When Cells Won't Sit Still: Mosaic Integration, EMT Trajectories, and Smarter Deconvolution

Introduction

Single-cell biology has moved well past "what cell types exist?" These days, the harder questions are about dynamic cell states, regulatory transitions, and how to pull a useful signal out of data that is, frankly, messy. Real experiments don't hand you the tidy multi-omic profile for every cell that benchmarking papers assume. Instead, you get imbalanced modality coverage, batch-confounded cohorts, and bulk tissues where every cell type is mashed together.

This journal club covers three 2026 papers that, despite tackling different problems, keep circling the same question: How do we reliably infer cell states and their gene regulatory programs from imperfect, heterogeneous data?

• Sheng et al. introduce Palette, a method for "mosaic" integration of single-cell multi-omic data with imbalanced modality coverage.
• Pérez-González et al. combine scRNA-seq + scATAC-seq with CRISPR to map the transcription factor networks driving EMT states in squamous cell carcinoma.
• Zhang et al. develop BayesPrism-DWLS, a deconvolution framework that recovers cell-type proportions and cell-type-specific expression from bulk and spatial data together.

Together, they trace a methodological and biological arc: from integration across modalities, to mechanistic dissection of tumor plasticity, to extending single-cell insights into cheaper bulk assays.

Paper 1: Palette — Robust Mosaic Integration for Imbalanced Modalities

Key Findings

• Palette outperforms state-of-the-art mosaic integration methods in systematic benchmarks, especially when modality composition varies dramatically across datasets (some cells profiled with RNA+ATAC, others with RNA+protein only).
• It preserves rare subpopulations and condition-specific cell states that competing methods tend to over-correct away, validated across cross-condition and cross-species scenarios.
• The framework disentangles technical noise from biological variation using a PCA-derived representation, then uses the topological structure (manifold and graph relationships between cells) to align datasets without forcing spurious mixing.
• Palette generalizes beyond its primary use case. The authors show it extends to other challenging integration scenarios, hinting that it can serve as a general-purpose backbone.

Methodology & Study Design

Palette sits on a variant of principal component analysis tailored for mosaic multi-omic data. The core idea: technical artifacts like missing modalities and batch effects often show up as low-rank perturbations, which can be modeled and removed while the geometric relationships that carry the biology are preserved. The method uses graph-based representations of cell neighborhoods to guide alignment, an approach that echoes manifold-learning frameworks but is reshaped for the mosaic setting. Benchmarks covered synthetic and real datasets with deliberately imbalanced modality compositions, with Palette compared against strong baselines including UINMF, scMoMaT, MIDAS, scVAEIT, mmAAVI, ACE, and liam.

Why does this matter?

Existing mosaic integration tools either over-correct, wiping out real biology, or fail to align when one modality dominates. Palette's PCA-based, topology-aware design offers a more interpretable alternative to deep learning methods while still handling the hard cases. For large consortia assembling disease atlases from heterogeneous data, a single flexible framework that works across modality combinations reduces the need for a bespoke pipeline per dataset.

Paper 2: EMT Multiomics — Transcription Factor Networks Driving Tumor States

Key Findings

• Using scRNA-seq + scATAC-seq in a mouse model of skin squamous cell carcinoma (SCC), the authors map the transcriptomic and chromatin landscapes of distinct EMT states and confirm EMT as a continuum rather than a binary switch.
• Early-EMT regulators Klf5 and Pitx1 were identified via CRISPR/Cas9 loss-of-function studies; both are essential for metastasis formation in vivo.
• Late-EMT regulators Nfatc1 and Creb3l1 act at later stages of the EMT trajectory, suggesting stage-specific regulatory control.
• The authors show that similar EMT states and regulatory patterns exist in mouse pancreatic adenocarcinoma and human cancers, supporting cross-tissue and cross-species conservation.

Methodology & Study Design

The study paired single-cell RNA-seq and single-cell ATAC-seq to profile gene expression and chromatin accessibility in the same cells. From this multi-omic map, the authors inferred TF activity using chromatin motif enrichment coupled with target-gene expression, then validated the top candidates with CRISPR/Cas9 knockouts, followed by in vitro assays (invasion, migration) and in vivo metastasis experiments. The approach fits a growing trend of using multi-omic data to infer causal regulatory relationships, moving past correlation-based TF-target predictions. The work lines up with broader efforts to classify EMT states (epithelial, hybrid, mesenchymal) and identify drivers like TCF7, ZEB1, and ETS-family TFs, but goes further by linking specific TFs to discrete transition stages.

Why does this matter?

Hybrid and partial EMT states are clinically critical because they show high plasticity, collective migration, and stem-like properties, which makes tumors resistant to therapy and prone to metastasis. By pinpointing stage-specific TF regulators, the study offers actionable candidate targets for blocking metastasis or breaking up the most aggressive EMT states. It also shows what multi-omic integration (here, shared-cell RNA+ATAC) can deliver beyond descriptive atlases: real mechanistic insight.

Paper 3: BayesPrism-DWLS — Integrated Deconvolution for Bulk and Spatial Data

Key Findings

• The authors systematically benchmarked multiple deconvolution approaches across pseudo-bulk mixtures and real bulk RNA-seq datasets from multiple tissues and uncovered critical underdeterminedness in existing tools.
• They developed BayesPrism-DWLS, which achieves single-sample, gene-level inference of cell-type-specific expression, a goal prior methods rarely meet.
• Applied to mouse colon bulk RNA-seq and spatial transcriptomics, the framework uncovered cell-type-specific genes and pathways undetectable at the bulk or spot level, effectively pulling hidden heterogeneity out of cheaper data.

Methodology & Study Design

BayesPrism-DWLS fuses two ideas: a Bayesian framework (BayesPrism) for estimating cell-type proportions, and a DWLS (Dampened Weighted Least Squares) step that sharpens gene-level expression estimates. The key move is treating proportion and expression inference as a joint optimization problem, instead of estimating them sequentially and letting errors propagate. The authors argue that sequential estimation is a major source of the underdeterminedness that plagues deconvolution; their integrated approach breaks that bottleneck. The tool is validated on pseudo-bulk controls, real bulk tissues, and spatial transcriptomics, showing broad applicability.

Why does this matter?

Single-cell profiling is powerful but expensive, and many valuable cohorts exist only as bulk RNA-seq or low-resolution spatial data. BayesPrism-DWLS stretches the reach of single-cell reference atlases, letting researchers run mechanistic studies informed by cell-type-specific transcriptional programs on cost-effective sequencing data. For translational projects with tight budgets or archived bulk samples, that opens up a lot of room.

Synthesis & Discussion

Common Threads

All three papers pivot on the same conceptual axis: how to extract cell-type- or state-specific programs from complex, heterogeneous data. Each addresses a different slice:

Paper	Data Setting	Primary Contribution
Palette	Multi-omic single-cell	Integration across imbalanced modalities
EMT multiomics	scRNA + scATAC	Mechanistic TF network for EMT states
BayesPrism-DWLS	Bulk + spatial	Deconvolution with joint expression inference

Where They Complement Each Other

• Palette integrates single-cell data with missing modalities; BayesPrism-DWLS then extends those single-cell insights to bulk and spatial data. A natural pipeline: integrate single-cell atlases with Palette, then feed the resulting atlas into BayesPrism-DWLS to deconvolve clinical bulk cohorts.
• The EMT study demonstrates the kind of mechanistic insight that such integrated, deconvolved data could enable at scale, for example tracking EMT state distributions across patient tumors by deconvolving bulk RNA-seq against an EMT-resolved single-cell reference.
• All three emphasize robustness to real-world data imperfections: Palette handles imbalanced modalities, the EMT study handles heterogeneous tumor samples, and BayesPrism-DWLS handles underdetermined bulk mixtures.

Tensions or Differences

• Granularity of "state": Palette zooms in on broad condition-specific states and rare subpopulations; the EMT paper resolves fine-grained trajectory stages; BayesPrism-DWLS recovers cell-type-level programs. These resolutions aren't contradictory, but they highlight the absence of a shared, quantitative metric for "cell state" across studies.
• Method-centric vs. biology-centric: Palette and BayesPrism-DWLS are method-development papers with biological demonstrations; the EMT paper is biology-driven, with integration as an enabler. Both flavors are needed, but they serve different audiences.

Emerging Patterns

• A clear shift toward hierarchical integration across assay types, from single-cell multi-omics through bulk and spatial data, is accelerating.
• Methods are increasingly co-developed with sophisticated biological questions (such as EMT-driven metastasis), rather than built in isolation.
• The field is moving from static cell-type catalogs toward dynamic state and trajectory models.

Unanswered Questions

• Can deconvolution frameworks incorporate regulatory information (ATAC, TF networks) to infer state-specific programs, not just expression?
• How well do mosaic integration methods scale to longitudinal or interventional clinical data?
• Are the EMT TFs identified in mouse SCC conserved in human patients, and can cross-species integration tools help validate this?

For Your Lab Meeting

1. Mosaic integration in the wild: If your lab generates multimodal data, how often do you hit imbalanced modality coverage? Would Palette's PCA-based, topology-aware approach slot into your pipeline better than deep-learning alternatives?
2. Causality in TF networks: The EMT paper infers TF activity from multi-omic data and validates with CRISPR. How generalizable is that workflow? What are the failure modes when motif enrichment doesn't correlate with expression changes?
3. Deconvolution validation: BayesPrism-DWLS claims gene-level accuracy. What ground-truth standards would you want before trusting deconvolved expression for downstream pathway analysis?
4. Combining tools: Could you build an end-to-end pipeline (Palette → EMT reference → BayesPrism-DWLS) to track EMT states across a bulk RNA-seq cohort? What are the practical bottlenecks?
5. Defining "state": Each paper operationalizes cell states differently (clusters, trajectories, deconvolved programs). Should the community adopt standardized state definitions, or is methodological pluralism a feature rather than a bug?

Key Terms

• Mosaic integration: Combining single-cell datasets where different cells or batches have different combinations of measured modalities (e.g., some with RNA+ATAC, others with RNA only).
• EMT (Epithelial-to-Mesenchymal Transition): A process by which epithelial cells lose cell-cell adhesion and take on mesenchymal, migratory traits; in cancer, linked to invasion, metastasis, and therapy resistance.
• scATAC-seq: Single-cell Assay for Transposase-Accessible Chromatin sequencing; profiles chromatin accessibility at single-cell resolution, revealing putative regulatory regions and TF binding sites.
• Deconvolution: Computational estimation of cell-type proportions and/or cell-type-specific expression from mixed (bulk or spatial) samples, usually with a single-cell reference.
• Manifold / topological structure: In single-cell data, the geometric shape formed by cells in high-dimensional gene-expression space; methods that respect this geometry (rather than Euclidean distances) often preserve biological relationships better.

---

References

1. Sheng Q, Zhou Y, Zhu F, Xu L, & Jin S (2026). Robust integration of single-cell datasets with imbalanced modality composition.. Nature communications. https://doi.org/10.1038/s41467-026-72933-4

2. Pérez-González A, Windels G, Bévant K, Lengrand J, Lemaire S, Moers V, et al. (2026). Single cell multiomics unravel the transcription networks controlling the different EMT tumor states.. Nature communications. https://doi.org/10.1038/s41467-026-75521-8

3. Zhang Z, Wang X, Hong F, Yu P, Suo S, & Chen YG (2026). Integrated inference of cellular compositions and gene expression programs by deconvolution.. Cell regeneration (London, England). https://doi.org/10.1186/s13619-026-00299-5