Automated microscopy can already turn a single cell into thousands of numbers. In the Cell Painting assay, five fluorescent stains light up the nucleus, mitochondria, endoplasmic reticulum, actin cytoskeleton, and other organelles, and software reduces each cell to a vector of shape, texture, and intensity measurements. This is cheap and scalable, but it has a documented blind spot: biologically different cells can look identical, and biologically important changes can leave shape untouched. Meanwhile, genetic lineage tracing solves a different problem: it writes permanent, heritable marks into the genome so that every descendant of a marked cell carries a family name, recoverable even after the original cell is gone. These two technologies have never been joined. Every method that models how cell states evolve along a family tree, such as Reconstructing unobserved cellular states from paired single-cell lineage tracing and transcriptomics data, works on RNA counts and therefore requires killing the cell; every method that reads rich morphology, such as Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion, treats cells as unrelated strangers.
This proposal argues that ancestry is the missing conditioning variable for morphological profiling. I propose to build a generative model, MorphoTree, that places the lineage tree directly into the prior over single-cell morphological profiles. The payoff is threefold: we can infer the morphology of ancestors we never imaged, we can separate cells that look alike but carry different heritable programs, and we can tell whether the cell-to-cell heterogeneity in a population is inherited from founders or imposed by the local environment. The last point is exactly the question that population-averaged measurements cannot answer — the same averaging that David Adams et al. show hides rare, high-consequence subpopulations, and that clinical deconvolution tools like Cross-modal transfer learning for mapping bulk transcriptomes at cellular level can only partially undo, because they resolve what is present but never how it got there.
Core Claim
The claim has three parts, in increasing order of ambition. First, morphological profiles are not noise around a population mean; they drift along lineage trees in a structured, heritable way, and a model with a tree-structured prior reconstructs that drift better than a model that ignores ancestry. Second, lineage context carries predictive signal that morphology alone lacks: conditioning an image-based classifier on clonal identity should improve predictions of molecular state and fate precisely in the regime where images are weakest — the "morphology ceiling" documented when drugs like Simvastatin barely change cell shape in Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion. Third, the same model cleanly partitions morphological variance into a between-clone (heritable) and within-clone (environmental or stochastic) component, which converts the standing tension between lineage-deterministic and niche-deterministic views of cell state into an estimable quantity.
The Model
MorphoTree is a variational autoencoder whose latent prior is a Gaussian random walk on the lineage tree, applied to continuous morphological feature vectors rather than RNA counts. The design deliberately adapts the machinery of Reconstructing unobserved cellular states from paired single-cell lineage tracing and transcriptomics data, which embeds a phylogenetic tree into a latent prior and marginalizes unobserved ancestral nodes with linear-time message passing, but swaps its negative-binomial count likelihood for a Gaussian one. This swap is not cosmetic: Count Bridges enable Modeling and Deconvolving Transcriptomic Data showed that treating integer RNA counts as continuous data violates the geometry of the problem, and the converse holds here — CellProfiler features are genuinely continuous, so Gaussian innovations are the right model, and the tree-structured prior transfers naturally.
Concretely, let $m_i \in \mathbb{R}^d$ be the log-transformed, batch-corrected morphological profile of cell $i$ (on the order of $d \approx 5{,}800$ features in the CPJUMP1 pipeline of Three million images and morphological profiles of cells treated with matched chemical and genetic perturbations), with latent state $z_i \in \mathbb{R}^q$. For a cell $c$ with parent $p(c)$ on the lineage tree, with $b_c$ the branch length measured in generations,
$$z_c \mid z_{p(c)} \sim \mathcal{N}\!\left(z_{p(c)},\; \tau\, b_c\, I\right),$$
where $\tau$ is a single learnable rate of heritable morphological drift. The decoder maps $z$ to the mean of $m$, and message passing over the tree yields posterior distributions at every internal node — inferred morphologies for ancestors that were never photographed, obtained without destroying any descendant. A second head maps $z$ to a 512-dimensional transcriptomic embedding, trained with the population-level supervision scheme of Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion, so the model also predicts the molecular state of unsequenced and unobservable cells. For classification tasks, a learned embedding of the clonal barcode is concatenated to $z$; this is the ingredient whose value we most need to isolate, because the analogous ablation in Deep Lineage: Single-Cell Lineage Tracing and Fate Inference Using Deep Learning — removing clonal context from a transcriptomic fate predictor — dropped accuracy from 77% to 52%, the strongest available evidence that ancestry carries signal independent of instantaneous state.
The variance partition falls out of the fitted tree: with $\hat{m}$ the model-implied clone means,
$$h = \frac{\mathrm{Var}_{\text{between clone}}(\hat{m})}{\mathrm{Var}_{\text{between clone}}(\hat{m}) + \mathrm{Var}_{\text{within clone}}(m)},$$
a heritability-style statistic for morphology, estimable per feature, per organelle channel, and per tissue.
Research Program
The program has three threads. The first builds and validates MorphoTree on simulation and on the one existing dataset with both lineage and live morphology, the zebrafish neuromast recordings of Live cell-lineage tracing and machine learning reveal patterns of organ regeneration, which tracked 106 founder cells and more than 100,000 descendant events through regeneration. The second generates the missing data: I will extend the scGeneScope protocol of Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion — matched Cell Painting images and single-cell RNA sequencing across 28 compounds — by adding lentiviral static barcodes read out through the destainable-dye, in-situ-sequencing strategy that A genome-wide atlas of human cell morphology proved works at scale (over 30 million cells, at a reported cost as low as $0.001 per cell profile). The third applies the model to the clinically sharpest open question: whether the rare, pre-existing drug-resistant lineages that Actionable biological programs to enhance EGFR-targeted therapy response unveiled by single-cell lineage tracing in clinically relevant lung cancer models discovered molecularly — subclones enriched for Hedgehog signaling and FOXD1 activity that survive osimertinib — already carry a detectable morphological fingerprint before any drug is applied. That paper's readouts were RNA sequencing and colony assays; nobody has imaged the seeds. If they are visible in shape, morphology becomes a non-destructive, thousand-times-cheaper proxy for resistance risk.
Baselines
The naive reference point is what the field does today: average morphological profiles per condition and compare them, ignoring both single-cell structure and ancestry. Its documented performance sets the floor. On the CPJUMP1 benchmark, Three million images and morphological profiles of cells treated with matched chemical and genetic perturbations report that matching a compound to a sister compound hitting the same target succeeds only 5% to 25% of the time — that is the number our single-cell, lineage-aware retrieval must beat on matched tasks to justify the added machinery.
The strongest existing method on the image-to-molecular-state question is PhenoSeq, the conditional diffusion model of Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion, which reports single-profile classification balanced accuracy of 0.293 using its synthetic RNA embeddings versus 0.270 for raw imaging, and a doubling of accuracy in the sparse regime (0.271 versus 0.157 for imaging alone at 32 cells per profile) on scGeneScope. I will run PhenoSeq's public pipeline on our barcode-augmented scGeneScope data with identical feature extraction, the same training cells, and the same tuning budget, and add our lineage embedding as its only modification. Beating it means higher balanced accuracy overall, but the decisive comparison is on the subset of morphologically quiet compounds — Simvastatin-like cases where PhenoSeq's own authors report failure — where I expect the lineage signal to matter most.
On the ancestry and fate questions, the strongest competitors are Deep Lineage: Single-Cell Lineage Tracing and Fate Inference Using Deep Learning, which reports 77% early fate-prediction accuracy on mouse hematopoiesis (versus 65% for CoSpar and 65% for WOT) and 94% on day-28 fibroblast reprogramming outcomes, and the tree-structured VAE of Reconstructing unobserved cellular states from paired single-cell lineage tracing and transcriptomics data, which reports latent reconstruction mean squared error of 1.89 versus 2.28 for a tree-ignoring VAE on Gaussian Process Factor Analysis simulations. Both are transcriptomic, so the fair comparison is to port each architecture onto our morphological feature streams — same trees, same cells, same compute — rather than compare across modalities. No reported result found on Lacuna for ancestral morphological reconstruction, so our simulation study will establish the first reference numbers rather than chase an existing one. Finally, the key ablation is MorphoTree with the tree prior replaced by an independent-Gaussian prior, which reduces the model to a standard image-profile autoencoder plus barcode label; if the tree structure adds nothing over the barcode alone, the core claim fails and the honest conclusion is that lineage helps as a label but not as a geometry.
Why not just…?
Why not just use bigger image-only models? This is the most serious alternative, and the evidence against it is convergent but circumstantial. Sister-compound matching sits at 5–25% in CPJUMP1 (Three million images and morphological profiles of cells treated with matched chemical and genetic perturbations), PhenoSeq fails outright on compounds that leave shape unchanged, and in Combining phenomics with transcriptomics reveals cell-type-specific morphological and molecular signatures of the 22q11.2 deletion the same genetic lesion is detectable from images in astrocytes (AUC 0.96) but invisible in neurons (AUC 0.54) — a capacity limit of the modality, not of model size. Still, capacity might recover some of this, so the quiet-compound subset of our benchmark is designed to settle it: if a vision transformer trained on images alone matches MorphoTree there, the lineage ingredient is not earning its keep.
Why not just sequence barcodes and transcriptomes together? That is the Deep Lineage and TreeVAE recipe, and for pure ancestral transcriptome inference it will likely remain superior. But it is destructive: as both Deep Lineage: Single-Cell Lineage Tracing and Fate Inference Using Deep Lineage and PhenoSeq stress, a sequenced cell cannot be watched afterwards, so every intermediate on the path from founder to resistant descendant is unmeasured rather than merely unmodeled. Morphology is observable in living cells across divisions, and the fusion evidence from Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion — imaging plus synthetic RNA recovers about 29% of the gap to real transcriptomics, and more than doubles accuracy when only 32 cells are available — shows the two signals are not redundant. Our design keeps sequencing as the supervision, not the observation, which is exactly what makes ancestral reconstruction and live tracking possible.
Why not just run standard deconvolution on the bulk sample? Tools like POPPY and Count Bridges enable Modeling and Deconvolving Transcriptomic Data estimate which cell types are present and in what proportions, but treat every cell in a mixture as exchangeable. The lineage-tracing literature shows this exchangeability assumption fails exactly where it matters clinically: Victor G. Puelles et al. proved kidney recovery after injury is hypertrophy of surviving cells, not regeneration, which a proportion-based method would misread; Priya H. Dedhia proved wound fibroblasts arrive partly from circulating bone-marrow precursors; Fabian Peisker et al. tracked cardiac fibroblasts switching phenotype while losing their original markers. Monoclonal expansion, recruitment, and transdifferentiation produce the same deconvolved proportions but demand opposite therapies. MorphoTree's heritability statistic $h$ is built to separate them, and the zebrafish and organoid experiments below test whether it does.
Decisive Experiments
The first experiment is a controlled simulation: I will extend the Gaussian Process Factor Analysis setup from Reconstructing unobserved cellular states from paired single-cell lineage tracing and transcriptomics data so that observations are continuous, CellProfiler-like vectors rather than counts, with known ground-truth ancestral latents, and measure reconstruction mean squared error for MorphoTree versus a tree-ignoring autoencoder and versus the ported TreeVAE architecture. The second experiment reanalyzes the zebrafish neuromast dataset, where the published Random Forest on 32 spatial and neighborhood features already predicts division outcomes at MCC = 0.91 (Live cell-lineage tracing and machine learning reveal patterns of organ regeneration); I will ask whether adding founder lineage raises that ceiling and estimate $h$ for the tracked morphological descriptors. I expect the niche to dominate here — the paper's own isotropic-regeneration finding points that way — and a low $h$ would be an informative validation of the statistic, not a failure. The third experiment is the new barcode-augmented scGeneScope collection, comparing against PhenoSeq's reported balanced accuracies of 0.293 (single profile) and 0.271 (sparse, 32-cell fusion) under matched conditions, with gains concentrated on the quiet-compound subset identified by a CPJUMP1-style signal-to-noise screen. The fourth and most consequential experiment applies the pipeline to EGFR-mutant lung cancer organoids under osimertinib, following the design of Actionable biological programs to enhance EGFR-targeted therapy response unveiled by single-cell lineage tracing in clinically relevant lung cancer models: image and barcode cells before drug exposure, then test whether pre-treatment morphology plus lineage predicts which clones survive, reported as AUROC for lineage-level resistance with an image-only classifier as the internal control. As a cross-check on heritability estimation, I will apply the partition to the Bitbow Drosophila data of Bitbow Enables Highly Efficient Neuronal Lineage Tracing and Morphology Reconstruction in Single Drosophila Brains, where barcodes demonstrably co-segregate with the eight classified neuronal morphological subtypes and $h$ should be high. If the pre-treatment organoid classifier beats image-only baselines, morphology becomes a live, cheap readout of clonal risk; if it does not, we will have cleanly bounded the morphology ceiling from above, which the field currently lacks.
References
- Reconstructing unobserved cellular states from paired single-cell lineage tracing and transcriptomics data
- Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion
- 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
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Three million images and morphological profiles of cells treated with matched chemical and genetic perturbations
- Deep Lineage: Single-Cell Lineage Tracing and Fate Inference Using Deep Learning
- Live cell-lineage tracing and machine learning reveal patterns of organ regeneration
- A genome-wide atlas of human cell morphology
- Actionable biological programs to enhance EGFR-targeted therapy response unveiled by single-cell lineage tracing in clinically relevant lung cancer models
- Combining phenomics with transcriptomics reveals cell-type-specific morphological and molecular signatures of the 22q11.2 deletion
- mTOR-mediated podocyte hypertrophy regulates glomerular integrity in mice and humans
- Single-cell analysis reveals fibroblast heterogeneity and myeloid-derived adipocyte progenitors in murine skin wounds
- Mapping the cardiac vascular niche in heart failure
- Bitbow Enables Highly Efficient Neuronal Lineage Tracing and Morphology Reconstruction in Single Drosophila Brains