The relationship between religious health fatalism, medication adherence, and health literacy among patients with hypertension: a cross-sectional study.

The rapid advancement of high-throughput sequencing technologies has enabled the simultaneous profiling of diverse molecular modalities at both bulk and single-cell levels, including transcriptomics, epigenomics, proteomics, and so on []. This has led to the accumulation of large-scale multi-omics datasets, offering unprecedented opportunities to dissect complex biological systems, identify disease subtypes, and characterize potential cell types []. However, multi-omics clustering remains challenging because omics profiles are often high-dimensional, sparse, heterogeneous, and noisy, with substantial variation in data quality across modalities and samples []. Bulk multi-omics data are often affected by batch effects, platform-specific biases, and cohort-level heterogeneity, whereas single-cell multi-omics data are further challenged by dropout events, low molecular capture efficiency, and uneven coverage across modalities. These factors can induce substantial reliability variation across features, modalities, and samples or cells, thereby compromising robust multi-omics clustering. Effective multi-omics clustering first requires a discriminative and robust embedding space that can preserve biologically meaningful structures while reducing the influence of technical noise and modality heterogeneity. To this end, many existing multi-omics embedding learning methods rely on reconstruction [], where neural networks are trained to recover the original multi-omics profiles from low-dimensional latent representations. Representative approaches include Autoencoders (AEs) [, ], Denoising Autoencoders (DAEs) [, ], and Variational Autoencoders (VAEs) [, ]. For single-cell data, reconstruction objectives based on the Zero-Inflated Negative Binomial (ZINB) distribution have further been introduced to account for dropout events and over-dispersion in single-cell RNA sequencing (scRNA-seq) count data []. Related studies have also demonstrated the value of deep learning, multi-omics integration, representation learning, and heterogeneity quantification in disease subtyping and cellular state analysis []. Recent advanced methods have further promoted multi-omics integration and clustering from different perspectives. MOFA and MOFA+ [, ] formulate multi-omics integration as a probabilistic factor analysis problem, learning interpretable low-dimensional factors that capture shared and modality-specific sources of variation across omics layers. MOSA [] adopts a deep generative strategy to synthesize missing molecular profiles and improve the completeness of multi-omics representations. For single-cell multi-omics data, scMDC [] combines ZINB-based deep representation learning with clustering optimization to jointly learn latent embeddings and cell clusters, while Matilda [] provides a neural multitask framework for multimodal single-cell omics analysis. More recently, scMHNN [] employs hypergraph neural networks and contrastive learning to model high-order relationships across transcriptomic, epigenomic, and proteomic modalities, and scMNMF [] introduces a joint nonnegative matrix factorization framework to couple cell clustering with feature selection.