Abstract
The analysis of spatial transcriptomics is hindered by high noise levels and missing gene measurements, challenges that are further compounded by the higher cost of spatial data compared to traditional single-cell data. To overcome this challenge, we introduce spRefine, a deep learning framework that leverages genomic language models to jointly denoise and impute spatial transcriptomic data. Our results demonstrate that spRefine yields more robust cell- and spot-level representations after denoising and imputation, substantially improving data integration. In addition, spRefine serves as a strong framework for model pre-training and the discovery of novel biological signals, as highlighted by multiple downstream applications across datasets of varying scales. Notably, spRefine enhances the accuracy of spatial ageing clock estimations and uncovers new aging-related relationships associated with key biological processes, such as neuronal function loss, which offers new insights for analyzing ageing effect with spatial transcriptomics.
Introduction
Spatial transcriptomics (ST) has revolutionized biology by enabling researchers to observe gene activity within its native tissue context. However, the technology is frequently limited by high noise levels and "dropout"—where many gene measurements are missing. Traditionally, mitigating these issues required a "reference" dataset, typically single-cell RNA-seq (scRNA-seq), to act as a gold-standard template. However, the availability of high-quality references is a major bottleneck, particularly for rare tissues, specific disease states, or non-model organisms.
To address these challenges, we introduce spRefine, a deep learning framework that eliminates the need for external references. Instead, it leverages the "internal logic" of the genome itself—captured by Genomic Language Models (gLMs)—to denoise and impute spatial transcriptomic data.
Methods
The innovation of spRefine lies in its reference-free architecture, which treats the problem as a matrix factorization task powered by a coupled Auto-Encoder.
The framework consists of several key components:
- Genomic Language Model (gLM) Priors: spRefine uses embeddings from Enformer, a gLM trained on DNA sequences. These embeddings capture fundamental regulatory relationships between genes based on their genomic context, serving as a universal prior.
- Decomposed Auto-Encoder: The model simultaneously learns two sets of embeddings: Spot/Cell Embeddings derived from measured spatial expression, and Dataset-Specific Gene Embeddings refined from the gLM priors to fit the specific biological context.
- Reconstruction: The final denoised and imputed matrix is generated by multiplying these two embeddings.
- Pre-training Mode: A masking and downsampling strategy allows the model to learn robust representations across massive, multi-sample spatial datasets.
Results
We validated spRefine across a diverse array of platforms, including Xenium, seqFISH, osmFISH, and 10X Visium.
Benchmarking and Clustering Accuracy
spRefine consistently outperformed state-of-the-art methods such as VISTA, ENVI, and Tangram. Across four different ST platforms, spRefine demonstrated improvements in clustering accuracy (Average Score) ranging from 0.4% to 14%. On the SpatialLIBD dataset, spRefine significantly reduced batch effects while preserving biological signals more effectively than both spatial-specific and single-cell-specific baselines.
Biological Discovery and Clinical Relevance
Beyond data cleaning, spRefine enables the discovery of complex biological signals. In a breast cancer dataset, spRefine uncovered 19 significant cell-cell interactions, compared to only 3 found in raw data. Furthermore, the framework allowed for the construction of a spatial aging clock, identifying key interactions between inhibitory neurons and microglia as drivers of brain aging—signals that were previously obscured by noise.
Clinical validation using TCGA datasets revealed that marker genes identified through spRefine imputation were superior predictors of patient survival. Notably, the spRefine pipeline was 37.2% faster than processing raw data for these downstream tasks.
Discussion
spRefine demonstrates that DNA sequences contain sufficient latent information to serve as a substitute for expensive scRNA-seq references in ST data refinement. While performance is currently tied to the quality of underlying gLMs, the framework represents a significant step toward a foundation model for spatial biology. Future work will focus on integrating larger multi-omic datasets and improving the availability of well-annotated spatial data to further enhance pre-training capabilities.
Summary updated