SUN Automates Louvain Clustering Resolution Selection for Single-Cell RNA Sequencing Benchmarks

A newly developed computational method named SUN provides an explicit, auditable framework for selecting Louvain-style clustering resolutions in single-cell RNA-sequencing workflows, according to benchmark findings published in BMC Bioinformatics. Developed by researchers including Junbo Wang, Qiaoling Liang, Maojin Liang, and their colleagues, the unsupervised software layer guides resolution choice on a fixed shared nearest-neighbour graph. In evaluations spanning synthetic simulations and eight real single-cell transcriptomic datasets, the framework demonstrated high concordance with curated reference annotations while converting what has traditionally been an undocumented trial-and-error procedure into a reproducible quantitative recommendation.

The study addresses a longstanding technical challenge in single-cell RNA-sequencing analysis, where clustering resolution is frequently chosen through ad hoc trial and error. Because the chosen resolution directly dictates the reported cellular groupings and shapes downstream biological interpretation, arbitrary selections can lead to over-splitting or over-merging of cell populations. These unstandardized decisions substantially degrade analytical reproducibility across research teams. Consequently, the authors emphasized the need for a practical and transparent strategy for resolution selection that can operate within established workflows, including practical scenarios where over-clustering must be systematically identified and corrected prior to interpretation.

To resolve these inconsistencies, the team introduced SUN, which stands for Stable-interval guided Unsupervised Neighborhood resolution selection. Designed as an unsupervised layer that integrates within standard Seurat-style computational pipelines, the method operates directly on a fixed shared nearest-neighbour graph. Rather than requiring investigators to manually inspect multiple arbitrary parameter settings, the algorithm systematically scans a defined resolution grid using computational caching.

Within the scanned resolution grid, SUN identifies stable intervals by locating maximal continuous runs of constant cluster counts along the cached resolutions, which the researchers describe as plateaus. To determine which plateau represents the most appropriate clustering resolution, the system evaluates and ranks the candidate intervals using a dedicated composite scoring metric. This quantitative ranking score combines an embedding-based approximate silhouette calculation with the physical width of the stable interval, balancing geometric separation in the embedding space against the mathematical stability of the identified partition across resolution changes.

The performance of the algorithm was initially established across controlled synthetic simulations encompassing 72 distinct testing conditions. To rigorously assess clustering agreement, the investigators employed multiple standard evaluation metrics, primarily relying on the adjusted Rand index, the Fowlkes–Mallows index, and normalised mutual information, supplemented by internal indices as complementary evidence. In these 72 simulated conditions, SUN achieved a mean adjusted Rand index of 0.860 with a standard deviation of 0.230, demonstrating robust recovery of underlying cluster structures across diverse parameterizations without relying on manual user supervision.

Following simulation testing, the researchers evaluated the framework across eight real-world single-cell RNA-sequencing benchmark datasets. At the dataset level, SUN achieved a mean adjusted Rand index of 0.864 with a standard deviation of 0.099. Across the same eight benchmark datasets, the method recorded a mean Fowlkes–Mallows index of 0.886 with a standard deviation of 0.083, alongside a mean normalised mutual information score of 0.873 with a standard deviation of 0.087. These empirical outcomes demonstrated generally high concordance with curated reference annotations while confirming that the approach remains competitive against alternative unsupervised partitioning methods.

The authors further evaluated SUN in a dedicated case study focusing on the correction of cortex over-clustering, benchmarked against the Manual Adjust evaluation target established in the scSHC literature. In this cortex case study, SUN achieved an adjusted Rand index of 0.996 relative to the reference manual adjustment partition. Notably, the unsupervised algorithm successfully recovered the expected UBC and MLI1 merge direction, illustrating its practical capacity to resolve over-clustered partitions and accurately guide cellular aggregation without requiring undocumented manual interventions from data analysts.

Despite its strong concordance across benchmarks, the authors outlined key practical caveats associated with the method, noting that seed sensitivity on certain graph structures and the heuristic design of the composite ranking score remain practical considerations rather than fully eliminated sources of variability. To facilitate adoption and independent auditability within standard single-cell bioinformatics routines, the SUN codebase has been made publicly accessible through a GitHub repository maintained by the BioinfoCenterSYSMH team, complemented by an active online trial deployment hosted on a public Shiny application server.

Subject of Research: Unsupervised Louvain resolution selection for single-cell RNA sequencing

Article Title: SUN for scRNA-seq: stable-interval and embedding-based approximate silhouette guided Louvain resolution selection

Article References: Wang, J., Liang, Q., Wang, J., Yang, B., Li, G., Pang, J., Liu, J., Liao, J., Zheng, Y., & Liang, M. (2026). SUN for scRNA-seq: stable-interval and embedding-based approximate silhouette guided Louvain resolution selection. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06766-9

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06766-9

Keywords: SUN, scRNA-seq, single-cell RNA-seq, Louvain resolution, clustering resolution, shared nearest-neighbour graph, SNN graph, stable intervals, approximate silhouette, adjusted Rand index, Fowlkes–Mallows index, normalised mutual information, Seurat-style workflows, over-clustering correction, cortex benchmark, scSHC, UBC/MLI1 merge, BMC Bioinformatics