Single-cell RNA sequencing has become one of the most powerful tools for studying complex biological systems. By measuring gene expression in individual cells, researchers can identify different cell types, uncover rare cell populations, and better understand how tissues function in health and disease.

A critical step in analyzing single-cell RNA sequencing data is clustering. Clustering groups similar cells together, helping researchers identify distinct cell populations for downstream analysis. However, this task is often challenging because single-cell datasets are highly complex.

Researchers from South China Agricultural University have developed a new artificial intelligence framework called scMVAF to improve clustering performance for single-cell RNA sequencing data.

Illustration of scMVAF framework

(A) The overall framework. The preprocessed  is first generated by multiple times down-sampling to generate  views, each of which is independently fed into the ZINB-based model for denoising projection to obtain the embedding features of a single view, followed by adaptive weighted fusion. A uniform target distribution  is obtained by the clustering algorithm on the fused features and  is the soft clustering distribution for each view. Finally, the uniform objective distribution  is used to guide the view of private soft distribution  for clustering optimization. (B) Adaptive weighted fusion Module.

One of the major difficulties in single-cell RNA sequencing analysis is that the data are high-dimensional and sparse. Many genes appear to have zero expression values in individual cells, but some of these zeros are caused by technical limitations rather than true biological absence. This can make it difficult to accurately determine relationships between cells.

Although many clustering methods have been developed, most analyze the data from a single perspective. As a result, they may fail to capture all of the biological information contained within the dataset.

The scMVAF framework addresses this limitation by using a multi-view learning strategy. Rather than examining the data through a single representation, the method creates multiple views of the same dataset by sampling different sets of features. Each view captures slightly different information about the cells.

To analyze these views, the researchers employed an autoencoder based on a denoising zero-inflated negative binomial model. Autoencoders are a type of deep learning algorithm that can compress complex data into lower-dimensional representations while preserving important information.

The model then combines information from the multiple cell views using a fusion module that integrates the embeddings into a unified feature space. This allows the algorithm to identify relationships between cells that may not be visible from any single view alone.

An additional feature of scMVAF is its iterative learning strategy. The model generates provisional cell labels, known as pseudo-labels, and uses them to continuously refine the learned cell representations. This process helps improve clustering accuracy over time.

The researchers evaluated scMVAF using 16 real-world single-cell RNA sequencing datasets and compared its performance with eight advanced clustering methods. Across these datasets, scMVAF consistently achieved superior clustering performance.

The results demonstrate how artificial intelligence and multi-view learning can improve the interpretation of increasingly large and complex single-cell datasets. As single-cell RNA sequencing continues to expand across biomedical research, tools such as scMVAF may help researchers more accurately identify cell populations and uncover new biological insights.

Availability – the code script can be obtained at https://github.com/LQXLE/scMVAF/