New Statistical Toolkit TranspaceR Accelerates Next Generation Spatial Transcriptomics Data Analysis

A new computational and statistical package called TranspaceR addresses major analytical bottlenecks in modern imaging-based spatial transcriptomics, enabling researchers to process tissue datasets containing hundreds of thousands to millions of cells in minutes on standard computing hardware. Developed by researchers Fatoumata Mangane and Pierre Bost and published in BMC Methods, the methodology targets the growing volume of multiplexed single-molecule fluorescence in situ hybridization data generated by platforms such as Xenium, CosMx, and MERSCOPE. The authors demonstrate that conventional analytical pipelines adapted directly from single-cell RNA sequencing fail to capture the physical and statistical realities of spatially resolved tissue measurements. By establishing rigorous quality controls, optimized autocorrelation calculations, fast Fourier transform-based variogram computations, and principled data normalization approaches, the package provides a tailored framework for next-generation spatial data.

The motivation for TranspaceR stems from an empirical analysis of historical trends documented in the Museum of Spatial Transcriptomic database through late 2024. The authors tracked how the field has shifted from small exploratory assays to industrial-scale experiments. While early spatial transcriptomics studies published around 2020 rarely tracked more than 1,000 cells per sample and primarily focused on healthy mouse tissues, studies published by 2023 exceeded 150 papers annually, and by 2024 recent datasets routinely averaged over 100,000 cells. Furthermore, commercial imaging instruments accounted for more than one-third of all datasets produced in 2024, with approximately half of the profiled tissues derived from human specimens or non-healthy disease conditions. Despite exponential gains in cell numbers and tissue coverage, commonly used statistical tools designed for single-cell genomics or small-scale spatial assays struggle with memory exhaustion and extended computational runtimes when confronted with contemporary million-cell cohorts.

To ensure high data integrity prior to downstream modeling, TranspaceR establishes structured protocols for filtering cells and lowly expressed genes. The framework implements cell filtering using tissue-specific morphological limits alongside scaled molecular thresholds. Cell radius boundaries are bounded between 2 micrometers and 15 micrometers, informed by reference literature data indicating that lymphoid tissue cells rarely exceed a radius of 10 micrometers while hepatocytes can exceed 15 micrometers. In parallel, the framework adapts the classic single-cell quality threshold of 1,000 unique molecular identifiers by scaling it proportionally to the targeted gene panel size relative to a full transcriptome baseline of 20,000 genes. For removing low-abundance transcripts, the package combines Otsu binarization on log-transformed counts with background estimation derived from negative control probes. Comparing the proportion of negative probes deemed unexpressed by Otsu thresholding serves as a diagnostic metric, where values nearing one verify high experimental quality.

For selecting biologically variable genes without spatial coordinates, the authors formulate parametric models grounded in the negative binomial distribution, avoiding arbitrary thresholds or unnecessary zero-inflated frameworks. Assuming a shared over-dispersion parameter across genes, TranspaceR estimates variance scaling through linear modeling and computes statistical significance for excess variance via asymptotic quantile sampling and Benjamini-Hochberg adjustments. A parallel test identifies genes with an excess proportion of zeros. When applied to a Xenium human lymph node sample, these methods effectively isolated established lineage markers, including MKI67 for cell division, GNLY for cytotoxic T cells, MZB1 for plasma cells, and LYVE1 for lymphatic endothelial cells. Concordant marker recovery occurred across CosMx prefrontal cortex and MERFISH human tonsil samples, detecting canonical markers for neurons, astrocytes, oligodendrocytes, and immune subtypes.

To evaluate spatial expression patterns at scale, the framework reformulates Geary’s C spatial autocorrelation index into a sparse matrix multiplication problem paired with Delaunay triangulation. This algebraic restructuring circumvents traditional computational bottlenecks on large spatial adjacency graphs. In a comparative benchmark on a 10,000-cell subset of the Xenium lymph node dataset against six existing algorithms—SpatialDE, SPARK, SPARK-X, MERINGUE, scBSP, and Squidpy’s Moran’s I—only TranspaceR’s Geary’s C, SPARK-X, and scBSP executed in less than one second without excessive memory overhead. However, scBSP detected only two spatially variable genes (MS4A1 and TRAC), exhibiting excessive stringency. SPARK-X detected 350 significant genes, including false-positive hits among negative control probes and a strong confounding correlation of 0.65 between mean expression and statistical significance. In contrast, TranspaceR identified 198 spatially variable genes with zero false-positive negative control probes and an expression-significance correlation of only 0.10.

Beyond proximity metrics, TranspaceR introduces a fast Fourier transform approach to compute empirical variograms for spatial transcriptomics data, enabling quantitative modeling of spatial correlation ranges. By rasterizing irregular cell coordinates into regular grid matrices and applying Fourier convolutions, pairwise distance computations are executed rapidly across large tissue regions. In performance tests across 10,000 cells, TranspaceR calculated empirical variograms for 436 genes in 4 seconds, compared to more than 2 minutes for the gstat package and 35 minutes for Seurat. When scaled to 300,000 cells, TranspaceR finished within 13 seconds, while gstat required over 8 hours. Variograms produced by the Fourier method achieved a median correlation of 0.74 against standard variograms and demonstrated spatial rotation invariance with a 0.95 median correlation under 45-degree rotations. Furthermore, non-linear model fitting categorized genes into flat, exponential, and fine-tuned exponential profiles, uncovering specific autocorrelation distances such as a 153 to 197 micrometer range for the B-cell marker CD20 in both Xenium and Visium HD data.

The study also challenges the widespread assumption that dividing gene expression counts by segmented cell volume serves as an effective normalization technique. Across four separate Xenium and CosMx datasets, the researchers observed minimal correlation between cell volume and individual gene expression. Mathematical modeling and gamma-distributed volume simulations revealed that inter-cellular variance is predominantly driven by Poissonian sampling noise rather than cell size fluctuations, except for extremely high-abundance transcripts. While dividing counts by cell volume adjusted mean expression values, it introduced artificial distortions into variance, skewness, and kurtosis. In differential expression benchmarks, applying rank-sum tests to volume-normalized data yielded an inflated false positive rate that escalated with sample size when volume distributions varied slightly. Conversely, generalized linear models using a Poisson framework with cell volume included as an explicit covariate maintained a well-controlled false positive rate and provided superior statistical power.

Investigating low-dimensional embeddings for cell clustering, the authors evaluated Singular Value Decomposition, Non-negative Matrix Factorization, Poisson NMF, and Poisson generalized principal component analysis across synthetic and experimental samples. When evaluated on the human lymph node dataset alongside PanglaoDB gene set annotations, SVD and NMF applied to unnormalized counts resolved between 21 and 23 biologically coherent cell types and produced the lowest incompatible expression scores, which quantify impossible biological co-expression. In contrast, Poisson NMF and glmPCA identified fewer cell types and yielded higher incompatible expression scores, while standard volume normalization tended to increase clustering errors.

Finally, the authors examined automated cell type annotation by benchmarking the deep-learning model Scimilarity against the penalized logistic regression framework CellTypist across Xenium and CosMx datasets. While both tools identified major immune compartments in the lymph node, CellTypist misclassified naive B cells within follicular centers and assigned an artificially high 14.9% of cells as regulatory T cells, of which fewer than 20% expressed the canonical transcription factor FOXP3. Scimilarity identified follicular cells as plasma cells and achieved over 80% FOXP3 expression among annotated regulatory T cells. In the human prefrontal cortex dataset, Scimilarity accurately demarcated cortical neuronal layers, assigning over 20% of cells to neuronal populations, whereas CellTypist identified only 2.6% neurons and over-represented oligodendrocytes at 70%. In computational benchmarks, Scimilarity demonstrated superior scaling speeds once dataset sizes surpassed 400,000 cells.

Despite its computational and statistical strengths, the authors emphasize several defined limitations in the current TranspaceR pipeline. The spatial metrics currently evaluate genes independently, leaving bivariate spatial correlations—such as the structural boundary formed between CD3 T cells and CD20 B cells—unmodeled at scale. Additionally, because the package relies on Poisson and negative binomial transcript distribution assumptions and assumes cell state is governed by RNA concentration rather than absolute count, its mathematical models cannot be directly transferred to antibody-based protein imaging methods such as CODEX, cyclic immunofluorescence, or imaging mass cytometry. Future developments will need to extend scalable spatial dependency tests across gene pairs and adapt theoretical protein distribution models to multiplexed protein imaging platforms.

Subject of Research: Computational methodology and statistical frameworks for high-resolution spatial transcriptomics data analysis

Article Title: Next generation statistical framework for next generation spatial transcriptomics data

Article References: Mangane, F., & Bost, P. (2026). Next generation statistical framework for next generation spatial transcriptomics data. BMC Methods, 3(1), Article 7. https://doi.org/10.1186/s44330-026-00055-w

Image Credits: AI Generated

DOI: 10.1186/s44330-026-00055-w

Keywords: Spatial Transcriptomics, Computational Biology, Bioinformatics, TranspaceR, Gene Expression, Cell Segmentation, Statistical Modeling, Biomedical Data Science, Next, generation, statistical, framework