion estimation can reach rather than serving as a direct target. Count Bridges is run as a second reference-based competitor; no reported result was found on Lacuna for Count Bridges on scGeneScope or on state-specific expression recovery.
Image-to-transcriptome generation. PhenoSeq generates expression from images alone. On scGeneScope it reports 0.293 versus 0.270 balanced accuracy for synthetic RNA versus raw imaging in single-profile classification, and the synthetic centroid is the nearest neighbor of the real centroid for 15 of 29 conditions (PhenoSeq). We reproduce its centroid-nearest-neighbor test with our unmixed state profiles and treat beating 15 of 29 on the held-out-experiment split as the bar.
Fairness conditions. All methods see the same wells and splits. They share the same encoder where one is needed, and each gets an equal hyperparameter search budget of the same number of configurations, selected on held-out wells from training experiments.
Ablations. Four ablations isolate the key ingredient, the measured fractions:
- Permuted fractions: $h$ shuffled across wells. If performance holds, the images contribute nothing.
- Mean-pooled image embeddings as regression covariates instead of state counts. This tests whether counting rather than averaging matters.
- State assignment with no batch correction, with ComBat, and with Harmony. This measures how much batch-induced bias in the fractions costs.
- Oracle fractions from the known mixing ratio. This gives the ceiling when images are perfect.
Why not just…?
Why not just generate transcriptomes from images, as PhenoSeq does? Generation is capped by what shape reveals, and PhenoSeq's own Simvastatin result shows the cap is real. Our method measures gene signal and borrows only the partition from images. On silent compounds it should therefore match bulk averaging instead of missing the effect. Our dedicated comparison on morphologically silent conditions, reported separately from visible ones, will show whether this holds.
Why not just deconvolve bulk with a single-cell reference atlas? This is the most serious threat. If drug-induced states resemble states already present in DMSO cells, a DMSO reference plus SCDC may recover them without images. The DMSO-only SCDC arm tests exactly this. We expect images to help mainly where treatment creates states or mixtures the reference lacks. If SCDC with DMSO references matches us on the cell-cycle and mixture tasks, the proposal's value reduces to avoiding the reference altogether, and we will report that honestly.
Why not just sequence single cells more cheaply with split-pool barcoding? Combinatorial indexing such as SCITO-seq (Byungjin Hwang et al.) reaches more than 100,000 cells per reaction at 10–100-fold lower cost. For moderate screens it may simply be the better choice. We settle this with a cost-accuracy curve. We subsample real single cells per condition and find how many are needed to match image-anchored accuracy. If the answer is a few dozen cells, the proposal loses on cost, and we will say so.
Why not just batch-correct image features with ComBat? The JUMP benchmark shows linear corrections fail across instruments (Evaluating batch correction methods). Over-strong correction can also absorb a treatment-specific state when batch and treatment are confounded. The correction ablation measures fraction bias directly under each option, including this over-correction failure.
Proposed Experiments
All experiments use scGeneScope. Its round with an unresolvable confound is excluded, as the dataset authors did. Pseudo-bulk is formed by summing single-cell RNA counts per well, and fractions come from the matched imaging arm of the same condition and experiment.
Controlled mixtures of real cells. For compound–DMSO pairs, we build pseudo-wells mixing treated cells at fractions from 1% to 50% in both modalities. The unmixed treated-state profile is scored against the true pure-compound pseudo-bulk by correlation of log fold changes over responsive genes. We report the smallest fraction at which correlation exceeds the bulk baseline's. Imaging cells come from held-out experiments, so fraction errors reflect real batch shift.
Real within-well heterogeneity via cell-cycle phase. Phase is assigned from nuclear morphology in images and, independently, from marker-gene scores in single-cell RNA, giving true phase-specific pseudo-bulk per compound. We then test whether image-anchored unmixing recovers phase-specific drug responses that pseudo-bulk averaging blends.
Decisive checks. Three results would decide the proposal:
- Recovery must beat both the bulk baseline and the DMSO-only SCDC arm at minority fractions near 5%, on held-out experiments rather than only within them.
- The silent-effect flag must fire on Simvastatin and similar compounds, while unmixed profiles there do no worse than bulk.
- The permuted-fraction ablation must collapse performance to baseline.
Failing the first check means morphology fractions are too batch-sensitive to anchor deconvolution. Failing the third means the gains come from the regularizer, not the images.
References
- Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties
- RNA-seq co-expression network analysis reveals anxiolytic behavior of mice with Efnb2 knockout in parvalbumin+ neurons
- 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
- SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion
- scGeneScope: A Treatment-Matched Single Cell Imaging and Transcriptomics Dataset and Benchmark for Treatment Response Modeling
- Morphology and gene expression profiling provide complementary information for mapping cell state
- Predicting cell health phenotypes using image-based morphology profiling
- Evaluating batch correction methods for image-based cell profiling
- Weakly Supervised Set-Consistency Learning Improves Morphological Profiling of Single-Cell Images
- Efficient Cell Painting Image Representation Learning via Cross-Well Aligned Masked Siamese Network