Abstract
Single-cell RNA-sequencing data are inherently sparse and high-dimensional, with dropout events introducing a large number of missing values that hinder downstream analyses. To address the dual challenges of data imputation and dimensionality reduction, we propose an integrated recovery framework that combines probabilistic modeling with sparse optimization. The proposed method captures latent distribution patterns underlying gene expression, enabling the reconstruction of nonlinear dependencies among genes. Furthermore, sparsity constraints and heuristic search strategies are employed to enhance recovery accuracy, particularly in regions with low expression, while maintaining global expression consistency. We evaluate the framework on in-domain breast-cancer cohorts, on mask-augmented datasets with prescribed masking ratios and known references, and on 5 heterogeneous external datasets spanning distinct protocols and scales; across these settings, the method outperforms representative baselines in reconstruction error, structural fidelity, and recovery of biologically critical genes. These results highlight the effectiveness of jointly modeling distributional structure and sparsity for the reliable restoration of single-cell expression data.
Funding Information
  • National Natural Science Foundation of China (12326377)
  • Natural Science Foundation of Jilin Province (YDZJ202301ZYTS401)

This publication has 47 references indexed in Scilit: