As single-cell RNA sequencing (scRNA-seq) becomes increasingly important for understanding cell diversity and disease, scientists are faced with a big challenge: how do you accurately group thousands, or even millions, of cells based on their gene expression profiles?
To do this, researchers use clustering algorithms, which group cells into populations that share similar gene activity. However, different clustering methods often give different answers because each one makes its own assumptions about the data. This inconsistency can lead to confusion or even missed discoveries, especially when studying complex tissues or diseases.
Researchers at the University of Bordeaux have developed a new tool to deal with this issue: scEVE (Single-cell Ensemble clustering based on Variation in Estimations). Unlike traditional ensemble approaches that try to “smooth out” differences between clustering results, scEVE takes a unique path, it embraces those differences. By analyzing where clustering methods disagree, scEVE gains insight into data complexity and uncertainty. This allows it to provide clustering results that not only make sense biologically but also include confidence scores and allow researchers to explore the data at different levels of detail.
Overview of the scEVE algorithm
(A) Base clusters with multiple clustering methods, and biological characterization of robust clusters. A single-cell count matrix is preprocessed and fed to multiple clustering methods to generate base clusters. From these, robust and leftover clusters are predicted (see panel C) and characterized. As a result, biologically distinct clusters are identified and returned. They will eventually be recursively subdivided, provided this increases the robustness of the clustering analysis. (B) Pairwise similarity between base clusters. Cells at the intersection of two base clusters x and y are used to quantify the similarity Sx,y. If this similarity is strong (Sx,y > Slim), it is leveraged downstream to identify robust clusters. (C) Robust clusters from pairwise similarities. Cells grouped together by multiple clustering methods form a robust cluster z. Its robustness Rz is quantified by exploiting the pairwise similarities computed previously (see panel B). If, and only if, this robustness is high (Rz > Rlim), the cluster will be characterized downstream. The remaining cells form a leftover cluster λ, with Rλ = 0, which will also be characterized downstream.
The team tested scEVE on 15 real-world datasets and up to 1,200 synthetic datasets, showing that it consistently outperforms existing methods. Most importantly, scEVE addresses two critical challenges in scRNA-seq data science: (1) quantifying uncertainty in how cells are grouped and (2) providing multi-resolution views that can reveal both broad cell types and subtle subpopulations.
For scientists working with scRNA-seq, especially in developmental biology, cancer research, or rare disease studies, scEVE offers a powerful new approach. By understanding why different algorithms disagree, it gives researchers a deeper, more flexible interpretation of the data and opens up new paths for discovery.
Availability – The scEVE algorithm is an instance of fEVE, our open framework for -omics ensemble clustering analyses. fEVE is implemented in an R package and is available in the following GitHub repository: https://github.com/yanisaspic/fEVE.
Asloudj Y, Mougin F, Thébault P. (2025) scEVE: a single-cell RNA-seq ensemble clustering algorithm capitalizing on the differences of predictions between multiple clustering methods. NAR Genomics and Bioinformatics 7(2): lqaf073. [article]
Related Posts
As single-cell RNA sequencing (scRNA-seq) becomes increasingly important for understanding cell diversity and disease, scientists are faced with a big challenge: how do you accurately group thousands, or even millions, of cells based on their gene expression profiles?
To do this, researchers use clustering algorithms, which group cells into populations that share similar gene activity. However, different clustering methods often give different answers because each one makes its own assumptions about the data. This inconsistency can lead to confusion or even missed discoveries, especially when studying complex tissues or diseases.
Researchers at the University of Bordeaux have developed a new tool to deal with this issue: scEVE (Single-cell Ensemble clustering based on Variation in Estimations). Unlike traditional ensemble approaches that try to “smooth out” differences between clustering results, scEVE takes a unique path, it embraces those differences. By analyzing where clustering methods disagree, scEVE gains insight into data complexity and uncertainty. This allows it to provide clustering results that not only make sense biologically but also include confidence scores and allow researchers to explore the data at different levels of detail.
Overview of the scEVE algorithm
(A) Base clusters with multiple clustering methods, and biological characterization of robust clusters. A single-cell count matrix is preprocessed and fed to multiple clustering methods to generate base clusters. From these, robust and leftover clusters are predicted (see panel C) and characterized. As a result, biologically distinct clusters are identified and returned. They will eventually be recursively subdivided, provided this increases the robustness of the clustering analysis. (B) Pairwise similarity between base clusters. Cells at the intersection of two base clusters x and y are used to quantify the similarity Sx,y. If this similarity is strong (Sx,y > Slim), it is leveraged downstream to identify robust clusters. (C) Robust clusters from pairwise similarities. Cells grouped together by multiple clustering methods form a robust cluster z. Its robustness Rz is quantified by exploiting the pairwise similarities computed previously (see panel B). If, and only if, this robustness is high (Rz > Rlim), the cluster will be characterized downstream. The remaining cells form a leftover cluster λ, with Rλ = 0, which will also be characterized downstream.
The team tested scEVE on 15 real-world datasets and up to 1,200 synthetic datasets, showing that it consistently outperforms existing methods. Most importantly, scEVE addresses two critical challenges in scRNA-seq data science: (1) quantifying uncertainty in how cells are grouped and (2) providing multi-resolution views that can reveal both broad cell types and subtle subpopulations.
For scientists working with scRNA-seq, especially in developmental biology, cancer research, or rare disease studies, scEVE offers a powerful new approach. By understanding why different algorithms disagree, it gives researchers a deeper, more flexible interpretation of the data and opens up new paths for discovery.
Availability – The scEVE algorithm is an instance of fEVE, our open framework for -omics ensemble clustering analyses. fEVE is implemented in an R package and is available in the following GitHub repository: https://github.com/yanisaspic/fEVE.
Asloudj Y, Mougin F, Thébault P. (2025) scEVE: a single-cell RNA-seq ensemble clustering algorithm capitalizing on the differences of predictions between multiple clustering methods. NAR Genomics and Bioinformatics 7(2): lqaf073. [article]
Related Posts
Recent Posts
Machine learning reveals how RNA splicing is regulated in individual cell types
Illumina releases SpliceAI2 to help advance rare disease research
Postdoctoral Fellow – Long-Read RNA Methods Development for Neurodegeneration
Novel biosensor uses DNA ‘barcodes’ to uncover disease signals in cells
GP-DHT uses single-cell RNA sequencing to map gene regulatory networks
Boehringer taps Envisagenics’ AI to mine RNA splicing for cancer targets
RNA-Seq Products & Services
Stay Connected