Answers are grounded in this paper and cite their sources.
MOTIVATION: Metacells have emerged as an important analytical strategy for addressing the computational barriers posed by the scale and complexity of large-scale single-cell omics data, while preserving biologically relevant information and improving analytical interpretability. However, existing metacell inference methods usually have limited biological interpretability, rely on deterministic assignments that insufficiently capture continuous state transitions, and may lose potential biological signals due to prior feature selection or dimensionality reduction. RESULTS: We propose TMCells, an interpretable metacell identification algorithm based on a probabilistic generative framework. TMCells introduces the concept of topic modeling and uses a Gaussian mixture model as the prior for topic distribution, defining metacells as statistical distribution centers in the latent topic space. By modeling the cell generation process, TMCells enables the simultaneous learning of topic representations and metacell partitions in an end-to-end manner, capturing the continuity and uncertainty in cell state transitions. TMCells effectively identifies high-quality metacells, eliminating technical noise while preserving the original biological heterogeneity, thereby enhancing downstream analyses such as differential expression, rare cell type identification, and trajectory inference. Furthermore, TMCells provides an intuitive interpretation of cellular heterogeneity through latent topics associated with biologically meaningful regulatory programs, offering a generative perspective for dissecting regulatory dynamics during cell-state transitions. AVAILABILITY AND IMPLEMENTATION: The TMCells along with detailed documentation is freely available at https://github.com/Jinsl-lab/TMCells with DOI: 10.5281/zenodo.24198840. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.
No takes yet. Share an insight, caveat, or question.
Qiu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: