Predictive and robust gene selection for spatial transcriptomics

Abstract

Fluorescence in situ hybridization (FISH) is a widely used method for visualizing gene expression in cells and tissues. A key challenge is determining a small panel of genes (typically less than 1% of the genome) to probe in a FISH experiment that are most informative about gene expression, either in general or for a specific experimental objective. We introduce predictive and robust probe selection (PROPOSE), a method that uses deep learning to identify informative marker genes using data from single-cell RNA sequencing (scRNA-seq). Using datasets spanning different brain regions, species, and scRNA-seq technologies, we show that our method reliably identifies gene panels that provide more accurate prediction of the genome-wide expression profile, thereby capturing more information while using fewer probes. Furthermore, PROPOSE can be readily adapted to meet specific experimental goals, such as classifying cell types or discerning neuronal electrical properties from scRNA-seq data. Finally, we demonstrate using a recent MERFISH dataset that PROPOSE’s binarization of gene expression levels enables models trained on scRNA-seq data to generalize with input data obtained via FISH, despite the complex domain shift between these technologies.

Multi-task learning from multimodal single-cell omics with Matilda

“…To evaluate the top features selected by Matilda across multiple data modalities and those selected from RNA modality using popular methods such as t -test and limma ( 7 ), and those specifically designed for scRNA-seq (e.g. MAST ( 8 ), ROC), and recently proposed deep learning feature selection methods, including PROPOSE ( 35 ) and scCapsNet ( 41 ), we compared their utility in classifying each cell type in each dataset. We found that cell-type-specific features selected by Matilda from multiple data modalities on average resulted in more accurate discrimination of their respective cell types as shown by the scatter plot and the overall rankings of methods in each dataset (Figure 5E and Supplementary Figure S8 ).…”

“…The PROPOSE procedure ( 35 ) was used for feature selection of RNA modality from all datasets. Following the author's pipeline ( https://github.com/iancovert/propose ), the raw count matrices were first binarized to {0,1} according to the sign of the values and then used for model training using the ‘PROPOSE’ function in propose package (Github version 41fd568) with the number of marker genes as 100 and other parameters as default.…”

stMMR: accurate and robust spatial domain identification from spatially resolved transcriptomics with multi-modal feature representation

“…The most crucial aspect is the integration of histological imaging information with gene expression data through spatial location. In SRT, gene expression data suffers from issues of sparsity and zero inflation, which are key factors that interfere with downstream analysis [33,77]. Previous research has shown that histological imaging information can predict gene expression data [30][31][32].…”

“…SRT data are characterized by high sparsity, discreteness, and variance greater than the mean, specifically manifested as a high number of genes expressed at zero (zero inflation) [33]. Previous research has found that the zero-inflated negative binomial (ZINB) distribution can effectively characterize gene expression in SRT [34].…”

Functional Diversity of Memory CD8 T Cells is Spatiotemporally Imprinted

“…To study the relationship of gene-expression differentiation programs of SI T RM to anatomical location in situ , we performed spatial transcriptomic profiling (Xenium,10X Genomics) 17,18 on SI from male mice adoptively transferred with female P14 CD8 T cells and analyzed over the course of LCMV infection (days 6, 8, 30, and 90). A 350-plex target gene panel was designed using a reference single nuclei RNAseq dataset to inform a prioritization algorithm based on predictive deep learning 19 ( Extended Data Fig. 1b ).…”