Abstract
Spatially Resolved Transcriptomics (SRT) provides critical insights into the spatial organization of gene expression within tissues. However, technical limitations often result in high levels of noise and data sparsity, such as "dropouts." We present DUSTED (Dual-Attention Enhanced Spatial Transcriptomics Denoiser), a self-supervised deep learning framework that leverages a dual-attention mechanism within a graph autoencoder to denoise SRT data. By simultaneously modeling gene-specific noise patterns and spatial dependencies, DUSTED recovers biological signals more accurately than existing methods without requiring auxiliary histological images.
Introduction
Spatially Resolved Transcriptomics (SRT) has revolutionized biological research by mapping gene expression to physical tissue locations. Despite its potential, SRT data is frequently compromised by technical noise, resulting in sparse matrices and blurred spatial patterns. Existing denoising tools often fall short by ignoring spatial context, assuming uniform noise across all genes, or relying heavily on high-resolution histological images (e.g., H&E stains) that may not be available or easily aligned.
Methods: The Dual-Attention Framework
DUSTED addresses these challenges using a Graph Autoencoder (GAE) architecture enhanced by a unique dual-attention mechanism:
- Gene Channel Attention (GCA): Inspired by computer vision, this module assigns unique weights to different genes. This recognizes that biological and technical factors (like PCR amplification bias) cause some genes to be naturally noisier or more variable than others.
- Graph Attention (GAT): This module utilizes the physical neighborhood of each cell or spot to interpolate expression levels, learning from spatially related neighbors to fill in missing data without over-blurring tissue boundaries.
The model employs probabilistic modeling, fitting data to Negative Binomial (NB) or Zero-Inflated Negative Binomial (ZINB) distributions, which are statistically better suited for genomic count data. As a self-supervised model, DUSTED learns directly from the noisy input, ensuring high flexibility across different datasets.
Key Results and Biological Validation
The efficacy of DUSTED was validated across synthetic and real-world datasets, including Human Ovarian Cancer (HOCWTA) and the Human Dorsolateral Prefrontal Cortex (DLPFC).
Superior Accuracy and Fidelity
On simulated datasets, DUSTED achieved a 69.6% reduction in Mean Square Error (MSE) compared to raw data, consistently outperforming established baselines like MAGIC, DCA, and STAGATE. In the ovarian cancer study, DUSTED improved the correlation between gene expression (PTPRC) and protein expression (CD45) by approximately 8.5% over the next best method, demonstrating its ability to recover true biological signals.
Enhanced Spatial Pattern Recovery
Using the DLPFC benchmark, DUSTED demonstrated superior performance in recovering distinct tissue structures. It achieved a median Adjusted Rand Index (ARI) of 0.515, the highest among tested methods, indicating more accurate identification of brain layers.
Visual analysis shows that DUSTED produces sharper, more biologically plausible boundaries between tissue regions compared to the over-smoothed or noisy outputs of other tools. Quantitative metrics like the Global Moran’s Index further confirm that DUSTED better preserves spatial autocorrelation. Ablation studies confirmed that the Gene Channel Attention (GCA) is critical; removing it significantly degraded performance, proving that accounting for gene-specific noise is essential for genomic denoising.
Conclusion
DUSTED represents a significant advancement in SRT data processing. By integrating gene-specific and spatial attention, it provides a robust, image-independent solution for denoising. Future developments will focus on improving computational scalability for massive datasets and automating hyperparameter selection to enhance user accessibility.
Summary updated