scIDPMs: Single-cell RNA-seq imputation using diffusion probabilistic models

Single-cell RNA sequencing (scRNA-seq) technology is a high-throughput sequencing analysis method that enables the sequencing of mRNA in individual cells, thereby facilitating a more precise understanding of cellular gene expression and metabolic products. This approach reveals cell function and characteristics, making it widely applicable in biological research. However, scRNA-seq data often suffers from false zero values known as dropout events due to limitations in sequencing technology. These dropout events not only mask true gene expression levels but also significantly impact downstream analysis accuracy and reliability. To address this challenge, numerous computational approaches have been proposed for imputing missing gene expression values. Nevertheless, existing imputation methods struggle to fully capture the distribution of dropout values due to the high sparsity of scRNA-seq data and the complexity and randomness associated with gene expression patterns. Recently, probabilistic diffusion models have emerged as deep generative models capable of accurately restoring probability density distributions in domains such as image and audio processing. In this paper, we propose a method called scIDPMs, which utilizes conditional diffusion probabilistic models to impute scRNA-seq data. scIDPMs first identifies dropout sites based on the characteristics of cellular gene expression and then infers the dropout values by conditioning on the available gene expression values, which provide context information for the dropout values. To effectively capture the global features of gene expression profiles, scIDPMs employs a deep neural network with an attention mechanism to optimize the objective function. The performance of scIDPMs was evaluated using both simulated and real scRNA-seq datasets, and compared with eight other imputation methods. The experimental results clearly demonstrated that, in comparison to alternative approaches, scIDPMs exhibited exceptional performance in recovering biologically meaningful gene expression values and enhancing various downstream analyses.