Abstract
Spatial transcriptomics links gene expression with tissue morphology,
however, current tools often prioritize genomic analysis, lacking
integrated image interpretation. To address this, we present Thor, a
comprehensive platform for cell-level analysis of spatial
transcriptomics and histological images. Thor employs an anti-shrinking
Markov diffusion method to infer single-cell spatial transcriptome from
spot-level data, effectively combining gene expression and cell
morphology. The platform includes 10 modular tools for genomic and
image-based analysis, and is paired with Mjolnir, a web-based interface
for interactive exploration of gigapixel images. Thor is validated on
simulated data and multiple spatial platforms (ISH, MERFISH, Xenium,
Stereo-seq). Thor characterizes regenerative signatures in heart
failure, screens breast cancer hallmarks, resolves fine layers in mouse
olfactory bulb, and annotates fibrotic heart tissue. In high-resolution
Visium HD data, it enhances spatial gene patterns aligned with
histology. By bridging transcriptomic and histological analysis, Thor
enables holistic tissue interpretation in spatial biology.
Subject terms: Transcriptomics, Computational platforms and
environments, Software, Bioinformatics, Tumour heterogeneity
__________________________________________________________________
Zhang, Chen, and colleagues present Thor, a platform that turns
spot-level spatial transcriptomics into single-cell gene maps using the
paired histology image, without using single-cell RNA-seq data. Thor
unveils fine tissue architectures, and it expands our knowledge on
fibrosis and vascular-regenerative gene expression.
Introduction
The complex organization of cells within tissues is profoundly
connected to their biological function. This underpins the widespread
utility of histological images in health and disease. The development
of computational methods empowered by deep learning on histological
images has drastically enhanced efficiency and accuracy in tissue
analysis in diverse applications^[59]1, including automated cancer
diagnosis^[60]2, survival prediction^[61]3, histopathology image
classification and retrieval^[62]4, tissue segmentation^[63]5,[64]6,
nucleus and cell segmentation^[65]7–[66]9, and in silico
staining^[67]10. Furthermore, rapid advancements in high-throughput
technologies such as RNA sequencing (RNA-seq) and whole genome
sequencing (WGS) are transforming the landscape of conventional
histological analysis, offering unprecedented insights beyond tissue
images. For example, recent research has demonstrated that the
integration of histological images with genomic biomarker mutations and
biological pathways leads to accurate predictions of survival across
diverse conditions^[68]3,[69]11. In the evolving landscape of
biological investigation, spatially resolved molecular technologies
have become a pivotal focus for unraveling cellular diversity, tissue
organization, and functions. Spatial omics data have been incorporated
and routinely acquired by programs such as the Human Cell Atlas (HCA)
and the Human Biomolecular Atlas Program (HuBMAP), advancing the
construction of comprehensive spatial maps featuring various
biomolecules, including RNA, proteins, and metabolites^[70]12,[71]13. A
widely adopted molecular technology is spatial transcriptomics (ST),
which involves slicing tissues into thin layers for hematoxylin and
eosin (H&E) staining and spatial sequencing, enabling simultaneous
investigation of tissue/cellular phenotype and molecular mechanism on
the same slide.
Recent efforts to advance ST analysis have focused on incorporating
spatial neighborhood information^[72]14, or integrating histology
images^[73]15–[74]17. However, these tools typically operate at subspot
or superpixel spatial scales, which do not correspond to individual
cells, hindering biologically relevant insights, particularly in
contexts requiring cell-level data, such as analyzing ligand-receptor
interactions. Another branch of ST analysis frameworks addresses
cellular heterogeneity by resolving cell-type compositions within
spatial spots^[75]18–[76]20. However, these approaches do not infer
cell-level gene expression and are further restricted by the quality
and availability of single-cell RNA-seq (scRNA-seq) reference data,
especially for formalin-fixed paraffin-embedded (FFPE) tissues where
transcriptomic data quality is often compromised. While emerging
methods enable cellular-level histological structure
analysis^[77]21,[78]22, similarly they do not generate single-cell
resolution gene expression matrices, thereby hindering downstream
functional or molecular analyses. Moreover, those platforms are mostly
tailored to specific tasks (e.g., deconvolution), whereas comprehensive
analysis platforms (e.g., Seurat) prioritize -omics analysis without
deeply analyzing histopathological images^[79]23–[80]25.
To meet the urgent need for jointly analyzing genomics and histology,
we present a multi-modal platform, Thor, for bridging and exploring
cellular phenotypes and molecular insights. Thor enhances the
incorporation of morphology and transcriptome data of individual cells
by inferring cell-resolution transcriptome from spot-level ST data
using an anti-shrinking Markov graph diffusion method. Moreover, Thor
features extensible modules for comprehensive genomic analyses, such as
immune response, functional pathway enrichment, transcription factor
(TF) activity, and copy number variation (CNV), alongside tissue
analyses such as semi-supervised tissue annotation and nucleus
detection. Additionally, we develop Mjolnir, a user-friendly web-based
platform for interactive exploration of cellular organization and
pathogenesis in tissues, on a laptop, with no coding required.
We elucidated the principles of Thor and rigorously assessed its
effectiveness and accuracy through simulations and various datasets,
obtained from high-resolution experimental methods, including in situ
hybridization (ISH), multiplexed error-robust fluorescence in situ
hybridization (MERFISH)^[81]26, spatio-temporal enhanced resolution
omics-sequencing (Stereo-seq)^[82]27, and Xenium^[83]28. Thor
outperformed state-of-the-art (SOTA) methods in predicting cell-level
ST on a breast cancer dataset using Xenium data as the ground truth. We
analyzed a mouse olfactory bulb (MOB) tissue, human breast cancer
tissues, and multi-sample heart failure patient tissues. Thor revealed
a refined layered structure in MOB and identified distinct gene
modules. In heart failure, Thor quantified fibrotic regions across
different heart zones. Furthermore, we collected in-house heart failure
samples from patients who received left ventricular assist device
(LVAD) implantation to study the signature genes in vascular
regeneration. We characterized regenerative signatures in heart failure
and validated them through immunofluorescence (IF) staining. In breast
cancer, Thor conducted an unbiased screening of breast cancer
hallmarks, uncovering the intricate heterogeneity of immune responses
in tumor regions. In summary, Thor enables comprehensive interpretation
of ST data at the single-cell and whole-transcriptome levels,
delivering advanced functional insights and providing an interactive
interface for in-depth analyses.
Results
Thor infers cell-resolution spatial transcriptome for multi-modal analysis
Histological images and high-throughput sequencing data are widely
adopted for various applications^[84]2,[85]29–[86]31. Despite their
significance, these two sources of information are often examined
independently with separate tools. Sequencing-based ST and the paired
histological whole slide image (WSI) capture inherent cellular
structures in the tissue at different resolutions, providing
complementary information. For example, in human heart tissues with
myocardial infarction (MI), we observed that the projection of
histological features onto principal components segregated tissues at
cellular resolution (Fig. [87]1a and Supplementary Note [88]1).
Similarly, spatial patterns can be discerned through marker gene
expression at a coarser resolution (spot level). Clustering results of
spots by using either source of features were consistent and
complementary, as demonstrated in the human MI samples, a human ductal
carcinoma in situ (DCIS) sample, and a MOB sample (See details in
Supplementary Note [89]1 and Supplementary Fig. [90]29). Previous
studies also indicated that spatial gene expression can be predicted or
refined based on histological images^[91]15–[92]17. Therefore, we
hypothesize that it is feasible to recover cell-level resolution
transcriptomics data by learning shared patterns from both the
histology and the transcriptome.
Fig. 1. Integrated analyses of histology and transcriptomics data at the in
silico cell level.
[93]Fig. 1
[94]Open in a new tab
a Histological images and high-throughput sequencing data capture
inherent cellular structures at different resolutions and share
complementary information. The projection of the histological features
to the first principal component highlights the tissue sections at cell
resolution; meanwhile, expression patterns of marker genes of the
cardiac smooth muscle cells (MYH11) and the fibroblast cells (RARRES2)
demonstrate consistent patterns at spot resolution. The cell–cell
network is constructed according to the distances in the combinatory
feature space of histology (including location) and transcriptomics. In
the illustration of cell–cell network, the nodes represent cells, edges
represent connections, and the colors indicate cell types. Thor infers
single-cell spatial transcriptome by utilizing an anti-shrinking Markov
graph diffusion model. The expression profile of the marker gene MYH11
in smooth muscle aligns with the texture of the H&E staining image, as
visualized by the Mjolnir web platform. b Thor adapts and implements a
diversity of modules for advanced single-cell analyses around the
inferred spatially resolved whole transcriptome of the in silico cells.
c The Mjolnir platform supports interactive multi-modal tissue
analysis.
Multi-modal analysis in Thor involves two key steps. First, elevating
spot-resolution ST data to single-cell resolution (Fig. [95]1a).
Second, in-depth genomics and tissue image analyses (Fig. [96]1b, c).
In the first step, Thor (i) applies deep learning methods to segment
cells/nuclei from the WSI, termed in silico cells; (ii) extracts
morphological and spot-level transcriptomic features into a combinatory
feature space to construct a cell–cell network; (iii) creates a Markov
transition matrix, representing the probabilities of transitioning from
a cell to every other cell in the system in one step; (iv) infers gene
expression of the in silico cells by data diffusion with the transition
matrix (Fig. [97]1a). Thor represents the cellular patterns using a
nearest neighbors graph, where cells are connected according to their
distances in the combinatory feature space reflecting the physical
separation, and the histological and genomic complexity. The Markov
transition matrix is constructed such that information from
“homogeneous” spots asymmetrically corrects information from
“heterogeneous” spots, where heterogeneity of a spot is determined by
the enclosed cells in the combinatory feature space (Fig. [98]1a). In
the second step, we establish a standardized genomics analysis
framework for in-depth research and clinical practice. The genomics
analysis encompasses a wide array of insights, including cell type
annotation, immune response analysis, biological functional pathway
analysis, differential gene expression analysis, spatially expressed
module detection, TF activity analysis, and CNV analysis (Fig. [99]1b).
Thor also includes tissue image analysis tools including nucleus
segmentation, region of interest (ROI) selection, and semi-supervised
annotation (SSA). To enhance accessibility and usability, we introduce
a web-based platform, Mjolnir, that seamlessly visualizes both
histological images and genomic analyses (Fig. [100]1c). Altogether,
Thor elevates tissue analysis by integrating image analysis and genomic
insights.
Thor demonstrates accuracy and robustness in simulation data
We systematically evaluated Thor’s accuracy and robustness in
simulations under realistic experimental conditions. We simulated
expression profiles for 1000 genes in 6579 cells, whose spatial
positions were extracted from a mouse cerebellum tissue as the ground
truth^[101]32; and based on those cells, we created “spots” by
aggregating gene expression levels in cells covered by a spot
(Supplementary Fig. [102]1a, see details in “Methods: Simulation
details”). We assessed Thor’s prediction accuracy by computing the
normalized root mean squared error (NRMSE) between the predicted and
the ground-truth gene expression values (see “Methods” for definition).
We first evaluated Thor’s performance under suboptimal histology
imaging conditions. Here we consider two primary issues that may impact
its accuracy: (i) missed detection of cell nuclei, which commonly
occurs in out-of-focus or high-density regions, and (ii) erroneous
cell–cell connections resulting from poor histological features. We did
not modify the spot-level gene expression profiles, thereby isolating
the specific challenge posed by cell segmentation or erroneous
cell–cell connections from gene expression dropouts, which would be
investigated separately. The performance was evaluated on the detected
cells between inferred and ground-truth gene expression. Under ideal
conditions with neither cell dropouts nor randomized connections,
Thor’s predicted gene expression closely matched the ground truth,
yielding a median NRMSE of 0.07 (Supplementary Figs. [103]1a and
[104]2). Introducing random “missouts” of cells (0–40%) led to a slight
increase in median NRMSE from 0.07 to 0.075 (Supplementary
Fig. [105]1b), while introducing randomized connections in 30–40% of
cells modestly increased the median NRMSE to 0.08 (Supplementary
Fig. [106]1b). To further highlight the advantages of our algorithm, we
compared Thor against two baseline methods: (a) nearest spot method,
assigning gene expression based on the nearest spot; (b) BayesSpace,
assigning gene expression based on local spatial neighborhoods of
sub-spots^[107]14. These findings suggest that Thor maintains robust
prediction accuracy in the presence of substantial missing cells and
disrupted cell connections, outperforming those baseline methods.
Next, we examined the spatial resolution, a critical factor in spatial
technologies ranging from subcellular scales to ~100 µm. Larger spots
lead to greater cell heterogeneity within each spot (Supplementary
Fig. [108]3). When we varied the spot diameter from 25 to 150 µm, Thor
accurately predicted single-cell gene expression for spots up to
~100 µm in diameter, although the median NRMSE increased to 0.08 at
150 µm. We further compared Thor with three baseline methods: (a)
nearest spot method; (b) k-nearest neighbors (KNN) smoothing method,
assigning gene expression by averaging over the nearest twenty cells;
and (c) BayesSpace. At 25 µm, both the nearest spot method and Thor
exhibited high accuracy (median NRMSE 0.06). The nearest spot method’s
performance declined sharply as spot size increased beyond 25 µm, while
Thor remained accurate with the spot size up to 100 µm. This suggests
that Thor’s superior performance is not solely due to incorporating
nucleus segmentation. By contrast, both the KNN smoothing method and
BayesSpace performed poorly across all spot sizes, with median NRMSE
values of ~0.2 (Supplementary Fig. [109]1c). The KNN smoothing method
consistently underperformed, underscoring the benefits of Thor’s shared
nearest neighbors (SNN) cell–cell graph and feature-preserving Markov
diffusion approach.
To quantitatively evaluate Thor’s performance under increasing spot
complexity, we plotted the mean absolute error (MAE) of each cell
against the Shannon entropy of cell type proportions. As spot
heterogeneity increased, the MAE for the nearest spot method rose
sharply; meanwhile, Thor accurately imputed gene expression for both
low (Supplementary Fig. [110]3c, “A”) and high (Supplementary
Fig. [111]3c, “B”, “C”) heterogeneity spots. Although a subset of cells
in highly heterogeneous spots showed a slight increase in MAE
(Supplementary Fig. [112]3c, “C”), Thor’s error remained much lower
than that of the nearest spot method.
Finally, to evaluate Thor’s imputation performance under varying
dropout levels, an important challenge in high-resolution ST, we
simulated 15 conditions with dropout ratios ranging from 5% to 60% and
categorized them into three regimes: low dropout (<15%), moderate
dropout (15–40%), and high dropout (>40%). We then measured cluster
separations in principal component analysis (PCA) space using
silhouette coefficients. As shown in the PCA plots (Supplementary
Fig. [113]1d), introducing dropouts severely diminished cluster
separations in the ground truth data, with silhouette coefficients
reduced from 0.8 to near 0. In contrast, Thor-imputed data maintained
the silhouette coefficient to 0.7–0.8 in the low-dropout regime,
outperforming the KNN smoothing method and BayesSpace. When dropout
ratios rose to the moderate regime, where the ground truth data’s
silhouette coefficients declined to 0.1–0.4, Thor-imputed data
recovered the cluster separation successfully (silhouette coefficients
0.5–0.6). Even under high-dropout conditions (>40%), Thor’s scores
remained substantially above those of KNN smoothing and BayesSpace.
Collectively, these analyses highlight Thor’s accuracy and robustness
in various conditions, including missing cells, disrupted cell
connections, varying spot sizes, and technical dropouts.
Thor infers accurate gene expression at single-cell resolution
Next, we evaluated Thor on a mouse brain receptor map data acquired by
MERFISH. The MERFISH data comprised 483 RNA targets from individual
cells (Supplementary Fig. [114]4a). We simulated Visium-like data
within the hippocampus region by creating a grid of evenly spaced “ST
spots”. The RNA molecule counts in a synthetic spot were aggregated
over the cells covered by the “ST spot”. These synthetic spots
contained a mixture of cells of different cell types, particularly
within the hippocampal subregions CA1/2/3 and the dentate gyrus (DG;
Supplementary Fig. [115]4a). Thor connected cells of the same cell
types by proximity in the morphological feature space and the spatial
space, as illustrated by the cell–cell network in CA1 and DG
(Supplementary Fig. [116]5a; note the cell type information was not
provided to Thor). Thor successfully predicted cell-level gene
expression in these heterogeneous regions evidenced by the profiles of
selected marker genes (Supplementary Figs. [117]4b and [118]5b). For
instance, Thor recovered Adra1d expression in CA1 and DG, which was
missing in the spot-level data and the BayesSpace result. Furthermore,
to gain a global view of the similarity between the in silico cells and
the MERFISH cells, we projected the high-dimensional gene expression
matrices to a joint uniform manifold approximation and projection
(UMAP) embedding. The in silico cells inferred by Thor seamlessly mixed
with the MERFISH cells on UMAP, and the distribution of cell type
clusters of the in silico cells matched the ground-truth cell types
(Supplementary Fig. [119]4c). As a baseline, mixtures of cell types
were aggregated in the spot-level data, resulting in a low silhouette
coefficient and Calinski-Harabasz index when mapped to the nearest
cells. Thor substantially improved the cell type separation, achieving
a silhouette score of 0.45 and a high Calinski-Harabasz index of
10,000, and outperformed BayesSpace by a large margin (Supplementary
Fig. [120]4d).
We further applied Thor to a Visium dataset of human breast cancer
tissue and compared the result against a Xenium reference dataset of
the adjacent tissue section^[121]28. Using transcriptome data from the
Visium dataset and the post-Xenium H&E image as input, Thor
successfully inferred in silico cell-level gene expression. Visually,
the spatial patterns of gene expression align closely with Xenium data
(Fig. [122]2a). To gain a global view, we clustered the in silico
cell-level gene expression using conventional scRNA-seq clustering. The
same major cell types were identified from the in silico cells as from
the Xenium data, evidenced by the spatial distribution of the cell
types and the mean expression heatmaps of the marker genes
(Fig. [123]2b). Additionally, integrating the predicted in silico cells
with the Xenium cells showed that cells from the same cell types
colocalize from both datasets (Supplementary Fig. [124]6), indicating
Thor’s ability to predict accurate and biologically meaningful
cell-level gene expressions.
Fig. 2. Thor accurately predicts single-cell spatial gene expression in human
breast cancer.
[125]Fig. 2
[126]Open in a new tab
a Spatial gene expression of in silico cells inferred from the Visium
data and the H&E staining image of a breast cancer tissue by Thor align
closely with Xenium data from the adjacent tissue section. The numbers
on the H&E staining image mark DCIS regions of interest. b
Thor-inferred spatial transcriptome of in silico cells demonstrates
consistent cell clusters with Xenium using scRNA-seq clustering. The
cluster annotations were adapted from the original study of the
dataset^[127]28. The mean expression levels (normalized) of
differentially expressed genes in each cluster were visualized using
heatmaps. c Thor outperforms other methods in spatial gene expression
prediction. The box plots summarize the similarity across 306 genes
included in the Xenium panel. The middle line in the box plot, median;
box boundary, interquartile range; whiskers, 5–95 percentile; minimum
and maximum, not indicated in the box plot. One-sided Mann–Whitney U
tests are used to compare Thor with the two next-best-performing tools;
corresponding p-values are shown in the plot. d Spatial expression
profiles of representative genes at the region of interest level are
compared between Thor, iStar, and Xenium. Thor-inferred spatial gene
expression closely aligns with the Xenium data, while iStar introduces
artifacts at segment boundaries (the red arrows) and in regions with
sparse cells (the blue arrow). Source data are provided in a
[128]Source Data file.
For a quantitative evaluation, we benchmarked Thor with three other
methods of enhancing ST to near-cell
resolution^[129]14,[130]15,[131]17. The spatial units vary among those
tools (Thor: cell, iStar: superpixel, BayesSpace: subspot, and TESLA:
superpixel), therefore, we calculated both image-centric and
cell-centric metrics to provide a more complete evaluation. On the one
hand, by converting spatial profiles of gene expression data into
images, we compared the similarities between the predicted spatial
patterns with the Xenium spatial patterns using the metrics structural
similarity index measure (SSIM) and root mean squared error (RMSE) of
pixel values. On the other hand, by mapping the pixel expression data
to the cells using the nearest neighbors approach, we compared the
deviations between the resulting cell-level gene expression with the
Xenium data using cell-wise RMSE as an additional metric. Thor achieved
the highest similarity with the Xenium data on all the metrics
(Figs. [132]2c and [133]S7). When using the cell-wise RMSE, the general
trend remains, yet the difference between the four methods became less
prominent. This is likely because all the gene expression levels,
including Thor, needed to be mapped to the common cell positions
(Xenium cells) using nearest neighbors before calculating cell-wise
RMSE, which might have smoothed out some intricate details in the
spatial pattern, as seen in Supplementary Fig. [134]7c, d. Overall,
Thor demonstrated significantly better agreement with the Xenium data.
To gain more insights into Thor’s unique advantage, we compared the
expression profiles of representative genes with second best performing
tool, iStar. Thor and iStar enhanced spatial resolution to (near) cell
resolution, iStar at times introduced artifacts, including excessive
fusion, for instance, at segment boundaries (Fig. [135]2d, red arrows),
and in regions with sparse cells (Fig. [136]2d, blue arrow). For
example, the spatial expression of myoepithelial marker DST inferred by
Thor accurately outlined the boundaries of three DCIS regions in ROI 5
(Fig. [137]2d), as confirmed by the Xenium data and the H&E staining
image. While Thor did not maintain the spatial gradient pattern due to
misdetection of flat nuclei around certain region boundaries, iStar
introduced excessive fusion in the tumor regions, as indicated by the
red arrows in Fig. [138]2d. Additional examples are provided in
Supplementary Figs. [139]8 and[140]9. These artifacts are likely due to
that iStar predicts the expression of super-pixel patches of the WSI,
rather than a cell. This approach may result in the omission of
valuable cellular morphology information. In contrast, Thor takes a
fundamentally different approach by considering a cell as the minimum
biological unit and can accurately infer single-cell gene expression
via a cell–cell network constructed from the transcriptomics and
histology data.
Thor unveils refined tissue structure in mouse olfactory bulb
We extended our evaluation of Thor-inferred gene expression levels on a
MOB dataset collected by Visium. We compared the inferred molecular
patterns with those acquired from high-resolution techniques, including
the ISH images^[141]33 and Stereo-seq data^[142]27. Results showed the
spatial patterns of gene expression levels inferred by Thor aligned
well with both data (Supplementary Figs. [143]10a and [144]11). For
example, Eomes is a marker gene of cells in the glomerular layer and
mitral layer^[145]34, as observed in the ISH and Stereo-seq data.
However, due to the limited spatial resolution, the spot-level Visium
data failed to adequately capture the pattern in the mitral layer and
exhibited discontinuities in the glomerular layer. By integrating the
high-resolution H&E image with the spot-resolution ST, Thor recovered
the spatial patterns marked by Eomes in glomerular and mitral layers
(Supplementary Fig. [146]10a). Detailed gene expression profiles from
Thor, ISH, Stereo-seq, and Visium were provided for comparison in
Supplementary Fig. [147]11.
At the whole-transcriptome level, the in silico cell clusters dissected
six main layers in the MOB, the subependymal zone, two granule layers,
the mitral layer, the glomerular layer, and the olfactory nerve layer
(Supplementary Fig. [148]10b). We further applied Cell-ID^[149]35 to
infer cell types (see “Method”; Signature genes are provided in
Supplementary Data [150]1). By integrating ST with spatial locations
and histological features, Thor resolved and refined neuron subtypes.
For example, Thor distinguished granule cells (GCs) between GC-1 and
GC-2 subtypes, with GC-1 concentrated in the internal plexiform layer
and GC-2 predominantly in the GC layer. Additionally, Thor separated
mitral cells (M/TCs) into M/TC-1 and M/TC-2 subtypes, with M/TC-2
concentrated in the mitral layer and M/TC-1 extending into the
glomerular layer. These results demonstrated Thor’s capability to
refine cell type classification by integrating histology and ST data.
Leveraging the cell-resolution spatial profiles, we next identified
genes with spatially dependent activation patterns and coordinated gene
modules using the package Hotspot^[151]36. The genes in the in silico
cells formed 8 gene modules reflecting the primary structure of MOB
(Supplementary Fig. [152]10c), with modules “2”, “4”, “7”, and “8”
capturing the glomerular layer, the mitral layer, the granule layers,
and the olfactory nerve layers, respectively (Supplementary
Fig. [153]10d). Remarkably, module “4” captured the thin mitral layer
(thickness <40 µm), indicating successful resolution enhancement by
Thor, enriching a thin layer of the M/TC-2 mitral cell subtype. The
gene ontology (GO) pathway enrichment analysis of the layer-specific
gene modules suggested a cascade of activities covering odor
information sensory, processing, signal transmission, and memory
formation in MOB layers. Together, Thor unveiled refined layers in the
MOB tissue by accurately inferring cell-level gene expression data,
aligning with various experimental measurements.
Thor supports semi-supervised annotation of fibrotic regions in human
myocardial infarction tissues
To better leverage the combinatory space of histological and
transcriptomic features, we developed a human-in-the-loop tool for
enhanced identification of tissue regions or spatial domains. The SSA
tool operates within Mjolnir, enabling researchers to annotate small
representative regions using marker gene expression and morphology of
cells in gigapixel resolution images. These transcriptome- and
morphology-guided annotations can then be quickly propagated across the
entire tissue section based on Pearson correlation of the combinatory
features, facilitating comprehensive tissue characterization.
We first quantitatively evaluated Thor’s SSA tool using a cohort of
heart tissue samples^[154]37, which included high-resolution H&E
images, high-quality ST data, and spot-level expert annotations for key
tissue types in heart, including vessel, nodal tissue, adipose tissue,
and fibrosis. Evaluated against spot-level expert annotations, SSA
achieved accuracy ranges of 0.94–0.99 for vessels, 0.92–0.98 for nodal
tissue, 0.84–0.92 for adipose tissue, and 0.92–0.93 for fibrosis
(Supplementary Fig. [155]12). In contrast, spot-level clustering, even
with optimized parameters, struggled to distinguish structures such as
vessels (enriched with smooth muscle cells) from certain myocardium
regions (Supplementary Fig. [156]13). These results suggest Thor
enhances spatial tissue annotation by integrating histology with
transcriptomics, surpassing spot-level clustering.
Next, we applied Thor to analyze six myocardial infarction patient
samples, comprising two ischemic zones (IZ), two unaffected remote
zones (RZ), and two late-stage fibrotic zones (FZ), to enable granular
characterization of these distinct tissue zones in heart failure. Using
the Mjolnir platform, we first defined an ROI based on fibroblast
marker gene expression (PDGFRA and FBLN2) and morphological patterns in
an H&E image. Thor then automatically extended the curated ROIs by
identifying similar cells in the entire tissue. The expression profiles
of representative genes, including fibroblast marker genes and cardiac
muscle-associated genes, displayed coherent patterns in the curated
ROIs and the discovered cells (Fig. [157]3a and Supplementary
Figs. [158]14–[159]17). SSA revealed dense fibrotic areas and shallow
areas which were otherwise difficult to identify manually (Fig. [160]3b
and Supplementary Fig. [161]16). The resulting fractions of fibrotic
areas in the six samples increased in the order of RZ, IZ, and FZ
(Fig. [162]3c).
Fig. 3. Thor detects fibrotic regions in multiple human heart tissues with
MI.
[163]Fig. 3
[164]Open in a new tab
a H&E staining images of tissues from a remote zone (RZ1), an ischemic
zone (IZ1), and a fibrotic zone (FZ1). Purple and green squares mark
curated ROIs and are annotated as fibrotic and non-fibrotic regions,
respectively. Close-up views of the cell morphology and inferred
cellular expression of the fibroblast marker gene PDGFRA are provided
for the curated ROIs. b Mjolnir-annotated fibrotic regions (blue) are
visualized on the H&E staining images. c Bar plot of the percentages of
the fibrotic regions in all six samples. d Heatmap of the GO pathway
enrichment based on the up-regulated DEGs (fold change >2, adjusted
p-value < 0.01 using two-sided Welch’s t-test) in the fibrotic region
compared to the non-fibrotic region in each sample, and the
up-regulated DEGs (fold change >2, adjusted p-value < 0.01 using
two-sided Welch’s t-test) in the non-fibrotic region compared to the
fibrotic region in each sample. e TF activity is inferred from the in
silico cell spatial transcriptome. We used RTN (R package) for the
transcriptional network inference and Cytoscape for network
visualization. Source data are provided in a [165]Source Data file.
The precisely annotated fibrotic regions then enabled unbiased
functional analysis. For each sample, we performed differential gene
expression analysis between cells in the fibrotic and non-fibrotic
regions (lists of differentially expressed genes (DEGs) are provided in
Supplementary Data [166]2), followed by GO pathway enrichment analysis.
Irrespective of sample zones, the fibrotic regions showed significant
enrichment of pathways such as fibroblast proliferation, stress fiber
assembly, and collagen fibril organization, whereas myocardium-related
pathways were enriched in non-fibrotic regions (Fig. [167]3a and
Supplementary Fig. [168]18a). Interestingly, the fibrotic regions of RZ
samples demonstrated more pronounced inflammation and fibrosis, likely
reflecting heterogeneous progression of ischemic injury among the
patient samples. After myocardial infarction, tissue in the immediate
infarct area often undergoes rapid cell death and necrosis, whereas RZs
may experience a delayed and prolonged inflammatory and fibrotic
response^[169]38,[170]39. While IZ and FZ contained the largest
proportions of fibrotic regions at the whole tissue level
(Fig. [171]3c), those findings demonstrate that functionally distinct
fibrotic domains can exist outside necrotic regions.
To identify regulatory factors influencing those fibrotic regions, we
estimated TF activities by utilizing a gene regulatory network
database^[172]40. Compared to non-fibrotic regions, the most
prominently activated TFs induced critical pathways, such as
epithelial-mesenchymal transition (TWIST2 and SNAI2) and immune
response (STAT4 and MYB; Fig. [173]3d and Supplementary Fig. [174]18b).
The detected top-regulating TFs agreed with existing studies: SMAD3 has
been identified as a principal mediator of the fibrotic response to
activate cardiac fibroblasts^[175]41; SPI1 has been reported as an
essential orchestrator of the pro-fibrotic gene expression program in
multiple human organs^[176]42.
Overall, Thor’s SSA tool refines fibrotic tissue boundaries, highlights
subtle variations in fibrotic progression, and facilitates functional
insights into the molecular drivers of post-infarction cardiac
fibrosis.
Thor characterizes signature gene expression in heart failure
Spot-level STs are often inadequate to reveal intricate patterns in
small or narrow regions due to limitations in spatial resolution. One
such example is to identify the regenerative signatures in vascular
regions. Thor allows for the exploration of gene expression in
cell-resolution spatial contexts by predicting gene expression in cells
detected from the histological images, thereby enhancing the ability to
uncover intricate patterns.
In a patient with advanced heart failure, an LVAD is commonly implanted
as a bridge to heart transplantation. Over the 6–12 months that the
LVAD provides cardiac support, it has been observed that the structure
and function of the heart improves to varying degrees^[177]43,[178]44.
Thus, we applied Thor to in-house heart tissues collected from
post-LVAD implantation patients to identify genes that may be driving
this regenerative remodeling. As the vasculature plays an important
role in cardiac recovery^[179]45, we prioritized our analysis in the
vascular regions. Blood vessels typically consist of three layers,
intima, media, and adventitia, from the lumen to the outer wall of the
vessel. The media layer is mostly comprised of smooth muscle cells. In
Mjolnir, based on the cell phenotypes and the expression levels of the
smooth muscle marker MYH11, we annotated 29 and 11 vessel regions on
two post-LVAD heart tissues (Fig. [180]4a and Supplementary
Fig. [181]19a–c). We extracted highly expressed genes in these vessels,
finding 56 genes common to both tissues (Fig. [182]4b and Supplementary
Fig. [183]19d). PLA2G2A stood out when excluding known smooth muscle
markers (such as TAGLN, ACTA2, MYH11, and MYLK). PLA2G2A was reported
to promote cell proliferation, angiogenesis, and tissue
regeneration^[184]46 in several tumor types. PLA2G2A was reported to be
preferentially expressed in donor heart fibroblasts in contrast to the
failing heart fibroblasts^[185]47. We have previously found a subset of
fibroblasts that are capable of transdifferentiating to endothelial
cells to support microvascular recovery of an ischemic
tissue^[186]48,[187]49. Accordingly, we were interested to determine if
PLA2G2A expression represents a signature of the fibroblast subtype. We
divided the vascular cells into PLA2G2A^+ and PLA2G2A^– groups based on
the PLA2G2A expression levels (Fig. [188]4c). Notably, the PLA2G2A^+
cells were enriched in GO pathways related to tube morphogenesis and
blood vessel development (Fig. [189]4d). Furthermore, IF staining of
tissues from these two post-LVAD patients confirmed preferential
PLA2G2A protein expression in microvascular regions, which might be
related to angiogenic transdifferentiation and expansion of the
microvasculature (Fig. [190]4e). Of note, the expression of PLA2G2A may
be dependent upon spatial context. We observed that in the conduit
vessels, PLA2G2A expression was increased in vessel sections
neighboring connective or adipose tissues, by comparison to those
neighboring myocardium (Supplementary Fig. [191]19e). This difference
could reflect an effect of the surrounding tissue on vascular wall gene
expression. For example, perivascular adipose tissue in conduit
arteries is associated with atherosclerosis, possibly due to the
generation of inflammatory and angiogenic factors by the adipose tissue
that may accelerate plaque neovascularization and plaque
expansion^[192]50,[193]51. Alternatively, differences in vascular
expression of PLA2G2A may be due to vessel size, as it is well known
that vascular gene expression changes along the course of a vessel and
its branches, possibly due to developmental or hemodynamic differences
in vascular segments^[194]52,[195]53. Altogether, Thor’s joint
histology-transcriptome analysis revealed cell-resolution expression
patterns of crucial molecular markers with spatial context in
cardiovascular tissues.
Fig. 4. Thor characterizes regenerative signatures in vessels in human heart
failure.
[196]Fig. 4
[197]Open in a new tab
a Thor infers cell-level gene expression, expression of the smooth
muscle marker MYH11 are visualized at the spot level and the cell level
on sample I tissue. Utilizing Mjolnir, vessel regions are annotated.
The expression levels of MYH11 in selected vessels (labeled “1”, “2”,
“3’, “4”) are recovered by Thor, where there exhibits low expression of
MYH11 at the spot resolution. b The upregulated genes in the vessels
shared by two samples are ranked according to the gene scores. The
green box marks the gene of interest. c Cells in the vessel regions are
divided into two groups according to PLA2G2A expression levels. The
dotted vertical line separates PLA2G2A^+ (green) and PLA2G2A^− (blue)
cells at log expression = 0.5. d GO pathway enrichment using the top
500 upregulated DEGs (which was determined by the lowest adjusted
p-values using two-sided Welch’s t-test) in the PLA2G2A^+ cells. e IF
staining views of protein level PLA2G2A expression. The experiment was
performed in two post-LVAD patient tissue samples. Source data are
provided in a [198]Source Data file.
Thor enables multi-layered investigation of hallmarks in DCIS data
Thor offers rich layers of information through streamlined multi-modal
analyses within a unified platform. To showcase Thor’s strengths and
functions, we analyzed a well-validated DCIS dataset that has been used
as benchmark widely^[199]18,[200]54. DCIS is a potential precursor to
invasive ductal carcinoma, a condition that can progress into a form
requiring surgical intervention and radiotherapy. Understanding the
heterogeneity of various DCIS regions is crucial for elucidating the
factors driving their diverse behavior. The DCIS dataset comprises 18
pathologist-annotated major tumor regions (T1–T18; Fig. [201]5a).
Histological features of segmented cells identified distinct clusters,
underscoring their ability to distinguish between tissue regions
(Fig. [202]5b and Supplementary Note [203]1). Through integrated
histological features and ST analyses, Thor enabled a multi-layered
investigation of breast cancer hallmarks.
Fig. 5. Thor provides unbiased screening of hallmarks in cancer.
[204]Fig. 5
[205]Open in a new tab
a H&E staining image of the DCIS tissue. The annotation of eighteen
major tumor regions (T1–T18) in the DCIS tissue is adapted from the
annotation by pathology experts (Agoko NV, Belgium). b Leiden clusters
of the segmented cells using morphological features. The list of image
features and details of Leiden clustering are provided in Supplementary
Note [206]1. Colors represent cell clusters. c The spatial distribution
of cell types. Cell types are obtained by Cell-ID using the
Thor-inferred spatial transcriptome of the in silico cells and refined
with cell type markers. d VEGFA gene expression pattern at tissue and
cell scales in tumor region T1. The boxes mark regions of interest. e
The tumor regions identified by high attention values in CLAM and
semi-supervised annotation in Mjolnir. The solid black boxes mark the
curated regions for semi-supervised annotation. The dotted black square
marks a high-attention region where adipocytes are predominantly
located. f Heatmap of the copy number profiles inferred by CopyKAT
based on the in silico cell-level transcriptome predicted by Thor.
Representative breast cancer-related genes are provided. g Aneuploid
(tumor) and diploid (non-tumor) regions inferred by CopyKAT show
consistent results between the in silico cell-level transcriptome and
the spot data. Source data are provided in a [207]Source Data file.
First, Thor facilitates cell type annotation at single-cell level. The
spatial distribution of annotated cell types aligned well with the
results from SOTA methods such as CytoSPACE and RCTD^[208]18,[209]55
(Figs. [210]5c and [211]S20; signature genes of each cell type are
provided in Supplementary Data [212]3 for reference). While these
methods require scRNA-seq reference data, Thor overcomes the limitation
by integrating the underused histological features with ST.
Additionally, Thor’s advantage lies in providing gene expression for
individual cells detected directly from the tissue image for additional
analysis, maintaining spatial arrangement of the cells.
Second, Mjolnir enables interactive exploration of the spatial profiles
of key molecules on the gigapixel histological images seamlessly at
various zoom levels spanning from the whole tissue to the cellular
scale. As an example, the visualization of VEGFA, a pivotal angiogenic
factor influencing tumor growth and metastasis, highlighted distinct
abundance levels within tumor subpopulations at the cellular resolution
(Fig. [213]5d). Additional gene expression profiles at both spot and in
silico cell levels were provided in Supplementary Fig. [214]21. A
closer examination of the tumor region T1 using Thor revealed the
morphological features and the nuanced expression patterns of the
cancer cells. VEGFA exhibited the highest expression at the center of
the tumor region T1, gradually decreasing in abundance towards the
boundary; and was minimally expressed in the myeloid cell population
outside of T1.
Third, Thor enables efficient search of similar cells in the
combinatory space of histological and transcriptomic features. We
curated a small set of tumor cells in T8 based on cell morphology and
the key gene expression profiles. Cells in most tumor regions were
successfully identified (Fig. [215]5e; accuracy: 0.83). Interestingly,
hardly any tumor cells in T7 matched the curated set, likely due to its
distinct immune microenvironment. Instead, tumor cells in T7 were
effectively identified using a separate set of curated cells within T7
(Supplementary Fig. [216]22). This demonstrates Thor’s precision in
identifying tumor cells through integrated analysis.
Using only the H&E image, the clustering-constrained-attention
multiple-instance learning (CLAM) method^[217]2 identified
high-attention regions (Fig. [218]5e) that broadly overlapped with
pathology-annotated tumor areas (Fig. [219]5a). However, CLAM also
identified adipose tissue as high-attention region, which was not
directly relevant to cancer (black box in Fig. [220]5e). These false
positives happen for patterns which are not strongly represented in the
negative samples^[221]2, and may require additional training of CLAM on
curated datasets of labeled WSIs for improved specificity. This
demonstrated the value of tissue image analysis for tumor detection
while highlighting the need for further multi-modal integration to
reduce false positives.
Fourth, Thor’s cell-level molecular signature and pathway enrichment
analysis provided deeper insights into the heterogeneity of tumor
progression. By examining spatial patterns of oncogenes and tumor
suppressors, we observed a marked contrast between ERBB2 (also known as
HER2; an oncogene) and ATM (a tumor suppressor)^[222]56: ERBB2 was
highly expressed across all tumor regions, whereas ATM was upregulated
exclusively in region T7 (Supplementary Fig. [223]23). An unbiased
investigation of cancer hallmark pathways further highlighted their
complexity across different tumor regions at the cell level, including
DNA repair, a crucial process for maintaining DNA integrity and
preventing mutations (Supplementary Fig. [224]24). Notably, despite the
low expression of ESR1 (Supplementary Fig. [225]22), the estrogen
response pathway still showed significant enrichment in tumor regions
(Supplementary Fig. [226]24), emphasizing the power of pathway-based
analyses to refine breast cancer classification.
Lastly, genomic CNV inference from Thor’s cell-level transcriptome
classified tumor and normal cells in DCIS. Thor uncovered genome-wide
CNV profiles (Fig. [227]5f) and, when evaluating the distribution of
aneuploid cells, achieved an F1 score of 0.78 and a Jaccard index of
0.64 (Fig. [228]5g), aligning closely with pathology-annotated tumor
regions and surpassed spot-level CNV analyses (F1 score: 0.73; Jaccard
index: 0.58; Fig. [229]5g). Unlike spot-level CNV, which averages all
cells within a spot, and can misrepresent regions containing both
aneuploid and diploid cells, Thor’s single-cell approach accurately
detected mixed populations, as seen in tumor region T7. While
spot-level analysis labeled the entire region as aneuploid,
Thor-inferred and CytoSPACE-mapped single-cell data identified a
mixture of aneuploid and diploid cells. These findings were further
supported by external CNV profiles, validated through paired WGS data
(Supplementary Fig. [230]25; see “Methods”). Moreover, Thor revealed
key copy number aberrations across all tumor cells, including gains in
1, 2q, 8q, 12p, and 18p and losses in 5, 8p, 11q, and 12q. These
aberrations highlighted well-known breast cancer-associated genes, such
as MDM4, ZNF595, FGFR4, HIST1H1B, TPD52, DECR1, GRB7, and JUP ^[231]57.
CNV analyses provide critical insights into the genomic alterations
that underpin tumor heterogeneity and progression, offering potential
biomarkers for prognosis and therapeutic targets. Altogether, as
a unified platform of integrated analyses of histology and
transcriptomics data, Thor offers an unbiased, multi-layered view of
breast cancer hallmarks.
Thor reveals heterogeneity of immune responses in tumor regions of DCIS
We further investigated cell-level immune responses in DCIS to
quantitatively capture local immune activity around the tumor
regions^[232]57,[233]58. The tumor regions were ranked based on the
median scores calculated from the expression levels of immune marker
genes (see “Methods”). Regions T7, T1, and T17 exhibited the highest
scores, indicative of robust immune activity (Fig. [234]6a, b).
To gain deeper insight into the molecular differences among these high-
and low-scoring tumor regions, we performed differential gene
expression analyses. Several immune-related genes exhibited marked
variation: for example, CD84 and SMAD3 were abundant in T7 but almost
undetectable in T15 (Fig. [235]6c and Supplementary Fig. [236]26a),
whereas KANK1, often relevant in cancer prognosis, was highly expressed
in T6 and T15 but absent in T7. We further examined functional
distinctions and interactions between each tumor region and its
immediate peritumoral neighbors (Supplementary Fig. [237]26b). T7 was
enriched in pathways related to immune responses and T cell
co-stimulation, whereas T15 showed enrichment for tumor-associated
pathways, including hypoxia response and cell adhesion. Finally, an
unbiased GO pathway enrichment analysis of genes upregulated in each
tumor region highlighted Thor’s ability to reveal immune-response
heterogeneity in DCIS. A global heatmap (Fig. [238]6d) showed that T7,
T1, T4, and T14 were strongly enriched for inflammatory and immune
pathways. Notably, high-scoring areas like T7 and T1 were enriched in
pathways involving B cell activation, pointing to a more robust immune
microenvironment with potential therapeutic relevance.
Fig. 6. Thor reveals mechanistic insights into the immune response of DCIS.
[239]Fig. 6
[240]Open in a new tab
a Spatial distribution of the cell-level scores based on 29 genes. The
red lines encircle tumor regions from pathology annotation. b The tumor
regions are ranked according to the median cell-level score. Source
data are provided in a [241]Source Data file. The middle line in the
box plot, median; box boundary, interquartile range; whiskers, 5–95
percentile; minimum and maximum, not indicated in the box plot. c
Zoom-in view of the tumor regions with highest/lowest (T7/T15) scores.
The dotted orange lines encircle tumor regions. The expression level of
one DEG, CD84, is visualized in the inner and perimetral parts of the
tumor regions (marked by the dotted black boxes). GO pathway enrichment
is based on 300 up-regulated (fold change >2, adjusted p-value < 0.01
using two-sided Welch’s t-test) and 300 down-regulated (fold change
<0.5, adjusted p-value < 0.01 using two-sided Welch’s t-test) DEGs
between T7 and T15. d Heatmap of the GO pathway enrichment based on the
up-regulated DEGs in each tumor region compared to the rest (fold
change >1.5, adjusted p-value < 0.05 using two-sided Welch’s t-test).
Source data are provided in a [242]Source Data file.
By mapping these immune landscapes at single-cell resolution, Thor
elucidated functional heterogeneity among tumor regions, thereby
refining our understanding of immune-tumor interactions in DCIS.
Thor enhances gene expression signals in high-resolution Visium HD data
Recent advances in ST technologies, such as Visium HD, offer cellular
or even subcellular resolution. However, these high-resolution
platforms still face technical challenges, including substantial
dropout, transcript diffusion, and high background noise. To
demonstrate Thor’s effectiveness under these conditions, we generated a
high-resolution dataset from an in-house bladder cancer sample using
Visium HD, which provides spatial resolution of up to 2 µm square bins
(aggregated into 8 µm square bins for analyses, per 10x Genomics
recommendations). In the Visium HD raw data, we observed high noise
levels: for example, PTPRC (a lymphoid marker) was sparsely expressed
in immune-rich areas, while SPINK1 (a urothelium-associated gene) was
erroneously detected in non-tissue regions (Supplementary
Fig. [243]26a).
We first applied Thor to infer cell-level expression profile from the
2 µm bin-level inputs by integrating ST with histology. Thor’s
cell-level imputation yielded more coherent expression patterns than
8 µm square bins. Thor correctly localized PTPRC to immune areas and
SPINK1 to the tumor boundary, aligning with pathology annotations.
Beyond single-gene assessments, Thor-imputed data captured distinct
cell populations more accurately. For instance, cluster 7 in Thor’s
results precisely matched the pathology-annotated immune cell regions,
whereas the raw bin-level data overestimated immune cell presence
(Supplementary Fig. [244]26b). Similar overestimation of certain cell
types was also reported recently in Visium HD data^[245]59.
To contextualize Thor’s performance, we further compared it with
Bin2Cell which assigns bin-level expression to cells based on overlaps
between transcriptomic bins and expanded nucleus masks^[246]59. Both
methods successfully reduced expression artifacts in non-tissue regions
and accurately localized SPINK1 to epithelial compartments
(Supplementary Fig. [247]26). However, Bin2Cell left approximately half
of the detected cells without assigned transcripts, whereas Thor
inferred expression profiles for a substantially larger cell population
(by 37%). This difference may reflect Thor’s use of histology-derived
features to inform expression patterns beyond direct bin overlap, and
Bin2Cell’s reliance on bin-to-nucleus mapping can lead to
under-assignment in dense or ambiguous regions.
While these proof-of-concept analyses demonstrate Thor’s promise for
refining gene expression signals and enhancing biological
interpretability in high-resolution ST datasets, we note that current
analyses remain preliminary. Future work will involve systematic
benchmarking across tissue types, staining protocols, and segmentation
pipelines to further validate and refine Thor’s performance.
Robustness of Thor to parameter settings
Thor is designed to be highly flexible, allowing customization of
various parameters that control the preprocessing of
image/transcriptome data, cell–cell graph construction, and the Markov
diffusion process. To evaluate Thor’s robustness, we conducted a
systematic sensitivity analysis of key parameters, including the
diffusion step size
[MATH: t :MATH]
, the number of cell neighbors
[MATH: k :MATH]
, and the number of principal components
[MATH: nPC :MATH]
of the transcriptome data. Thor constructs a SNN cell–cell graph based
on the KNN in the combinatory space. First, we tested a range of
[MATH: k :MATH]
values on the MOB dataset while keeping other parameters fixed (
[MATH: t :MATH]
= 40 and
[MATH: nPC :MATH]
= 10). To reduce bias from highly expressed genes, we applied z-score
normalization for each gene. We then calculated the Pearson correlation
coefficients (
[MATH: r :MATH]
) across each pair of
[MATH: k :MATH]
settings. Thor demonstrated strong robustness for
[MATH: k :MATH]
values between 4 and 10, with a mean
[MATH: r :MATH]
= 0.88 and standard deviation (std) = 0.09. However, very small
[MATH: k :MATH]
values (<3) may produce disconnected cell graphs, whereas very large
[MATH: k :MATH]
values (40–100) may lead to over-smoothing and weaker correlations with
the results of other
[MATH: k :MATH]
values (mean
[MATH: r :MATH]
= 0.56, std = 0.27). Second, we evaluated the impact of varying
[MATH: nPC :MATH]
values while fixing
[MATH: k :MATH]
= 5 and
[MATH: t :MATH]
= 40. As shown in Supplementary Fig. [248]27a, Thor remains highly
robust when
[MATH: nPC :MATH]
>= 8 (mean
[MATH: r :MATH]
= 0.94, std = 0.05). In contrast,
[MATH: nPC :MATH]
< 4 fails to capture sufficient complexity in the data, leading to
lower correlations with high
[MATH: nPC :MATH]
values. Third, we also evaluated a range of diffusion time
[MATH: t :MATH]
while keeping
[MATH: nPC :MATH]
= 10 and
[MATH: k :MATH]
= 5 fixed. Thor converged after ~10 diffusion steps, achieving a mean
[MATH: r :MATH]
= 0.90 (std = 0.10) for
[MATH: t :MATH]
= 10. However, large
[MATH: t :MATH]
values (e.g., t > 50) may notably increase run time without significant
performance gains (Supplementary Fig. [249]27b). Overall, our analyses
show that Thor is robust to a broad range of
[MATH: t :MATH]
,
[MATH: k :MATH]
, and
[MATH: nPC :MATH]
values. These findings indicate that minor adjustments within
reasonable parameter ranges have minimal effect on Thor’s results,
which justifies keeping a common set of parameters across all case
studies.
Moreover, variational autoencoder (VAE) is widely used for RNA-seq data
analysis^[250]60–[251]62. Thor can utilize the latent representation in
VAE for faster predictions. In the fast mode, the Markov diffusion is
conducted on the VAE latent embeddings. The hyperparameter tuning, such
as adjusting the input and latent dimensions of VAE can affect the
results of Thor and contributes to generalizability. The input
dimension should depend on the genes of interest, such as highly
variable genes or spatially variable genes. Moreover, a proper latent
dimension should sufficiently capture the biological complexity in the
data. For instance, a latent dimension of 10 is set by default in
scvi-tools^[252]61, with 20 or 30 being appropriate for more complex
scRNA-seq datasets. We evaluated Thor’s performance on the MOB dataset
by varying the latent dimensions in separate VAE models (8, 16, 20, 32,
64, and 128), while keeping other parameters fixed (
[MATH: nPC :MATH]
= 10,
[MATH: k :MATH]
= 5, and
[MATH: t :MATH]
= 40). Thor-predicted gene expressions remained highly consistent with
Pearson’s
[MATH: r :MATH]
> 0.85 across all settings (Supplementary Fig. [253]27c). These
results indicate that Thor is robust to a broad range of parameter
settings.
Discussion
Thor is an extensible and customizable platform detailed in the
following aspects. First, the cell-level ST broadens the spectrum of
downstream analyses to those originally designed for scRNA-seq data.
Outputs from Thor interoperates with libraries such as Squidpy^[254]24
and stLearn^[255]25, and can be easily adapted for use with scRNA-seq
tools. Thor has included submodules such as cell-specific pathway
enrichment^[256]63, inference of genomic CNV profiles^[257]64, and
ligand-receptor analysis^[258]65. Second, Thor supports customized cell
features for building the cell–cell network. In this work, we highlight
Thor’s performance using intensity-based morphological features such as
color intensities of the staining image patches. The inclusion of more
task-relevant features elevates the quality of the cell–cell network.
For example, research has shown that spatial cellular graphs built from
multiplexed IF data enable better modeling of disease-relevant
microenvironments^[259]66. In addition, Thor supports direct input of a
cell–cell network adjacency matrix. Last, beyond ST, emerging omics
technologies such as spatial metabolomics and proteomics are
increasingly adopted to capture local metabolic or protein-level
processes that underlie key tissue functions and disease mechanisms.
While our current work focuses on applying Thor to ST, we envision that
its underlying framework, which constructs a cell–cell graph from
spot-level data, cell coordinates, and histological features, and then
refines those data through graph diffusion, could be adapted for
spatial metabolomics or proteomics as well. By substituting
transcriptomic values with metabolomic or proteomic intensities, Thor
could enable a more comprehensive, multi-omics view of tissue biology
at single-cell resolution. We anticipate that future developments will
provide deeper insights into complex tissue characterization by
integrating these additional modalities. In heart tissues from
post-LVAD patients, we observed PLA2G2A expression in vessels embedded
in myocardium as well as in those adjacent to connective or adipose
tissues. While expression in the former may support beneficial
regeneration^[260]47, expression in conduit vessels near pericoronary
adipose tissue may instead reflect pathological vascular inflammation
linked to plaque formation and atherosclerosis^[261]67. These
observations highlight the value of Thor’s detailed spatial resolution
in facilitating spatially aware analysis and supporting further
investigation into spatially regulated processes in cardiovascular
biology.
Thor integrates histological and transcriptomic features by inferring
cell-level ST. Notably, Thor does not require any additional scRNA-seq
data as a reference. This not only reduces the sequencing cost but
is also practically advantageous in FFPE tissues. FFPE tissues serve as
the most abundant specimens for longitudinal studies with preserved
tissue morphological details, yet RNA-seq profiling encounters hurdles
due to RNA crosslinking, modifications, and degradation. The Visium
platform offers a solution for profiling mRNA levels in both
fresh-frozen and FFPE tissues, employing a de-crosslinking
process^[262]68. Nevertheless, it falls short of providing
cellular-level resolution. In contrast, commonly used methods like
chromogenic immunohistochemistry (IHC) for assessing in situ biomarker
expression in FFPE tissues are limited by the number of analytes,
non-linear staining intensity, and the subjective nature of
quantitative analysis^[263]69. Thor strategically leverages the
advantages of Visium and overcomes these challenges by delivering
cell-level whole-transcriptome analysis, reducing cost and workload.
To better leverage the combinatory space of histological and
transcriptomic features, we developed a human-in-the-loop tool SSA for
targeted exploration. In Mjolnir, researchers select small
representative regions guided by marker gene expression and cellular
morphology; SSA then searches the joint feature space of geometry,
histology, and transcriptomics to retrieve cells with matching
profiles. Unlike unsupervised, global tissue segmentation tools which
uncover large-scale tissue architecture without user
input^[264]17,[265]70, SSA focuses on high-resolution, localized
annotation. It is powered by Thor-inferred gene expression and
histological features extracted from the high-resolution histology
image, supporting sub-regional or sub-cell-type level search and
labeling. Rather than replacing global tissue segmentation methods, SSA
complements them by offering precise, user-guided analysis within
complex tissues.
Thor offers several advantages over existing frameworks for studying
histological structures. PROST uses spatial relationships and
transcriptomics data to identify spatially variable genes and to
cluster spatial domains, but it does not enhance the resolution of the
original ST data^[266]21. Thus, with Visium data, PROST operates at the
Visium-spot level and does not utilize histology images. By contrast,
Thor integrates cell-level features from histology images with
spot-level transcriptomics, enabling inference of gene expression at
the single-cell level and providing a more granular analysis of
histological structures from Visium data. METI, meanwhile, is an
end-to-end framework tailored to cancer ST data, mapping tumor cells
and the surrounding microenvironment primarily in oncology-focused
contexts^[267]22. Thor, on the other hand, was conceived as a more
generalizable approach applicable across various tissues, disease
states, and organisms. Moreover, neither PROST nor METI directly
outputs single-cell gene expression. In contrast, Thor integrates
histological and transcriptomic data in a task-agnostic manner to infer
spatially resolved single-cell gene expression. This capability
supports a broad range of downstream analyses and comes bundled with
extensive analytical modules, including pathway enrichment, spatial
gene module identification, differential gene expression, TF activity
estimation, and interactive whole-slide data visualization. Taken
together, these features allow Thor to complement and extend the
capabilities of these frameworks by offering deeper spatial and
molecular insights into tissue architecture.
Several cutting-edge tools released recently have advanced the analysis
of spatial-omics data^[268]70–[269]72 in directions complementary with
Thor’s aims. Spotiphy, for example, generates cell-level gene
expression from spot-level data by combining matched scRNA-seq data
with cell locations segmented from histology images. In contrast, Thor
achieves single-cell-level expression without external references by