Spatial transcriptomics has given biologists something they have wanted for decades: the ability to look at a slice of tissue and see, with genome-wide precision, which genes are turned on at every location. The technology is transformative for fields from developmental biology to cancer research, revealing how cells organize themselves into tissues, how tumors remodel their microenvironment, and how organs form during development.
But there is a catch. Most spatial transcriptomics platforms do not capture single cells. They capture spots — small circular or square regions of tissue, each containing anywhere from one to a dozen cells. The RNA measured at a single spot is a mixture of signals from whatever cell types happen to be sitting there. A spot that falls on the boundary between a tumor and a blood vessel, for example, might contain cancer cells, endothelial cells, and immune cells all mixed together. To understand the biology, you need to untangle that mixture — a process called deconvolution.
For years, the standard approach to deconvolution has been simple in concept: bring in a separate reference dataset, usually from single-cell RNA sequencing (scRNA-seq), that tells you what each cell type looks like transcriptionally. Then use that reference to unmix the spatial data, assigning proportions of each cell type to every spot. It works well when you have a high-quality, matched reference from the exact same tissue and condition.
But that is a big “when.”
The reference problem
Matched single-cell references are expensive and time-consuming to generate. They require dissociating the tissue into individual cells, a process that destroys spatial context and can itself introduce biases. Many important tissue samples — clinical biopsies, archived FFPE specimens, rare developmental samples — simply cannot be paired with a fresh single-cell dataset from the same specimen.
Even when a reference exists, it may not be a good match. Single-cell and spatial transcriptomics platforms have different capture efficiencies, gene detection rates, and biases. A reference from another donor, disease stage, or lab introduces biological variation that is hard to separate from genuine tissue biology. Marker gene lists are a common workaround, but they are coarse, missing subtle cell states and failing altogether in unfamiliar tissues.
This creates something of a chicken-and-egg problem. You need to know what cell types are present to analyze your spatial data, but you need spatial data to understand what cell types are present. The reference, rather than being a helpful addition, becomes a prerequisite that limits the scope of what spatial transcriptomics can study.
RETROFIT: deconvolution without the reference
A new method published today in Nature Communications breaks that dependency. Called RETROFIT (Reference-Free Deconvolution of Cell-Type Mixtures in Spatial Transcriptomics), the method is a Bayesian framework that works directly from the sequencing measurements themselves, without requiring any external reference at the deconvolution stage.
Developed by a team at Penn State University — Roopali Singh, Xi He, Xinyue Wang, Adam Keebum Park, Ross Cameron Hardison, Xiang Zhu, and Qunhua Li — RETROFIT treats the reference-free nature of the problem not as a limitation to work around, but as the central design constraint.
The core insight is mathematical. Every spot in a spatial transcriptomics dataset records a vector of gene expression values — thousands of genes measured simultaneously across hundreds or thousands of spots. If you have enough spots and enough genes, the structure of the data itself contains the information needed to separate cell types. Different cell types express different combinations of genes, and those combinatorial patterns leave statistical signatures in the raw data matrix. The challenge is extracting them.
RETROFIT frames this as a Bayesian non-negative matrix factorization problem. It decomposes the spatial transcriptomics data matrix into two smaller matrices: one that captures the cell-type-specific gene expression profiles (what each cell type looks like), and one that captures the proportion of each cell type at each spot (where each cell type is located). The Bayesian approach gives the model a principled way to handle uncertainty — spots with ambiguous mixtures are flagged rather than overconfidently assigned — and the non-negativity constraint ensures that the solutions are biologically interpretable: no negative expression values or negative proportions.
The method uses a structured stochastic variational inference algorithm to scale efficiently to the size of typical spatial transcriptomics datasets, which can include thousands of genes and tens of thousands of spots.
Crucially, RETROFIT’s architecture separates the deconvolution step from the cell-type-labeling step. The method first learns latent cell-type components from the spatial data alone. Only after those components are identified does it optionally use external information — either a single-cell reference or a marker gene list — to assign biological names (e.g., “CD4+ T cell,” “hepatocyte”) to the components it found. This two-stage design means that the core decomposition is never contaminated by reference bias. The reference, when used, is only a label, not a crutch.
How it performs
The Penn State team tested RETROFIT extensively against both reference-free and reference-based alternatives. In controlled simulations where ground-truth cell-type composition was known, RETROFIT outperformed existing reference-free methods. More importantly, it matched or exceeded reference-based methods even when those methods were given perfect, matched references.
That last result is striking. It suggests that reference-based deconvolution, for all its popularity, may be introducing error rather than removing it — error from platform effects, incomplete cell-type coverage, or mismatched biological states. RETROFIT avoids all of that by staying grounded in the data it is analyzing.
The advantages became clearer with imperfect references — the real-world cases where reference-based methods struggle most. When the team introduced mismatched references (simulating the common scenario where a public reference from a different donor is used instead of a matched one), reference-based performance dropped substantially. RETROFIT’s performance was unchanged.
Seeing fine structure without a map
Perhaps the most compelling demonstration came from Visium HD data. Visium HD, released by 10x Genomics, represents a leap in spatial transcriptomics resolution, replacing the standard 55-micrometer spots with continuous 2-micrometer features that approach single-cell scale. But even at this higher resolution, individual bins can still contain mixtures of cell types, and Visium HD’s finer structure introduces new data sparsity challenges that test the limits of deconvolution methods.
RETROFIT performed effectively on Visium HD data, recovering fine-grained spatial patterns without requiring single-cell references or marker genes. The method identified spatially coherent cell-type domains — regions where particular cell types dominate — that matched known tissue anatomy, all derived purely from the structure of the spatial expression data.
This is particularly valuable for the growing number of researchers adopting high-resolution spatial platforms. The ability to analyze Visium HD data without needing a matched scRNA-seq reference removes a major barrier to entry and accelerates the turnaround from tissue section to biological insight.
Broad applicability
RETROFIT is not limited to any single platform. The team demonstrated its effectiveness across multiple spatial transcriptomics technologies, including Slide-seq (which uses barcoded beads) and the 10x Visium platform. The method handles varying spot sizes, different capture efficiencies, and the range of gene detection rates that characterize different platforms, all without retuning.
The method is available now as a Bioconductor R package called `retrofit`, with full documentation and vignettes for both simulated and real data. This makes it accessible to the broad community of computational biologists who work within the Bioconductor ecosystem.
A new path forward
The publication of RETROFIT does not make reference-based deconvolution obsolete. When a high-quality, matched single-cell reference is available, reference-based methods remain a powerful tool. But RETROFIT changes the default assumption. Instead of asking “Do I have a good enough reference to analyze my spatial data?” researchers can now ask “What does my spatial data reveal on its own?” and add references later for biological interpretation.
This shift is timely. Spatial transcriptomics is expanding rapidly — into clinical laboratories, into developmental biology, into fields like plant biology and microbiology where matched single-cell references are rare or nonexistent. A reference-free approach that works across platforms and resolutions makes spatial transcriptomics a more democratic technology, accessible to any lab with a tissue section and a sequencer.
For the growing community of scientists mapping the spatial organization of gene expression, RETROFIT offers something simple and powerful: the ability to see what is there without needing to know what you are looking for first.
Reference: Singh, R., He, X., Wang, X., Park, A. K., Hardison, R. C., Zhu, X., & Li, Q. RETROFIT: Reference-free deconvolution of cell-type mixtures in spatial transcriptomics. Nature Communications (2026). DOI: 10.1038/s41467-026-74928-7