Integrative Biomedical Research

1. Introduction

There is a particular kind of frustration in molecular pathology that anyone who has worked with bulk tissue samples will recognize: you sequence an entire tumor biopsy, get back a single, tidy expression profile, and then have to remind yourself that this number is, in a sense, a fiction — an average smeared across thousands of cells that were never actually behaving the same way (Khoury et al., 2025; Lee, 2025; Qiao et al., 2025). Precision medicine, almost by definition, needs something more honest than that average. It needs to resolve the actual cellular and molecular state of an individual patient’s disease closely enough to act on it (Khoury et al., 2025; Lee, 2025; Qiao et al., 2025). Next-generation sequencing has done remarkable work identifying oncogenic drivers and therapeutic vulnerabilities over the past two decades (Qiao et al., 2025; Sweileh, 2026), but it has mostly done so from that same bulk, homogenized starting point — and homogenization has a cost. It masks cellular heterogeneity, dilutes the signal from rare but clinically decisive subpopulations such as cancer stem cells or pre-metastatic clones, and obscures cell-type-specific changes that may be exactly what matters for a given patient (Lee, 2025; Qiao et al., 2025). Because disease progression and drug resistance are so often driven by small, minority cell states rather than the bulk average, this limitation is not a minor technical footnote; it genuinely caps how far bulk diagnostics can take precision oncology (Lee, 2025; Nesari et al., 2026).

Single-cell RNA sequencing (scRNA-seq) was, in many ways, the field’s answer to this problem. Droplet-based platforms — the 10x Genomics Chromium system being the most familiar example, with its microfluidic partitioning and Gel Bead-in-Emulsion chemistry — made it possible to profile thousands of individual cells in parallel, with reasonably high capture efficiency and manageable doublet rates (Luo et al., 2025; Nesari et al., 2026). This unlocked something genuinely new: international consortia began assembling healthy and disease reference atlases, tracing lineage trajectories, and identifying specific pathogenic cell states — exhausted T cells, chemoresistant tumor subclones — that bulk sequencing had simply never been able to see (Nesari et al., 2026; Qiao et al., 2025).

But scRNA-seq bought that resolution at a real cost, and it is worth being honest about what that cost actually is. Isolating single cells requires enzymatic and mechanical tissue dissociation, and dissociation, almost by necessity, destroys the spatial context those cells previously occupied (Lee, 2025; Qiao et al., 2025; Sweileh, 2026). This matters more than it might sound like it should. Where a cell sits, and which neighbors it happens to be touching, governs a great deal of tissue biology — local signaling gradients, physical proximity effects, the coordinated behavior of cellular neighborhoods that keep tissue either healthy or push it toward disease (Olokede et al., 2026; Qiao et al., 2025). Dissociation erases all of that information in a single step. It also, somewhat perversely, introduces its own artifacts: the physical and enzymatic stress of the process can trigger artificial transcriptional stress responses, meaning some of what looks like biology in a dissociated dataset may actually just be the memory of the dissociation itself (Lee, 2025; Qiao et al., 2025).

Spatial transcriptomics (ST) emerged, more or less directly, as an answer to this second problem. Rather than pulling cells out of their tissue to sequence them, ST maps gene expression directly onto intact tissue sections — marrying conventional histological imaging with barcoded capture arrays or multiplexed hybridization probes (Sweileh, 2026; Weiderman et al., 2025). Because it preserves native spatial organization, ST lets researchers and, potentially, clinicians visualize tumor-stroma boundaries, trace ligand-receptor signaling gradients across real physical distances, and dissect the coordinated cellular niches that drive immune evasion or therapeutic resistance (Olokede et al., 2026; Sweileh, 2026; Weiderman et al., 2025).

And yet, for all that biological richness, ST has had almost no impact on routine clinical diagnostics so far (Olokede et al., 2026; Qiao et al., 2025). It remains, largely, an academic discovery and pharmaceutical target-identification tool — which is a strange place for a technology this powerful to be stuck (Olokede et al., 2026; Qiao et al., 2025). The reasons are not mysterious, even if they are stubborn. Cost is one: high-resolution in situ sequencing instruments can run past $350,000, with per-sample consumable costs around $1,000 — numbers that are simply incompatible with the cost-sensitive, high-volume reality of a clinical pathology lab (Olokede et al., 2026; Weiderman et al., 2025). Dimensionality is another: ST typically profiles hundreds to thousands of genes at once, while pathology labs are built around small, robust, easily interpretable panels of a handful of markers, not gigabyte-scale genomic matrices (Olokede et al., 2026). Sample compatibility is a third and perhaps the most practically stubborn: most clinical biospecimens exist as formalin-fixed, paraffin-embedded (FFPE) blocks, and formalin fixation, however useful for preservation, causes RNA fragmentation and chemical modification that whole-transcriptome ST methods — largely optimized for fresh-frozen tissue — were never really designed to handle (Olokede et al., 2026; Weiderman et al., 2025). And underlying all of this is a simple lack of standardization: no consensus SOPs yet exist for sample preparation, panel selection, normalization, computational deconvolution, or regulatory clearance, leaving the field technically fragmented and vulnerable to batch effects (Olokede et al., 2026; Weiderman et al., 2025).

What is beginning to change this picture — and this is really the throughline of what follows — is a shift in how the field thinks about what ST is actually for. Rather than treating high-dimensional spatial transcriptomics as an endpoint clinical assay, researchers have begun reframing it as an upstream discovery engine (Olokede et al., 2026). The central insight behind this “compression-to-clinic” paradigm is that spatial-omics data are, in a sense, biologically redundant: many genes carry overlapping signal because they mark the same cell states, pathways, or localized niches (Olokede et al., 2026). Using biologically informed feature reduction, stability-aware machine learning tools such as Stabl, and deep learning deconvolution methods such as cell2location, it becomes possible to compress these sprawling datasets into minimal, information-efficient panels — typically just five to ten markers — that retain the essential spatial signal of tumor microenvironment architecture while being fully convertible into standard, low-cost pathology assays such as immunohistochemistry or RNA in situ hybridization, applied directly to archived FFPE tissue (Lee, 2025; Olokede et al., 2026).

This review takes stock of that trajectory. It traces the technological evolution from single-cell to spatially resolved profiling, examines the computational machinery — deconvolution algorithms, graph neural networks, transformer-based foundation models — that has grown up around it, catalogs where spatial biology has already demonstrated clear clinical relevance in oncology, and asks, as concretely as the current evidence allows, what a realistic path from molecular microscope to bedside diagnostic actually looks like.

2. Spatially Resolved Omics Technologies and Clinical Biomarker Translation

2.1. Beyond Bulk Averages: Toward Spatially Resolved Tissue Ecosystems

Tissues, whatever else they are, are organized things — structured arrangements of diverse cell types and extracellular matrix rather than simple bags of similar cells (Choe et al., 2023; Weiderman et al., 2025). Physiological homeostasis and pathological progression both emerge from coordinated multicellular networks and short-range paracrine signaling, not from isolated cellular decisions made in a vacuum (Olokede et al., 2026; Wang et al., 2024). For most of its history, clinical diagnostics leaned on bulk transcriptomic profiling to characterize these molecular states (Nesari et al., 2026; Lee, 2025), and while that approach genuinely advanced disease subclassification, it did so by averaging expression across every cell type present in a tissue homogenate — masking heterogeneity, missing rare but pivotal subpopulations such as circulating tumor cells, and losing spatial context entirely (Choe et al., 2023; Luo et al., 2025).

Single-cell RNA sequencing addressed the heterogeneity half of that problem directly, sequencing thousands of individual cells to map distinct states and transitional dynamics, and in doing so reshaped how tumor microenvironments and developmental biology are understood (Nesari et al., 2026; Lee, 2025). But its reliance on mechanical and enzymatic dissociation means the spatial half of the problem persists — physical cellular interactions, localized receptor-ligand pairing, and native microenvironmental organization are all lost the moment cells go into suspension (Choe et al., 2023; Olokede et al., 2026; Pentimalli et al., 2025). The dissociation process itself can also induce artificial transcriptional stress responses, introducing noise that can be mistaken for genuine biology (Choe et al., 2023; Lee, 2025).

Spatial transcriptomics resolves this “spatial dissociation paradox” by mapping gene expression directly onto intact sections, functioning as something close to a molecular microscope — one that can chart cellular neighborhoods, signaling pathways, and even host-microbe interactions within native tissue architecture (Mondal et al., 2025; Pentimalli et al., 2025; Weiderman et al., 2025). The technology’s significance was recognized fairly early: Nature Methods named spatially resolved transcriptomics its Method of the Year in 2020 (Zhu et al., 2026).

2.2. The Spatial-Omics Technological Landscape: Resolution versus Breadth

Spatial platforms broadly split into two families, and each carries a fairly fundamental trade-off (Olokede et al., 2026; Pentimalli et al., 2025). Sequencing-based methods capture polyadenylated RNA locally using spatially barcoded capture arrays (Pentimalli et al., 2025; Lee, 2025). The original 10x Genomics Visium platform used 55 µm hexagonal capture spots — enough to aggregate transcripts from thirty to fifty adjacent cells, which is useful but not single-cell resolution (Choe et al., 2023; Weiderman et al., 2025). Later platforms pushed resolution down considerably: Slide-seq and Slide-seqV2 use randomly deposited 10 µm DNA-barcoded beads (Stur et al., 2022; Lee, 2025); high-definition spatial transcriptomics tightens spot spacing further with 2 µm beads in etched microwells (Weiderman et al., 2025); and Stereo-seq, using patterned DNA nanoball chips spaced roughly 500 nm apart, reaches genuinely nanoscale, subcellular resolution (Chen et al., 2022; Weiderman et al., 2025). Open-ST, an open-source approach, achieves 0.6 µm resolution using standard, commercially available sequencing flow cells as the capture surface — a notable move toward democratizing access (Pentimalli et al., 2025). The trade-off with all of these sequencing-based approaches is that transcripts can laterally diffuse during tissue permeabilization, displacing signal and lowering effective capture efficiency (Choe et al., 2023; Lee, 2025).

Imaging-based platforms take the opposite approach, visualizing target transcripts directly within tissue using automated fluorescence microscopy rather than physically extracting them (Pentimalli et al., 2025; Lee, 2025). In situ hybridization methods such as MERFISH and seqFISH+ use combinatorial probe hybridization across multiple imaging and stripping rounds to detect up to ten thousand genes at roughly 500 nm resolution, with high sensitivity and low false-discovery rates (Pentimalli et al., 2025; Lee, 2025). Commercial platforms like Xenium and CosMx SMI use cyclic fluorescent reporter hybridization of padlock probes to achieve comparable subcellular resolution directly on pathology slides (Weiderman et al., 2025; Lee, 2025). The catch, unsurprisingly, is that these are closed systems built around pre-designed, targeted probe panels rather than the whole transcriptome (Weiderman et al., 2025; Lee, 2025).

The field has also grown multimodal. Spatial proteomics platforms such as CODEX and imaging mass cytometry (IMC) use heavy metal- or oligonucleotide-tagged antibodies to map dozens of proteins on the same slide (Weiderman et al., 2025; Lee, 2025), while spatial epigenomics tools like Spatial-ATAC-seq and Spatial-CUT&Tag apply microfluidic barcoding and Tn5 transposition to profile chromatin accessibility and histone modifications in situ — revealing the regulatory logic underlying localized transcriptional programs (Deng et al., 2022; Lee, 2025).

2.3. Computational Analysis: Deconvolution and Spatially Aware Deep Learning

The complexity of spatial data has forced rapid development of specialized bioinformatic tools (Nesari et al., 2026; Weiderman et al., 2025). Because array-based platforms such as Visium or Slide-seq capture multi-cellular mixtures within each spot, computational deconvolution is essentially a prerequisite for meaningful biological interpretation (Zhu et al., 2026; Weiderman et al., 2025). Algorithms like cell2location, RCTD, and SPOTlight use matched scRNA-seq reference datasets and statistical modeling — regularized negative binomial regression, in cell2location’s case — to estimate cell-type proportions and absolute abundances at each spatial coordinate (Zhu et al., 2026; Weiderman et al., 2025).

Cells organize into localized functional microenvironments, or spatial domains, and traditional clustering algorithms — Louvain, Leiden — were never built to account for that, since they operate purely in expression space and discard physical relationships entirely (Zhu et al., 2026; Nesari et al., 2026). Spatially aware alternatives correct for this by explicitly integrating expression with physical coordinates and tissue morphology. SpaGCN builds a weighted graph via deep graph convolutional networks to model spatial neighborhoods in Visium data (Zhu et al., 2026; Weiderman et al., 2025); BayesSpace applies a Bayesian framework with spatial neighborhood priors to reach sub-spot resolution (Zhu et al., 2026; Weiderman et al., 2025); and GraphST, arguably the most sophisticated of the three, pairs graph neural networks with self-supervised contrastive learning — minimizing embedding distance between spatially adjacent spots while maximizing it for distant ones — to perform domain segmentation and batch correction simultaneously (Zhu et al., 2026). Tools like CellChat and SpaTalk extend this logic to signaling, inferring ligand-receptor networks constrained by physical proximity (Zhu et al., 2026; Weiderman et al., 2025).

More recently, the field has begun moving toward foundation models pretrained on atlas-scale single-cell data (Zhu et al., 2026; Nesari et al., 2026). Transformer architectures like scGPT and Nicheformer ingest millions of single-cell transcriptomic vectors alongside physical coordinates to generate highly transferable embeddings, which in turn substantially improve cell annotation accuracy, trajectory inference, and imputation of missing gene values (Zhu et al., 2026).

2.4. Clinical Translational Bottlenecks: Cost, Sample Processing, and FFPE Constraints

Despite the biological richness these tools unlock, moving from academic discovery platform to routine clinical diagnostic remains genuinely restricted (Olokede et al., 2026; Pentimalli et al., 2025). Cost is the first and most obvious obstacle — specialized scanners, microfluidic chips, complex library preparation, and deep sequencing add up quickly, and a single spatial section can run into the thousands of dollars, which is simply not viable for routine testing at scale, particularly in resource-limited health systems (Olokede et al., 2026; Pentimalli et al., 2025).

Sample preservation is the second, and arguably more stubborn, obstacle. Most spatial platforms were originally optimized for fresh-frozen tissue to preserve RNA integrity (Pentimalli et al., 2025; Weiderman et al., 2025), but clinical archives are overwhelmingly FFPE (Olokede et al., 2026; Pentimalli et al., 2025). Formalin fixation causes chemical cross-linking and RNA fragmentation that historically made high-quality spatial sequencing difficult; newer probe-based platforms such as Xenium, Visium CytAssist, and CosMx have been engineered around this constraint, but processing FFPE material for these assays remains technically demanding and highly sensitive to pre-analytical handling (Olokede et al., 2026; Pentimalli et al., 2025).

The third obstacle is simply interpretive load. Pathology labs are built for rapid turnaround and simple, often binary or qualitative readouts — immunohistochemistry scoring, for instance — not gigabyte-scale genomic matrices requiring centralized bioinformatic infrastructure, batch correction, and machine-learning-based interpretation (Olokede et al., 2026; Pentimalli et al., 2025).

2.5. The “Compression-to-Clinic” Paradigm: Engineering Sparse Spatial Biomarkers

Olokede and colleagues (2026) proposed a way through this impasse: treat high-dimensional spatial transcriptomics not as the clinical assay itself, but strictly as an upstream discovery engine that feeds a much simpler downstream product. The logic rests on molecular redundancy — many genes serve as co-expressed markers of the same localized cell state or niche, so a whole-transcriptome signature is rarely necessary to capture the diagnostically relevant signal (Olokede et al., 2026). The goal, instead, is to computationally compress high-dimensional datasets down to the smallest, most information-efficient panel that still captures that signal (Olokede et al., 2026).

A central tool in this workflow is Stable Biomarker Learning (Stabl), a machine learning pipeline that evaluates feature-selection stability under repeated data perturbation and resampling, rather than relying on standard cross-validation approaches that tend to overfit noisy single-cell data (Lee, 2025). By retaining only features that remain stable across independent patient cohorts, Stabl identifies sparse, reproducible, generalizable biomarker sets (Lee, 2025). Within the compression-to-clinic pipeline, this stability-driven reduction typically distills a spatial dataset down to five to ten key markers, chosen for their association with clinically meaningful, spatially localized phenotypes — immune exclusion at a tumor-stroma boundary, for instance, or stromal remodeling at an invasive front (Olokede et al., 2026). Once identified, these markers translate into standard, scalable pathology assays — multi-color immunohistochemistry, multiplexed immunofluorescence, or multiplex RNA in situ hybridization such as RNAscope — that pathologists can integrate directly alongside routine H&E evaluation, without needing specialized spatial sequencing instrumentation or centralized bioinformatics at all (Olokede et al., 2026; Pentimalli et al., 2025).

3. Methods

3.1. Review Design

This article was constructed as a narrative literature synthesis, informed by the reporting structure of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework, without formal PROSPERO registration, to support independent reproducibility of the search-and-selection process. Given the deliberately broad, multi-domain scope of the review — spanning platform engineering, computational methodology, and clinical oncology translation — a narrative rather than fully systematic meta-analytic design was chosen, consistent with prior high-impact narrative syntheses in this space (Olokede et al., 2026; Pentimalli et al., 2025).

3.2. Data Sources and Search Strategy

Literature was identified through structured searches of PubMed/MEDLINE, Scopus, and Web of Science, supplemented by manual screening of recent tables of contents from Nature Methods, Nature Biotechnology, Nature Genetics, Cell, and Briefings in Bioinformatics, given the concentration of high-impact spatial-omics methodology in these venues. Search terms combined the core concept “spatial transcriptomics” (and synonyms: spatially resolved transcriptomics, spatial-omics) and “single-cell RNA sequencing” with domain-specific secondary terms: (i) platform engineering (e.g., “in situ hybridization,” “in situ sequencing,” “barcoded capture array,” “FFPE-compatible”); (ii) computational methods (e.g., “deconvolution,” “graph neural network,” “spatial domain detection,” “foundation model,” “transformer”); and (iii) clinical translation (e.g., “tumor microenvironment,” “biomarker panel,” “immunohistochemistry,” “clinical validation,” “regulatory clearance”). Searches were restricted to records published between January 2016 (coinciding with the first description of array-based spatial transcriptomics; Ståhl et al., 2016) and February 2026, with foundational methodology papers retained outside this window where they establish a platform or algorithm still in active clinical or research use.

3.3. Eligibility Criteria

Records were eligible if they (a) were published in English in a peer-reviewed journal or a recognized preprint server for very recent methodology; (b) reported primary platform development, computational methodology, or clinical/translational application of single-cell or spatial transcriptomics; and (c) provided sufficient detail to extract at least one of: technical platform specification, computational algorithm and validation metric, or clinical study design and outcome. Conference abstracts without full-text availability and opinion pieces lacking a primary data or reference basis were excluded. Where multiple publications described the same platform or algorithm, the primary methodological reference was prioritized for technical extraction, and the most clinically detailed source was prioritized for outcome extraction.

3.4. Study Selection and Data Extraction

Screening proceeded in two stages: title/abstract screening for topical relevance, followed by full-text review of records meeting initial criteria. Data extraction was organized around the four domains structuring this review’s synthesis: (1) spatial transcriptomics platforms, extracted with resolution, genomic capacity, specimen compatibility, and detection chemistry; (2) computational and analytical software tools, extracted with algorithmic approach, input modality, and reported performance or limitation; (3) deep learning and foundation-model architectures, extracted with model paradigm, validation metric, and reported biological or clinical insight; and (4) clinical oncology translation studies, extracted with cancer type, specimen preparation, cohort size, spatially resolved findings, and identified biomarkers.

3.5. Synthesis Approach

Because performance metrics vary substantially across studies in units, reference standards, and validation cohorts, quantitative pooling via meta-analysis was not undertaken; results were synthesized narratively and organized into comparative tables (Tables 1–4) and summary figures (Figures 1–2). Quantitative metrics (e.g., resolution in micrometers, reported accuracy or AUC) are reproduced as originally published.

3.6. Reproducibility Statement

The search string architecture, database coverage, and inclusion/exclusion logic described above are provided in sufficient detail to permit independent replication; the complete reference list (Section 7) documents every source contributing extracted data to Tables 1–4 and the narrative synthesis.

4. High-Dimensional Landscape and the Translation Path of Spatial Transcriptomics

4.1. Technical Benchmarking of Spatial Platforms: Navigating the Resolution-Breadth Trade-Off

Comparing commercial and emerging spatial platforms side by side (Table 1; Figure 1) makes the field’s central engineering trade-off unmistakable: resolution and transcriptomic breadth pull in opposite directions, and no platform currently offers both at once (Sweileh, 2026). Planar barcoded array sequencing platforms — 10x Genomics Visium, Curio Seeker, Slide-seqV2, BGI’s Stereo-seq — enable unbiased, genome-wide discovery but have historically faced resolution ceilings from lateral transcript diffusion during permeabilization (Lee, 2025). Standard Visium arrays, at 55 µm capture spots, aggregate transcripts from dozens of adjacent cells (Cable et al., 2022; Ståhl et al., 2016), while more recent microfluidic approaches such as DBiT-seq reach 10–50 µm resolution and add multi-omic co-profiling capability, at the cost of requiring customized microfluidic chips. Newer platforms have pushed physical resolution limits substantially further: Visium HD reaches 2 µm bin resolution, and Stereo-seq, using DNA nanoball-patterned chips, achieves roughly 220 nm resolution with 500 nm center-to-center spacing — genuinely subcellular mapping (Chen et al., 2022) — a span visualized directly in Figure 1, which plots resolution across nine representative platforms on a logarithmic scale.

Imaging-anchored platforms — Vizgen MERSCOPE, 10x Genomics Xenium, NanoString CosMx SMI — take the opposite strategy, bypassing sequencing entirely to

Figure 1. Spatial resolution benchmarking of representative spatial transcriptomics platforms. Bar chart (logarithmic scale) comparing the reported spatial resolution, in micrometers, of nine representative sequencing-based and imaging-based platforms, spanning Visium’s 55 µm capture spots down to Stereo-seq’s 220 nm nanoball spacing. The figure visually reinforces the resolution-versus-breadth trade-off discussed narratively in Section 4.1 and Table 1, clarifying why platform choice should be matched to discovery versus validation use cases (Chen et al., 2022; Lee, 2025).

Figure 2. Reported performance of representative deep learning and foundation-model architectures in single-cell and spatial omics. Bar chart summarizing reported validation performance (accuracy, AUC, or equivalent metric, expressed as a percentage) for four representative computational models, each verified against its primary-literature source: spatial domain detection (SpaGCN), cell-type annotation (scGPT), multi-sample domain alignment (GraphST), and cross-dataset transfer performance (Nicheformer). The figure demonstrates that current spatially aware and foundation-model architectures consistently exceed 85–90% validation performance across diverse computational tasks, underscoring their suitability as upstream discovery engines within the compression-to-clinic paradigm (Cui et al., 2024; Hu et al., 2021).

achieve single-molecule, subcellular resolution below 1 µm, which makes them well suited to precise spatial cell-typing (Janesick et al., 2023). Their limitation is target capacity: measurements are restricted to predefined panels of several hundred to five thousand genes (Janesick et al., 2023; Lee, 2025). This divergence has a fairly clean practical implication: whole-transcriptome sequencing-based platforms suit upstream discovery work, while targeted imaging-based platforms suit diagnostic validation and eventual clinical deployment (Pentimalli et al., 2023).

Clinical translation adds a further constraint that Table 1 makes explicit: compatibility with archival FFPE tissue, which dominates clinical biobanks (Janesick et al., 2023). Fresh-frozen sections preserve RNA integrity best, but managing FF material in a routine clinical workflow is impractical (Janesick et al., 2023; Lee, 2025). Platforms such as Xenium, CosMx, and Visium FFPE have been specifically adapted to read degraded RNA from archival blocks via probe-based hybridization, which meaningfully expands their clinical feasibility relative to earlier FF-only methods (Janesick et al., 2023; Villacampa et al., 2021).

4.2. The Computational and Analytical Ecosystem: Deconvolution and Spatial Cellular Neighborhoods

As spatial datasets have scaled, Seurat and Scanpy have remained the dominant general-purpose toolkits (Table 2), with Scanpy offering substantially greater scalability — five- to ninety-fold faster processing on datasets exceeding one million cells (Nesari et al., 2026; Wolf et al., 2018). Because multicellular-resolution platforms like Visium mix transcriptomes within a single spot, deconvolution represents a genuine computational bottleneck (Long et al., 2023; Palla et al., 2022). Probabilistic algorithms such as cell2location, RCTD, SPOTlight, and SpatialDWLS address this by integrating matched scRNA-seq reference signatures to estimate cell-type proportions within each spot (Cable et al., 2022; Kleshchevnikov et al., 2022); cell2location specifically applies a regularized negative binomial regression model to build fine-grained reference profiles (Kleshchevnikov et al., 2022; Long et al., 2023).

Beyond deconvolution, the toolbox has shifted toward modeling spatially constrained cell-cell communication. Tools including CellChat, COMMOT, SpaTalk, and MISTy map localized receptor-ligand pairing across adjacent neighborhoods, effectively recovering the physical context that dissociation-based scRNA-seq discards (Dries et al., 2021; Shao et al., 2022). This has direct clinical relevance: whether an immune cell sits physically adjacent to a malignant cell, or is instead excluded behind a fibrotic border, can determine immunotherapy response (Schürch et al., 2020).

4.3. The Ascendancy of Deep Learning and Artificial Intelligence Foundation Models

A clear pattern across the synthesized literature is the rapid absorption of deep representation learning into standard spatial bioinformatics pipelines (Table 3; Figure 2). Convolutional and graph neural networks now treat tissue slides essentially as mathematical coordinate networks (Hu et al., 2021). SpaGCN, for instance, fuses gene expression, physical coordinates, and H&E histological features via graph convolution to map coherent biological domains and locally variable genes (Hu et al., 2021), and related architectures such as STAGATE and GraphST apply self-supervised contrastive learning and graph attention autoencoders to refine domain boundaries and correct batch effects across multi-sample datasets (Dong & Zhang, 2022; Long et al., 2023).

More recently, the field has moved toward atlas-scale, generative pretrained “foundation models.” Architectures like scGPT, Geneformer, and Nicheformer, trained on tens of millions of single-cell profiles, learn transferable molecular representations (Cui et al., 2024; Tejada-Lapuerta et al., 2025). scGPT specifically, pretrained on more than 33 million single-cell profiles, achieves over 94% zero-shot cell-type annotation accuracy (Cui et al., 2024; Nesari et al., 2026) — one of several performance benchmarks summarized in Figure 2 alongside SpaGCN’s domain-detection accuracy and GraphST’s and Nicheformer’s domain-alignment and cross-dataset transfer performance.

Deploying these architectures directly inside a clinical pathology laboratory, however, remains constrained by cost and turnaround requirements (Olokede et al., 2026). This is precisely the gap the compression-to-clinic paradigm is built to close: using foundation models and stability-aware feature selection tools like Stabl strictly as upstream discovery engines, then compressing the resulting high-dimensional signal into panels of five to ten markers convertible into standard immunohistochemistry or RNA-ISH assays fully compatible with routine FFPE workflows (Hédou et al., 2024; Olokede et al., 2026).

Table 1. Technical and Performance Comparison of Representative Spatial Transcriptomics (ST) Platforms. This table compares nine major commercial and open-source spatially resolved transcriptomic platforms, detailing technical resolution, genomic capacity, chemical principle, and specimen requirements. The comparison illustrates a consistent inverse relationship between spatial resolution and transcriptomic breadth across sequencing-based and imaging-based platform classes (Lee, 2025; Weiderman et al., 2025).

Table 2. Analytical and Computational Software Toolbox for Single-Cell and Spatial Transcriptomics. This table catalogues ten leading bioinformatic software tools used to normalize, cluster, deconvolve, and interpret single-cell and spatial data, specifying algorithmic approach, input modality, programming environment, and known limitations. The table shows a clear progression from general-purpose single-cell tools toward spatially aware, stability-driven methods purpose-built for translational biomarker discovery (Zhu et al., 2026; Hédou et al., 2024).

4.4. Clinical Oncology Translation: Delineating the Tumor-Immune Architecture in Solid Tumors

The clinical value of this technology stack is most concretely demonstrated in oncology (Table 4), where spatial context reveals tumor microenvironment architecture that bulk or dissociated single-cell data simply cannot resolve (Walsh & Quail, 2023). In breast cancer, spatial profiling across molecular subtypes has identified nine distinct spatial archetypes, mapping hypoxic cores, invasive margins, and tumor-stroma interfaces (Croizer et al., 2024; Wang et al., 2024). In triple-negative breast cancer specifically, spatial transcriptomics has mapped immunosuppressive niches formed by TREM2-positive tumor-associated macrophages sitting adjacent to PD-1-positive exhausted lymphocytes, a spatial configuration associated with poorer overall survival (Croizer et al., 2024; Wu et al., 2021).

In colorectal cancer, spatial-omics has helped resolve a longstanding classification problem: bulk transcriptomic profiling frequently misclassifies specimens under the Consensus Molecular Subtypes (CMS) system because stromal content confounds the signal, whereas spatially resolved sequencing shows that distinct microenvironmental neighborhoods within a single tumor can map to genuinely different CMS classes (Lee, 2025; Xiao et al., 2024). Fibrotic, immune-excluded border zones tend to carry CMS4-like signal directly adjacent to classical epithelial islands aligning with CMS2 or CMS3 — a spatial juxtaposition invisible to bulk sequencing and directly relevant to preventing diagnostic misclassification (Feng et al., 2024; Xiao et al., 2024).

In pancreatic ductal adenocarcinoma, spatial profiling using Stereo-seq and CosMx has decoded the classical-to-basal lineage spectrum, identifying transitional epithelial populations with hybrid classical-basal phenotypes; the more aggressive, drug-resistant basal-like states appear consistently associated with severe tissue hypoxia and a “fibroblast-high” surrounding stroma (Hwang et al., 2022; Khaliq et al., 2024). Within these basal niches, spatial cell-cell interaction analysis has mapped elevated macrophage migration inhibitory factor driving M2-like macrophage polarization, alongside CXCR4-CXCL12 signaling axes that actively exclude cytotoxic CD8-positive T lymphocytes — a fairly clear molecular map of localized immune evasion (Chen et al., 2024; Hwang et al., 2022).

5. Discussion

5.1. A Technology That Has Proven Its Biology, Not Yet Its Workflow

Read together, the platform, computational, and clinical-translation data assembled here (Tables 1–4; Figures 1–2) point to a field that has convincingly answered its central biological question — does spatial context matter clinically? — while leaving its operational question largely open. The oncology findings in Section 4.4 are not marginal; spatially resolved tumor-immune architecture predicts outcomes and reclassifies tumors in ways bulk and dissociated single-cell data cannot (Croizer et al., 2024; Xiao et al., 2024). What remains unresolved is how that biology gets delivered reliably, cheaply, and reproducibly to an actual pathology bench.

5.2. The Resolution-Breadth Trade-Off Is Not Going Away Soon

Figure 1 and Table 1 make a point worth dwelling on: despite a decade of platform innovation, the resolution-versus-breadth trade-off has not been engineered away, only shifted. Stereo-seq and Visium HD have pushed sequencing-based resolution into genuinely subcellular territory (Chen et al., 2022), and imaging-based platforms like Xenium and CosMx now reach comparable resolution from the opposite direction, at the cost of panel size (Janesick et al., 2023). For clinical purposes, this trade-off is arguably a feature rather than a bug: whole-transcriptome discovery platforms and targeted, FFPE-compatible imaging platforms serve genuinely different roles across the diagnostic pipeline, and the compression-to-clinic paradigm essentially formalizes that division of labor (Olokede et al., 2026; Pentimalli et al., 2023).

5.3. Computational Sophistication Has Outpaced Clinical Infrastructure

The progression documented in Table 2, Table 3, and Figure 2 — from Louvain clustering, to spatially aware graph neural networks, to atlas-scale transformer foundation models achieving over 90% annotation accuracy — represents genuinely fast-moving methodological progress (Cui et al., 2024; Zhu et al., 2026). But this progress has, if anything, widened rather than narrowed the gap with clinical infrastructure. A pathology laboratory built around rapid, qualitative IHC scoring has no obvious place to run a GPU-intensive transformer model, regardless of how accurate that model is (Olokede et al., 2026). This is precisely why the compression step matters as much as the discovery step:

Table 3. Deep Learning Architectures and Machine Learning Milestones in Single-Cell and Spatial Omics. This table synthesizes four advanced deep learning models developed to resolve data sparsity, model spatial neighborhoods, and generate clinically relevant predictions, specifying architecture, algorithmic paradigm, target biological insight, and reported validation performance, each verified directly against its primary-literature source. Three additional models (scDrug, CellOT, GeneFormer) originally included from a secondary survey's case-study summaries were removed after their specific reported findings could not be confirmed against the primary literature (Cui et al., 2024; Hu et al., 2021).

Table 4. Clinical Translation and Oncology Applications of Near Single-Cell Spatial Transcriptomics. This table summarizes seven clinical studies applying single-cell and spatial transcriptomics to decode the tumor-immune microenvironment, identify prognostic biomarkers, and predict therapeutic response across breast, colorectal, pancreatic, hepatocellular, endometrial, and glial malignancies. The hepatocellular carcinoma row has been verified against its primary source (Wu et al., 2023). The endometrial and glioma rows remain unverified and are marked for author action prior to submission (Andersson et al., 2021; Hwang et al., 2022; Xiao et al., 2024).

a foundation model’s value to clinical medicine, on the evidence assembled here, lies almost entirely in what it can distill down to a five- or ten-marker panel, not in what it can compute at inference time inside a hospital.

5.4. Oncology Case Studies Suggest Where Compression Will Work Best

The clinical translation studies in Table 4 — breast, colorectal, hepatocellular, and pancreatic cancer — share a common structure worth noting: in each case, the clinically decisive information is a spatial relationship (macrophage-lymphocyte adjacency in TNBC; stromal-epithelial boundary structure in CRC; basal-niche hypoxia and stromal density in PDAC) rather than a long gene list (Croizer et al., 2024; Hwang et al., 2022; Xiao et al., 2024). This is encouraging for the compression-to-clinic approach specifically, because relationships of this kind are often reducible to a handful of well-chosen markers scored by adjacency or co-localization on a standard multiplex IHC panel, rather than requiring whole-transcriptome resolution at the point of clinical testing (Olokede et al., 2026).

5.5. Toward a Realistic Roadmap

Taken together, Sections 4.1 through 4.4 suggest that the path to clinical adoption runs through standardization more than invention. The individual pieces — high-resolution discovery platforms, FFPE-compatible targeted assays, stability-aware compression algorithms, and oncology evidence that spatial architecture carries independent prognostic value — already exist (Hédou et al., 2024; Olokede et al., 2026). What is missing is a validated, regulator-facing pipeline connecting them: standardized SOPs for panel selection and FFPE processing, multi-center validation of compressed panels against existing diagnostic standards, and a clear CLIA/FDA clearance pathway of the kind Olokede and colleagues (2026) have begun to outline as a fourth study objective.

5.6. Limitations

As a narrative rather than fully systematic review, source selection here was structured but not exhaustive, and reported performance metrics (resolution figures, accuracy percentages, AUC values) were reproduced from their original sources rather than independently verified or meta-analytically pooled. Several cited foundation-model and compression-pipeline studies remain early-stage, and some clinical oncology findings summarized in Table 4 derive from single-center cohorts awaiting multi-center replication.

6. Conclusion

Single-cell and spatial transcriptomics have fundamentally changed how tissue biology is understood, replacing the blunt average of bulk sequencing with a genuinely cellular, and now spatially situated, view of disease. The clinical case for this resolution is no longer really in question — spatially resolved tumor-immune architecture already predicts outcomes that bulk and dissociated data cannot. What stands between this evidence base and routine diagnostic use is not further biological discovery but translational engineering: standardized, FFPE-compatible, cost-effective assay conversion. The compression-to-clinic paradigm, which treats high-dimensional spatial-omics strictly as an upstream discovery engine feeding sparse, validated marker panels, offers the most credible route from molecular microscope to bedside test within the next several years.

Author Contributions

M.R.S. contributed to the conception and design of the review, literature search, analysis and synthesis of the relevant evidence, and drafting of the manuscript. H.Y.K. contributed to the literature search, interpretation of the findings, and critical revision of the manuscript. Both authors reviewed and approved the final version of the manuscript and agreed to be accountable for all aspects of the work.

Acknowledgements

The authors would like to acknowledge the Department of Biology, Faculty of Medicine, Universitas Baiturrahmah, Padang, Indonesia, and the Faculty of Applied Sciences, Universiti Teknologi MARA, Sarawak Branch, Malaysia, for their academic and institutional support. The authors also acknowledge the researchers whose published studies contributed to the scientific foundation of this review.

References