scUnified: AI-Ready Standardized scRNA-seq

• scUnified is a curated, AI-ready collection of 13 uniformly processed scRNA-seq datasets spanning human and mouse across nine tissue types.
• It applies consistent quality control, normalization, and annotation in a uniform AnnData (.h5ad) format to support tasks like clustering, cell type classification, and marker discovery.
• Benchmarking with methods such as Leiden, scMAE, and GeneCompass demonstrates its ability to enable reproducible multi-dataset analysis and fair comparisons.

scUnified is an AI-ready standardized resource for single-cell RNA sequencing analysis that consolidates 13 high-quality scRNA-seq datasets spanning two species, human and mouse, and nine tissue types. Its central purpose is to provide standardized, analysis-ready datasets for tasks such as cell clustering, cell type annotation, marker gene identification, and classification, thereby reducing the reproducibility problems created by heterogeneous data formats, preprocessing workflows, and annotation strategies. All datasets undergo standardized quality control and preprocessing and are distributed in a uniform .h5ad format based on AnnData, so they can be used directly in computational pipelines without additional data cleaning (Xu et al., 30 Sep 2025).

1. Concept and design objectives

scUnified is defined as a curated, uniformly processed collection of scRNA-seq datasets intended for direct use in modern machine learning and AI workflows. In practical terms, it is a standardized resource of 13 publicly available datasets, all preprocessed in the same way and stored in a uniform AnnData representation. The resource is described as “AI-ready” because the data are already quality controlled, normalized, log-transformed, scaled, and consistently annotated, allowing direct use in classical clustering methods such as Leiden, deep clustering methods such as scMAE and scCDCG, foundation models such as GeneCompass and scGPT-like models, and downstream analyses including marker gene discovery and cell type classification (Xu et al., 30 Sep 2025).

The design objective is not to create a new atlas from raw sequencing experiments, but to create a standardized layer over existing public scRNA-seq data. This standardization has four explicit components: no extra data munging, consistent preprocessing, annotations included, and multi-task support. Uniform feature representation and data structure are provided through AnnData; the same preprocessing and quality-control logic are applied across datasets; every cell carries a ground-truth cell type label; and each dataset is intended to support clustering, cell-type annotation, marker gene identification, and classification. In this sense, scUnified is a benchmarking and development corpus as much as a data repository.

2. Dataset composition and quantitative profile

scUnified integrates 13 datasets drawn from Tabula Muris, Tabula Sapiens, Muraro et al., and MacParland et al. The included tissues are pancreas, liver, ear (Crista ampullaris and Utricle), lung, testis, trachea, limb muscle, brain, and kidney. The collection spans CEL-seq2, 10X Genomics, and Smart-seq2 protocols, with all mouse datasets coming from Tabula Muris Smart-seq2 (Xu et al., 30 Sep 2025).

Dataset	Species / tissue	Cells; genes; clusters; protocol
Mauro Pancreas	Human / pancreas	2,122; 19,046; 9; CEL-seq2
Sonya Liver	Human / liver	8,444; 4,999; 11; 10X Genomics
Sapiens Liver	Human / liver	2,152; 61,759; 15; Smart-seq2
Sapiens Ear Crista Ampullaris	Human / ear	2,357; 61,759; 7; Smart-seq2
Sapiens Ear Utricle	Human / ear	611; 61,759; 5; Smart-seq2
Sapiens Lung	Human / lung	6,530; 61,759; 25; Smart-seq2
Sapiens Testis	Human / testis	7,494; 61,759; 8; Smart-seq2
Sapiens Trachea	Human / trachea	22,592; 61,759; 20; Smart-seq2
Muris Limb Muscle	Mouse / limb muscle	3,855; 21,609; 6; Smart-seq2
Muris Brain	Mouse / brain	13,417; 21,609; 2; Smart-seq2
Muris Kidney	Mouse / kidney	1,817; 21,609; 9; Smart-seq2
Muris Liver	Mouse / liver	2,859; 21,609; 11; Smart-seq2
Muris Lung	Mouse / lung	5,167; 21,609; 25; Smart-seq2

The quantitative range is intentionally broad. Cells per dataset range from 611 to 22,592, with 3 datasets exceeding 10,000 cells. Genes per dataset range from 4,999 to 61,759, and 6 datasets are described as “high-dimensional” with more than 60,000 genes. Annotated clusters range from 2 to 25 per dataset, while the text also notes up to 39 in aggregate; 4 datasets have at least 20 clusters. Sparsity, defined as the percentage of zero entries in the cell-by-gene count matrix, ranges from 73.02% to 95.42%, and 12 of 13 datasets have sparsity greater than 80%. Formally, for raw count matrix $X \in \mathbb{R}^{N \times G}$,

$$\text{sparsity}(X) = \frac{\#\{(i,j): x_{ij}=0\}}{N \times G} \times 100\%.$$

The selection criteria emphasize high-quality public references with curated annotations, adequate cell numbers, coverage of major cell types, and standard scRNA-seq technologies. Low-quality, poorly annotated, or idiosyncratic datasets are explicitly avoided. This choice suggests that scUnified is optimized for fair method comparison rather than maximal breadth of biological conditions.

3. Harmonized preprocessing and storage model

All datasets are passed through a unified preprocessing pipeline. The workflow is: raw data are loaded into AnnData objects; the presence or absence of library size normalization, log-transformation, and scaling is inspected; and for datasets lacking any of these operations, library size normalization, log1p transformation, and z-score scaling are applied. The paper emphasizes harmonization of these steps rather than specifying low-level filtering thresholds such as minimum genes per cell or mitochondrial cutoffs (Xu et al., 30 Sep 2025).

Library size normalization is defined per cell. For cell $i$ and gene $j$, with raw counts $x_{ij}$, the library size is

$s_i = \sum_{k=1}^{G} x_{ik}.$

Counts are then rescaled to a common library size, with a typical formulation

$$x'_{ij} = \frac{x_{ij}}{\sum_{k=1}^{G} x_{ik}} \times 10^4.$$

This is followed by log1p transformation,

$y_{ij} = \log(1 + x'_{ij}),$

and gene-wise z-score scaling. Let

$$\mu_j = \frac{1}{N}\sum_{i=1}^{N} y_{ij}, \qquad \sigma_j = \sqrt{\frac{1}{N}\sum_{i=1}^{N}(y_{ij}-\mu_j)^2},$$

then

$z_{ij} = \frac{y_{ij}-\mu_j}{\sigma_j}.$

The storage model is equally central. All datasets are stored as .h5ad files containing AnnData objects with a standardized internal layout. adata.X contains the main expression matrix, typically the log-normalized and scaled representation used downstream. adata.obs stores per-cell metadata, including the standardized cell_type field. adata.var stores gene-level metadata, including feature_name. adata.uns can contain unstructured dataset-level metadata such as color maps or embedding parameters, and adata.obsm can hold derived embeddings such as PCA or UMAP. This common representation supports direct interoperability with Scanpy, PyTorch-based models, TensorFlow workflows, and Seurat through format conversion.

4. Annotation regime and analytical scope

scUnified uses the original study annotations as its primary ground truth. Cell types are imported from the source publications and standardized into a consistent per-cell field, adata.obs["cell_type"]. The resource is described as providing “multi-level annotations,” but the explicit minimum guarantee is a dataset-specific cell type label for every cell and a dataset-level count of annotated clusters. No formal ontology, such as Cell Ontology identifiers, is imposed, and no global cross-dataset vocabulary is enforced. Instead, label consistency is achieved by standardized field names and uniform storage conventions, while preserving the source-specific label names (Xu et al., 30 Sep 2025).

This annotation strategy has two consequences. First, the datasets are immediately usable for supervised evaluation, because each cell carries a ground-truth label. Second, cross-dataset biological equivalence is not forced. A plausible implication is that scUnified is best suited to per-dataset benchmarking and transfer-learning experiments rather than ontology-level harmonization across atlases.

The paper also demonstrates a reproducible cluster annotation procedure based on marker overlap. Marker genes for each annotated cluster are derived with Scanpy’s rank_genes_groups, producing a “gold-standard marker set” for each reference cluster. For a predicted cluster $$\text{sparsity}(X) = \frac{\#\{(i,j): x_{ij}=0\}}{N \times G} \times 100\%.$$0 with top 100 marker genes $$\text{sparsity}(X) = \frac{\#\{(i,j): x_{ij}=0\}}{N \times G} \times 100\%.$$1 and a reference cluster $$\text{sparsity}(X) = \frac{\#\{(i,j): x_{ij}=0\}}{N \times G} \times 100\%.$$2 with top 100 marker genes $$\text{sparsity}(X) = \frac{\#\{(i,j): x_{ij}=0\}}{N \times G} \times 100\%.$$3, the overlap score is

$$\text{sparsity}(X) = \frac{\#\{(i,j): x_{ij}=0\}}{N \times G} \times 100\%.$$4

Each predicted cluster is assigned the label of the reference cluster with maximal overlap. This gives a reproducible annotation scheme that complements purely combinatorial cluster-to-label matching.

The resource is explicitly intended to support unsupervised clustering, cell type classification and annotation, marker gene identification, and multi-dataset benchmarking. It does not define a fixed train/validation/test split. Instead, the demonstration protocol relies on repeated runs—5 runs per method per dataset—and reporting mean performance under consistent preprocessing and evaluation metrics.

5. Benchmarking framework and representative findings

The paper evaluates scUnified with three clustering methods and one classification method: Leiden, implemented via Seurat with automatic cluster number; scMAE, a masked-autoencoder deep clustering method; scCDCG, a deep graph clustering method with autoencoder and structural embedding; and GeneCompass, a knowledge-informed cross-species foundation model for classification. Clustering performance is measured with Accuracy (ACC), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI). Classification performance is measured with ACC, Precision (PRE), and Recall (REC). All methods are run 5 times per dataset, and mean values with small standard deviations are reported (Xu et al., 30 Sep 2025).

The reported results show that the standardized corpus can resolve meaningful differences among methods. On Mauro Pancreas, Leiden achieves approximately 92.1% ACC, 90.0% NMI, and 93.6% ARI; scMAE reaches approximately 95.6% ACC and 92.4% ARI; and scCDCG reaches approximately 92.7% ACC and 91.4% ARI. On Sonya Liver, scMAE clearly outperforms Leiden and scCDCG, with ACC of 80.7% versus 69.8% for Leiden and ARI of 88.9% versus 54.9% for Leiden. On more complex tissues such as Sapiens Lung, Muris Kidney, and Muris Lung, the deep learning methods often obtain higher ARI and ACC than Leiden, which the paper interprets as reflecting better capture of complex structure.

GeneCompass yields consistently high classification accuracy across the datasets. ACC is often above 95%, and Muris Brain reaches 100% ACC, PRE, and REC. Mauro Pancreas reaches approximately 98.35% ACC, 97.29% PRE, and 98.26% REC, while Sonya Liver reaches approximately 98.58% ACC. The paper interprets these results as evidence that the standardized preprocessing and annotations provide clean inputs for foundation-model-based classification.

A detailed case study is provided on Muris Limb Muscle using scCDCG. The learned embeddings show strong intra-cluster coherence in cosine-similarity heatmaps, and two-dimensional t-SNE projections show well-separated clusters. Marker identification with rank_genes_groups selects the top 100 differentially expressed genes per cluster; dot plots display the top 3 marker genes per cluster, including Chodl, Des, and Cd82 for Cluster 1, and Gpx3, Cd63, and Sdc4 for Cluster 3. Predicted clusters are then annotated by marker overlap, yielding labels such as endothelial cell and mesenchymal stem cell. Sankey-diagram comparisons between marker-overlap annotation and best-mapping via the Hungarian algorithm indicate that marker-overlap produces more biologically coherent associations than purely combinatorial matching.

6. Reproducibility, field position, and limitations

scUnified is publicly available through a GitHub repository at https://github.com/XPgogogo/scUnity-AI, with the qualification that the URL string uses “scUnity-AI” and the exact project naming should be checked in the repository. The resource includes .h5ad files for each dataset, and the accompanying description indicates likely scripts or notebooks for downloading and loading data, reproducing experiments, and running baseline methods. The experiments use recommended default parameters from the original methods with minimal adjustments for stability, which is consistent with the resource’s role as a standardized benchmarking substrate rather than a task-specific software package (Xu et al., 30 Sep 2025).

Within the broader single-cell ecosystem, scUnified occupies a distinct position. Tabula Muris and Tabula Sapiens are large atlases, but are characterized here as mainly raw or lightly processed data with differing formats, preprocessing choices, and annotation conventions. Existing benchmarks often focus on a single task such as clustering, imputation, or batch correction, and frequently use inconsistent preprocessing across methods. scUnified differs by standardizing multiple high-quality datasets from different sources into one coherent collection with uniform preprocessing and format, explicitly designed for clustering, classification, annotation, and marker discovery.

The limitations are also clearly defined. Species coverage is currently restricted to human and mouse. Biological coverage is limited to 9 tissue types, excluding many organs, disease states, tumors, and developmental time series. Protocol coverage is primarily Smart-seq2, supplemented by one 10X Genomics dataset and one CEL-seq2 dataset; newer modalities such as spatial transcriptomics, multiome data, and other omics layers are not part of this version. The preprocessing design assumes that a single normalization-plus-log1p-plus-scaling pipeline is adequate across datasets, although the text notes that some methods may prefer raw counts or alternatives such as SCTransform. scUnified does not perform explicit batch integration into a single corrected atlas, and it does not impose a formal cross-dataset ontology. For these reasons, it is best understood as a standardized benchmarking corpus rather than an integrated biological reference.

The stated future directions are expansion to more species, additional tissues, and complementary omics data such as ATAC and multi-omics. A possible source of confusion is nomenclature: in the provided literature, a descriptively similar use of “scUnified” appears in connection with a generic query-classification framework for e-commerce, but that usage is distinct from the single-cell RNA-seq resource discussed here (Yuan et al., 26 Jun 2025).