SpaLSTF: Diffusion-based generative model with BiLSTM and XCA-Transformer for spatial transcriptomics imputation.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatial transcriptomics (ST) technologies offer insights into gene expression patterns within tissues.
• Current ST methods face limitations in gene detection and expression coverage.
• Existing computational imputation methods using scRNA-seq data do not fully capture cellular temporal dependencies or gene regulatory mechanisms.

Purpose of the Study:

• To develop a novel computational method, SpaLSTF, for enhancing spatial transcriptomics gene expression data.
• To address limitations in current ST data imputation by incorporating temporal dependencies and gene regulatory insights.
• To improve the accuracy and completeness of spatial gene expression analysis.

Main Methods:

• SpaLSTF employs a conditional diffusion model guided by scRNA-seq data.
• A dual Markov process is used to capture gene expression relationships via noise perturbation and denoising.
• Bidirectional long short-term memory (BiLSTM) networks model cell state dependencies, and a cross-covariance attention Transformer (XCA-Transformer) computes attention coefficients.
• A variational lower bound (VLB) objective with KL divergence and mean squared error loss ensures accurate noise distribution.

Main Results:

• SpaLSTF demonstrated superior performance across twelve cross-platform datasets compared to seven state-of-the-art methods.
• The method achieved higher accuracy in gene expression imputation.
• SpaLSTF improved cell population identification and preserved spatial structures effectively.

Conclusions:

• SpaLSTF significantly enhances spatial transcriptomics data imputation and analysis.
• The proposed method offers a more comprehensive approach to understanding spatial gene expression.
• SpaLSTF represents a advancement in computational tools for spatial biology research.