Tissue architecture refers to the organized arrangement of cell types, layers, and functional zones within a tissue. In healthy organs this organization is precise, but in diseases such as cancer and Alzheimer's disease the architecture becomes disrupted, producing cytoarchitectural abnormalities that are central to understanding pathology and guiding treatment.
Spatially resolved transcriptomics technologies like the 10x Genomics Visium platform simultaneously measure gene activity at thousands of spatial locations within a tissue slice, preserving the positional context that bulk RNA sequencing discards. This makes it possible to map not just what genes are active, but exactly where in the tissue each activity pattern occurs.
Existing computational tools for spatial transcriptomics analysis, including Seurat, BayesSpace, SpaGCN, and stLearn, can cluster tissue spots into spatial domains but struggle to visualize the resulting tissue heterogeneity intuitively and often miss fine-grained architectural boundaries due to limited spatial representation capacity.
The authors identified a key opportunity: three-dimensional numerical embeddings learned from spatial transcriptomics data can be directly mapped to the red, green, and blue channels of an RGB color image, converting abstract gene expression data into a human-eye-distinguishable visual representation of tissue structure that can then be segmented using standard computer vision techniques.
RESEPT (REconstructing and Segmenting Expression mapped RGB images based on sPatially resolved Transcriptomics) is a deep learning framework with two main stages: first, constructing an RGB image from spatial transcriptomics data; second, segmenting that image to identify tissue architecture using a supervised convolutional neural network.
In the first stage, gene expression data from each tissue spot is compressed into a three-dimensional embedding using a positional variational autoencoder that incorporates the spatial coordinates of each spot alongside gene expression values. This embedding is then refined by a graph autoencoder that models the spatial neighborhood relationships between spots, producing a final three-dimensional representation that captures both gene expression patterns and tissue topology.
The three-dimensional embedding is mapped linearly to the Red, Green, and Blue color channels of an RGB image, where each tissue spot becomes a pixel. Spots with similar gene expression profiles and spatial context receive similar colors, making tissue layer boundaries and cell-type transitions visually apparent. The human eye can distinguish roughly 2.3 million colors, meaning the RGB representation can intuitively encode enormous molecular complexity.
The second stage applies a pre-trained image segmentation deep learning model based on ResNet101, a deep convolutional network with 33 residual blocks, to the RGB image. This computer vision approach identifies boundaries between spatial domains and classifies each spot into a specific tissue segment, enabling systematic identification of tissue architecture across samples.
RESEPT was benchmarked against seven established spatial transcriptomics analysis tools, including Seurat, BayesSpace, SpaGCN, stLearn, STUtility, HMRF, and Giotto, using 16 human brain cortex samples with expert-annotated ground truth tissue layers.
RESEPT achieved a mean Adjusted Rand Index (ARI) of 0.706 across the 16 samples, surpassing all seven comparison tools across all four evaluation metrics: ARI, Rand Index, Fowlkes-Mallows Index, and Adjusted Mutual Information. The ARI measures how closely the predicted tissue architecture matches expert annotations, with 1 being a perfect match.
The framework was tested for robustness under simulated low sequencing depth conditions by progressively reducing the number of sequencing reads. Even at substantially reduced depths, RESEPT maintained stable ARI scores, demonstrating practical resilience to the read depth variation that often occurs across real clinical and research samples.
RESEPT also accepted RNA velocity as an alternative input to gene expression, capturing the dynamic directionality of gene activity rather than just its level. RNA velocity-based images achieved a Moran's I spatial coherence score of 0.920 compared to 0.787 for standard gene expression, suggesting that RNA velocity carries richer spatial structure information for certain tissue types.
To verify that the RGB images encode real biology rather than arbitrary colors, the researchers performed a pixel correlation analysis on 836 spatially variable genes (SVGs) identified in a cortex sample, clustering them into three groups and comparing each group's spatial pattern to the R, G, and B channels.
The Red channel showed a Pearson correlation of 0.726 with one cluster of 60 SVGs enriched in ATP and ribonucleotide metabolic processes, functions associated with cortical layers 4 and 5. The Green channel correlated at 0.916 with 594 SVGs enriched in synaptic vesicle biology reflecting layers 2-4, and the Blue channel correlated at 0.88 with 179 SVGs linked to synaptic transmission and protein targeting to the endoplasmic reticulum.
This channel-to-biology mapping demonstrates that each color in the RESEPT RGB image is not decorative but represents a distinct biological program, making the images directly interpretable. Researchers can read tissue layer identity from the image colors without needing to run separate differential gene expression analyses for each spatial region.
The framework also demonstrated cross-species generalizability by successfully mapping mouse brain cortex architecture after training on human brain data, and performed best among tested tools on a triple-negative human breast cancer sample, pointing toward applicability in diverse cancer tissue types beyond the brain.
Using two postmortem human brain samples from Alzheimer's disease patients profiled with 10x Genomics Visium, RESEPT successfully identified the main cortical architecture of the middle temporal gyrus and distinguished individual cortical layers 2 through 6 using layer-specific marker genes.
RESEPT localized excitatory neuron distribution in cortical layers 2-6 by embedding five well-defined excitatory neuron marker genes. Module scoring confirmed that these markers were statistically significantly enriched in the predicted regions (p-value less than 0.0001), validating the spatial accuracy of the cell-type localization.
Critically, RESEPT was able to identify regions enriched with amyloid-beta plaques, the hallmark pathological lesions of Alzheimer's disease. By embedding a module of 20 upregulated genes associated with amyloid-beta deposition, RESEPT produced RGB images in which plaque-enriched spots displayed consistent colors, distinguishing them from surrounding tissue. This was validated against adjacent-section immunofluorescence staining of amyloid-beta.
The amyloid-beta regions showed statistically tighter color dispersion than non-plaque regions (p-value less than 0.0001 by F-test), confirming that RESEPT RGB images can objectively quantify pathological regional boundaries with precision approaching laboratory staining methods.
Glioblastoma is a grade IV brain cancer with a median survival of only 15 months. It is characterized by extreme tissue heterogeneity, with highly dense tumor regions, areas of necrosis, and zones where individual tumor cells infiltrate the surrounding brain tissue, making accurate spatial mapping critical for prognosis and treatment planning.
Applied to a public glioblastoma sample from the 10x Visium platform, RESEPT identified eight distinct tissue segments and successfully distinguished tumor-enriched regions, non-tumor tissue, and neuropil with infiltrating tumor cells. These mapped regions showed strong correspondence to known histological patterns and secondary structures observed in glioblastoma pathology.
The ability to detect infiltrating tumor cells is particularly significant because these single-cell invasions into surrounding tissue are notoriously difficult to identify by standard H&E staining alone, yet they are critical determinants of surgical resection completeness and patient prognosis.
These results establish that RESEPT is not limited to healthy brain tissue mapping but has direct clinical and prognostic value for characterizing cancer architecture, with implications for guiding surgical boundaries, identifying biopsy targets, and understanding the spatial distribution of treatment-resistant tumor subpopulations.
RESEPT establishes a novel paradigm by framing tissue architecture identification as an image segmentation problem, bringing the full power of modern computer vision to bear on spatial transcriptomics data. This reframing allows RESEPT to leverage large pretrained CNN models and benefits directly from advances in the deep learning computer vision field.
The RGB image output is a key innovation because it is simultaneously a high-quality data visualization, a biologically interpretable molecular map, and a direct input to standard image analysis tools. Each color captures a distinct gene expression program, making the images readable at a glance by researchers and potentially by automated clinical systems.
Demonstrated applications across human brain layers, Alzheimer's disease pathology, mouse brain cortex, human breast cancer, and glioblastoma confirm that RESEPT generalizes across tissues, diseases, and species. As more spatially annotated datasets become available, the segmentation model will continue to improve.
In cancer research specifically, the ability to reliably map the spatial boundaries between tumor and non-tumor tissue, identify infiltrating cell populations, and link spatial patterns to specific gene programs opens new avenues for understanding tumor progression, identifying prognostic biomarkers, and guiding precision surgical and therapeutic interventions.