2.1 STARIT distinguishes cell states with subcellular molecular heterogeneity in simulated spatial transcriptomics data

To demonstrate the utility of STARIT for delineating cell states with subcellular molecular heterogeneity, beyond what conventional gene count clustering can capture, we first simulated imSRT data comprising of three transcriptionally distinct cell-types, each uniquely expressing one of three genes. For one cell-type, we further simulated cells such that they can be distinguished into two cell states based on nuclear versus perinuclear subcellular localization with no variation in expression magnitude between the two states. Specifically, we simulated cells of Group 1 to express Gene 1 uniquely, cells of Group 2 to express Gene 2 uniquely in a nuclear manner, cells of Group 3 to express Gene 2 uniquely in a perinuclear manner, and cells of Group 4 to express Gene 3 uniquely (Fig 1C). When we represent this simulated imSRT data as conventional gene counts and perform PCA dimensionality reduction and Louvain graph-based clustering, we are only able to identify three clusters of cells that correspond to the three simulated transcriptionally distinct cell-types (Fig 1D). This lack of correspondence between identified clusters and ground-truth groups can be further quantified via adjusted rand index (ARI) (score: 0.706) and the Jaccard similarity (mean: 0.833). While Cluster 2 corresponds perfectly to Group 1 cells and Cluster 1 to Group 4 cells, Cluster 3 contains a mixture of Group 2 and Group 3 cells. This overlap demonstrates that conventional representations of imSRT data and analyses are unable to distinguish between these two cell states, despite their distinct subcellular heterogeneity (Fig 1E). In contrast, when we apply STARIT, using the ResNet101 model for feature extraction followed by the same dimensionality reduction and clustering, we are able to identify four clusters of cells corresponding to the four simulated groups of cells representing cell-types and cell states with distinct subcellular heterogeneity (Fig 1F). This correspondence can be further quantified via ARI (score: 1.0) and in the on-diagonal Jaccard similarity (mean: 1.0) (Fig 1G). We further perform the Wilcoxon rank-sum test on the extracted features from ResNet101 and STARIT between cells in the four identified clusters as an analogous analysis to conventional differential expression testing. We identified significant features associated with the expected genes differentiating the four identified clusters, suggesting that the extracted features are capturing aspects of both cellular and subcellular molecular heterogeneity (Fig 1H).

Of note, while we used ResNet101, a feature extractor originally developed for natural images, for this analysis, STARIT is compatible with any downstream computer vision model for feature extraction. Although STARIT does not require the use of any specific feature extractor, feature extractor choice can influence which forms of spatial heterogeneity are emphasized. To demonstrate this, we generated an additional simulated dataset in which Groups 1 and 4 again uniquely expressed Gene 1 and Gene 3, respectively, while Groups 2 and 3 uniquely expressed Gene 2, now with left- versus right-polarized intracellular organization (S1A Fig, Methods). Applying STARIT with ResNet101, we resolved four clusters corresponding to the four simulated groups, indicating that cell states defined by left- versus right-polarized transcript localization can be resolved (S1B Fig). However, when we applied STARIT with UNI2-h, a feature extractor developed from a histopathology foundation model based on a vision transformer (ViT) encoder (ViT‑H/14), to the same dataset, the analysis recovered three clusters, with reduced separation of the left- and right-polarized Groups 2 and 3 cells (S1C Fig) [9]. Similar behavior was observed for CTransPath, another feature extractor for pathology images. In contrast, additional evaluated ResNet architectures including ResNet18, ResNet50, and ResNet152 more closely matched ResNet101 performance (S1 Table). This inability to distinguish left- and right- polarized groups may be expected because CTransPath, for example, is trained using frameworks that incorporate data augmentations such as random rotations and flips in learned visual representations [10]. Likewise, for UNI2‑h, augmentation follows DINOv2 and other self‑supervised ViT frameworks, which apply similar geometric transforms [9]. This result highlights that feature extractor choice may determine whether rotated and flipped spatial patterns such as left- versus right-polarization are preserved or diminished. Such sensitivity may be desirable when left- versus right-polarized spatial patterns are biologically meaningful, in the case of apical-basal organization, but undesirable when orientation reflects technical or arbitrary differences in cell positioning, for example. Thus, feature extractor selection should be guided by the biological question and expected sources of spatial variation in the dataset.

The simulated datasets thus far evaluate STARIT in settings where cell states are represented as discrete. However, cell states may vary continuously across developmental, environmental, or disease-associated trajectories [11,12]. We therefore sought to evaluate whether STARIT could capture continuous variation in subcellular organization. To model a continuous localization trajectory, we simulated cells with molecules sampled from either a circular nuclear region or a surrounding perinuclear annulus where the fraction of molecules assigned to the annulus increased monotonically from 0% to 100% across cells (S2A Fig, Methods). As such, this simulation generated a gradual transition from nuclear to perinuclear transcript localization while holding total expression magnitude constant. STARIT preserved this gradual progression in various dimensional embedding spaces, suggesting that STARIT can capture continuous trajectories represented in a gradual transition of intracellular transcript organization (S2B Fig).

Finally, a common challenge with current imSRT technologies, particularly in a tissue-setting, is that a whole cell may not always be captured in the associated thin tissue section. When a partial cell volume is captured, proportionally fewer transcripts may be detected. As such, we sought to evaluate whether STARIT remains robust when variable fractions of the cell volume are captured. We therefore generated additional simulated imSRT datasets to approximate partial cell volume capture in half the cells by systematically reducing molecule counts to represent 50%, 25%, and 12.5% volumetric capture variants, respectively (S3A-S3D Fig, Methods). STARIT recovered the expected localization-defined groups in the 50% and 25% volumetric capture simulations, indicating robustness to moderate volumetric capture variation (S3E-S3F Fig). However, performance degraded in the 12.5% volumetric capture setting, where sparse molecular sampling likely limited reliable recovery of subcellular localization patterns (S3G Fig). These results suggest that STARIT can tolerate moderate reductions in transcript detection due to decreased volumetric capture. Based on these simulations, we recommend appropriate quality control filtering to remove cells based on low transcript abundance or segmented cell volume prior to applying STARIT.

In general, these simulation results demonstrate that STARIT enables an orthogonal approach for analyzing imSRT data to identify and characterize cell-states with diverse subcellular transcript localization patterns.

2.2 STARIT is sensitive to misassigned molecules, but can be remedied with gene-expression feature weighting

Having demonstrated the utility of STARIT on simulated data, we next sought to apply STARIT to real imSRT data. We applied STARIT to real imSRT data of the mouse cortex assayed by osmFISH, comprised of cell segmentation information and transcript spatial coordinates for 33 genes in 4,833 cells representing 31 cell-types (S4 Fig) [13]. Applying STARIT with ResNet101, we derived a 4,833 cell by 67,584 feature matrix and followed with PCA dimensionality reduction, Louvain graph-based clustering, and UMAP visualization (S5A Fig). We identified 34 clusters using this approach and applied agglomerative clustering to obtain 31 clusters in an effort to recapitulate the original 31 cell-types identified by Codeluppi et al. (S5B-S5C Fig). However, we observed poor correspondence between our clusters to the original 31 cell-types (S5D Fig). This lack of correspondence between identified clusters and ground-truth cell-types can be further quantified via ARI (score: 0.138) and in the on-diagonal Jaccard similarity (mean: 0.102). We hypothesized that this discrepancy could be driven by minor segmentation errors, noisy gene detection, or other technical artifacts, whereby misassignment of a few molecules would substantially impact our extracted image features.

To illustrate this effect, we used a simulated imSRT dataset comprising once again of four groups of cells, now augmented with a few misassigned non-cell-type-specific molecules representing minor segmentation errors and noisy gene detection (Methods, Fig 2A). When we apply STARIT to this simulated noisy imSRT data and perform PCA, Louvain graph-based clustering, and visualize in UMAP space, we can still identify four clusters of cells. However, these four clusters do not correspond to our four simulated groups of cells representing distinct cell-types and cell states but are rather impacted by the misassigned molecules (Fig 2B). This lack of correspondence can be further quantified via ARI (score: 0.582) and in the on-diagonal Jaccard similarity (mean: 0.616). As such, this confirms how the presence of misassigned mRNA molecules, such as from minor segmentation errors or noisy gene detection, can have an impact on STARIT.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. STARIT with feature weighting and application to osmFISH mouse cortex dataset.

(A) Simulated imSRT data comprises four groups of cells (Groups 1–4) that exhibit distinct subcellular molecular heterogeneity. Left: Example simulated cells from each group. Groups are defined by distinct transcript expression and localization patterns across three genes (Gene 1, red; Gene 2, green; Gene 3, yellow). Groups 2 and 3 differ by nuclear versus perinuclear localization of Gene 2. All groups contain simulated noise across all genes. (B) STARIT analysis of simulated imSRT data with introduced noise, without feature weighting. Left: UMAP visualization of cells colored by ground-truth group annotation. Middle: UMAP visualization of cells colored by Louvain cluster assignment. Right: Jaccard similarity heatmap comparing ground-truth group annotations to Louvain clusters. (C) STARIT analysis of simulated imSRT data with introduced noise, with gene expression–based feature weighting. Left: UMAP visualization of cells colored by ground-truth group annotation. Middle: UMAP visualization of cells colored by Louvain cluster assignment. Right: Jaccard similarity heatmap comparing ground-truth group annotations to Louvain clusters. (D) Ground-truth cell-type annotations of osmFISH mouse somatosensory cortex dataset (Codeluppi et al., 2018) in tissue space. (E) UMAP embedding of ResNet101-extracted STARIT features from the osmFISH mouse cortex dataset, colored by ground-truth cell-types. (F) STARIT-derived Louvain clustering of osmFISH mouse somatosensory cortex dataset in tissue space. (G) UMAP visualization of STARIT-derived Louvain clusters of osmFISH mouse somatosensory cortex from ResNet101-extracted STARIT features. (H) Jaccard similarity heatmap comparing ground-truth osmFISH mouse somatosensory cortex annotations to STARIT-derived Louvain clusters.

https://doi.org/10.1371/journal.pcbi.1014845.g002

To address this issue, we introduced feature weighting to downweigh the contribution of such lowly detected genes in our STARIT analysis. Briefly, we multiply the extracted ResNet101 feature values for each gene in each cell by the min-max normalized expression magnitude of that gene in that cell. Such feature weighting effectively incorporates the prior expectation that more highly expressed genes are more important in defining a cell-type or state, downweighing lowly expressed genes that may have been detected due to minor segmentation errors and noisy gene detection. Using feature weighting with our simulated noisy imSRT data, we are able to identify four clusters of cells corresponding perfectly to the four simulated groups of cells representing distinct cell-types and cell states as quantified via ARI (score: 1.0) and on-diagonal Jaccard similarity (mean: 1.0) (Fig 2C). We further confirm that conventional gene counts clustering again identifies three clusters of cells, demonstrating that gene count clustering is more robust to such misassigned molecules but is still unable to resolve subcellular heterogeneity (S6 Fig). This establishes that while STARIT may be sensitive to minor segmentation errors and noisy gene detection, augmentation strategies such as weighting extracted features can be used to compensate to recapitulate expected results.

2.3 STARIT with feature weighting recapitulates expected cell-types

Having demonstrated that feature weighting can be an effective strategy to handle potential minor segmentation errors and/or noisy gene detection, we incorporate such feature weighting into our analysis of the osmFISH data. With feature weighting, we observe improved correspondence between our identified clusters and the original cell-type annotations, as reflected in the similar patterns in tissue and UMAP space as well as improved ARI with original cell types (score: 0.299) and higher on-diagonal Jaccard similarities (mean: 0.272) (Fig 2D-2H). This improved correspondence is consistent across anatomical regions (S7 Fig).

We do observe some new clusters unique to our STARIT analysis. For example, mature oligodendrocyte cells from the original osmFISH annotation are now split into clusters 3 and 24. As shown with our simulated data, such clusters may reflect distinct subcellular heterogeneity uniquely distinguishable through our STARIT analysis. We again used the Wilcoxon rank-sum test on the features between these two clusters to identify significantly differential features (Bonferroni-corrected p-value < 0.001) corresponding to the Anln, Ctps, Gfap, Itpr2, and Plp1 genes’ STARIT representations. To evaluate whether our newly identified clusters are solely due to STARIT or could be attributable to inherent instability in gene count clustering analysis, we performed normalization, PCA dimensionality reduction, followed by Louvain graph-based and agglomerative clustering to identify 31 clusters from re-clustering on the original gene counts matrix. We were able to observe similar clustering results for the gene counts re-clustering as from our weighted STARIT analysis, reflected in the similar patterns in tissue and UMAP space (S8 Fig), with a comparable ARI score of 0.342 and on-diagonal Jaccard similarity of 0.295 (S8D Fig). Again, this comparable correspondence is consistent across anatomical regions (S7 Fig).

These clustering results also included a splitting of mature oligodendrocyte cells (S8D Fig). Consistent with this, when we used the Wilcoxon rank-sum test on the gene counts between the cells in the mature oligodendrocyte subclusters (clusters 3 and 24) from STARIT, we were able to identify the same genes (Anln, Ctps, Gfap, Itpr2, Plp1) as significantly differentially expressed (Bonferroni-corrected p-value < 0.001). These results suggest that STARIT identified new clusters that split annotated mature oligodendrocytes, potentially reflecting distinct cell states. However, re-clustering of gene counts could produce similar results, potentially indicating how subcellular localization changes can accompany subtle magnitude changes. Nonetheless, STARIT captures comparable aspects of transcriptional heterogeneity as conventional gene count analysis. Furthermore, analysis of STARIT-derived image features identified genes driving cluster differences comparable to differential expression with gene count clustering analysis.

2.4 STARIT distinguishes bacterial states with molecular heterogeneity in bacterial-MERFISH data

We next sought to apply STARIT to real imSRT data of cultured cells representing one cell-type to identify possible subcellular heterogeneity suggestive of distinct cellular states. We applied STARIT to imSRT data of Escherichia coli (E. coli) K-12 MG1655 bacteria, assayed by bacterial-MERFISH [14,15]. We focused on the 1000-fold volumetric expansion dataset, where the high expansion enabled reliable segmentation of bacteria confirmed by visual inspection. We performed quality control filtering to obtain 463 bacterial cells (Methods). As a proof-of-concept, we focused on one assayed operon, fliC-fliX. fliX is a post-transcriptional regulatory small RNA processed from the 3′ untranslated region (UTR) of the fliC transcript that base-pairs with fliC mRNA to regulate flagellar gene expression. The fliC gene encodes flagellin, a major flagellar filament subunit involved in bacterial motility [16]. Because fliC and fliX are co-transcribed as a single polycistronic mRNA molecule, they are detected as a single spatial unit within the bacterial-MERFISH probe design workflow [14,16]. We applied STARIT with ResNet101 to create a 463 cell by 2,048 feature matrix, followed by PCA dimensionality reduction and Louvain graph-based clustering. We identified 3 clusters of bacteria using this approach (S9A Fig). Notably, these clusters of bacteria are not obviously distinguished based on fliC-fliX expression magnitude (S9B Fig). Therefore, they would not have been readily found via conventional count-based analysis. When visualized in physical space, these 3 clusters appear distinct in their spatial rotation of fliC-fliX molecules (S9C Fig). However, when visualizing the physical bacterial coordinates of these 3 clusters in culture, we observe that cells from different clusters exhibit distinct spatial rotations (S10 Fig). Therefore, in this example, the distinct spatial organization of fliC-fliX molecules is trivially driven by the rotation of the entire bacterium. Nonetheless, these results demonstrate that STARIT can resolve cells with distinct spatial heterogeneity including forms of rotations from well-segmented imSRT data of cultured cells.

To characterize potential biologically meaningful molecular spatial organizational heterogeneity that is not driven by the trivial rotation of the entire bacterium, we next sought to control for this rotational variation. To accomplish this, we standardized bacterial orientation across all 463 bacterial cells expressing fliC-fliX prior to rasterization (Methods). Briefly, for each bacterium, we centered all molecular coordinates and performed PCA in order to obtain the axis of maximal spatial variance, represented by the first principal component, to compute a rotation angle (θ) aligning the bacterium’s major axis to the x-axis. A corresponding counterclockwise 2-D affine transformation was then applied to all molecular coordinates to standardize orientation across cells (S11 Fig). To further minimize the impact of rotational variation, we applied STARIT with the UNI2-h feature extractor based on its previously demonstrated invariance to rotations via our simulations (S1C Fig). We applied STARIT to these rotationally standardized molecular coordinates with UNI2-h to create a 463 cell by 1,536 feature matrix, followed by PCA and Louvain graph-based clustering. We identify 4 clusters of bacteria (Fig 3A). Again, these 4 clusters of bacteria are not obviously distinguished based on fliC-fliX magnitude (Fig 3B). When visualizing the physical bacterial coordinates of these 4 clusters in culture, we no longer observe that cells from different clusters exhibit distinct spatial rotations as expected (S12 Fig). We further perform the Wilcoxon rank-sum test on the UNI2-h features between cells in the 4 identified clusters to identify significant extracted features differentiating the 4 identified clusters (S13 Fig). However, because these extracted features are abstract, high-dimensional embeddings produced by the pretrained UNI2-h model rather than human-interpretable measurements, it remains unclear what exact aspects of subcellular organization each feature captures, lacking direct biological interpretability.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. STARIT performance on standardized bacterial-MERFISH 1000X data.

(A) Uniform Manifold Approximation and Projection (UMAP) of STARIT-derived UNI2-h features from rotationally standardized 463 E. coli cells based fliC-fliX operon localization, imaged with bacterial-MERFISH at 1000X volumetric expansion. (B) UMAP embedding of cells colored by log-transformed fliC-fliX gene expression levels. (C) Distribution of spatial spread in X of fliC-fliX across bacterial clusters. Boxplots display median spread (line), interquartile range (box), and outliers, with colors denoting cluster. Statistical significance determined by Mann-Whitney U test (Bonferroni corrected). (D) Distribution of total fliC-fliX signal across bacterial clusters. Boxplots display median intensity (line), interquartile range (box), and outliers, with colors denoting cluster. Statistical significance determined by Mann-Whitney U test (Bonferroni corrected). (E) Distribution of variation in normalized fliC-fliX gene expression across bacterial clusters. Expression levels normalized to total molecular counts per 1000 and log-transformed. Boxplots display median normalized expression (line), interquartile range (box), and outliers, with colors denoting cluster. Statistical significance determined by Mann-Whitney U test (Bonferroni corrected). (F) Select examples of rasterized fliC-fliX transcript distributions for cells in clusters 1-4.

https://doi.org/10.1371/journal.pcbi.1014845.g003

Therefore, to interpret the differences between the clusters identified by STARIT, we applied a supervised approach to extract a set of interpretable magnitude, morphological, and spatial organizational metrics from each bacterium (Methods, S2 Table). Specifically, we computed each bacterium’s normalized fliC-fliX expression magnitude and total operon expression from the original gene counts, the spatial fliC-fliX spread (in x) and cell length from the associated transcript coordinates, and the total fliC-fliX image intensity from STARIT where higher total intensity is associated with a more diffuse spatial distribution while lower total intensity is associated with a more punctated spatial organizational pattern (Methods, S2 Table). Pairwise two-sided Wilcoxon rank-sum tests on these supervised features revealed distinct combinations of magnitude, morphological, and spatial organizational properties among the 4 clusters (Figs 3C-3E; S14 Fig). For example, Cluster 3 and Cluster 4 do not significantly differ in their normalized fliC-fliX expression magnitude (Fig 3E) and thus likely would not have been distinguished by conventional gene count clustering analysis. Cluster 3 does exhibit significantly greater fliC-fliX total intensity compared to Cluster 4 but also greater spread, cell length, and total operon count (Figs 3C-3D; S14A-S14B Fig). As such, the spatial heterogeneity in fliC-fliX distinguishing these two clusters could be explained by cell size differences. In contrast, Clusters 2 and 3 neither significantly differ in their normalized fliC-fliX expression magnitude, spread, cell length, nor total operon count (Figs 3C and 3E; S14A-S14B Fig). Yet, they differ significantly in fliC-fliX total intensity, with Cluster 2 cells exhibiting higher total intensities than Cluster 3 cells (Fig 3D). We can thus interpret Cluster 2 cells as having a more diffuse distribution of fliC-fliX transcripts throughout the bacterium and Cluster 3 cells as having a more punctated distribution of fliC-fliX transcripts (Fig 3F). Again, these differences in subcellular spatial organization would be missed by conventional gene count clustering and highlights the ability of STARIT to resolve cells with distinct subcellular spatial heterogeneity.

We note that these supervised metrics are highly correlated with some STARIT-derived UNI2-h extracted features (spread maximum Spearman = 0.61; intensity maximum Spearman = 0.64). This correlation suggests that the STARIT-derived UNI2-h extracted features, to some extent, capture these magnitude, morphological, and spatial organizational properties. Still, while fliC-fliX spread and total intensity are significantly different between our 4 identified clusters, these two supervised metrics do not as clearly delineate all cells across the 4 clusters as well as the UNI2-h extracted features (S15 Fig), suggesting that the bacterial subpopulations may be distinguished by a multivariate combination of features rather than by any two interpretable metrics.

Prior work has shown that bacterial mRNAs are not necessarily uniformly distributed within cells: mRNAs, ribosomes, and RNA polymerase can occupy distinct subcellular regions, and E. coli transcripts, including motility-related transcripts, can be asymmetrically enriched across cytosolic, membrane, and polar areas [17–19]. This polarization has been proposed to reflect localized translation of flagellar proteins, coordinating transcript localization with spatially constrained flagellar assembly [19]. In other bacteria, fliC translation is similarly controlled by intrinsic mRNA sequence or structure features that may be paired with secretion or assembly [20]. Given that fliC encodes the major structural component of the bacterial flagellar filament and fliX post-transcriptionally regulates flagellar gene expression, differences between dispersed and punctate fliC-fliX transcript distributions identified by STARIT may reflect proximity to flagellar assembly-associated machinery, differences in translational engagement, or other transcript organizational states. Because the bacterial-MERFISH data measures mRNA position but not ribosomes or flagellar proteins directly, there is no evidence of active localized translation directly. Nevertheless, the distinct fliC-fliX spatial localization patterns present in these clusters of bacterial cells despite overall similar normalized fliC-fliX expression magnitude suggests that STARIT can identify cell-states with biologically meaningful subcellular transcript localization patterns.

4.1 Spatial Transcriptomics As Rasterized Image Tensors (STARIT)

STARIT is available as an open-source Python toolkit with additional documentation and tutorials available at https://jef.works/STARIT/ and source code at https://github.com/JEFworks-Lab/STARIT.

STARIT converts transcript-level spatial coordinates from imSRT data into rasterized density image representations that preserve subcellular molecular organization for each segmented cell. STARIT takes three primary inputs: (i) an axis-aligned bounding box and the (ii) and (iii) coordinate vectors of all detected transcripts to be rasterized for a given cell. To maintain the original cell aspect ratio, a bounding box helper function outputs the calculated minimum and maximum spatial extrema of the molecular coordinates assigned to a cell .The box bounded by these four values is then expanded by a factor called expand (default 1.1) to provide a small margin for the rasterized image. The bounding box image grid is then defined at a desired pixel resolution size (default 1.0, in the same units as the input coordinates). From there, each gene molecule is then rendered to the canvas grid by convolving a normalized Gaussian kernel, blur (default 1.0), centered at (𝑥,𝑦). The pixelated image is created by summing contributions from all gene molecules in the cell. The result is STARIT producing one PNG image per gene per cell. When a gene has no molecules, a blank black image of the canonical size is saved to preserve the tensor shape.

4.2 Image preprocessing and feature extraction workflow

Prior to model input, each rasterized gene image was converted to RGB before the image was resized to 224x224 pixels with aspect-ratio preservation and padding. Each image was proportionally scaled to fit within a target bounding box of 224x224 pixels, determined by the limiting dimension:

Images were downscaled using Lanczos interpolation. The resulting resized images were then symmetrically padded with black pixels (zero-padding) to fulfill the target dimensions, ensuring the original scaled images remained centered.

Images were normalized using ImageNet preprocessing parameters, with channel means of [0.485, 0.456, 0.406] and standard deviations of [0.229, 0.224, 0.225]. Unless otherwise specified, feature extraction was performed using pretrained computer vision models with the final classification layer removed.

For ResNet-based analyses, each gene-specific image was passed independently through the pretrained model. ResNet18 produced a 512-dimensional embedding per gene image, whereas ResNet50, ResNet101, and ResNet152 produced 2,048-dimensional embeddings per gene image. These per-gene embeddings were concatenated across all genes to generate one cell-level feature vector. For example, in simulated datasets with three genes, ResNet101 generated a 6,144-dimensional feature vector per cell. For the osmFISH dataset with 33 genes, ResNet101 generated a 67,584-dimensional feature vector per cell.

We also evaluated non-ResNet feature extractors, including CTransPath and UNI2-h. CTransPath produced 768 features per gene image whereas UNI2-h produced 1,536 features per gene image in the simulated analyses, with the resulting feature matrix dimensions reported in S1 Table. For all feature extractors, the resulting cell-by-feature matrix was used as input for dimensionality reduction, clustering, and visualization.

4.3 Gene expression-based feature weighting

To reduce the influence of image features derived from sparsely detected or potentially noisy genes, we implemented gene expression-based feature weighting. For each dataset, the corresponding cell-by-gene count matrix was min–max scaled per gene across cells. For each cell and gene, the image-derived feature vector for that gene was multiplied by the corresponding scaled gene count value. Weighted gene-specific feature vectors were then concatenated across genes to form a weighted cell-by-feature matrix.

This strategy preserves the image-derived spatial features for highly detected genes while downweighing features from genes with low or absent expression in a given cell. Feature weighting was applied to selected noisy simulated analyses and to the osmFISH analysis, as indicated in S1 Table.

4.4 Simulated imSRT datasets

All simulated datasets were generated using Python scripts with fixed random seeds. Each simulated dataset was exported as a transcript-level coordinate table containing x-position, y-position, gene identity, cell identity, and group identity where applicable. A corresponding cell-by-gene count matrix was also generated for conventional gene-count analysis. Unless otherwise specified, each dataset consisted of 80 cells organized into four groups of 20 cells each, with three possible genes: gene 1, gene 2, and gene 3. Simulated cells were represented on a 224x224 coordinate canvas centered at (112, 112), with a circular cell radius of 75 pixels.

4.5 Applying STARIT to simulated imSRT datasets

STARIT was applied to each simulated dataset to generate one rasterized image per cell-gene pair with a pixel resolution size (dx) of 1 and a Gaussian kernel (blur) of 1. For datasets with three genes, three images were produced for each cell. If a cell had zero molecules for a given gene, a blank black 224x224 image was generated for that gene. For ResNet101-based simulated analyses, the three gene-specific embeddings were concatenated to form an 80 x 6,144 cell-by-feature matrix. For ResNet18, ResNet50, ResNet152, CTransPath, and UNI2-h comparisons, the same image preprocessing and concatenation strategy was used, with feature dimensions determined by the corresponding model architecture and reported in S1 Table.

4.6 Downstream analysis of simulated datasets

For STARIT-based analyses of the main simulated datasets, PCA was first applied to the cell-by-feature matrix. Unless otherwise specified, the top 5 principal components were retained. UMAP was then applied to the PCA representation for visualization. For the main ResNet101 nuclear-perinuclear simulation, UMAP used 30 nearest neighbors, a minimum distance of 0.1, and random state 42. For most additional simulated STARIT analyses, UMAP used 30 nearest neighbors, a minimum distance of 0.2, and random state 42, unless otherwise indicated in S1 Table, keeping all other parameters at their defaults.

Graph-based clustering was performed by constructing a k-nearest neighbors (kNN) graph from the PCA representation using cosine distance, connectivity mode, and include self to true. Louvain clustering was then applied to the resulting graph. The number of neighbors and Louvain resolution were selected for each analysis as reported in S1 Table. For the main ResNet101 nuclear-perinuclear simulation, 18 neighbors were used. For feature-weighted noisy simulated analyses, PCA retained 5 PCs, UMAP used 10 nearest neighbors and a minimum distance of 0.5, and Louvain clustering was applied to a kNN graph constructed with the number of neighbors specified in S1 Table.

To compare STARIT-derived clusters with ground-truth simulated groups, we computed the adjusted rand index (ARI) and a Jaccard similarity matrix. ARI was used to summarize agreement between the complete inferred clustering and ground-truth annotations while correcting for chance agreement. Jaccard similarity was computed between each ground-truth group and each inferred cluster. When the number of inferred clusters differed from the number of ground-truth groups, mean on-diagonal Jaccard similarity was computed from the best one-to-one matching between groups and clusters, such as the case for certain models with the left-right polarized simulated dataset.

For conventional gene-count analysis of simulated datasets, PCA was applied to the cell-by-gene count matrix rather than the STARIT feature matrix. Unless otherwise specified, the top 3 PCs were retained. UMAP and Louvain clustering were then performed using the parameters reported in S1 Table. Cluster agreement with ground truth was quantified using the same ARI and Jaccard similarity approaches used for STARIT-derived clusters, where again when the number of inferred clusters differed from the number of ground-truth groups, mean on-diagonal Jaccard similarity was computed from the best one-to-one matching between groups and clusters.

For selected analyses, extracted image features were visualized using heatmaps. Feature values were standardized across cells by z-score normalization. Hierarchical clustering was performed using Ward’s linkage and Euclidean distance.

4.7 Differential feature testing and differential expression analyses

To identify STARIT-derived features that differed between selected clusters, we applied two-sided Wilcoxon rank-sum tests to extracted image features. P-values were adjusted for multiple testing using Bonferroni correction. For the simulated datasets, differential testing was performed on STARIT-derived image features to identify features associated with group-specific gene images and localization patterns.

For the osmFISH analysis, differential testing was performed between selected STARIT-derived clusters using both STARIT image features and conventional gene counts. For STARIT feature testing, Wilcoxon rank-sum tests were applied to the image-derived feature matrix. For conventional expression testing, the same tests were applied to the cell-by-gene count matrix. For the bacterial-MERFISH analysis, differential testing was also performed between the selected STARIT-derived clusters using STARIT-derived image features.

Genes or features with Bonferroni-corrected p-values below the specified threshold of 0.001 were considered significant.

4.8 osmFISH mouse cortex data

The osmFISH mouse somatosensory cortex dataset was obtained from the original publication by Codeluppi et al. [13]. Cell regions annotated as “Excluded” in the original dataset were removed. We also removed six additional cells with no detected molecules for any of the 33 assayed genes (cells 7825, 1524, 3085, 5619, 7823, 4323). After filtering, the dataset contained 4,833 cells, 33 genes, 11 curated anatomical regions, and 31 curated cell-type annotations.

For each cell, subcellular transcript coordinates were centered at the cell origin. STARIT was then applied to all 4,833 cells and all 33 genes using a pixel resolution size of 1 (dx = 1) and a Gaussian kernel of 1 (blur = 1). This generated 159,489 gene-specific rasterized images. For genes with zero detected molecules in a given cell, a blank black 224x224 image was generated.

4.9 Downstream STARIT analysis of osmFISH data

Each STARIT image was converted to RGB, resized to 224x224 pixels with aspect-ratio preservation and padding, and normalized with ImageNet preprocessing. A pretrained ResNet101 with the final fully connected layer removed was used to extract a 2,048-dimensional feature embedding for each gene image. Because there are thirty-three gene markers, 33 corresponding images were produced. The 2048-dimensional ResNet features extracted from each gene-specific image were concatenated to form a 67,584-dimensional feature vector per cell (33 genes x 2,048 features). Feature vectors from all 4,833 cells were then stacked into a 4,833 x 67,584 matrix, which served as the input for all downstream analyses.

For the unweighted STARIT analysis, PCA was applied to the feature matrix, and the top 30 PCs were retained. UMAP was applied using cosine distance, 50 nearest neighbors, a minimum distance of 0.15, and random state 42. For clustering, a kNN graph was constructed in PCA space using cosine distance, connectivity mode, include self to true, and 5 nearest neighbors. Louvain clustering was applied with the parameters reported in S1 Table. To compare with the 31 curated osmFISH cell types, hierarchical agglomerative clustering was performed on Louvain community centroids in PCA space to obtain 31 meta-clusters. Agreement between inferred clusters and curated cell-type annotations was quantified using ARI and Jaccard similarity.

For the feature-weighted osmFISH STARIT analysis, the original osmFISH cell-by-gene count matrix was min–max scaled per gene across cells. During feature extraction for each cell and gene, the corresponding 2,048-dimensional ResNet101 embedding was multiplied by the scaled count value for that gene in that cell. Weighted gene-specific embeddings were concatenated to generate a 4,833 x 67,584 weighted cell-by-feature matrix. PCA, UMAP, Louvain clustering, and agglomerative clustering were then performed as described above, except that the kNN graph for Louvain clustering used 6 nearest neighbors, as reported in S1 Table. Agreement between inferred clusters and curated cell-type annotations was once again quantified using ARI and Jaccard similarity.

4.10 Conventional gene-count analysis of osmFISH data

Using the original osmFISH cell-by-gene count matrix, gene counts were filtered identically and min–max scaled per gene across cells (4,833 cells x 33 genes). We first performed PCA and retained the top 30 PCs. These PCs were then used as input to UMAP using 50 nearest neighbors, a minimum distance of 0.15, and random state 42. A kNN graph was constructed from the PCA representation using 6 nearest neighbors, cosine distance, include self to true, and connectivity mode, keeping all other settings at their defaults. Louvain clustering was performed with a resolution of 1.2 and random state 42. Hierarchical agglomerative clustering was then applied to Louvain community centroids in PCA space to obtain 31 meta-clusters. Agreement with curated osmFISH cell-type labels was quantified using ARI and Jaccard similarity. Agreement with ground-truth osmFISH cell-type labels was quantified using ARI and Jaccard similarity.

4.11 Bacterial-MERFISH data

The 1000-fold volumetric expansion bacterial-MERFISH dataset was obtained from the original publication by Sarfatis et al. [14,15]. For the 1000-fold bacterial-MERFISH dataset, we filtered for bacterial cells with more than 250 total mRNA molecules detected, resulting in 463 cells. For each retained cell, molecular coordinates were subsetted to those corresponding to the fliC–fliX operon. Subcellular positional coordinates were centered at each cell’s origin prior to rasterization.

STARIT was applied to these centered fliC-fliX molecular coordinates, with a pixel resolution size of 1 (dx = 1) and a Gaussian kernel of 1 (blur = 1), producing one rasterized STARIT image per cell. A total of 463 operon image representations of fliC-fliX were generated.

4.12 Downstream analysis of unstandardized bacterial-MERFISH data

For the unstandardized bacterial-MERFISH analysis, each STARIT image was converted to RGB, resized to 224x224 pixels with aspect-ratio preservation and padding, and normalized using ImageNet preprocessing. A pretrained ResNet101 with the final fully connected layer removed was used to extract a 2,048-dimensional feature embedding for each operon image to form a 2,048-dimensional feature vector per cell (1 operon x 2,048 features). Feature vectors from all 463 cells were then stacked into a 463 x 2,048 matrix.

PCA was applied to the feature matrix and the top 30 PCs were retained. UMAP was performed using 15 nearest neighbors, a minimum distance of 0.1, and random state 42. For clustering, a kNN graph was constructed in PCA space using cosine distance, connectivity mode, include self to true, and 100 nearest neighbors. Louvain clustering was then applied using random state 42, keeping all other settings at their defaults. Cluster-specific variation in fliC–fliX expression was visualized using PCA embeddings colored by raw and log-transformed counts.

4.13 Rotation standardization of bacterial-MERFISH cells

To reduce clustering driven by whole-cell orientation, we generated a rotation-standardized bacterial-MERFISH data representation. For each bacterium, fliC-fliX molecular coordinates were centered translating the origin to the global centroid, calculated from all detected operon molecules within the cell.

PCA was then applied to the centered fliC-fliX coordinates to identify the first principal component eigenvector, , corresponding to the major axis of spatial variation. The orientation angle between this first principal component and the x-axis was computed as, using the four-quadrant inverse tangent:

A two-dimensional counterclockwise rotation transformation was then applied to align each cell’s major axis with the x-axis prior to rasterization, using the angle :

This standardization was used to reduce variation associated with arbitrary bacterial orientation.

4.14 Downstream analysis of rotation-standardized bacterial-MERFISH data

STARIT was applied to the rotation-standardized fliC–fliX coordinates to generate one rasterized operon image per cell. Images were converted to RGB, resized to 224x224 pixels with aspect-ratio preservation and padding, and normalized using ImageNet preprocessing.For the standardized bacterial-MERFISH analysis, UNI2-h was used for feature extraction, producing a 463 x 1,536 cell-by-feature matrix. PCA was applied and the top 30 PCs were retained. UMAP was performed using 15 nearest neighbors, a minimum distance of 0.1, and random state 42. A kNN graph was constructed in PCA space using cosine distance, connectivity mode, include self to true, and 100 nearest neighbors. Louvain clustering was then applied using random state 42, keeping all other settings at their defaults.

4.15 Interpretable image-feature analysis of bacterial-MERFISH clusters

To characterize the STARIT-derived bacterial-MERFISH clusters using interpretable measurements, we extracted supervised morphological and intensity metrics from each cell. Metrics included spatial spread of fliC-fliX gene expression in the x-direction, total sum pixel intensity, total normalized expression of fliC-fliX, cell length, and total operon expression.

In the absence of experimentally provided cell segmentation from the Bacterial-MERFISH data, we defined a proxy cell segmentation mask for each cell using the molecular coordinate data provided from Sarfatis et al [14,15]. Specifically, we used STARIT to generate a cell segmentation image per cell using all detected molecules assigned to all operons from their spatial coordinates. The resulting image was converted to grayscale and thresholded (values greater than 0) to generate a binary cell segmentation mask. This proxy cell mask was used to approximate the cell region over which intensity-based metrics were quantified, such as total sum intensity, summing up all the pixel intensities from each fliC-fliX image that are inside the cell mask. Calculations for spatial spread of fliC-fliX gene expression in the x-direction and cell length per cell were derived from the actual x,y molecular coordinate information rather than the images themselves. A full list defining each of these metrics and how they are calculated is provided in S2 Table.

For each metric, pairwise differences between clusters were evaluated using two-sided Wilcoxon rank-sum tests. P-values were adjusted for multiple testing using Bonferroni correction. These supervised metrics were used to summarize interpretable differences between bacterial clusters, while recognizing that STARIT-derived cluster structure may reflect multivariate combinations of spatial and intensity features rather than a single supervised metric.

Spearman correlations were also computed between extracted image features and each supervised metric.

4.16 Computational specifications

All STARIT rasterization and downstream feature extraction runtimes reported were measured in series, without any parallelization across cells or genes, on a Mac Studio (2023) equipped with an Apple M2 Ultra system-on-chip (24-core CPU: 16 performance cores and 8 efficiency cores; 60-core GPU), 192 GB of unified memory, running macOS Sequoia 15.6.1.