Abstract
A prominent trend in single-cell transcriptomics is providing spatial context alongside a characterization of each cell’s molecular state. This typically requires targeting an a priori selection of genes, often covering less than 1% of the genome, and a key question is how to optimally determine the small gene panel. We address this challenge by introducing a flexible deep learning framework, PERSIST, to identify informative gene targets for spatial transcriptomics studies by leveraging reference scRNA-seq data. Using datasets spanning different brain regions, species, and scRNA-seq technologies, we show that PERSIST reliably identifies panels that provide more accurate prediction of the genome-wide expression profile, thereby capturing more information with fewer genes. PERSIST can be adapted to specific biological goals, and we demonstrate that PERSIST’s binarization of gene expression levels enables models trained on scRNA-seq data to generalize with to spatial transcriptomics data, despite the complex shift between these technologies.
Introduction
Spatial transcriptomics (ST) technologies, such as Fluorescence In Situ Hybridization (FISH), allow researchers to map gene activity within tissue architecture. However, unlike whole-genome sequencing, most FISH-based methods are restricted to a small "panel" of genes—typically 10 to 250—due to technical and economic constraints. Current gene selection methods are often ad hoc, relying on well-known markers or high expression levels. These approaches frequently ignore gene-gene redundancies and fail to account for the "domain shift" between reference single-cell RNA sequencing (scRNA-seq) data and the target spatial data.
Methods
We present PERSIST (PredictivE and Robust gene SelectIon for Spatial Transcriptomics), a deep learning framework that treats gene selection as a differentiable feature selection problem.
The core of PERSIST involves several technical innovations:
- Binary Mask Layer: Utilizing a continuous relaxation of discrete variables (Concrete distribution), PERSIST uses gradient descent to "turn off" uninformative genes, converging on a user-specified panel size.
- Expression Binarization: To mitigate the domain shift between scRNA-seq and FISH, PERSIST binarizes expression levels during training. This focuses the model on the presence or absence of transcripts, a signal more stable across platforms than raw counts.
- Hurdle Loss Function: To address "dropouts" in scRNA-seq data, a hurdle loss separately predicts the probability of expression and the expression level itself.
- Flexible Training Modes: The framework supports unsupervised reconstruction of the genome-wide profile or supervised prediction of specific targets like cell types or electrophysiological properties.
The model was validated using diverse datasets, including SmartSeq v4 (SSv4) mouse neurons, 10X Genomics human neurons, and multimodal Patch-seq data.
Results
PERSIST consistently identifies more informative gene panels than existing methods. In genome-wide reconstruction tasks, it outperformed standard tools like Seurat and GeneBasis, particularly for small panel sizes.
Key findings include:
- Efficiency: PERSIST achieves high classification accuracy with minimal genes. For example, with only 64 genes, it reached 74% accuracy in classifying 113 distinct mouse cell types.
- Robust Generalization: Binarization proved essential for transferring models from scRNA-seq to spatial data. When applied to MERFISH data, PERSIST-selected genes enabled high-accuracy imputation of the remaining genome.
- Functional Prediction: PERSIST successfully selected genes predictive of non-transcriptomic properties, such as a cell's electrical firing patterns, demonstrating its utility for cross-modal mapping.
Discussion
While PERSIST offers significant improvements in gene selection, it has limitations. The stochastic nature of gradient descent means selected panels can vary slightly between runs, although performance remains consistent. Additionally, the framework's effectiveness is tied to the quality and representativeness of the reference scRNA-seq data. Future developments may include incorporating expert biological constraints and optimizing for other modalities like chromatin accessibility. PERSIST represents a robust, automated solution for designing the next generation of spatial transcriptomics experiments.
Summary updated