A systematic review of algorithmic challenges in noise modelling for multimodal single-cell

Multimodal single-cell profiling technologies generate heterogeneous and high-dimensional molecular measurements across multiple cellular modalities. Although several reviews have discussed multimodal single-cell integration broadly, technical noise modelling strategies have not been systematically synthesized. These datasets are affected by substantial technical noise, including sparsity, dropout, batch effects, and ambient contamination, which complicate multimodal integration and downstream biological interpretation. Consequently, biological and technical sources of variation become statistically confounded, complicating latent representation learning and multimodal data integration. This systematic review evaluates algorithmic approaches that explicitly model or mitigate technical noise in multimodal single-cell datasets. A total of twenty-four latest articles discussing multimodal datasets, including RNA-ATAC, RNA-protein, and spatial transcriptomics integration, published from 2018 to 2026 were considered for the analysis under PRISMA 2020 guidelines. The inclusion criteria cover articles proposing new algorithms that included probabilistic models of noise or latent variable inference; however, studies discussing only unimodal data, bulk omics, or not specifying how noise was handled were excluded from the study. Probabilistic generative models and hybrid deep learning architectures combining variational inference with attention-based or transformer-derived components were the most reported methodological frameworks. Variational autoencoder-based approaches were frequently associated with improved denoising, clustering consistency, and latent representation learning across benchmark datasets. However, these improvements may not fully reflect performance under biologically realistic conditions such as rare cell populations and compositional batch effects.