Computational Deconvolution Review Benchmarks Twenty Spatial Transcriptomics Methods Across Diverse Technologies

Spatial transcriptomics technologies have reshaped modern biological inquiry by recording gene expression profiles directly across intact tissue sections, preserving the critical spatial coordinates necessary for deciphering tissue architecture, developmental progression, and complex disease microenvironments. However, a major computational bottleneck persists because many high-throughput sequencing platforms operate at capture spot dimensions that exceed individual cell sizes, leaving capture spots populated by heterogeneous mixtures of multiple cells. To resolve this resolution limit, computational deconvolution algorithms have emerged to infer the precise cellular composition and underlying cell-type proportions within each spatial feature. In a comprehensive review published in Genomics and Informatics, researchers Jahanzeb Saqib and Junil Kim systematically examine twenty distinct deconvolution algorithms, evaluating their underlying mathematical frameworks, computational paradigms, and operational trade-offs. The review organizes these modern algorithms into five overarching methodological families: probabilistic and stochastic models, non-negative matrix factorization techniques, graph theory-driven architectures, optimal transport frameworks, and deep learning or transformer systems, establishing a detailed technical resource for analyzing complex spatial datasets.

The review details the core technological divide in spatial profiling, contrasting imaging-based approaches with next-generation sequencing approaches that dictate spatial resolution and transcriptome coverage. Imaging-based methods rely on fluorescence microscopy to detect mRNA molecules in situ, providing single-cell or subcellular resolution at the expense of comprehensive gene capacity. Foundational single-molecule fluorescence in situ hybridization utilizes targeted fluorescent probes, whereas sequential hybridization expands profiling across dozens of genes, and combinatorial barcoding enables detection across hundreds to thousands of transcripts through intricate optical decoding systems. In contrast, sequencing-based platforms capture polyadenylated transcripts from permeabilized tissue sections using spatially barcoded oligonucleotide arrays, delivering unbiased transcriptome-wide coverage with reduced spatial precision. Platforms such as original Spatial Transcriptomics and commercial 10x Genomics Visium capture mixed cellular contents within multi-cellular spots. Although innovations such as Slide-seq microbeads, Seq-Scope patterned flow-cell arrays, and Stereo-seq DNA nanoball arrays push NGS capture toward submicron or nanometer scales, deconvolution remains essential for deciphering mixed cellular components.

Probabilistic and stochastic formulations represent the first major algorithmic category, modeling raw sequencing count data through negative binomial or Poisson distributions to account for inherent overdispersion and technical variance. Within this family, the AdRoit algorithm utilizes an adaptive learning framework originally designed for bulk RNA sequencing to estimate gene-wise technology biases, cross-sample variability, and cell-type specificity. AdRoit applies maximum likelihood estimation to determine mean and dispersion parameters, incorporates variance-stabilizing transformations to mitigate noise, and implements regularized non-negative least squares regression to minimize collinearity among transcriptionally similar cell subtypes. Adopting a different probabilistic strategy, the STRIDE framework utilizes Latent Dirichlet Allocation to discover latent gene expression topics from single-cell reference atlases. STRIDE applies an online variational Bayes algorithm to construct cell-type-by-topic distributions, optimizing topic numbers based on single-cell re-annotation accuracy before projecting these pre-trained topics onto spatial coordinates to infer cell proportions and reconstruct three-dimensional tissue structures.

Additional probabilistic tools emphasize count modeling and continuous transcriptional variation rather than relying on predefined marker gene sets. The stereoscope algorithm operates directly on unnormalized count matrices, fitting negative binomial distributions to single-cell data through maximum likelihood estimation and deconvolving spatial mixtures via maximum a posteriori estimation of gene expression rates and observation probabilities. The RCTD method, initially engineered for high-resolution Slide-seq data, extends Latent Dirichlet Allocation by incorporating platform-specific effects and random overdispersion terms. RCTD approximates expression through Poisson distributions, employing Sequential Quadratic Programming to optimize model log-likelihood under strict convergence thresholds while providing doublet and multimode operating configurations. Moving beyond static cell categories, DestVI implements multi-resolution Bayesian modeling through conditional deep generative networks. DestVI establishes two distinct latent variable models: a single-cell model capturing discrete cell-type signatures and a spatial model mapping continuous sub-cell-type variations, allowing investigators to quantify cell-type proportions alongside spatial state gradients and localized differential expression.

Hierarchical Bayesian structures and specialized regression paradigms further expand probabilistic deconvolution capabilities across complex tissues and variable capture depths. The cell2location framework deploys a Bayesian graphical model that decomposes spatial expression counts into predefined reference types by accounting for technology sensitivity, additive background noise, and location-specific RNA detection efficiencies. A key architectural element of cell2location is its factorized hierarchical Gamma prior on cell abundance, which captures spatial co-localization patterns across latent cell groups, enhances sensitivity for rare cell types, and regulates spot sparsity using image-derived cell counts. Cell2location utilizes GPU-accelerated variational inference via PyMC and Pyro within the scvi-tools environment to jointly model multiple experimental batches. Alternatively, SpatialDecon employs log-normal regression alongside explicit background modeling to outperform standard least-squares approaches. SpatialDecon incorporates an automated outlier detection mechanism, calculates formal confidence intervals, and translates relative fractions into absolute cell numbers whenever corresponding nuclei counts become accessible from imaging technologies such as GeoMx.

Spatial correlation modeling and reference-free inference define another branch of probabilistic algorithms designed for settings where high-quality single-cell atlases are unavailable. The CARD algorithm integrates single-cell references with spatial transcriptomics through a Conditional Autoregressive model combined with non-negative matrix factorization, applying spatial smoothing across neighboring locations via constrained maximum likelihood optimization. CARD can also impute cell compositions and unmeasured gene patterns across tissue gaps, whereas its reference-free extension, CARDfree, requires only curated marker lists to delineate overall tissue domains. Unsupervised reference-free deconvolution is also exemplified by the Berglund algorithm, which utilizes Poisson factor analysis to decompose expression counts into gene factors and spatial activity maps without single-cell data. Berglund applies Markov Chain Monte Carlo sampling, kernel-weighted smoothing, and hierarchical clustering to dissect complex prostate cancer sections. Similarly, STdeconvolve applies Latent Dirichlet Allocation directly to spatial count matrices after filtering invariant or uninformative genes, estimating optimal cell-type counts and generating deconvolved transcriptional profiles without external reference datasets.

The second major category centers on non-negative matrix factorization and constrained linear regression, offering computationally efficient and interpretable additive decompositions of capture spots into reference profiles. The NMFreg method maps single-cell identities onto Slide-seq beads by deriving a reduced-space basis using non-negative matrix factorization on highly variable reference genes. NMFreg evaluates factor loadings through non-negative least squares regression, assigning a cell type when its loading exceeds fifty percent of the total L2 norm, a process demonstrated on adult mouse cerebellum and hippocampus datasets where quality thresholding removed low-yield beads. SpatialDWLS improves deconvolution accuracy by introducing a two-phase workflow that first identifies candidate cell types at each spot using Parametric Analysis of Gene Set Enrichment. It then estimates fractional abundances using dampened weighted least squares with cross-validated damping constants to stabilize variance across clusters. Meanwhile, SPOTlight combines unit-variance normalization with seeded non-negative matrix factorization initialized with Seurat marker genes, utilizing non-negative least squares to ensure stable topic deconvolution even with compact single-cell reference sizes.

Graph-based architectures constitute the third paradigm, constructing topological networks where nodes represent spatial spots or individual cells while edges encode spatial adjacency, physical distance, or transcriptional similarity. The DSTG algorithm uses semi-supervised graph convolutional networks, training multi-layer convolutional structures on synthetic pseudo-spatial data derived from single-cell transcriptomes combined with real spatial spots. DSTG applies canonical correlation analysis to reduce dimensionality and learns spot link graphs to predict real tissue composition across heteroskedastic data. Addressing technical dropout effects, the SD2 framework combines an autoencoder with a three-layer graph convolutional network activated by rectified linear units and normalized by softmax functions, integrating dropout-based gene features and spatial coordinates to map cell distributions. Shifting toward probabilistic graphical representations, the SpiceMix model integrates non-negative matrix factorization with a hidden Markov random field. SpiceMix simultaneously learns intuitive metagenes, node latent states, and spatial affinities via alternating maximum a posteriori optimization, directly linking continuous cell states to tissue microenvironments.

Optimal transport theory provides a distinct mathematical framework for spatial deconvolution by formulating cell allocation as a constrained distribution-matching problem between single-cell and spatial domains. The SpaOTsc method implements structured, unbalanced optimal transport to match single-cell RNA sequencing data to restricted spatial marker genes, mapping single cells into a continuous spatial metric space. SpaOTsc then uses probability distributions of ligand and receptor expression to reconstruct intercellular communication networks, employing random forest models to determine the spatial distance boundaries of localized signaling events. In parallel, the novoSpaRc algorithm operates on the structural correspondence hypothesis, assuming that cells situated near each other exhibit correlated gene expression profiles. NovoSpaRc optimizes an objective function balancing Gromov-Wasserstein discrepancies between cell-cell expression space and physical coordinate space with Euclidean distances to reference marker patterns. Regulated by weighting and entropy regularization hyperparameters, novoSpaRc outputs a probabilistic transport matrix that assigns single cells to physical coordinates deterministically without requiring extensive reference atlases.

The final algorithmic class harnesses deep learning and transformer architectures to integrate spatial transcriptomics with morphological and multimodal data. The Tangram method aligns single-cell or single-nucleus profiles with voxelized spatial coordinates through non-convex optimization, evaluating alignment via Kullback-Leibler divergence for cell density and cosine similarity for gene expression. Tangram incorporates computer vision modules with Siamese neural networks and U-Net architectures to align histological sections directly to anatomical atlases, while also projecting multi-omic chromatin accessibility patterns from SHARE-seq data into spatial space. In contrast, TransformerST provides reference-free super-resolution deconvolution by combining vision transformers with graph transformers, processing histological image patches alongside expression data via multi-head attention and variance-aware training. Across all twenty methods, Saqib and Kim emphasize that method selection involves distinct trade-offs: regression and matrix factorization methods excel in computational speed; probabilistic frameworks deliver rigorous uncertainty estimation and rare-cell detection despite higher resource demands; and graph, transport, and deep learning models effectively incorporate spatial architecture and multimodal images at the cost of higher hardware requirements and reduced interpretability.

Subject of Research: Computational deconvolution methods and algorithmic frameworks for spatial transcriptomics data

Article Title: From pixels to cell types: a comprehensive review of computational methods for spatial transcriptomics deconvolution

Article References: Saqib, J., & Kim, J. (2025). From pixels to cell types: a comprehensive review of computational methods for spatial transcriptomics deconvolution. Genomics & Informatics, 23(1), Article 22. https://doi.org/10.1186/s44342-025-00055-2

Image Credits: AI Generated

DOI: 10.1186/s44342-025-00055-2

Keywords: spatial transcriptomics, computational biology, cell deconvolution, machine learning, probabilistic modeling, optimal transport, deep learning, bioinformatics, pixels, cell, types, comprehensive