Journal: bioRxiv

Article Title: Impact of Data Quality on Deep Learning Prediction of Spatial Transcriptomics from Histology Images

doi: 10.1101/2025.09.04.674228

Figure Lengend Snippet: A. Acquisition of paired breast cancer spatial transcriptomics datasets and histology images from 10x Visium and Xenium. B. Co-registration of Visium and Xenium histology slides into a common coordinate system. The green box highlights the overlapping region retained between the two technologies. C. Rasterization of gene counts onto a uniform grid matched to Visium spot resolution, followed by extraction of the overlapping tissue region. Expression is visualized as patches. D. Training of deep learning models to predict per-patch gene expression from histology image patches. E. Performance evaluation on held-out replicates, comparison across technologies, and ablation experiments of inputs.

Article Snippet: Here, we investigate how variation in molecular and image data quality stemming from differences in imaging (Xenium) versus sequencing (Visium) spatial transcriptomics technologies impact deep learning-based gene expression prediction from histology images.

Techniques: Extraction, Expressing, Gene Expression, Comparison

Journal: bioRxiv

Article Title: Impact of Data Quality on Deep Learning Prediction of Spatial Transcriptomics from Histology Images

doi: 10.1101/2025.09.04.674228

Figure Lengend Snippet: Histogram showing the distribution of Pearson correlation coefficients for gene expression predictions using Visium and Xenium data. The dotted vertical line denotes the mean PCC, and the solid curved line traces the density estimate. Results are computed on the held-out test set and represent the average performance across five independently trained models. B. Scatterplot comparing the Pearson correlation coefficients of predictions from Visium and Xenium data. The gray dotted line denotes x=y, and select genes corresponding to (C) labeled. C. Representative examples of ground truth and predicted gene expression for HDC , ANKRD30A , AHSP , and GZMK in both the Visium and Xenium datasets. Predicted gene expressions are visualized for the full dataset, while the performance metrics (PCC and normalized rMSE) are computed from the held-out test set only.

Article Snippet: Here, we investigate how variation in molecular and image data quality stemming from differences in imaging (Xenium) versus sequencing (Visium) spatial transcriptomics technologies impact deep learning-based gene expression prediction from histology images.

Techniques: Gene Expression, Labeling

Journal: bioRxiv

Article Title: Impact of Data Quality on Deep Learning Prediction of Spatial Transcriptomics from Histology Images

doi: 10.1101/2025.09.04.674228

Figure Lengend Snippet: A. Histogram showing the distribution of normalized rMSE for gene expression predictions using Visium and Xenium data. The dotted vertical line denotes the mean rMSE, and the solid curved line traces the density estimate. Results are computed on the test set and represent the average performance across five independently trained models. B. Scatterplot comparing the normalized rMSE of predictions from Visium and Xenium data, based on the test set and averaged over five models. The gray dotted line denotes x=y.

Article Snippet: Here, we investigate how variation in molecular and image data quality stemming from differences in imaging (Xenium) versus sequencing (Visium) spatial transcriptomics technologies impact deep learning-based gene expression prediction from histology images.

Techniques: Gene Expression

Journal: bioRxiv

Article Title: Impact of Data Quality on Deep Learning Prediction of Spatial Transcriptomics from Histology Images

doi: 10.1101/2025.09.04.674228

Figure Lengend Snippet: A. Histogram of Pearson correlation coefficients for gene expression predictions using Visium and Xenium data with the Visium image. The dotted vertical line denotes the mean PCC, and the solid curved line traces the density estimate. B. Scatterplot comparing PCC values from Visium and Xenium data with the Visium image on the test set, averaged across five models. The gray dotted line denotes x=y. C. Histogram of PCC values for predictions using Visium and Xenium data with the Xenium image. The dotted vertical line denotes the mean PCC, and the solid curved line traces the density estimate. D. Scatterplot comparing PCC values from Visium and Xenium data with the Xenium image on the test set, averaged across five models. The gray dotted line denotes x=y. E. Scatterplot comparing PCC values between Xenium, an increasing amount of sparsity in the Xenium dataset, and the Visium results on the test and replicate 2 Xenium data. The dotted line indicates the dataset used, and error bars represent the standard error across five runs. The histogram below denotes the total number of genes used to calculate the mean PCC. F. Scatterplot comparing PCC values between Xenium, an increasing amount of Poisson noise in the Xenium dataset, and the Visium results on the test and replicate 2 Xenium data. The dotted line indicates the dataset used, and error bars represent the standard error across five runs. G. Scatterplot comparing PCC values between Visium, various imputation methods on the Visium dataset, and the Xenium results on the test and replicate 2 Xenium data. The dotted line indicates the dataset used, and error bars represent the standard error across five runs.

Article Snippet: Here, we investigate how variation in molecular and image data quality stemming from differences in imaging (Xenium) versus sequencing (Visium) spatial transcriptomics technologies impact deep learning-based gene expression prediction from histology images.

Techniques: Gene Expression

Journal: bioRxiv

Article Title: Impact of Data Quality on Deep Learning Prediction of Spatial Transcriptomics from Histology Images

doi: 10.1101/2025.09.04.674228

Figure Lengend Snippet: Scatterplots of normalized rMSE for models trained on varied molecular inputs, evaluated on the held-out test set and averaged across five independent runs, using (A) the Visium histology image and (B) the Xenium histology image. The gray dotted line denotes x=y.

Article Snippet: Here, we investigate how variation in molecular and image data quality stemming from differences in imaging (Xenium) versus sequencing (Visium) spatial transcriptomics technologies impact deep learning-based gene expression prediction from histology images.

Techniques:

Journal: bioRxiv

Article Title: Impact of Data Quality on Deep Learning Prediction of Spatial Transcriptomics from Histology Images

doi: 10.1101/2025.09.04.674228

Figure Lengend Snippet: Violin plots of the per-patch fraction of zero counts in Visium and Xenium molecular data. The shape of each violin reflects the density of values along the y-axis, and the overlaid boxplot indicates the median and the 25th and 75th percentiles.

Article Snippet: Here, we investigate how variation in molecular and image data quality stemming from differences in imaging (Xenium) versus sequencing (Visium) spatial transcriptomics technologies impact deep learning-based gene expression prediction from histology images.

Techniques:

Journal: bioRxiv

Article Title: Impact of Data Quality on Deep Learning Prediction of Spatial Transcriptomics from Histology Images

doi: 10.1101/2025.09.04.674228

Figure Lengend Snippet: A. Histogram showing the distribution of Pearson correlation coefficients for gene expression predictions using Visium data with the Visium and Xenium images. The dotted vertical line denotes the mean PCC, and the solid curved line traces the density estimate. Results are computed on the test set and represent the average performance across five independently trained models. B. Scatterplot comparing the Pearson correlation coefficients of predictions from Visium data with the Visium and Xenium images, based on the test set and averaged over five models. The gray dotted line denotes x=y. C. Histogram showing the distribution of Pearson correlation coefficients for gene expression predictions using the Xenium data with the Visium and Xenium image. The dotted vertical line denotes the mean PCC, and the solid curved line traces the density estimate. Results are computed on the test set and represent the average performance across five independently trained models. D. Scatterplot comparing the Pearson correlation coefficients of predictions from Xenium data with the Visium and Xenium image, based on the test set and averaged over five models. The gray dotted line denotes x=y. E. Scatterplot of mean Pearson correlation coefficients on both the test set and the Replicate 2 Xenium section, comparing the Xenium, Xenium images with increasing Gaussian blur, and Visium results (all applied with the same blur levels). The dotted line indicates the dataset used, and error bars represent the standard error of the mean across five independent model runs. F. Grad-CAM heatmaps for two select genes: CD4 (T-cell marker) and PDGFRA (fibroblast marker).

Article Snippet: Here, we investigate how variation in molecular and image data quality stemming from differences in imaging (Xenium) versus sequencing (Visium) spatial transcriptomics technologies impact deep learning-based gene expression prediction from histology images.

Techniques: Gene Expression, Marker

Journal: bioRxiv

Article Title: Impact of Data Quality on Deep Learning Prediction of Spatial Transcriptomics from Histology Images

doi: 10.1101/2025.09.04.674228

Figure Lengend Snippet: Scatterplots of normalized RMSE for models trained on varied image inputs, evaluated on the held-out test set and averaged across five independent runs, using (A) the Visium molecular data and (B) the Xenium molecular data. The gray dotted line denotes x=y.

Article Snippet: Here, we investigate how variation in molecular and image data quality stemming from differences in imaging (Xenium) versus sequencing (Visium) spatial transcriptomics technologies impact deep learning-based gene expression prediction from histology images.

Techniques:

Journal: bioRxiv

Article Title: Impact of Data Quality on Deep Learning Prediction of Spatial Transcriptomics from Histology Images

doi: 10.1101/2025.09.04.674228

Figure Lengend Snippet: A. Histogram showing the distribution of Pearson correlation for gene expression predictions using Visium and Xenium data. The dotted vertical line denotes the mean rMSE, and the solid curved line traces the density estimate. Results are computed on the test set and represent the average performance across five independently trained models. B. Scatterplot comparing the Pearson correlation of predictions from Visium and Xenium data, based on the test set and averaged over five models. The gray dotted line denotes x=y. C. Histogram showing the distribution of normalized rMSE for gene expression predictions using Visium and Xenium data. The dotted vertical line denotes the mean rMSE, and the solid curved line traces the density estimate. Results are computed on the test set and represent the average performance across five independently trained models. B. Scatterplot comparing the normalized rMSE of predictions from Visium and Xenium data, based on the test set and averaged over five models. The gray dotted line denotes x=y.

Article Snippet: Here, we investigate how variation in molecular and image data quality stemming from differences in imaging (Xenium) versus sequencing (Visium) spatial transcriptomics technologies impact deep learning-based gene expression prediction from histology images.

Techniques: Gene Expression

Journal: Nature Communications

Article Title: Spatial transcriptomic analysis drives PET imaging of tight junction protein expression in pancreatic cancer theranostics

doi: 10.1038/s41467-024-54761-6

Figure Lengend Snippet: A Overview of the project. Parts created with BioRender.com. B H&E of PDAC tumor excised from four patients (bottom row) and their corresponding spatial transcriptomic spots (top row) arranged based on increasing CLDN4 expression. The spatial transcriptomes between the PDAC slices were integrated based on canonical correlation analysis (Seurat), clustered, projected on the Uniform Manifold Approximation and Projection (UMAP) dimension in ( C ), and projected onto their histology slices. Scale bar = 3 mm. C Leiden clustering of PDAC transcriptome projected on UMAP space and heatmap of Leiden clusters with genes selected from the PDAC data set of the cancer cell surfaceome with a specification of expression on at least 80% of cells. Scale represents z-score of log-normalized gene counts. Cluster identities were determined with one-sided hypergeometric t tests of all clusters in PDAC with respect to cell type signatures detailed in refs. , , which were derived from scRNA and plotted as a dotplot. D Differential expression of Cluster 1 versus all remaining clusters and versus samples of unaffected pancreas surrounding precancerous IPMN. E Cancer surfaceome score based on average normalized and scaled gene expression of all cells and all cancer cell surfaceome pancreatic cancer genes in each cluster. F Pearson’s correlation similarity matrix of select cancer cell surfaceome genes based on the PDAC spatial transcriptome. G Pearson’s correlation of all spots for S100P vs key markers of interest. H KEGG and REACTOME pathway enrichment of Cluster 1. Each calculation was based on n = 7 samples. For box plots, the center is the median and the lower and upper bound of the box are 25% and 75% of the distribution, respectively. The lower whisker is the lower 25% −1.5 x interquartile range (IQR). The upper whisker is the upper 75% + 1.5 x IQR. Differential expression analysis in ( D ) and ( E ) were based on non-parametric Wilcoxon rank sum test, which is a default setting in Seurat’s FindMarkers function. Gene enrichment analysis in ( H ) was done with 1-sided hypergeometric t test based on clusterProfiler package. Source data are provided as a Source Data file.

Article Snippet: Formalin-fixed, paraffin-embedded (FFPE) human pancreatic cancer tissue sections were placed on spatial transcriptomic slides for spatial transcriptomic sequencing based on the 10x FFPE workflow (10x Genomics, Pleasanton, CA).

Techniques: Expressing, Derivative Assay, Quantitative Proteomics, Gene Expression, Whisker Assay

Journal: Nature Communications

Article Title: Spatial transcriptomic analysis drives PET imaging of tight junction protein expression in pancreatic cancer theranostics

doi: 10.1038/s41467-024-54761-6

Figure Lengend Snippet: A Summary of the IPMN study. B H&E slices of IPMN tumor excised from three patients (bottom two rows) and their corresponding spatial transcriptomic spots (top two rows) arranged based on increasing CLDN4 expression. The spatial transcriptomes between the IPMN slices were merged, clustered, and projected on the Uniform Manifold Approximation and Projection (UMAP) dimension in ( C ) and projected to their H&E slices. Scale bar = 3 mm. C Leiden clustering of IPMN transcriptome projected on UMAP space. Dot plot of one-sided hypergeometric t test results of all clusters in IPMN based on cell type signatures. D Heatmap of Leiden clusters with genes selected from the PDAC cell surfaceome atlas with greater than 80% expression across sequenced PDAC cells. Scale represents z-score of log-normalized gene counts. E – G Differential expression of Cluster 6 ( E ) and Cluster 12 ( F ) versus all remaining clusters and G ) Cluster 12 versus Cluster 6. H Cancer surfaceome score based on average normalized and scaled gene expression of all cells and all cancer cell surfaceome PDAC genes in each cluster. I Pearson’s correlation similarity matrix of select cancer cell surfaceome genes based on the IPMN spatial transcriptome. J Pearson’s correlation of all spots for S100P vs key markers of interest. K KEGG pathway enrichment of Cluster 6 and Cluster 12. Each calculation is based on n = 18 samples. For box plots, the center is the median and the lower and upper bound of the box are 25% and 75% of the distribution, respectively. The lower whisker is the lower 25% − 1.5 x interquartile range (IQR). The upper whisker is the upper 75% + 1.5 x IQR. Differential expression analysis in ( E ), ( F ) and ( G ) were based on non-parametric Wilcoxon rank sum test, which is a default setting in Seurat’s FindMarkers function. Gene enrichment analysis in ( K ) was done with hypergeometric t test based on clusterProfiler package. Source data are provided as a Source Data file.

Article Snippet: Formalin-fixed, paraffin-embedded (FFPE) human pancreatic cancer tissue sections were placed on spatial transcriptomic slides for spatial transcriptomic sequencing based on the 10x FFPE workflow (10x Genomics, Pleasanton, CA).

Techniques: Expressing, Quantitative Proteomics, Gene Expression, Whisker Assay