NUS CellScope: Breakthrough Single-Cell Atlas Workflow Using Manifold Fitting in Nature Communications

Revolutionizing Single-Cell Research: NUS CellScope Delivers Unprecedented Resolution

Researchers at the National University of Singapore (NUS) have introduced CellScope, a groundbreaking framework that transforms how scientists construct high-resolution cell atlases from single-cell RNA sequencing (scRNA-seq) data. Published in the prestigious journal Nature Communications on December 30, 2025, this innovation addresses longstanding challenges in analyzing cellular heterogeneity. Led by Associate Professor Zhigang Yao from NUS's Department of Statistics and Data Science, the tool promises to accelerate discoveries in biology and medicine by enabling multi-level exploration of cell types, subtypes, and lineages.

Single-cell RNA sequencing has revolutionized biology by providing gene expression profiles for thousands to millions of individual cells, revealing diversity within tissues that bulk methods obscure. However, noise from technical artifacts, dropout events, and irrelevant housekeeping genes often hampers accurate clustering and visualization. CellScope tackles these issues head-on, offering superior performance over industry standards like Seurat and Scanpy.

Understanding the Challenges in scRNA-seq Data Analysis

scRNA-seq data resides in high-dimensional space where true biological signals lie on a low-dimensional manifold—a smoother, intrinsic structure capturing cell states. Yet, observed data is distorted by two noise types: (1) housekeeping genes constant across cells, inflating noise space, and (2) technical noise like mRNA loss in signal space. Traditional pipelines struggle with scalability on large datasets (>50,000 cells), rare cell detection (<5% abundance), and hierarchical relationships, leading to suboptimal clustering (Adjusted Rand Index or ARI around 0.65-0.68).

In Singapore's vibrant biotech ecosystem, NUS researchers recognized these gaps, building on prior work like scAMF (published in PNAS, 2024), which first applied manifold fitting for denoising. CellScope extends this into a complete workflow for atlas construction.

The Core Innovation: Two-Stage Manifold Fitting in CellScope

CellScope's power stems from its two-stage manifold fitting process, mathematically grounded in density-distance metrics to separate signal from noise.

• Stage 1: Signal Gene Selection Identifies "manifold seeds"—high-density, distant cells in PCA space—using local density (ρ) and relative distance (δ), combined into γ score. "Highly reliable cliques" form from seed neighbors. ANOVA selects top 500 genes showing low intra-clique variance but high inter-clique differences, filtering housekeeping genes.
• Stage 2: Technical Noise Reduction Projects low-density outliers (bottom 5%) onto high-density submanifolds via estimation (t=0.9), refining positions to emphasize biology.

Post-fitting, UMAP builds a neighborhood graph (Gaussian kernel), enabling agglomerative clustering with average linkage. This yields a tree-structured representation, visualizing hierarchies from broad lineages to fine subtypes.

Benchmark Superiority: Outperforming Seurat and Scanpy Across 36 Datasets

Tested on 36 diverse scRNA-seq datasets (90 to 265,767 cells, human/mouse brain, pancreas, PBMC, etc.), CellScope achieved average ARI of 0.88 ± 0.014—topping Seurat (0.65) and Scanpy (0.68) on 32/36 sets (p < 10⁻⁵). It excelled in accuracy (ACC), Normalized Mutual Information (NMI), Jaccard Index (JI), and Silhouette Scores, especially for rare cells.

Runtime was fastest or competitive, scaling to massive data where others faltered. Gene selection beat Disp, VST, HVG via ASW and purity metrics. Hybrids (CellScope genes + other clustering) confirmed its edge.Read the full benchmarks in Nature Communications.

Photo by Bing Hui Yau on Unsplash

Tree-Structured Visualization: Unlocking Multi-Level Biological Insights

CellScope's hallmark is its intuitive tree: UMAP embeds fitted data, with hierarchical clustering overlaid, showing clusters merging upward. Genes gain dynamic identities—housekeeping (stable), moderately/strongly cell-type-related—via Wasserstein distances between siblings.

In human midbrain (Siletti-1 dataset), it resolved Oligodendrocyte subtypes (OL1/OL2) with markers like RBFOX1 (myelination regulator), unseen by others. Sankey diagrams traced transitions, e.g., from progenitors to mature cells. In COVID-19 PBMC, it separated monocyte-dendritic subsets, revealing severity-linked markers (IFIT1/OAS2) for antivirals.

Practical Advantages: Scalable, Parameter-Light, and Interpretable

Unlike hyperparameter-heavy tools, CellScope is robust across PCA dims, gene counts, seeds (ARI stable). It runs on laptops for <50k cells, subsets for larger. Open-source Python package on GitHub includes tutorials for clustering, trees, markers.

For researchers, this means faster, clearer atlases without tuning—ideal for brain, cancer, immunology studies.

NUS's Growing Leadership in Single-Cell Bioinformatics

Under Prof. Yao's group, NUS bridges stats and biology. Preceding scAMF (PNAS 2024) pioneered manifold denoising; CellScope builds hierarchical atlases. Singapore's NUS ranks top globally (QS 2026: 8th), fueling biotech via RIE2030 investments.Explore Singapore higher ed opportunities.

NUS hosts Single-Cell Res/volution events, contributes to Asian immune atlases—positioning it as Asia's hub for computational biology.

Real-World Applications: From Neuroscience to Infectious Diseases

CellScope unveils hierarchies in development (embryos), aging (Alzheimer's brain), cancer (tumor microenvironments). In COVID, it pinpointed immune shifts, aiding vaccine design. Future: personalized medicine via patient atlases, rare disease subtype discovery.

Stakeholders praise its interpretability: "Reveals insights others miss," per benchmarks.

Photo by Scribbling Geek on Unsplash

Career Opportunities in Singapore's Booming Biotech Sector

This breakthrough highlights demand for bioinformaticians. NUS trains via stats-data science programs; Singapore offers research jobs at A*STAR, CSI. Aspiring pros can leverage tools like CellScope for theses, publications.

Check Rate My Professor for NUS faculty insights; explore academic CV tips.

Future Outlook: Transforming Global Cell Atlas Initiatives

CellScope integrates with Human Cell Atlas, accelerating multi-omics. Yao group eyes multimodal (spatial+RNA) extensions. For Singapore, it bolsters biomedical sovereignty amid R&D surge (S$37B RIE2030).

Download code, try tutorials—join the revolution in single-cell analysis. Opportunities abound in university jobs and higher ed careers.