ScEnsemble: weighted hypergraph ensemble clustering for single-cell RNA sequencing Authors: Beste Uncu, Idil Yavuz Published: 2026 ABSTRACT AbstractBackgroundSingle-cell RNA sequencing enables detailed profiling of cellular heterogeneity, with clustering serving as a critical step for identifying distinct cell populations. However, no single clustering algorithm consistently outperforms others across diverse datasets, creating uncertainty in robust cell population identification. Existing ensemble methods either treat all algorithms equally or employ simple filtering strategies, failing to account for varying solution quality across different datasets.ResultsWe present ScEnsemble, a weighted hypergraph ensemble clustering framework that integrates multiple base algorithms through quality-based weighting. ScEnsemble constructs a hypergraph where edges represent cluster co-assignments, weighted by internal validation indices including Silhouette coefficient, Calinski-Harabasz index, Davies-Bouldin index, and Dunn index. Multiple consensus algorithms partition the weighted hypergraph to produce final clusters, including CSPA variants with hierarchical clustering and community detection methods, MCLA with multiple consensus strategies, and hypergraph spectral clustering. Benchmarking across five scRNA-seq datasets demonstrates that the best ensemble configuration matched or exceeded the best individual algorithm in 23 of 25 metric–dataset combinations (92%). Quality-based weighting further improved upon unweighted ensembles in 21 of 25 combinations (84%). Biological validation on a breast cancer tumor microenvironment dataset demonstrated that ScEnsemble clusters correspond to known cell types as confirmed by independent gene-set scoring.ConclusionsScEnsemble provides a principled solution for scRNA-seq clustering that leverages algorithmic diversity rather than forcing selection of a single method. The framework enables researchers to optimize for either mathematical cluster quality or biological interpretability based on their analytical priorities, addressing a fundamental challenge in single-cell data analysis.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12859-026-06525-w.Keywords: Single-cell RNA sequencing, Ensemble clustering, Weighted hypergraph partitioning, Quality-based weighting, Multi-metric optimization Background Single-cell RNA sequencing has revolutionized our understanding of cellular heterogeneity, enabling researchers to profile gene expression at unprecedented resolution [1, 2]. Unlike bulk RNA sequencing, which averages gene expression across populations of cells, scRNA-seq reveals the transcriptomic landscape of individual cells, providing critical insights into cell type identification, developmental trajectories, and cellular responses to various stimuli [3, 4]. The rapid advancement of scRNA-seq technologies has generated massive datasets containing expression profiles for thousands to millions of cells. However, this wealth of data presents significant computational challenges, particularly in identifying distinct cell populations within heterogeneous samples [5]. The inherent complexity of scRNA-seq data, characterized by high dimensionality, dropout events, technical noise, and sparse expression patterns, necessitates sophisticated analytical approaches that can effectively navigate these challenges while preserving biological signal [6, 7]. Clustering analysis has emerged as a fundamental step in scRNA-seq data processing pipelines, serving as the foundation for downstream analyses including cell type annotation, trajectory inference, and differential expression analysis [8]. The primary objective is to group cells with similar transcriptomic profiles into biologically meaningful clusters corresponding to distinct cell types, developmental stages, or functional states. Numerous algorithms have been developed, each employing distinct mathematical frameworks. Graph-based methods construct cell-cell similarity networks and apply community detection algorithms; Seurat implements k-nearest neighbor graphs with Jaccard similarity refinement combined with Louvain or Leiden clustering [9, 10], demonstrating scalability to millions of cells through modularity optimization [11, 12]. Recent graph-based advances incorporate deep learning, with scGNN introducing graph neural network frameworks modeling heterogeneous gene expression patterns [13], and multi-view approaches like scMCGF integrating multiple graph representations to capture diverse cellular relationships [14]. Recent advances also focus on integrating multi-omics data using deep learning architectures to improve clustering resolution, as seen in models like scMultiNet [15]. Notably, hypergraph-based representations have emerged to address limitations of traditional pairwise networks by preserving higher-order information [16]. Hierarchical approaches organize cells into tree-like structures revealing multi-resolution relationships; CIDR addresses dropout events through Epanechnikov kernel-based imputation followed by principal coordinates analysis and hierarchical clustering [17], while SC3 employs consensus clustering across multiple transformations and parameters to achieve robust identification [18]. BackSPIN implements divisive hierarchical clustering iteratively partitioning data along principal components [19], and DendroSplit utilizes feature selection-based hierarchical clustering with modified Welch t-tests to determine optimal splitting points [20]. Contemporary hierarchical developments include self-supervised deep clustering for rare cell detection [21]. Partitioning methods divide cells by optimizing specific objective functions; SIMLR learns similarity matrices using multiple kernels followed by dimensionality reduction and k-means clustering [22], while specialized methods like pcaReduce integrate PCA with iterative k-means to create hierarchical structures [23], and RaceID combines k-medoids clustering with outlier detection mechanisms employing noise models to distinguish biological heterogeneity from technical artifacts [24]. Advanced partitioning algorithms like scAIDE facilitate rare cell type identification on global population scales [25]. Model-based approaches assume data originate from probability distribution mixtures; BISCUIT implements hierarchical Bayesian models simultaneously performing normalization, imputation, and clustering [26], TSCAN combines Gaussian mixture models with minimum spanning trees for cluster identification and pseudotemporal ordering [27], and Monocle integrates density-based clustering with trajectory inference through reverse graph embedding [28, 29]. Population-scale Bayesian models like BAMM-SC address droplet-based data from multiple individuals [30], while deep learning integration exemplified by scGMAI employs autoencoder-based imputation with Gaussian mixture models [31], and model-based deep embedding methods incorporate domain knowledge for constrained clustering [32]. Density-based methods identify high-density regions without geometric assumptions; GiniClust utilizes the Gini index focusing on highly variable genes and applies DBSCAN for rare cell detection [33], DensityCut estimates densities using directed k-nearest neighbor graphs with random walk refinement [34], and Galapagos provides straightforward workflows performing clustering directly on t-SNE embeddings [35]. Despite this algorithmic sophistication, comprehensive benchmarking studies demonstrate that no single method consistently outperforms others across diverse experimental conditions and dataset characteristics [36, 37]. Different methods optimize distinct aspects of cluster quality, with systematic evaluations showing substantial variation in cell type estimates and cluster assignments across methods [38, 39]. This algorithmic diversity reflects fundamental differences in mathematical frameworks rather than inadequacy of individual approaches, creating a persistent methodological challenge for robust cell population identification. Ensemble clustering approaches have emerged to address these limitations by integrating multiple algorithms. Early hypergraph-based methods like SAFE-clustering pioneered hypergraph partitioning for scRNA-seq data, integrating solutions from multiple base methods [40], while mixture model approaches like SAME-clustering employ expectation-maximization procedures to integrate algorithms and selectively filter poor performers based on Adjusted Rand Index [41]. Memory-efficient strategies exemplified by SHARP utilize random projection with weighted meta-clustering [42], and consensus approaches like conCluster combine t-SNE-based dimensionality reduction with k-means across varying parameters [43]. Recent developments emphasize sophisticated combination strategies; feature selection-based ensemble methods like scEFSC demonstrate that strategic preprocessing enhances performance on high-dimensional data [44], and multi-scale frameworks like scMSCF combine multi-dimensional PCA with weighted metaclustering to handle multi-resolution cellular heterogeneity [45]. However, these approaches share fundamental limitations: most treat all base algorithms equally or employ simple binary filtering, failing to systematically weight algorithms based on comprehensive quality assessment. We present ScEnsemble, a weighted hypergraph ensemble clustering framework addressing these limitations through three key innovations (Fig. 1). (i) We develop a hypergraph representation where edges are scaled based on a comprehensive quality assessment, ensuring that high-performing algorithms contribute substantially while poor performers are minimized. (ii) We implement a multi-criteria internal validation framework integrating four complementary metrics (Silhouette coefficient, Calinski-Harabasz, Davies-Bouldin, and Dunn indices). These metrics capture cluster compactness, separation, and connectivity to establish both individual and combined (averaged) weighting schemes. (iii) We integrate diverse consensus approaches spanning hierarchical clustering, community detection, meta-clustering, and hypergraph spectral methods enabling researchers to optimize for either mathematical cluster quality or biological interpretability based on their analytical priorities. Overview of the ScEnsemble methodology. Single-cell RNA-seq data is processed by base clustering algorithms (SC3, CIDR, Seurat, SIMLR, Monocle, RaceID). Quality metrics (Silhouette, Calinski-Harabasz, Davies-Bouldin, Dunn) are computed for each partition and used to construct a weighted hypergraph. Different weighting schemes (Standard, Average, CH, Silhouette, DB) determine edge weights. Consensus methods (CSPA-HC, CSPA-Leiden, CSPA-Louvain, MCLA, MCLA-Leiden, MCLA-Louvain, HGSC) integrate the weighted partitions to generate the final consensus clustering Methods Datasets We evaluated ScEnsemble using five publicly available benchmark scRNA-seq datasets that represent diverse biological contexts and sequencing technologies. All datasets were retrieved using the scRNAseq Bioconductor package[46], ensuring reproducible data access and consistent preprocessing across experiments. The datasets include Pollen [47], Kolodziejczyk [48], Muraro [49], Tasic [50], and Zeisel [19]. These datasets vary in size, complexity, and biological context, providing a comprehensive evaluation framework for clustering methods. Base clustering algorithms We employed six state-of-the-art scRNA-seq clustering algorithms as base methods: SC3, CIDR, Seurat, SIMLR, Monocle, and RaceID. These methods were selected based on their widespread adoption in the community and their complementary algorithmic approaches (graph-based, hierarchical, model-based, and density-based clustering paradigms). Each algorithm was run with parameters as recommended by the original authors, however, the number of clusters was estimated automatically in each algorithm for fair comparison. Specific implementation details: SC3 (version 1.36.0, Bioconductor package 3.21), CIDR (version 0.1.5), Seurat (version 5.3.1, resolution parameter optimized via silhouette score), SIMLR (version 1.34.0), Monocle (version 1.4.25), and RaceID (version 0.3.9). Data preparation and algorithm configuration Each base clustering algorithm received the raw count matrix as input and applied its own internal preprocessing pipeline, consistent with its original implementation and the recommendations of its authors. This design preserves the methodological integrity of each algorithm: imposing a common external transformation prior to clustering would override the preprocessing assumptions embedded in each method and potentially distort the structure it was designed to recover. Internal validity indices (Silhouette, Calinski-Harabasz, Davies-Bouldin, Dunn) were subsequently computed within the native output representation of each algorithm using Euclidean distance, following a model-intrinsic evaluation strategy. SC3 produces a consensus probability matrix across subsampled runs; CIDR derives a dropout-corrected dissimilarity matrix via its scDissim procedure; Seurat embeds cells in the top ten principal components of the scaled, variable-feature expression matrix; SIMLR learns a multi-kernel similarity matrix through its iterative optimisation; Monocle retains the top fifty principal components computed during preprocess_cds; and RaceID constructs a Pearson correlation-based distance matrix via compdist. Validity indices are computed within each of these representations so that each algorithm is assessed on how coherently it partitioned the data within the geometric structure it was designed to exploit, rather than within an externally imposed projection. Applying a single common reference space would systematically disadvantage algorithms whose clustering objectives are defined over nonlinear manifolds or kernel-induced similarity spaces, constituting a representational bias that would render the resulting weights unfair across methods. Raw metric values obtained from these heterogeneous spaces are normalised to a common scale prior to hypergraph weighting. This normalisation pipeline ensures that scale differences arising from different geometric spaces do not introduce systematic bias in the final weighting. Algorithm-specific configurations were as follows. CIDR cluster number was selected by evaluating and retaining the value that maximised the mean Silhouette coefficient computed on CIDR’s own dissimilarity matrix. Seurat community detection was performed via the Louvain algorithm at resolution 0.7, with cluster number emerging from graph modularity optimisation. SIMLR was run with internal feature normalisation enabled (normalize = TRUE), and cluster number was estimated using SIMLR_Estimate_Number_of_Clusters over [22]. Monocle was preprocessed with num_dim = 50 principal components followed by Leiden community detection [28]. RaceID applied quality filtering (mintotal = 1000, minexpr = 5, minnumber = 1), Pearson-based distance computation, and k-medoids clustering with clustnr = 30 candidate values and bootnr = 50 bootstrap iterations [24]. Prior to hypergraph construction, any partition yielding fewer than two distinct clusters was excluded to prevent undefined metric values. External and internal validation metrics and normalization framework A critical contribution of this work lies in the systematic integration of multiple internal clustering validation metrics to weight the contribution of each base clustering algorithm. We employ four complementary metrics: Silhouette coefficient, Calinski-Harabasz index, Davies-Bouldin index, and Dunn index. Since these metrics operate on different scales and optimization directions, we develop a comprehensive normalization framework to enable fair comparison and integration. We assessed clustering performance using both external and internal validation metrics. External validation quantifies agreement between predicted clusters and known cell type labels, while internal validation evaluates cluster quality based solely on data structure without requiring ground truth annotations. External validation metric To quantitatively assess the agreement between inferred clustering results and ground truth annotations, we employed the Adjusted Rand Index (ARI) [51], which corrects the standard Rand Index () for chance: ARI yields a score between -1 and 1, where 1 indicates perfect agreement and 0 denotes random cluster assignment. Internal validation metrics and normalization framework Silhouette Coefficient The silhouette coefficient [52] evaluates the quality of cluster assignments by measuring both intra-cluster cohesion and inter-cluster separation: where a(i) represents the average distance from point i to all other points within its cluster, and b(i) denotes the minimum average distance from point i to points in any other cluster. The silhouette score ranges from [-1,1] and is mapped to [0, 1]: Calinski-Harabasz Index The Calinski-Harabasz (CH) index [53] quantifies clustering quality through the ratio of between-cluster variance to within-cluster variance: where represents the centroid of cluster i, c is the global centroid, and denotes the number of points in cluster i. Since CH is unbounded and positively scaled, relative normalization is applied: Davies-Bouldin Index The Davies-Bouldin (DB) index [54] measures clustering quality by evaluating the ratio of within-cluster similarity to between-cluster dissimilarity: where represents the average scatter within cluster i, and denotes the distance between cluster centroids. As DB is a lower-is-better metric, its reciprocal is used for conversion to higher-is-better form: Dunn Index The Dunn index [55] evaluates clustering by measuring the ratio of minimum inter-cluster distance to maximum intra-cluster distance: where represents the distance between clusters and , and denotes the diameter of cluster . Average Score: The unified quality score combines all normalized metrics: Ensemble clustering problem formulation The ensemble clustering problem addresses the fundamental challenge of combining multiple clustering solutions to produce a more robust and accurate consensus partition [56, 57]. Given a dataset with n data points, ensemble clustering integrates r base clustering solutions to derive a consensus clustering that maximizes the shared information among individual partitions. Each base clustering assigns cluster labels to data points, where and denotes the cluster assignment of object by the q-th clustering algorithm producing clusters. To address label inconsistencies across different algorithms, canonical form adjustments are applied: The consensus clustering objective is formulated as an optimization function Γ : The quality of clustering solutions is typically assessed using Normalized Mutual Information (NMI) [56], defined through Mutual Information I(X, Y) and entropy H(X): The NMI between two clustering solutions and is computed as: where and represent the number of elements in the h-th and l-th clusters of methods a and b, respectively, and denotes the number of elements belonging to both clusters simultaneously. The consensus clustering that maximizes average mutual information with individual clusterings is defined as: Weighted hypergraph construction framework The core contribution of this work is the development of a weighted hypergraph framework that systematically integrates multiple clustering results based on their internal validation quality scores. This approach ensures that high-quality clustering solutions contribute more significantly to the final consensus. Each clustering algorithm produces a label vector: where denotes the cluster assignment of sample by algorithm q, and represents the number of clusters produced by that algorithm. Each clustering result is transformed into a binary incidence matrix : In this hypergraph representation, each cluster functions as a hyperedge connecting the data points that share the same cluster assignment. The conventional approach constructs a unified hypergraph by concatenating all incidence matrices horizontally: The sample-to-sample similarity matrix is computed as: This matrix quantifies how frequently pairs of samples are co-clustered across all algorithms, normalized by the total number of algorithms. Our novel contribution extends the baseline approach by incorporating quality-based weighting. For each internal validation metric , clustering solution is associated with a normalized weight . The weighted incidence matrix for clustering solution under metric s is defined as: The combined weighted incidence matrix becomes: The corresponding weighted sample similarity matrix is calculated as: This formulation ensures that clustering algorithms with superior internal consistency contribute more substantially to the resulting hypergraph structure, while maintaining proper normalization across different weighting schemes. Ensemble clustering algorithm implementation We implement three primary ensemble clustering approaches: Weighted Cluster-based Similarity Partitioning Algorithm (W-CSPA), Weighted Meta-CLustering Algorithm (W-MCLA), and Weighted HyperGraph Spectral Clustering (W-HGSC), each enhanced with weighted hypergraph representations and multiple consensus generation methods. Weighted cluster-based similarity partitioning algorithm (W-CSPA) based ensemble methods Given weighted hypergraph matrices derived from different internal validation metrics, the similarity matrix is converted to a dissimilarity matrix for hierarchical clustering [58]: Hierarchical clustering with average linkage is applied to D, and the optimal number of clusters is selected by maximizing the silhouette score: For community detection-based consensus using the Louvain algorithm [11], the similarity matrix S defines an undirected weighted graph G = (V,E), where the modularity Q is maximized: where represents the adjacency matrix derived from , denotes the degree of node i, and δ is the Kronecker delta function. The Leiden algorithm [12] extends this approach with resolution parameter γ : The optimal resolution parameter is selected with a bootstrap consensus stability procedure [59] that operates exclusively on algorithmic outputs. For each candidate resolution γ , the procedure is as follows: Run Leiden on the full co-association matrix at resolution γ to obtain the reference partition . Run Leiden on the full co-association matrix at resolution γ to obtain the reference partition . For i = 1, … , N: draw a random index set of size ; run Leiden on the induced submatrix at the same γ to obtain sub-partition . For i = 1, … , N: draw a random index set of size ; run Leiden on the induced submatrix at the same γ to obtain sub-partition . Compute the mean partition stability: 33 where denotes the ARI computed between two algorithmic partitions. Ground-truth labels are not accessed at any stage. Compute the mean partition stability: 33 where denotes the ARI computed between two algorithmic partitions. Ground-truth labels are not accessed at any stage. Weighted meta-clustering algorithm (W-MCLA) based ensemble methods The MCLA approach [56] performs meta-clustering by clustering hyperedges, where each hyperedge corresponds to a cluster produced by a base algorithm. In the weighted hypergraph incidence representation, each column of represents a hyperedge, and meta-clustering is carried out by clustering these hyperedge vectors using the generalized Jaccard similarity. Specifically, the similarity between hyperedges i and j is defined as: The distance matrix is used for hierarchical clustering [58] to partition hyperedges into k meta-clusters. For each meta-cluster , defined as the set of hyperedges , the association score of sample x is computed using the weighted hypergraph as: Each sample is assigned to the meta-cluster with maximum association score: Similar to W-CSPA, W-MCLA variants employ Louvain [11] and Leiden [12] algorithms for meta-cluster identification, utilizing the Jaccard similarity matrix as the basis for graph construction and subsequent community detection. Weighted hypergraph spectral clustering (W-HGSC) based ensemble methods W-HGSC operates directly on the weighted hypergraph structure using spectral methods [60], providing a third complementary approach that preserves the native hypergraph topology throughout the consensus generation process. From the weighted hypergraph matrix , we construct the vertex and hyperedge degree matrices: where and denote the all-ones vectors for vertices and hyperedges, respectively. The normalized hypergraph Laplacian [60] is constructed as: This formulation ensures proper normalization while preserving the weighted contributions from high-quality base clustering algorithms. The final clustering is obtained by computing the smallest eigenvectors of L, constructing the embedding matrix , and applying k-means clustering [61] to the row-normalized embedding: The optimal number of clusters is determined using the eigengap heuristic [62]: where represents the k-th smallest eigenvalue of the hypergraph Laplacian. Algorithm integration and evaluation framework Our comprehensive ensemble clustering framework integrates three complementary strategies, each tailored to different structural assumptions and clustering characteristics. W-CSPA is particularly effective for approximately spherical cluster structures and scenarios where interpretable pairwise sample similarities are essential. W-MCLA is better suited for capturing overlapping or heterogeneous cluster configurations, as it operates at the hyperedge level and enables detailed meta-cluster analysis. Finally, W-HGSC is designed to handle complex, non-convex cluster shapes by operating directly on hypergraph-derived similarity structures, making it especially advantageous when preserving the native topology of the hypergraph is critical for accurate consensus formation. All methods benefit from our weighted hypergraph construction framework, ensuring that reliable clustering solutions dominate the consensus generation process while poor-performing algorithms contribute minimally to the final result. This systematic integration of internal validation metrics provides a principled mechanism for automated algorithm selection and adaptive ensemble weighting. The performance of our weighted hypergraph ensemble methods is evaluated primarily using internal validation metrics for both hyperparameter optimization and weighting scheme determination. The comprehensive evaluation framework ensures robust assessment of the proposed methodology across diverse clustering scenarios and datasets. All algorithms are implemented with systematic hyperparameter optimization, including automatic cluster number selection for hierarchical methods and resolution parameter tuning for community detection approaches. The integration of multiple internal validation metrics in the weighting scheme provides a principled approach to ensemble clustering that adapts to the intrinsic characteristics of different clustering algorithms and datasets. Mathematical notations for all ensemble clustering methodologies described in this section are centralized in the Symbol Explanation table (Supplementary Table S8). Results We evaluated ScEnsemble across five benchmark scRNA-seq datasets representing diverse biological contexts: Pollen neural progenitor cells (301 cells, 5 cell types), Kolodziejczyk embryonic stem cells (704 cells, 3 cell types), Muraro pancreatic islets (3,072 cells, 11 cell types), Tasic adult mouse cortex (1,809 cells, 8 cell types), and Zeisel mouse brain (3,005 cells, 7 cell types). The evaluation employed both external validation using the Adjusted Rand Index to measure agreement with known cell type annotations and internal validation metrics including Silhouette coefficient, Calinski-Harabasz index, Davies-Bouldin index, and Dunn index to assess cluster quality independent of ground truth labels. Complete performance results for all 175 configurations are provided in Supplementary Tables S1–S5. Performance of individual clustering algorithms The base clustering algorithms exhibited complementary performance patterns across datasets and evaluation metrics, confirming that no single method consistently outperforms others across all conditions (Table 1). Table 1Performance of individual clustering algorithms across multiple single-cell RNA-seq datasets based on normalized internal and external validation metricsDatasetAlgorithmARISilhouetteCHDBDunnAverage PollenSC30.4820.8260.7290.4370.2950.572CIDR0.6270.5190.0430.0670.2000.208Seurat0.6840.6470.0900.1700.0750.245SIMLR0.1320.5530.0430.1160.3010.253Monocle0.7180.5230.0230.0880.0480.170RaceID0.2270.5170.0730.1220.0800.198KolodziejczykSC30.6980.9950.9940.8790.6480.879CIDR0.5460.5330.0010.0190.0970.162Seurat0.7220.5980.0020.0310.0130.161SIMLR0.2930.5300.0000.0150.1720.179Monocle0.8340.5480.0010.0270.0530.157RaceID0.2710.5670.0020.0280.0180.154MuraroSC30.4680.7500.2100.2860.1750.355CIDR0.3750.6980.5220.2630.3030.446Seurat0.7180.6710.1990.2440.0290.286SIMLR0.5340.5010.0050.0420.3350.221Monocle0.9200.7140.0640.1640.1590.275TasicSC30.2330.8330.4450.3510.2840.478CIDR0.9020.5960.2780.1230.2080.301Seurat0.2660.6780.1130.2090.0190.255SIMLR0.8280.5090.0080.0450.3880.238Monocle0.2960.6190.0240.1190.0520.203RaceID0.4120.6030.1330.1510.0500.234ZeiselSC30.4170.6950.1260.2370.3560.354CIDR0.3470.5880.2600.1430.0780.267Seurat0.4440.6640.1190.2530.0330.267SIMLR0.3850.4870.0060.0480.4300.243Monocle0.8790.5960.0500.1420.0840.218RaceID0.6620.6870.4380.1780.0180.330Bold values indicate the best result per metric within each dataset Performance of individual clustering algorithms across multiple single-cell RNA-seq datasets based on normalized internal and external validation metrics Bold values indicate the best result per metric within each dataset Monocle achieved the highest ARI scores on four of five datasets: 0.920 on Muraro, 0.879 on Zeisel, 0.834 on Kolodziejczyk, and 0.718 on Pollen. CIDR outperformed all methods on the Tasic dataset with an ARI of 0.902, demonstrating that certain data structures favour specific algorithmic approaches. Internal validation metrics revealed a markedly different performance landscape. SC3 dominated Silhouette scores across all five datasets, achieving the highest values on Pollen (0.826), Kolodziejczyk (0.995), Muraro (0.750), Tasic (0.833), and Zeisel (0.695). The composite average score identified SC3 as the overall best performer on four of five datasets, including Pollen (0.572), Kolodziejczyk (0.879), Tasic (0.478), and Zeisel (0.354), while CIDR led on Muraro (0.446). This divergence between external and internal validation performance underscores a fundamental challenge in scRNA-seq clustering: algorithms optimising mathematical cluster properties do not necessarily produce partitions that best match biological annotations, and vice versa. The systematic variation in optimal algorithm choice across datasets and metrics motivated our ensemble approach, which leverages this algorithmic diversity through quality-based weighting. Weighted hypergraph-based ensemble performance ScEnsemble outperformed the best individual algorithm in 23 of 25 metric-dataset combinations (92%), demonstrating broad and consistent superiority. Table 2 presents the best-performing configurations for each ensemble architecture and dataset, with the optimal weighting scheme indicated in parentheses. Figure 2 visualises the comparative performance patterns. Table 2Weighted ensemble performance metrics across datasetsDatasetAlgorithmARISilhouetteCHDBDunn Pollencspa_hc0.7160(Std)0.9173(CH)5852.48(CH)0.2339(CH)2.6893(CH)cspa_leiden0.7192(DB)0.9173(CH)5852.48(CH)0.2339(CH)2.6893(CH)cspa_louvain0.7228(Std)0.8143(CH)992.73(CH)0.4590(CH)0.7289(CH)hgsc0.4858(CH)0.7337(CH)641.50(CH)0.5330(CH)0.7289(CH)mcla0.6847(Std)0.6297(CH)443.08(Sil)0.6182(Avg)0.1704(Sil)mcla_leiden0.7002(Std)0.9173(CH)5852.48(CH)0.2339(CH)2.6893(CH)mcla_louvain0.6978(CH)0.5359(Sil)469.65(Sil)0.7380(Sil)0.1667(Std)Kolodziejczykcspa_hc0.7260(Std)0.9983(CH)59900780.99(CH)0.0022(CH)206.6107(CH)cspa_leiden0.8378(Avg)0.9983(CH)59900780.99(CH)0.0022(CH)206.6107(CH)cspa_louvain0.7265(Std)0.9983(CH)59900780.99(CH)0.0022(CH)206.6107(CH)hgsc0.8378(Std)0.8804(DB)17734.47(DB)0.3860(Sil)0.6751(Avg)mcla0.8075(Avg)0.9983(CH)59900780.99(CH)0.0022(CH)206.6107(CH)mcla_leiden0.8031(Std)0.9983(CH)59900780.99(CH)0.0022(CH)206.6107(CH)mcla_louvain0.8313(Sil)0.7150(Sil)2181.35(Sil)0.3886(Sil)0.9955(CH)Murarocspa_hc0.9101(DB)0.6683(CH)16115.47(CH)0.3190(CH)1.0899(CH)cspa_leiden0.9202(Std)0.9722(DB)23347.50(DB)0.3190(CH)1.3378(Sil)cspa_louvain0.3908(Avg)0.6683(CH)16115.47(CH)0.3190(CH)1.0899(CH)hgsc0.6218(DB)0.6470(CH)11557.84(CH)0.4257(CH)0.4166(CH)mcla0.3800(CH)0.6269(CH)9786.13(CH)0.3961(CH)0.2000(Std)mcla_leiden0.9187(Avg)0.8008(DB)16115.47(CH)0.3190(CH)1.0899(CH)mcla_louvain0.9360(DB)0.7518(Avg)5124.75(Avg)0.3911(DB)0.2000(Std)Tasiccspa_hc0.6294(Std)0.6523(CH)1393.00(Avg)0.4364(CH)0.6158(CH)cspa_leiden0.9041(Sil)0.7643(DB)2264.17(Std)0.5825(DB)0.5749(Avg)cspa_louvain0.9041(Sil)0.4960(Std)2263.70(Std)0.6748(Std)0.5943(DB)hgsc0.7286(DB)0.5187(CH)2269.48(Std)0.5957(Std)0.1667(Std)mcla0.8760(Std)0.3406(CH)961.31(Std)0.8057(Std)0.1667(Std)mcla_leiden0.3305(Avg)0.6935(CH)2036.10(Std)0.4972(CH)0.1667(Std)mcla_louvain0.7311(DB)0.5166(CH)1298.20(CH)0.7245(Avg)0.1667(Std)Zeiselcspa_hc0.8570(Avg)0.6421(CH)2347.71(CH)0.9266(CH)0.2603(CH)cspa_leiden0.8736(Std)0.6743(DB)2198.05(Std)0.6629(CH)0.1667(Std)cspa_louvain0.8333(Std)0.4567(Std)2621.49(CH)0.8276(Sil)0.1667(Std)hgsc0.6166(CH)0.4659(CH)2069.68(CH)0.8737(Std)0.1667(Std)mcla0.5286(Sil)0.3180(Std)1384.36(CH)1.1392(CH)0.1667(Std)mcla_leiden0.7661(Avg)0.6425(CH)2345.70(CH)0.9291(CH)0.7800(CH)mcla_louvain0.7702(CH)0.4832(CH)1501.62(CH)0.9714(Std)0.1667(Std)Values in parentheses indicate the weighting scheme that produced each result, Std, standard/unweighted, Sil, Silhouette; CH, Calinski-Harabasz; DB, Davies-Bouldin; average, combined weighting Weighted ensemble performance metrics across datasets Values in parentheses indicate the weighting scheme that produced each result, Std, standard/unweighted, Sil, Silhouette; CH, Calinski-Harabasz; DB, Davies-Bouldin; average, combined weighting Weighted ensemble performance across datasets and validation metrics. A ARI scores comparing individual algorithms, best unweighted ensemble, and best weighted ensemble methods across datasets. B Internal validation metric (Silhouette coefficient) improvements achieved through quality-based weighting relative to the best individual algorithm baseline. C Heatmap of improvement percentages across all metric-dataset combinations comparing best ensemble versus best individual algorithm For biological concordance, ensemble methods achieved the highest ARI scores in four of five datasets (Supplementary Table S6). The most notable gain occurred on the Muraro dataset, where MCLA-Louvain with DB weighting achieved an ARI of 0.936 compared to 0.920 for Monocle, a 1.7% improvement. On the Tasic dataset, CSPA-Leiden with average weighting matched the best individual algorithm, achieving an ARI of 0.904 compared to 0.902 for CIDR. On Pollen, CSPA-Louvain achieved 0.723 compared to 0.718 for Monocle, a 0.7% improvement. The Zeisel dataset represented the single case where Monocle marginally outperformed ensemble approaches, with an ARI of 0.879 versus 0.874 for CSPA-Leiden, a difference of only 0.6%. Weighted configurations further matched or exceeded unweighted configurations on ARI in four of five datasets, confirming that quality-based weighting is beneficial even for external validation (Supplementary Table S7). Internal validation metrics revealed more pronounced advantages. On Pollen, CSPA-HC with CH weighting achieved a Silhouette coefficient of 0.917 compared to 0.652 for SC3 (40.7% improvement), a Calinski-Harabasz index of 5,852.5 versus 689.7 (748.6% increase), and a Dunn index of 2.689 versus 0.208for SIMLR (1,192.6% improvement). The Kolodziejczyk dataset demonstrated a notable convergence phenomenon: multiple ensemble configurations including CSPA-HC, CSPA-Leiden, CSPA-Louvain, MCLA, and MCLA-Leiden all reached an identical optimal clustering with Silhouette coefficient of 0.998, Davies-Bouldin index of 0.002, and Dunn index of 206.611. This convergence reflects the unambiguous biological structure of embryonic stem cell states, where quality-based weighting guides diverse algorithmic paths toward the same globally optimal partition. Across all datasets, weighted configurations outperformed their unweighted counterparts in 21 of 25 combinations (84%), confirming that quality-based weighting represents a robust and broadly effective enhancement (Supplementary Table S7). Integrative analysis of weighting strategies and ensemble architectures The choice of weighting scheme influenced optimisation outcomes, with different schemes favouring distinct aspects of cluster quality (Fig. 3; Supplementary Table S7). Impact of weighting schemes on cluster quality across datasets. A Radar plots comparing normalized internal validation metrics between unweighted (standard) and best weighted ensemble methods for each dataset. B Improvement percentages of quality-based weighted approaches over unweighted baselines across all internal validation metrics and datasets CH-based weighting was the most frequently selected scheme for internal validation optimisation, achieving the best weighted result in 13 of 25 metric–dataset combinations. On Kolodziejczyk, CH weighting produced a Calinski-Harabasz index of 59,900,781 compared to 2,252 for the best unweighted configuration, while on Pollen, the Silhouette coefficient improved from 0.577 (best unweighted) to 0.917 (+58.9%) and the Dunn index from 0.333 to 2.689 (+706.8%). DB-based weighting was also effective, achieving the best Silhouette score on three of five datasets (Muraro: 0.972, Tasic: 0.764, Zeisel: 0.674). These improvements suggest an alignment between the CH criterion, which favours high between-cluster variance relative to within-cluster variance, and the ensemble integration goal of identifying structurally coherent partitions. On the Muraro dataset, DB weighting achieved the best ARI (0.936) via MCLA-Louvain, while CH weighting produced the best Davies-Bouldin index (0.319) through CSPA-HC. On Pollen, DB-weighted CSPA-Leiden achieved an ARI of 0.719, closely trailing the best configuration (CSPA-Louvain, standard, 0.723), indicating that different weighting schemes can yield comparable external concordance. Average weighting, which combines all four internal metrics, achieved competitive biological concordance on datasets with complex cellular hierarchies. On the Tasic dataset, featuring hierarchically organised cortical neuronal subtypes, average-weighted CSPA-Leiden achieved an ARI of 0.904, comparable to the best individual algorithm (CIDR, 0.902). On Kolodziejczyk, average-weighted CSPA-Leiden produced the best ensemble ARI (0.838), slightly exceeding Monocle (0.834). These results indicate that balanced multi-metric optimisation can be effective for datasets where no single geometric criterion fully captures the underlying cluster structure. Among ensemble architectures, CSPA variants showed consistent performance across datasets. MCLA variants performed well on biologically complex datasets, with MCLA-Louvain distinguishing alpha, beta, delta, and gamma pancreatic cell subtypes on Muraro (ARI = 0.936). The meta-clustering strategy, which operates on cluster prototypes rather than individual cells, appears suited to datasets where cell types exhibit gradual transcriptional transitions. W-HGSC achieved competitive Davies-Bouldin indices on Tasic (0.596 vs. 0.795 for SC3, + 25.0%) and Zeisel (0.874 vs. 1.074 for Seurat, + 18.6%), suggesting that spectral decomposition of the hypergraph Laplacian can offer advantages for datasets with complex community structure. Table 3 summarises the best-performing ensemble configuration per dataset under two complementary criteria: external validity (ARI) and internal validity (Silhouette width). For internal quality, CH-based and DB-based weighting were selected across all five datasets, indicating that aligning base-clusterer influence with geometric cluster structure is generally beneficial. For external concordance, weighted configurations produced the highest ARI on three of five datasets (Muraro: DB weighting, ARI = 0.936; Tasic and Kolodziejczyk: average weighting, ARI = 0.904 and 0.838, respectively); on Pollen and Zeisel, the standard (unweighted) configuration achieved comparable ARI while weighted schemes still led on internal metrics. Table 3Best ensemble configuration per dataset based on ARI and SilhouetteBest by ARIBest by SilhouetteDatasetMethodWeightARIKMethodWeightSilKPollenCSPA-LOUVAINStd0.7233CSPA-HCCH0.9178MuraroMCLA-LOUVAINDB0.9369MCLA-LOUVAINAvg0.7529TasicCSPA-LEIDENAvg0.9048MCLA-LEIDENCH/Sil0.87518KolodziejczykCSPA-LEIDENAvg0.8384CSPA-HCCH0.9985ZeiselCSPA-LEIDENStd0.8746MCLA-LEIDENCH0.6429Std, standard/unweighted; Sil, Silhouette; CH, Calinski-Harabasz; DB, Davies-Bouldin; average, combined weighting; K, number of cluster Best ensemble configuration per dataset based on ARI and Silhouette Std, standard/unweighted; Sil, Silhouette; CH, Calinski-Harabasz; DB, Davies-Bouldin; average, combined weighting; K, number of cluster These results suggest two complementary principles for weighting scheme selection. CH-based weighting tends to maximise geometric cluster quality and may be preferred when cluster structure optimisation is the primary goal, while average or DB weighting can be more effective when biological concordance with known cell type annotations is the objective, particularly for datasets with hierarchical cellular organisation. This practical guidance enables researchers to align their weighting strategy with specific biological questions rather than relying on a single approach. Biological validation on a complex tumor microenvironment dataset ScEnsemble weights base clusterers using internal validity indices, including the Calinski-Harabasz, Davies-Bouldin, and silhouette criteria, which measure geometric properties of a partition without requiring ground-truth labels. To examine whether optimizing for these internal criteria produces biologically interpretable structure, we applied ScEnsemble to a real-world dataset and assessed cluster identity using independent gene-set scoring that relies on neither internal indices nor ground-truth annotations. We used the Wu et al. (2021) breast cancer scRNA-seq dataset [63], which profiles the tumor microenvironment (TME) across 28 annotated cell type categories. The full dataset exceeds 100,000 cells, which is beyond the practical scale of several base algorithms in ScEnsemble. We therefore constructed a stratified subsample of n = 2,000 cells by sampling each cell type proportionally to its abundance in the full dataset, with a minimum of 10 cells per type to retain all 28 categories. This design preserves the cell type composition of the original dataset and does not introduce selection bias toward any particular clustering outcome; subsampling of this scale is standard practice in single-cell clustering benchmarks [59]. The CSPA-HC variant with CH weighting, which achieved the highest ARI across the benchmark datasets, was used for all downstream analyses. The ScEnsemble UMAP embedding closely mirrors the ground-truth cell type structure: immune populations, stromal cells, and epithelial clusters occupy distinct regions of the embedding, consistent with the annotated celltype_major labels (Fig. 4, Panel A). To evaluate biological coherence independently of the ground-truth annotation, we applied AUCell gene-set scoring [64] using curated marker gene signatures for 15 TME cell type categories. AUCell ranks genes by expression within each cell and computes an enrichment score for a given gene set without consulting either the clustering result or the cell type labels. Aggregating AUCell scores per predicted cluster shows that each ScEnsemble cluster selectively enriches for a single dominant TME signature: clusters containing predominantly immune cells scored highest on T cell (CD3D, CD8A), B cell (CD19, MS4A1), and NK cell (GNLY, NKG7) gene sets; stromal clusters scored highest on CAF (FAP, COL1A1) and endothelial (PECAM1, VWF) signatures; and epithelial clusters were distinguished by Cancer Epithelial (EPCAM, KRT8) and Normal Epithelial (KRT14, KRT5) scores (Fig. 4, Panel B). The one-to-one correspondence between predicted clusters and dominant AUCell signatures was obtained without any reference to ground-truth labels at any step. Biological validation on the Wu et al. (2021) breast cancer TME dataset. A UMAP embedding of 2,000 stratified cells coloured by ScEnsemble predicted clusters and ground-truth celltype_major annotations. B AUCell gene-set enrichment heatmap showing the dominant TME signature per predicted cluster. C Expression of canonical marker genes across predicted clusters, confirming one-to-one correspondence with known cell types At the individual gene level, expression of two canonical markers per cell type across predicted clusters confirms the same pattern: marker genes are concentrated in the clusters that score highest for the corresponding AUCell signature, and show low expression in unrelated clusters (Fig. 4, Panel C). The marker enrichment is consistent across all 15 cell type categories represented in the subsample. The CH-weighted CSPA-HC variant, which yielded the highest Silhouette coefficient and Calinski-Harabasz scores on Pollen (Sil = 0.917, CH = 5,852.5) and Kolodziejczyk (Sil = 0.998, CH = 59,900,781) among the benchmark datasets, also produced the clearest TME cell type separation in the Wu dataset. This parallel outcome indicates that maximizing within-cluster compactness relative to between-cluster separation, as captured by the CH index, is a practical criterion for selecting base clusterers in the absence of ground-truth labels, and that doing so yields partitions with interpretable biological structure. Comparison to existing ensemble approaches To evaluate the performance of ScEnsemble against established single-cell clustering methods, we conducted a comprehensive benchmark across the five aforementioned scRNA-seq datasets. For each dataset, 2,000 Highly Variable Genes (HVGs) were selected to optimize biological signal detection. ScEnsemble was compared against SHARP and SAME, two representative ensemble and consensus clustering methods. Performance was assessed using three complementary metrics: ARI, NMI, and Silhouette. ScEnsemble demonstrated consistently strong performance across the majority of datasets. On the Kolod dataset, multiple ScEnsemble variants achieved similar scores (ARI = NMI = 1.000), outperforming SHARP (ARI = 0.697) and SAME (ARI = 0.723). Notably, in the Kolod dataset, the standard and silhouette weightings in the CSPA_Louvain variant, the std, sil, and avg weightings in the CSPA_Leiden variant, and the std, sil, and avg weightings in the HGSC variant simultaneously yielded the best results across ARI, NMI, and Silhouette evaluation indices (Table 4). Table 4Performance comparison of ScEnsemble against SHARP and SAME-clustering across five benchmark datasetsDatasetMethod (Weight)ARINMISilhouette KolodziejczykScE_best_ARI MCLA-Louvain (db)0.91040.88700.1865ScE_best_NMI MCLA-Louvain (db)0.91040.88700.1865ScE_best_Sil MCLA-Louvain (db)0.91040.88700.1865SHARP0.69700.73860.1740SAME0.72320.70670.1131MuraroScE_best_ARI CSPA-Hc (ch)0.93930.89050.3746ScE_best_NMI MCLA-Louvain (ch)0.93890.89420.3753ScE_best_Sil MCLA-Louvain (ch)0.93890.89420.3753SHARP0.51620.4947−0.0010SAME0.61770.70910.1911PollenScE_best_ARI CLA-Leiden (db)0.76620.68570.0822ScE_best_NMI MCLA-Leiden (sil)0.71430.70050.0943ScE_best_Sil CSPA-Leiden (std)0.45480.40840.2129SHARP0.00530.0139−0.0446SAME0.71900.69360.1480TasicScE_best_ARI HGSC (std)0.73540.74500.2456ScE_best_NMI HGSC (sil)0.73540.74580.2459ScE_best_Sil HGSC (db)0.65200.63920.2540SHARP0.90620.81030.2310SAME0.40830.46910.1639ZeiselBrainScE_best_ARI CSPA-Leiden (std)0.89310.83740.2431ScE_best_NMI CSPA-Leiden (std)0.89310.83740.2431ScE_best_Sil CSPA-Louvain (ch)0.47660.42640.3747SHARP0.60570.63620.2351SAME0.79090.74010.2534For each dataset, the best ScEnsemble configuration is reported for each evaluation criterion (ARI, NMI, Silhouette). Ensemble architecture and weighting scheme are shown in parentheses. Bold values indicate the best result per metric within each dataset Performance comparison of ScEnsemble against SHARP and SAME-clustering across five benchmark datasets For each dataset, the best ScEnsemble configuration is reported for each evaluation criterion (ARI, NMI, Silhouette). Ensemble architecture and weighting scheme are shown in parentheses. Bold values indicate the best result per metric within each dataset On the Muraro pancreatic dataset, ScEnsemble achieved a peak ARI of 0.939, representing a 42–82% improvement over SHARP (0.516) and SAME (0.618). Similarly, ScEnsemble attained the highest ARI (0.893) and NMI (0.837) on the ZeiselBrain dataset. While SHARP showed higher label agreement on the Tasic dataset, ScEnsemble maintained a comparable Silhouette score (0.254 vs. 0.231), suggesting it produces geometrically compact and well-delineated clusters. On the Pollen dataset, ScEnsemble demonstrated competitive performance, recording an ARI of 0.766 and a Silhouette score of 0.213 (Table 4). To further assess the biological coherence of ScEnsemble’s partitioning, we visualized the clustering results on the Muraro dataset using UMAP and overlaid the expression of known cell-type marker genes (Fig. 5). The ScEnsemble UMAP embedding (ARI = 0.939) closely recapitulates the ground-truth cell-type structure, with a visual concordance superior to SAME (ARI = 0.618) and SHARP (ARI = 0.390; Fig. 5, Panel A). Consistent with the clustering assignments, high INS (insulin) expression was concentrated in the predicted beta-cell cluster, GCG (glucagon) expression marked the alpha-cell population, SST (somatostatin) delineated delta cells, and KRT19 expression was enriched in ductal cells (Fig. 5, Panel B). Additional markers, including GHRL, REG1A, PECAM1, and COL1A1, confirmed that each ScEnsemble-predicted cluster corresponds to a biologically distinct cell type. This concordance between predicted clusters and known marker gene expression profiles indicates that ScEnsemble recovers biologically meaningful cell-type boundaries (Fig. 5). Biological validation on the Muraro dataset. A Comparison of ensemble clustering approaches. B Expression profiles of established cell-type marker genes, demonstrating that clusters identified by ScEnsemble are consistent with known biological cell types. C UMAP visualization of the best-performing configurations under different weighting strategies Discussion Single-cell RNA sequencing enables researchers to study gene expression at single-cell resolution, but choosing the right clustering algorithm remains a practical challenge. With a growing number of clustering methods available for scRNA-seq data [65], selecting an appropriate approach requires careful consideration. Our weighted hypergraph ensemble framework offers a structured approach by combining multiple algorithms through quality-based weighting, rather than requiring researchers to select a single method. ScEnsemble introduces two complementary design choices. First, it integrates a multi-criteria internal validation framework directly into the hypergraph edge weighting process, so that base partitions with higher structural quality contribute proportionally more to the consensus. Second, it diversifies the consensus partitioning step by applying hierarchical clustering, Louvain, and Leiden community detection on the similarity and meta-cluster matrices derived from the quality-weighted hypergraph, ensuring that both the edge weights and the partitioning strategy reflect clustering quality. Together, these choices enable a systematic examination of architecture–weighting interactions. These design choices address limitations shared by current scRNA-seq ensemble methods. In our direct benchmark (Sect. 4.5), SAME-clustering [41], which selectively filters algorithms via binary ARI thresholding, and SHARP [42], which employs weighted meta-clustering based on random projection features, were both outperformed by ScEnsemble on four of five datasets. ScEnsemble addresses the gap left by these approaches by providing continuous quality-based weighting and multiple consensus architectures within a single framework. Our results illustrate a systematic trade-off between biological concordance and geometric cluster quality. Monocle achieved the highest ARI on four of five datasets, yet SC3 and CIDR consistently produced better-separated clusters by internal metrics. This is consistent with the benchmark by Duò et al. [36], who found that no single method dominated across all conditions and that pairwise ensemble combinations often produced results between the better and worse individual methods. ScEnsemble mitigates this averaging effect: across the 25 metric–dataset combinations (after excluding degenerate configurations), the best ensemble configuration matched or exceeded the best individual algorithm in 23 cases (92%; Supplementary Table S6). Quality-based weighting further improved upon unweighted ensembles in 21 of 25 combinations (84%; Supplementary Table S7). The weighting scheme determines what the ensemble optimises, and our results indicate that this choice can be aligned with specific analytical priorities. CH-based weighting was the most frequently optimal scheme for internal validation, improving Silhouette coefficients by up to 58.9% (from 0.577 to 0.917), Dunn indices by up to 706.8% (from 0.333 to 2.689), and Davies-Bouldin indices by up to 56.3% (from 0.731 to 0.319) over unweighted baselines (Supplementary Table S7). Average weighting, which integrates all four internal metrics simultaneously, achieved competitive biological concordance on datasets with hierarchically organised cellular populations. On the Tasic cortical neuron dataset, average-weighted CSPA-Leiden achieved an ARI of 0.904, comparable to the best individual algorithm (CIDR, 0.902). Average weighting similarly produced the best ensemble ARI on Kolodziejczyk (0.838 vs. 0.834 for Monocle). These observations are consistent with the view that integrating complementary geometric criteria can be a useful strategy for datasets with complex cellular organisation, and align with a practical need identified in recent reviews [66]: a single consensus strategy may not serve all analytical purposes equally well. Dataset characteristics influenced which configuration performed best. The Kolodziejczyk embryonic stem cell dataset exhibited a convergence phenomenon: five architectures independently reached an identical optimal partition with a Silhouette coefficient of 0.998 and a Dunn index of 206.611 under CH weighting. This convergence suggests that when the underlying biological structure is well-separated, quality-guided weighting can guide diverse algorithmic paths toward the same solution. The Muraro pancreatic dataset, where cell types share transcriptional programmes while maintaining distinct identities, performed best with DB-weighted MCLA-Louvain (ARI = 0.936), suggesting that prototype-based consensus may be suited to datasets with overlapping transcriptional signatures. A related concern is the potential divergence between internal validation metrics and biological correctness. As our results show, Monocle achieved the highest ARI on four of five datasets, yet SC3 and CIDR dominated internal metrics. The quality-based weighting mechanism therefore optimises geometric compactness, which does not always coincide with the biologically most meaningful partition. ScEnsemble mitigates this through averaged weighting, which balances multiple geometric criteria, and through diversified consensus architectures. Importantly, however, our biological validation analyses on both the Wu tumor microenvironment and Muraro pancreatic datasets demonstrate that even when geometric optimization diverges from raw label agreement, the resulting partitions still preserve robust biological meaningfulness. The clusters recovered by ScEnsemble consistently align with distinct functional marker gene signatures, confirming that the structural compactness optimized by internal metrics translates into biologically coherent subpopulations. Nevertheless, users should be aware that internal-metric-guided weighting may not be optimal for all biological questions. Several additional limitations should be acknowledged. The weighting approach assigns quality scores based on the full dataset partition and does not adapt locally: an algorithm performing well overall but poorly on a specific cell subpopulation will receive a uniformly high weight, potentially biasing the consensus for that subpopulation. Our failure-case analysis (Supplementary Fig. S2) confirms that populations representing less than 10% of the dataset may experience reduced F1 scores compared to specialised base algorithms; iterative or region-adaptive weighting strategies could address this in future work. Additionally, ScEnsemble produces deterministic hard partitions without per-cell confidence scores; soft membership outputs indicating assignment uncertainty would be informative for cells near cluster boundaries. Evaluation was conducted on five benchmark datasets with well-defined cell type labels; extension to spatial transcriptomics and multi-modal single-cell assays, where ground truth is less well-defined, remains an important direction. Regarding computational scalability, all experiments were conducted on a workstation equipped with a 12th Gen Intel Core i7-12650H processor (2.30 GHz) and 32 GB RAM running Windows 1164-bit. A systematic profiling using subsamples of (500, 1000, 1500, 2000, 3000) cells from a breast cancer scRNA-seq dataset confirmed that the ScEnsemble consensus step accounts for a minor fraction of total runtime relative to the base clustering algorithms (Supplementary Fig. S1). The computational cost is dominated by the base algorithms, particularly SIMLR and SC3, rather than the hypergraph construction and partitioning steps. In summary, ScEnsemble provides principled control over what the ensemble optimises through quality-proportional weighting and diversified consensus partitioning, offering a practical tool for single-cell clustering applications. Conclusions Clustering is a cornerstone of single-cell RNA sequencing analysis, yet the persistent diversity of methods reflects a genuine truth: no single algorithm captures all aspects of biological heterogeneity equally well. ScEnsemble transforms this diversity from a source of uncertainty into a methodological advantage by integrating multiple clustering algorithms through quality-proportional weighting, ensuring that the ensemble is guided by the most informative partitions rather than treating all contributions as equal. The framework allows researchers to select a weighting scheme that aligns with their specific biological question, whether prioritising geometric cluster quality or biological concordance. This addresses a practical limitation of existing ensemble approaches and provides a principled basis for method selection in varied analytical contexts. The systematic characterisation of interactions between weighting scheme and consensus architecture offers concrete guidance for single-cell clustering practice. As the scale and complexity of single-cell data continue to grow, approaches that combine algorithmic diversity with principled integration will be essential for producing robust, reproducible, and biologically meaningful cell type discoveries. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Author contributions B.U. contributed in conceptualization, methodology, validation, formal analysis, and writing both the original draft and reviewing–editing; I.Y. contributed in conceptualization, methodology, supervision, and writing and reviewing–editing the manuscript. Funding This research was supported by the Scientific and Technological Research Council of Turkey (TUBITAK) BİDEB 2211—National Scholarship Programme for Graduate Students awarded to Beste Uncu. Data availability All datasets used in this study are publicly available. The five benchmark scRNA-seq datasets (Pollen, Kolodziejczyk, Muraro, Tasic, and Zeisel) were retrieved from the scRNAseq Bioconductor package (version 2.22.0). Code Availability The source code for ScEnsemble can be found at https://github.com/besteuncu/ScEnsemble. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no conflict of interest. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement All datasets used in this study are publicly available. The five benchmark scRNA-seq datasets (Pollen, Kolodziejczyk, Muraro, Tasic, and Zeisel) were retrieved from the scRNAseq Bioconductor package (version 2.22.0). The source code for ScEnsemble can be found at https://github.com/besteuncu/ScEnsemble.