Single-cell RNA-seq data feature extraction using dual-depth model

In the realm of single-cell transcriptomic sequencing, deep generative models have proven invaluable in capturing gene expression features. Nevertheless, technical challenges have introduced a notable presence of missing values in the data, leading to the observed “dropout” phenomenon within the gene expression matrix. This phenomenon is characterized by numerous technical zero values, potentially stemming from data noise. To address this issue, interpolation algorithms leverage known values to infer and fill in these “dropout” occurrences, effectively mitigating data incompleteness and aiding in the preservation of biological information within the samples. Relevant studies suggest that interpolation algorithms play a crucial role in enhancing the reliability and completeness of data in the context of feature extraction within deep models. To contribute to this area, this research introduces scDIVAE, a framework encompassing two deep generative models. The first model is dedicated to interpolating gene data, sharing information among similar cells to eliminate noise and the “dropout” phenomenon. The second model employs a natural language topic model for data feature extraction. This methodology not only improves the clustering accuracy of deep generative models but also effectively eliminates batch effects.