SingleCellMQC: A comprehensive quality control workflow for single-cell multi-omics Authors: Daihan Ji, Mei Han, Shuting Lu, Jiaying Zeng, Wen Zhong Published: 2026 ABSTRACT SummaryAs single-cell multi-omics studies scale in size and complexity, comprehensive and modality-aware quality control (QC) is essential to ensure data integrity. Here, we develop SingleCellMQC, an open-source R package that provides a unified QC framework for single-cell RNA sequencing (scRNA-seq), surface proteome profiling (antibody-derived tags, ADTs), and immune repertoire (T cell receptors [TCRs]/B cell receptors [BCRs]) data. SingleCellMQC implements multi-level QC across sample, cell, feature, and batch levels, integrating empirical thresholds, tissue-specific reference ranges, and data-driven outlier detection. Built on Seurat and BPCells, SingleCellMQC supports common preprocessing outputs and generates interactive hypertext markup language (HTML) reports with visual summaries and automated QC flags. Its modular architecture allows flexible integration with existing workflows, and the implementation is optimized for scalability on standard computing environments. The performance and reliability of SingleCellMQC were demonstrated in three datasets: an in-house peripheral blood mononuclear cells (PBMCs) multi-omics dataset (28,498 cells), a public PBMC scRNA-seq dataset (137,214 cells), and a large-scale breast tissue scRNA-seq dataset (> 1 million cells). Introduction Single-cell omics technologies have revolutionized our understanding of cellular complexity, providing multi-dimensional insights into molecular mechanisms, lineage differentiation, and phenotypic heterogeneity at single-cell resolution.1,2,3,4,5,6,7,8 In biomedical and precision medicine research, a common strategy for single-cell multi-omics integration involves combining single-cell RNA sequencing (scRNA-seq), the most mature and widely used single-cell technology, with additional omics layers, such as surface protein expression, T cell receptors (TCRs) and B cell receptors (BCRs) sequencing.9,10,11,12,13,14 For example, the cellular indexing of transcriptomes and epitopes by sequencing (CITE-seq) and RNA expression and protein sequencing assay (REAP-seq) enable simultaneous measurements of proteins and RNAs within the same cell.11,12,15,16 Additionally, the expanded CRISPR-compatible cellular indexing of transcriptomes and epitopes by sequencing (ECCITE-seq) integrates 5′ capture-based scRNA-seq methods, allowing the detection of V(D)J regions of rearranged TCR and BCR mRNAs alongside surface proteins and scRNA-seq data, enabling systematic analysis of cellular phenotypes and immune repertoires.17,18,19 In parallel, additional single-cell modalities, including single-cell assay for transposase-accessible chromatin sequencing (scATAC-seq) and single-cell DNA methylation profiling, provide valuable epigenetic insights into gene regulatory landscapes.20,21 While these approaches continue to advance rapidly, they are often applied in more focused or mechanistic studies. Together, these integrative approaches provide a comprehensive view of the cellular landscape and immune repertoire diversity, allowing for the precise identification of rare and effective cell populations, thereby providing insights for precision medicine. Comprehensive quality control (QC) is crucial to ensure the reliability of single-cell data and enable meaningful comparisons across samples and modalities. Technical variation can be introduced at multiple stages, including sample collection, cell dissociation, library preparation, and sequencing, resulting in artifacts such as mislabeling, cell damage, doublets, ambient RNA contamination, inconsistent antibody staining, V(D)J misassignment, and batch effects.22,23,24,25,26,27,28,29,30,31,32 The integration of multiple omics modalities further increases the complexity that requires careful alignment, normalization, and harmonization to address varying data quality and technical biases.31 Numerous tools have been developed to address specific QC challenges,23,27,28,29,33,34,35,36,37,38,39,40,41,42,43 including low-quality cell filtering methods such as miQC,35 ddqc,39 and median absolute deviation (MAD)-based thresholds23; doublet detection tools such as scDblFinder33 and DoubletFinder28; and background contamination correction approaches such as CellBender,44 DecontX,29,41 and single-cell Contamination Detection and Correction (scCDC).40 Several integrated QC tools have also been developed to streamline multiple QC processes. For example, SCTK-QC supports empty droplet filtering, doublet detection, ambient RNA correction, and interactive report generation for scRNA-seq data.45 scQCEA combines QC with enrichment analysis for scRNA-seq.46 scRNABatchQC evaluates systematic errors, batch effects, and outlier samples by comparing technical factors and biological metrics across datasets.36 Despite these advances, as single-cell multi-omics datasets continue to grow in scale and complexity, there remains a pressing need for unified and scalable QC tools that enable multi-level quality assessment across modalities for large-scale single-cell multi-omics datasets. To address this gap, we developed SingleCellMQC, an open-source R package that provides a modular, integrative QC framework for single-cell multi-omics datasets. SingleCellMQC supports scRNA-seq, single-cell surface proteomics, and single-cell TCR/BCR sequencing data, implementing QC at four analytical levels: sample, cell, feature, and batch. The package generates interactive hypertext markup language (HTML) reports with comprehensive visual summaries, providing a flexible and user-friendly platform for data quality assessment before downstream analysis. SingleCellMQC is freely available in the GitHub repository at https://github.com/WenZhong-Lab/SingleCellMQC. Results Overview of SingleCellMQC workflow The SingleCellMQC workflow consists of four major QC modules: sample-level QC, cell-level QC, feature-level QC, and batch-level QC, which together provide a systematic framework for evaluating data quality at multiple levels (Figures 1 and S1A). SingleCellMQC supports multiple commonly used single-cell data formats (e.g., h5, mtx, csv, standard matrices, and BPCells matrices) as input. These data are automatically converted into Seurat objects internally for downstream analyses. All functions within the SingleCellMQC package are built upon the Seurat and BPCells frameworks and are compatible with commonly used single-cell analysis pipelines.47,48,49 Functions used for each module are listed in Table S1. Each QC module can be executed independently, allowing users to perform QC tasks for specific aims. Moreover, interactive HTML reports summarizing QC results can be generated via the “RunReport” function. A comparison of SingleCellMQC with existing QC tools for single-cell studies is summarized in Table 1. In addition, the workflow is optimized for large-scale datasets, and all QC tasks can be executed on a standard laptop without requiring high-performance computing.Figure 1Overview of the SingleCellMQC frameworkSchematic overview of the SingleCellMQC framework, an open-source R package for comprehensive QC of single-cell multi-omics data. The pipeline processes single-cell RNA-seq, surface proteome profiling, and immune repertoire (TCR/BCR) data through four hierarchical modules: (1) sample-level QC, (2) cell-level QC, (3) feature-level QC, and (4) batch-level QC. Built on Seurat and BPCells frameworks, each module implements specialized functions for QC metrics calculation, quality assessment, outlier detection, and artifact identification. The package generates both static and interactive visualizations with downloadable data tables, culminating in HTML reports with automated quality flagging across all modalities. The figure was created using BioRender (https://biorender.com/).Table 1Comparison of SingleCellMQC with other single-cell QC toolsSingleCellMQCscQCEASCTKSeuratscRNABatchQCscaterSupported input data structures CellRanger outputs (h5/mtx)✓✓✓✓✓✓ Seurat object (R)✓–✓✓–– BPCells sparse matrix (disk-backed)✓––✓–– Generic count matrices (csv, txt, and mtx)✓✓✓✓✓✓Supported data modalities scRNA-seq (gene expression)✓✓✓✓✓✓ ADT (surface protein)✓✓–✓–– V(D)J (TCR/BCR repertoire)✓✓––––Sample-level QC Multi-omics sample QC metrics✓✓–✓–– Cell-type composition-based outlier detection✓––––– Tissue-specific reference-based QC✓––––– Data-driven outlier detection✓––––– Sample identity validation✓–––––Cell-level QC Multi-omics cell QC metrics✓––✓–– Empty droplet detection✓–✓––– Doublet detection RNA-based detection✓–✓––– ADT-based detection✓––––– V(D)J-based detection✓––––– Doublet detection benchmarking✓––––– Low quality cell detection method Empirical threshold-based QC✓–✓✓✓✓ Data-driven QC✓––––✓ V(D)J-based QC✓–––––Feature-level QC Multi-omics feature QC metrics✓––✓–– Ambient/background contamination correction Global correction✓–✓––– Targeted correction✓–––––Batch-level QC Sample-level evaluation✓––––– Cell-level evaluation✓–✓✓✓✓ Feature-level evaluation✓–––✓✓Empirical QC thresholds✓–––––Tissue-specific reference database✓–––––Interactive HTML QC report✓✓✓––– Overview of the SingleCellMQC framework Schematic overview of the SingleCellMQC framework, an open-source R package for comprehensive QC of single-cell multi-omics data. The pipeline processes single-cell RNA-seq, surface proteome profiling, and immune repertoire (TCR/BCR) data through four hierarchical modules: (1) sample-level QC, (2) cell-level QC, (3) feature-level QC, and (4) batch-level QC. Built on Seurat and BPCells frameworks, each module implements specialized functions for QC metrics calculation, quality assessment, outlier detection, and artifact identification. The package generates both static and interactive visualizations with downloadable data tables, culminating in HTML reports with automated quality flagging across all modalities. The figure was created using BioRender (https://biorender.com/). Table 1Comparison of SingleCellMQC with other single-cell QC toolsSingleCellMQCscQCEASCTKSeuratscRNABatchQCscaterSupported input data structures CellRanger outputs (h5/mtx)✓✓✓✓✓✓ Seurat object (R)✓–✓✓–– BPCells sparse matrix (disk-backed)✓––✓–– Generic count matrices (csv, txt, and mtx)✓✓✓✓✓✓Supported data modalities scRNA-seq (gene expression)✓✓✓✓✓✓ ADT (surface protein)✓✓–✓–– V(D)J (TCR/BCR repertoire)✓✓––––Sample-level QC Multi-omics sample QC metrics✓✓–✓–– Cell-type composition-based outlier detection✓––––– Tissue-specific reference-based QC✓––––– Data-driven outlier detection✓––––– Sample identity validation✓–––––Cell-level QC Multi-omics cell QC metrics✓––✓–– Empty droplet detection✓–✓––– Doublet detection RNA-based detection✓–✓––– ADT-based detection✓––––– V(D)J-based detection✓––––– Doublet detection benchmarking✓––––– Low quality cell detection method Empirical threshold-based QC✓–✓✓✓✓ Data-driven QC✓––––✓ V(D)J-based QC✓–––––Feature-level QC Multi-omics feature QC metrics✓––✓–– Ambient/background contamination correction Global correction✓–✓––– Targeted correction✓–––––Batch-level QC Sample-level evaluation✓––––– Cell-level evaluation✓–✓✓✓✓ Feature-level evaluation✓–––✓✓Empirical QC thresholds✓–––––Tissue-specific reference database✓–––––Interactive HTML QC report✓✓✓––– Comparison of SingleCellMQC with other single-cell QC tools To demonstrate the functionality of SingleCellMQC, we applied the workflow to three single-cell datasets: (1) dataset 1: an in-house single-cell multi-omics dataset generated using the 10× Genomics Chromium 5′ Gene Expression platform, which included matched scRNA-seq, surface proteomics (antibody-derived tags, ADTs), and TCR/BCR sequencing across four peripheral blood mononuclear cell (PBMC) samples (totaling 28,498 cells, Table S2). Three samples (TP1, TP2, and TP3) were collected from the same donor at three-day intervals, and one sample (TP3) that failed initial QC was re-measured (TP3-rep); (2) dataset 2: a public scRNA-seq dataset (10× Genomics) comprising 17 PBMC samples (137,214 cells)13; and (3) dataset 3: a large-scale scRNA-seq dataset (10× Genomics) comprising 167 breast tissue samples and over one million cells.50 Each functional module of SingleCellMQC is described in detail in the following text. Sample-level QC The main tasks of sample-level QC include: (1) evaluation of sample QC metrics; (2) detection of outlier samples with abnormal cell-type compositions; and (3) verification of sample identity based on marker gene expression (Figure 1). These tasks are designed to assess technical quality, biological composition, and potential sample mislabeling in a complementary manner. To implement the SingleCellMQC sample-level QC workflow, count matrix or contig annotation files generated from single-cell preprocessing tools, such as Cell Ranger and dnbc4tools, are required. The “CalculateMetrics” function generates a total of 54 QC metrics (Table S3), covering scRNA-seq, surface proteomics (ADT), and TCR/BCR sequencing data. This design allows sample quality to be evaluated from multiple complementary perspectives across transcriptomic, ADT, and TCR/BCR modalities. Sequencing and library-level QC metrics from the “metrics_summary.csv” file generated by Cell Ranger can also be incorporated into the workflow. For scRNA-seq data, key QC metrics include the total number of cells (nCell), the number of detected genes (nGene_RNA), the distribution of per-cell unique molecular identifier (UMI) counts (nCount_RNA), the distribution of per-cell non-zero expression genes (nFeature_RNA), and the distribution of per-cell mitochondrial gene expression (percent.mt) (Figures 2A–2C). For ADT data, QC metrics include the number of detected surface proteins (nPro_ADT), the distribution of per-cell ADT UMI counts (nCount_ADT), and the distribution of per-cell isotype control reads (percent.isotype) (Figures 2A, 2D, and 2E). For TCR/BCR sequencing, QC metrics include the number of cells with detected TCR or BCR chains (nCell_TCR and nCell_BCR), the proportion of cells assigned to specific TCR or BCR subtypes, the proportion of cells with particular receptor groups (e.g., TCR, BCR, TRA+TRB, single IGH, and multichain_T), the per-cell distribution of CDR3 length, and the percentage of unique clonotypes (Figures 2A, 2F–2H, and S1B). Samples falling below empirically defined QC thresholds (summarized in Table S4) or exhibiting significant deviations via MAD analysis are automatically flagged as warning samples (Figure 2I). We combined empirically defined thresholds with MAD-based outlier detection to enable more reliable identification of low-quality samples.Figure 2Sample-level QC(A) A bar plot showing key QC metrics for each sample in dataset 1.(B) Violin plots showing per-sample distributions of the total genes detected per cell (nFeature_RNA, left) and the proportion of mitochondrial reads per cell (percent.mt, right) in dataset 1.(C and D) Interactive tables summarizing per-sample statistics of percent.mt (C) or total UMI counts associated with ADT per cell (nCount_ADT) (D) in dataset 1, including mean, median, quartiles, minimum, and maximum values.(E) Violin plots showing per-sample distributions of nCount_ADT (left) and the proportion of isotype control reads per cell (percent.isotype, right) in dataset 1.(F) A scatterplot showing the distribution of the proportion of detected TCR chains per sample (TCR%) versus BCR% across samples in dataset 1. The solid black line indicates equal proportions of TCR% and BCR%.(G) Distribution of the proportion of receptor subtypes across samples in dataset 1.(H) Density plot showing the distribution of TCR CDR3 length in each sample in dataset 1.(I) Violin plots showing the distribution of sample-level QC metrics: median UMI counts per cell (nCount_RNA), median nFeature_RNA, and median percent.mt in dataset 1. Red lines indicate thresholds based on median absolute deviation (MAD), with red dots marking potential outlier samples.(J) A bar plot showing average sample-level cell-type proportions within each tissue in the reference cell proportion database across 2,299 samples.(K) PCA of cell-type proportions across samples in dataset 1. Outliers were identified by DBSCAN clustering. The blue ellipse indicates the 95% confidence region of the core population.(L) Interactive table showing the computationally inferred biological sex for each PBMC sample in dataset 1. The values indicate the percentage of cells expressing sex-specific genes. 3 male-specific genes, DDX3Y, UTY, and RPS4Y1, and 1 female gene, XIST were evaluated in each sample.(M) A bar plot showing the distribution of cell-type proportions across samples in dataset 1.(A), (C)–(D), (F)–(I), and (K)–(M) are based on four independent PBMC samples. (B), (E), and (H) show distributions across 28,498 cells in four samples. For the boxplots overlaid on violin plots, the center line indicates the median, boxes indicate the interquartile range (IQR), and whiskers extend to 1.5 × IQR. Sample-level QC (A) A bar plot showing key QC metrics for each sample in dataset 1. (B) Violin plots showing per-sample distributions of the total genes detected per cell (nFeature_RNA, left) and the proportion of mitochondrial reads per cell (percent.mt, right) in dataset 1. (C and D) Interactive tables summarizing per-sample statistics of percent.mt (C) or total UMI counts associated with ADT per cell (nCount_ADT) (D) in dataset 1, including mean, median, quartiles, minimum, and maximum values. (E) Violin plots showing per-sample distributions of nCount_ADT (left) and the proportion of isotype control reads per cell (percent.isotype, right) in dataset 1. (F) A scatterplot showing the distribution of the proportion of detected TCR chains per sample (TCR%) versus BCR% across samples in dataset 1. The solid black line indicates equal proportions of TCR% and BCR%. (G) Distribution of the proportion of receptor subtypes across samples in dataset 1. (H) Density plot showing the distribution of TCR CDR3 length in each sample in dataset 1. (I) Violin plots showing the distribution of sample-level QC metrics: median UMI counts per cell (nCount_RNA), median nFeature_RNA, and median percent.mt in dataset 1. Red lines indicate thresholds based on median absolute deviation (MAD), with red dots marking potential outlier samples. (J) A bar plot showing average sample-level cell-type proportions within each tissue in the reference cell proportion database across 2,299 samples. (K) PCA of cell-type proportions across samples in dataset 1. Outliers were identified by DBSCAN clustering. The blue ellipse indicates the 95% confidence region of the core population. (L) Interactive table showing the computationally inferred biological sex for each PBMC sample in dataset 1. The values indicate the percentage of cells expressing sex-specific genes. 3 male-specific genes, DDX3Y, UTY, and RPS4Y1, and 1 female gene, XIST were evaluated in each sample. (M) A bar plot showing the distribution of cell-type proportions across samples in dataset 1. (A), (C)–(D), (F)–(I), and (K)–(M) are based on four independent PBMC samples. (B), (E), and (H) show distributions across 28,498 cells in four samples. For the boxplots overlaid on violin plots, the center line indicates the median, boxes indicate the interquartile range (IQR), and whiskers extend to 1.5 × IQR. To detect anomalies in cell-type composition across samples, SingleCellMQC integrates ScType51 for automated annotation of nine selected cell types based on curated marker genes (Table S5 and STAR Methods). Reference distributions for these nine cell-type proportions were derived from 2,299 single-cell samples across ten human tissues in the Deeply Integrated Human Single-Cell Omics (DISCO) dataset52 (Figures 2J and S1C and Table S6, and STAR Methods). Using the “FindCommonPCTOutlier” function, users can identify warning samples, in which cell-type proportions fall outside the reference ranges for at least three cell types. Users can also optionally specify custom expected ranges for cell-type composition. This reference-based strategy provides an empirical baseline for detecting deviations from expected compositional patterns. In addition, SingleCellMQC supports an unsupervised, data-driven outlier detection method based on principal component analysis (PCA) of cell-type compositions. By combining PCA with Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and confidence ellipses, SingleCellMQC identifies samples with distinct cell-type compositions (Figures 2K, S1D, and S1E). Users can also group samples by biological or experimental context to allow the detection of group-specific outliers that deviate in cell-type composition. Besides, SingleCellMQC incorporates a gene expression-based strategy for sample identity verification, enabling detection of potential sample mislabeling. Specifically, SingleCellMQC allows users to verify sample identity by assessing the proportion of cells expressing genes of interest. The “PlotSampleLabel” function enables the visualization of expression patterns for user-defined gene sets across samples to confirm the biological identity. By default, sex-specific genes, including XIST, DDX3Y, UTY, and RPS4Y1, are included to assess the consistency between predicted and recorded sex information (Figure 2L). In dataset 1, sample TP3 was automatically flagged as low quality due to multiple QC metrics falling below empirical thresholds, such as a low fraction of confidently mapped reads in RNA sequencing (58.3% vs. recommended 70.0%), low mapped reads to V(D)J regions (36.98% vs. 40%), and a low fraction of usable ADT reads (10.83% vs. 20%) (Table S7). The “FindSampleMetricsWarning” function further confirmed TP3 as an outlier, identifying notable deviations in nCount_RNA, nFeature_RNA, and percent.mt compared to other samples in the dataset (Figure 2I). Also, cell-type composition analysis revealed that TP3 exhibited a markedly increased proportion of B cells and a decreased proportion of T cells (Figure 2M). These deviations were detected by both the data-driven DBSCAN method and by comparison to tissue-specific reference ranges derived from single-cell blood datasets (Figure 2K and Table S8). All four samples were correctly annotated as male based on the expressions of sex-specific genes (XIST, DDX3Y, UTY, and RPS4Y1) (Figure 2L). In dataset 2, 15/17 (88.2%) samples passed sample-level QC with expected cell-type composition for blood samples (Table S8), while sample F013 was flagged as a potential outlier due to deviations in multiple QC metrics or cell-type composition (Tables S8 and S9). In dataset 3, 161 of 167 samples (96.4%) passed sample-level QC with expected cell-type composition for breast tissue (Table S8). Six samples, such as hbca_c10 and hbca_c11, were flagged as potential outliers based on deviations in both QC metrics and cell-type composition (Figure S1F and Tables S8 and S9). Across all three datasets, samples flagged as outliers exhibited consistent abnormalities across multiple independent QC dimensions, underscoring the robustness of the integrated sample-level QC framework. Cell-level QC The cell-level QC in SingleCellMQC includes the identification of empty droplets, doublets, and low-quality cells. The detection of empty droplets is implemented by incorporating the “emptyDrop” function from the “DropletUtils” package.27 For doublet detection, SingleCellMQC integrates five widely used RNA expression-based methods (DoubletFinder, scDbFinder, cxds, bcds, and hybrid),28,33,42 along with two complementary multi-omics strategies: (1) an ADT-based method that identifies heterotypic doublets by detecting co-expression of mutually exclusive surface markers, and (2) a V(D)J-based method that flags cells containing multiple TCR and/or BCR chains (see STAR Methods). Because doublets may present different molecular signatures across data modalities, these complementary strategies improve detection coverage. To aid in method selection, SingleCellMQC evaluates the influence of doublet removal on clustering performance using five metrics, including Average Silhouette Width (ASW), Calinski-Harabasz index (CH), Davies-Bouldin Index (DBI), Adjusted Rand Index (ARI), and Normalized Mutual Information (NMI), to identify the approach that produces optimal clustering results (see STAR Methods). To detect low-quality cells, SingleCellMQC utilizes cross-modal QC features derived from RNA, ADT, and V(D)J data. Cells that are damaged, dead, or lysed typically exhibit high mitochondrial gene abundance, low UMI counts, and a reduced number of detected genes in single-cell studies.22 At the protein level, such cells exhibit reduced total ADT UMI counts (nCount_ADT) and elevated isotype background (percent.isotype). In V(D)J sequencing, cells with only a single detected chain are most likely the results of stochastic capture or sequencing dropouts rather than true biology. For each cell, SingleCellMQC computes a comprehensive set of QC metrics across RNA, protein and V(D)J modalities (Table S3). RNA-level QC metrics include the total UMI count (nCount_RNA), the number of detected genes (nFeature_RNA), and the proportions of mitochondrial (percent.mt), ribosomal (percent.rb), and hemoglobin (percent.hb) genes. ADT-level QC metrics include nCount_ADT, the number of detected ADTs (nFeature_ADT), and percent.isotype. In addition, SingleCellMQC infers the receptor subtype for each cell from the V(D)J data using the “CalculateMetricsPerCell” function, enabling the identification of incomplete or ambiguous chain profiles. Based on these QC metrics, SingleCellMQC provides empirical QC thresholds (see STAR Methods) and also allows users to define customized thresholds. In addition, given the substantial variability of QC metrics across tissues, sample types, and experimental conditions,39 SingleCellMQC further integrates three data-driven methods (miQC, ddqc, and MAD) to identify low-quality cells in a dataset-specific and modality-aware manner. This combined strategy improves robustness while reducing dependence on fixed thresholds alone. We applied these methods to three datasets to validate their performance. In dataset 1, which contained 28,498 cells, RNA-based methods identified 1,838 (6.45%) doublets, while the ADT and VDJ-based approaches identified 480 (1.68%) and 764 (2.68%) doublets, respectively (Table S10). In dataset 2, RNA-based methods identified 11,090 (8.08%) doublets and the VDJ method identified 5,792 (4.22%) doublets (Table S10). In dataset 3, 146,464 (8.79%) doublets were identified by RNA-based methods (Table S10). In dataset 1, the ADT-based doublet detection method effectively identified T-B doublets using two mutually exclusive surface marker pairs (CD19/CD3 and CD20/CD3), and the most ambiguous doublets flagged by V(D)J-based method overlapped with those identified by the ADT-based approach (Figures S2A–S2D). However, many doublets detected by ADT- or V(D)J-based methods were missed by RNA-based methods, indicating the importance of integrating modality-specific QC methods for comprehensive doublet detection (Figure 3A). We next ranked RNA-based doublet detection methods using five clustering-based metrics (ASW, CH, DBI, ARI, and NMI). scDbFinder achieved the highest overall ranking, outperforming other methods across all three datasets (Figures 3B–3D).Figure 3Cell-level QC(A) An UpSet plot displaying the intersection patterns among different doublet detection methods in dataset 1.(B–D) Bar plots showing the distribution of the overall clustering performance ranks for each sample across five RNA doublet detection methods in dataset 1 (4 samples) (B), dataset 2 (17 samples) (C), and dataset 3 (167 samples) (D).(E) An UpSet plot displaying the intersection patterns among different low-quality cell detection methods in all samples in dataset 1.(F) Violin plots showing the distributions of percent.mt, nFeature_RNA, and nCount_RNA between “fail” and “pass” cells in samples from dataset 1 (28,498 cells). Within each sample, cells are categorized as fail (low-quality, identified by at least two detection methods) or pass (non-low-quality).(G) Violin plots showing the distributions of percent.mt, nFeature_RNA, and nCount_RNA between fail and pass cells in samples from dataset 2. Data are shown for three randomly selected samples (14,698 cells) within dataset 2.(H) Violin plots showing the distributions of percent.mt, nFeature_RNA, and nCount_RNA between fail and pass cells in samples from dataset 3. Data are shown for three randomly selected samples (21,769 cells) within dataset 3.(A), (B), (E), and (F) are based on dataset 1 (4 PBMC samples). For the boxplots overlaid on violin plots, the center line indicates the median, boxes indicate the interquartile range (IQR), and whiskers extend to 1.5 × IQR. Cell-level QC (A) An UpSet plot displaying the intersection patterns among different doublet detection methods in dataset 1. (B–D) Bar plots showing the distribution of the overall clustering performance ranks for each sample across five RNA doublet detection methods in dataset 1 (4 samples) (B), dataset 2 (17 samples) (C), and dataset 3 (167 samples) (D). (E) An UpSet plot displaying the intersection patterns among different low-quality cell detection methods in all samples in dataset 1. (F) Violin plots showing the distributions of percent.mt, nFeature_RNA, and nCount_RNA between “fail” and “pass” cells in samples from dataset 1 (28,498 cells). Within each sample, cells are categorized as fail (low-quality, identified by at least two detection methods) or pass (non-low-quality). (G) Violin plots showing the distributions of percent.mt, nFeature_RNA, and nCount_RNA between fail and pass cells in samples from dataset 2. Data are shown for three randomly selected samples (14,698 cells) within dataset 2. (H) Violin plots showing the distributions of percent.mt, nFeature_RNA, and nCount_RNA between fail and pass cells in samples from dataset 3. Data are shown for three randomly selected samples (21,769 cells) within dataset 3. (A), (B), (E), and (F) are based on dataset 1 (4 PBMC samples). For the boxplots overlaid on violin plots, the center line indicates the median, boxes indicate the interquartile range (IQR), and whiskers extend to 1.5 × IQR. Across datasets, RNA-, ADT-, and V(D)J-based methods identify distinct sets of low-quality cells (Figure 3E). In dataset 1, for example, the ddqc method identified 1,314 cells (4.61%) as low-quality, the MAD-based mitochondrial percentage method (RNA_mt_3mad) identified 3,378 cells (11.85%), ADT_umi_3mad identified 121 cells (0.43%), and VDJ method based on chain types identified 1,568 cells (5.50%) (Table S10). In dataset 2, for example, the ddqc method identified 6,179 cells (4.50%) as low-quality, the RNA_mt_3mad identified 8,227 cells (6.00%), and V(D)J method identified 20,542 (14.97%) (Table S10). These differences reflect that each omics modality is sensitive to different experimental conditions (Figure 3E). We further quantified the proportion of low-quality cells detected by at least two methods. Across three datasets, the proportion of low-quality cells ranged from 9.16% to 27.73% in dataset 1, 0.76% to 17.69% in dataset 2, and 2.48% to 26.74% in dataset 3 (Figure S2E). As expected, these cells exhibited hallmark signatures of low-quality cells, including a higher percentage of mitochondrial gene expression and reduced number of detected genes and UMI counts (Figures 3F–3H). Feature-level QC Feature-level QC is a critical yet often overlooked component in single-cell analysis, essential for accurate interpretation of cell functions and reliable cross-sample comparisons. Inadequate feature-level QC may interfere with downstream analyses, leading to biased clustering annotations or confounding effects. In SingleCellMQC, feature-level QC mainly involves systematic background noise correction, which aims to detect and remove contamination caused by ambient RNA and nonspecific antibody binding. Feature-level QC metrics are also provided to aid in the detection of features with irregular or unexpected distributions. To address contamination from sources such as ambient RNA or nonspecific antibody binding, SingleCellMQC employs the “RunCorrection” function, which supports global, targeted, or combined background correction strategies. For global correction of RNA contamination, DecontX is applied to model and remove ambient RNA signals.41 For targeted correction, scCDC is applied to identify and remove highly contaminated RNAs or proteins40 (Figures 4A and 4B). In dataset 1, several classic surface protein markers, such as CD4 and CD3 in sample TP1, CD8 in TP2, lgM in TP3, and CD8 in TP3-rep, exhibited high contamination levels (Figures 4A and S3A–S3C). Although scCDC was originally developed based on RNA data, it also performed well on our protein dataset (Figures 4B and S3A–S3C).Figure 4Feature-level QC(A and B) Dot plots showing the highly contaminated proteins identified by the scCDC package comparing their expression levels before (A) and after (B) scCDC correction in TP1 sample in dataset 1.(C) Bar plots (5% bins) showing the distribution of gene detection rates in TP1 RNA data in dataset 1. Corresponding the right bar plots highlight the top 10 genes with the highest detection rates.(D) Bar plots (5% bins) showing the distribution of protein detection rates in TP1 ADT data in dataset 1. Corresponding the right bar plots highlight the top 10 proteins with the highest detection rates.(E) Mean-variance plot for RNA in TP1 cells (dataset 1). The x axis represents the mean of log-normalized counts, and the y axis represents the variance of log-normalized counts for each gene. Each blue dot denotes a single gene. The top 10 genes with the highest variance in log-normalized counts are highlighted in red.(F) Mean-variance plot for protein (ADT) in TP1 cells (dataset 1). The x axis represents the mean of centered log-ratio (CLR) transformed counts, and the y axis represents the variance of CLR transformed counts for each antibody-derived tag (ADT). Each blue dot denotes a single protein ADT. The top 10 ADTs with the highest variance in CLR transformed counts are highlighted in red.(G) Mean-variance plots for dataset 1 (RNA, left) and dataset 1 (ADT, right, 4 samples). Each colored line represents the generalized additive model (GAM)-fitted mean-variance trend for an individual sample (raw RNA or ADT count data).(H) Mean-variance plot for dataset 2 (RNA, 17 samples). Each colored line represents the GAM-fitted mean-variance trend for an individual sample (raw RNA-seq count data).(I) Mean-variance plot for dataset 3 (RNA, 167 samples). Each colored line represents the GAM-fitted mean-variance trend for an individual sample (raw RNA-seq count data).(C)–(F) analyses were performed using the TP1 sample from dataset 1, comprising 36,601 genes and 137 ADTs. Feature-level QC (A and B) Dot plots showing the highly contaminated proteins identified by the scCDC package comparing their expression levels before (A) and after (B) scCDC correction in TP1 sample in dataset 1. (C) Bar plots (5% bins) showing the distribution of gene detection rates in TP1 RNA data in dataset 1. Corresponding the right bar plots highlight the top 10 genes with the highest detection rates. (D) Bar plots (5% bins) showing the distribution of protein detection rates in TP1 ADT data in dataset 1. Corresponding the right bar plots highlight the top 10 proteins with the highest detection rates. (E) Mean-variance plot for RNA in TP1 cells (dataset 1). The x axis represents the mean of log-normalized counts, and the y axis represents the variance of log-normalized counts for each gene. Each blue dot denotes a single gene. The top 10 genes with the highest variance in log-normalized counts are highlighted in red. (F) Mean-variance plot for protein (ADT) in TP1 cells (dataset 1). The x axis represents the mean of centered log-ratio (CLR) transformed counts, and the y axis represents the variance of CLR transformed counts for each antibody-derived tag (ADT). Each blue dot denotes a single protein ADT. The top 10 ADTs with the highest variance in CLR transformed counts are highlighted in red. (G) Mean-variance plots for dataset 1 (RNA, left) and dataset 1 (ADT, right, 4 samples). Each colored line represents the generalized additive model (GAM)-fitted mean-variance trend for an individual sample (raw RNA or ADT count data). (H) Mean-variance plot for dataset 2 (RNA, 17 samples). Each colored line represents the GAM-fitted mean-variance trend for an individual sample (raw RNA-seq count data). (I) Mean-variance plot for dataset 3 (RNA, 167 samples). Each colored line represents the GAM-fitted mean-variance trend for an individual sample (raw RNA-seq count data). (C)–(F) analyses were performed using the TP1 sample from dataset 1, comprising 36,601 genes and 137 ADTs. The feature-level QC metrics in SingleCellMQC include the distribution of detected counts, detection rate, average expression level, and variance across cells. These QC metrics are visualized to facilitate the identification of outlier RNA or ADT features across samples (Figures 4C–4F). Across the three datasets, we observed a large number of lowly expressed RNAs with 18,029 to 18,557 genes detected in fewer than 0.1% of cells in dataset 1, 17,831 to 20,414 in dataset 2, and 11,794 to 16,016 in dataset 3 (Table S11). These results indicate the consistently high prevalence of lowly expressed genes in scRNA-seq data. In contrast, nearly all ADT features were detected across cells in all samples in dataset 1, likely due to ambient or nonspecific antibody binding, which generates pervasive background signal in CITE-seq datasets (Table S11). Moreover, in raw count data, the mean and variance of features generally showed a positive correlation in all datasets (Figures 4G–4I). Batch-level evaluation Batch-to-batch variation is a significant challenge in single-cell analysis, where differences in protocols, reagents, and experimental conditions can introduce noise that obscures true biological signals and leads to misleading interpretations. To address this, SingleCellMQC provides batch-effect evaluation at the pseudobulk sample, single-cell and feature levels, and assesses the contributions of batch-associated covariates to capture both global and specific technical variation. At the pseudobulk sample level, SingleCellMQC aggregates single-cell counts to generate pseudobulk profiles and performs PCA to evaluate the batch effects across samples. Variance decomposition of principal component (PC) scores further quantifies the contribution of each covariate (e.g., platform, donor, and library batch) to sample-level variation (Figures 5A and 5B). In addition, SingleCellMQC applies a linear mixed model to each feature in the pseudobulk data to estimate per-feature variance attributable to each covariate, enabling insights into the main drivers of batch effects (Figures 5C and 5D).Figure 5Batch-level QC(A) The PCA plot of pseudobulk data generated from single-cell RNA data in dataset 3 (167 samples), colored by platform.(B) Bar plots showing the proportions of variance explained by four variables across PCs in dataset 3 pseudobulk data, as estimated using a linear mixed model.(C) Violin plots summarizing variance partitioning analysis of pseudobulk gene expression counts in dataset 3, showing the proportion of expression variance for each of the 32,320 gene explained by four variables, as estimated using a linear mixed model.(D) Violin plots of the proportions of gene expression variance explained by four variables across different cell types in dataset 3 pseudobulk data, as estimated using a linear mixed model. The analysis included 32,320 genes.(E) UMAP plots of cell clustering based on single-cell ADT data in dataset 1 (4 samples, 28,498 cells), with colors indicating distinct cell clusters. AB-1 and AB-2 indicate experimental antibody batches.(F) UMAP plots of cell clustering for single-cell RNA-seq data in dataset 3 (167 samples, 1,666,717 cells) across three sequencing platforms, with colors indicating distinct cell clusters.(G) Bar plot showing the top 10 genes ranked by GTE score for platform in dataset 3.(C) and (D) for the boxplots overlaid on violin plots, the center line indicates the median, boxes indicate the interquartile range (IQR), and whiskers extend to 1.5 × IQR. Batch-level QC (A) The PCA plot of pseudobulk data generated from single-cell RNA data in dataset 3 (167 samples), colored by platform. (B) Bar plots showing the proportions of variance explained by four variables across PCs in dataset 3 pseudobulk data, as estimated using a linear mixed model. (C) Violin plots summarizing variance partitioning analysis of pseudobulk gene expression counts in dataset 3, showing the proportion of expression variance for each of the 32,320 gene explained by four variables, as estimated using a linear mixed model. (D) Violin plots of the proportions of gene expression variance explained by four variables across different cell types in dataset 3 pseudobulk data, as estimated using a linear mixed model. The analysis included 32,320 genes. (E) UMAP plots of cell clustering based on single-cell ADT data in dataset 1 (4 samples, 28,498 cells), with colors indicating distinct cell clusters. AB-1 and AB-2 indicate experimental antibody batches. (F) UMAP plots of cell clustering for single-cell RNA-seq data in dataset 3 (167 samples, 1,666,717 cells) across three sequencing platforms, with colors indicating distinct cell clusters. (G) Bar plot showing the top 10 genes ranked by GTE score for platform in dataset 3. (C) and (D) for the boxplots overlaid on violin plots, the center line indicates the median, boxes indicate the interquartile range (IQR), and whiskers extend to 1.5 × IQR. At the single-cell level, SingleCellMQC performs single-cell clustering analysis on the original single-cell data matrices to visualize batch-related variability and supports both single-omics and multi-omics datasets (Figures 5E and S4A). SingleCellMQC also integrates Harmony32 as an optional batch correction method. At the feature level, SingleCellMQC integrates the group technical effects (GTEs)53 scores to assess the influence of batch covariates on individual features, enabling the identification of features that are influenced by technical variations. As dataset 2 lacked batch metadata, batch-level evaluations were carried out only for datasets 1 and 3. In dataset 1, TP3-rep exhibited marked batch effects in ADT measurements, with notable deviations from other samples in uniform manifold approximation and projection (UMAP) plot (Figure 5E and Table S2). These effects were likely attributed to the utilization of different antibody manufacturing batches.54 In the dataset 3, the pseudobulk sample PCA plot and variance decomposition analysis revealed that sequencing platform and sample source were the major contributors to sample-level variation (Figures 5A, 5B, and S4B). Per-feature variance decomposition analysis further confirmed that sequencing platform was the strongest driver of expression variability, accounting for a median of 34.0% of variance, which was substantially greater than biologically relevant variables such as age (median 0.1%) or procedure group (median ∼0%) (Figure 5C). We also performed variance decomposition within individual cell types and found that, in most cases, sequencing platform and sample source remained the dominant batch-associated covariates (Figure 5D). Further analysis at the cellular level revealed significant heterogeneity between different sequencing platforms, suggesting that sequencing platform may be the primary source of batch effects (Figures 5F and S4C). We further calculated the platform-level GTE scores for each gene and observed that the top 10 genes are predominantly mitochondrial (MTRNR2L12, MT-ND3, and MT-ATP6), ribosomal (RPS17, RPL31, and RPL36A), and dissociation/stress-related (JUND) genes (Figure 5G). HTML report SingleCellMQC provides automated, interactive HTML-based QC reports using R Markdown, offering users with a comprehensive overview of data quality before downstream analysis (Figure 6). Users can perform basic preprocessing with the “RunPreprocess” function or selectively execute individual QC modules. The RunReport function then compiles all QC results into an HTML document containing interactable plots, tables, and QC warnings across sample, cell, feature, and batch levels (Figure 6). To support large-scale datasets, SingleCellMQC integrates BPCells to reduce memory usage and improve computational efficiency. The HTML QC reports for all three datasets are provided (https://doi.org/10.5281/zenodo.15120930). Notably, for the large-scale public breast tissue single-cell RNA-seq dataset containing 167 samples and over one million cells (dataset 3), SingleCellMQC produced a complete, navigable report using RunReport function in approximately 22 min without the need for high-performance computing.Figure 6HTML reports for quality assessmentThe HTML reports generated by SingleCellMQC summarizing QC results for the sample-level QC module (A), cell-level QC module (B), feature-level QC module (C), and batch-level QC module (D) in dataset 1 (4 samples). HTML reports for quality assessment The HTML reports generated by SingleCellMQC summarizing QC results for the sample-level QC module (A), cell-level QC module (B), feature-level QC module (C), and batch-level QC module (D) in dataset 1 (4 samples). Comparison of SingleCellMQC with other QC tools A comparison of SingleCellMQC with existing QC tools, including scQCEA, SCTK, Seurat, scater, and scRNABatchQC, is summarized in Table 1. SingleCellMQC implements a hierarchical QC framework spanning sample-, cell-, feature-, and batch-level assessments and supports multiple single-cell modalities, whereas existing tools are generally limited to specific QC tasks or single data modalities. Within this framework, SingleCellMQC integrates multiple QC strategies, including empirical thresholds, tissue-specific reference-based QC, and data-driven approaches, enabling evaluation of data quality from multiple perspectives within each level. In addition, it generates interactive HTML reports that integrate QC metrics, visualizations, and automated warning flags to facilitate interpretation of QC results. SingleCellMQC also supports large-scale data analysis through BPCells-based on-disk computation. Comparative results based on doublet detection, which is a commonly used and computationally intensive QC task, indicated improved computational efficiency and scalability of SingleCellMQC compared with SCTK (Table S12). For example, when processing 30 samples from dataset 3, SingleCellMQC completed the doublet detection task in 32.33 min using 2.34 GB of memory, whereas SCTK-QC required 754.65 min and 37.78 GB. In addition, SingleCellMQC successfully processed the full dataset 3 (> 1 million cells) with a maximum memory usage of around 10 GB, while SCTK could not complete the analysis at this scale. Discussion In this study, we report SingleCellMQC, a single-cell multi-omics QC tool that provides an integrated workflow for comprehensive QC across multiple single-cell modalities. SingleCellMQC addresses several practical limitations in existing QC approaches through four key features: (1) multi-omics QC integration across scRNA-seq, surface protein sequencing, and immune repertoire data; (2) systematic quality assessment at sample, cell, feature, and batch levels; (3) data-driven, tissue-specific reference ranges for various QC tasks; and (4) computational efficiency optimized for large-scale datasets. By integrating a series of both in-house developed and publicly available QC strategies within a unified pipeline, SingleCellMQC addresses compatibility issues between different QC tools and provides a streamlined solution that connects independent QC processes into a cohesive and efficient workflow. One of the key aspects of SingleCellMQC is its emphasis on a hierarchical sample-to-feature-level QC framework, which is critical yet often overlooked for large-scale single-cell studies. At the sample level, by leveraging both data-driven approaches and a curated reference database derived from 2,299 single-cell samples spanning 10 widely studied human tissues, SingleCellMQC provides a quantitative framework for detecting warning or outlier samples based on unexpected QC metric deviations and abnormal cell-type compositions, both within the analyzed dataset and relative to empirical tissue-specific cell composition references. Through the implementation of data-driven analysis, SingleCellMQC successfully identified anomalies in key QC metrics in dataset 1 and precisely flagged TP3 as a low-quality sample, detecting abnormal cell-type proportions compared with other samples in the dataset and recognizing cell-type distributions that fell outside the expected ranges for PBMC samples. At the cell level, SingleCellMQC integrates multiple doublet detection strategies across RNA, ADT, and V(D)J data modalities. This multi-omics approach improves the identification of both heterotypic and homotypic doublets compared with RNA-based methods alone. As demonstrated in our PBMC dataset, the integration of ADT-based and V(D)J-based strategies successfully captured doublets that were missed by RNA-based methods, particularly T-B heterotypic doublets identified through mutually exclusive surface markers, CD19/CD3 and CD20/CD3, and cells with multiple TCR or/and BCR chain expression. The minimal overlap observed between these approaches highlights the limitations of relying solely on transcriptomic data for doublet identification. For low-quality cell detection, SingleCellMQC utilized data-driven methods to identify outlier cells based on dataset distributions. Furthermore, SingleCellMQC allows the use of five clustering evaluation metrics, including ASW, CH, DBI, ARI, and NMI, to assess the clustering performance after doublet removal for each sample. The results showed that the scDbFinder method performed best in the majority of samples, which is consistent with conclusions from other benchmarking studies on doublet detection methods.55,56,57 For feature-level QC, SingleCellMQC integrates modality-specific background decontamination methods alongside targeted correction of highly contaminated features, providing an effective solution for ambient contamination removal without over-correction. This is particularly important for surface protein profiling. Notably, the scCDC method, originally designed based on RNA data, performed well in our protein dataset, suggesting its potential applicability in protein analysis. This is particularly valuable for CITE-seq datasets, where ambient antibody binding can produce widespread low-level background that inflates the detection of ADT features. By incorporating detection rate, mean expression, and variance-based QC metrics, SingleCellMQC further enables users to identify outlier features that may bias downstream analysis. Batch effects remain a major source of technical variation in large-scale or multi-center single-cell studies, particularly when datasets are generated across different platforms or processing sites. SingleCellMQC assesses batch variation at three complementary levels: pseudobulk, to capture sample-level variation; single-cell, to visualize cell-level batch effects driven by technical covariates; and feature-level, to identify genes or proteins disproportionately affected by specific batch factors. By quantifying the contribution of various batch factors at different levels, SingleCellMQC provides insights into primary sources of batch variation, facilitating informed decisions about correction strategies. In dataset 3, for example, SingleCellMQC identified sequencing platform as the dominant source of batch-to-batch variation, explaining a median of 34% of gene expression variability, which far exceeded biological factors such as age or procedure group. Feature-level analysis using GTE scores further highlighted mitochondrial, ribosomal, and stress-responsive genes as particularly sensitive to platform differences, consistent with known technical artifacts.53 Single-cell clustering analyses also revealed strong platform-dependent separation, highlighting the importance of systematic batch evaluation prior to downstream integration or interpretation. Moreover, the QC strategies within each level are designed to complement each other, allowing data quality to be evaluated from multiple perspectives rather than relying on a single metric or method. This contrasts with many existing QC tools, which typically focus on specific QC tasks or individual data modalities, and therefore provide more limited or fragmented assessments. By integrating multiple QC strategies within a unified framework, SingleCellMQC supports more comprehensive and flexible QC evaluation across different analytical levels. While the use of multiple QC approaches may increase the number of detected quality signals, these generally reflect distinct aspects of data quality rather than redundant findings. Systematically identifying and interpreting these signals is important, as low-quality samples, cells, or features can introduce bias and may be misinterpreted as biological variation.22,58 At the same time, there is an inherent trade-off between removing low-quality data and retaining sufficient data for downstream analysis. In this context, SingleCellMQC allows users to adapt QC criteria to their specific datasets and analysis goals, while its multi-level design facilitates evaluation of QC results from different perspectives. This enables cross-validation of quality signals and supports more balanced filtering decisions. SingleCellMQC supports a wide range of commonly used single-cell data formats, including CellRanger outputs, Seurat objects, BPCells-backed matrices, and generic count matrices, enabling direct use of data from diverse preprocessing pipelines. Built on Seurat and BPCells frameworks, it is compatible with standard single-cell analysis tools and supports efficient analysis of large datasets through on-disk computation, including those exceeding typical in-memory limits, without requiring high-performance computing resources. This is particularly advantageous compared with existing tools that rely primarily on in-memory processing and may encounter scalability limitations for large datasets. In addition, the integrated HTML reporting system provides a structured overview of QC results with automated warning flags, facilitating interpretation and reducing the need for manual inspection. In conclusion, SingleCellMQC provides a standardized, scalable, and comprehensive QC solution for single-cell multi-omics studies, enabling systematic assessment of data quality and ensuring robust downstream analyses in large-scale single-cell research. By providing QC at the sample, cell, feature, and batch levels, this workflow will enhance the reliability and reproducibility of single-cell analyses, ultimately facilitating meaningful cross-sample and cross-study comparisons. Limitations of the study Despite these capabilities, SingleCellMQC has several limitations. First, the input data formats are restricted to specific sequencing platforms. Second, the current version of SingleCellMQC supports transcriptomic, surface protein, and V(D)J data, but does not yet extend to additional single-cell omics layers such as ATAC-seq or DNA methylation, which provide critical epigenetic insights. Third, the reference ranges for tissue-specific cell-type composition remain incomplete, and broader coverage of tissue types, disease conditions, and cell classification categories is needed. Future development will focus on incorporating additional single-cell modalities, expanding compatibility with additional single-cell technologies and extending reference ranges to encompass a wider variety of cell types, tissues, and disease states. Resource availability Lead contact Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Wen Zhong (zhong_wen@gzlab.ac.cn). Materials availability This study did not generate new unique reagents. Data and code availability The original Human PBMC ECCITE-seq data generated in this study are available in the Genome Sequence Archive for Human database under accession code HRA010489 (available at https://ngdc.cncb.ac.cn/gsa-human/). The processed data and HTML QC reports in this study are available in Zenodo under accession code https://doi.org/10.5281/zenodo.15120930. All original codes have been made publicly available on GitHub as of the date of publication at https://github.com/WenZhong-Lab/SingleCellMQC. Acknowledgments This study was supported by the Major Project of Guangzhou National Laboratory (grant no. GZNL2024A03002), the 10.13039/501100014219National Science Foundation for Young Scientists of China (grant no. 8230102287), and the Major Program of National Clinical Research Center for Children and Adolescents’ Health and Diseases in Children’s Hospital of Chongqing Medical University, Chongqing, China (grant no. NCRCCHD-2023-MP-01). Author contributions W.Z. conceptualized, designed, and supervised the study. D.J., M.H., S.L., J.Z., and W.Z. designed the software architecture and analytical framework. D.J., S.L., and J.Z. constructed the package and conducted comprehensive testing on both in-house and public datasets. D.J., M.H., and W.Z. wrote the original draft. W.Z revised and finalized the manuscript with input from all authors. All authors contributed to data interpretation, manuscript revision, and approved the final version of the manuscript. Declaration of interests W.Z., D.J., and M.H. declare that they hold issued patents related to the work described in this study. The authors declare no other competing interests. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERAntibodiesTotalSeq™-C Human Universal Cocktail V1.0BioLegendCat. No. 399905; RRID: AB_2876728Biological samplesHuman PBMCs, see Table S2This studyN/ADeposited dataRaw pbmc single-cell multi-omics sequencingThis studyGSA: HRA010489Breast single cell dataKumar et al.50GEO: GSE2353262,299 single cell dataLi et al.52DISCO (https://disco.bii.a-star.edu.sg/)Public PBMC single cell dataLuo et al.13GEO: GSE157007Software and algorithmsSingleCellMQC (0.7.1)This studyhttps://github.com/WenZhong-Lab/SingleCellMQCADT doublet detection methodThis studyhttps://github.com/WenZhong-Lab/SingleCellMQCSeurat (5.0.1 & 5.1.0)Hao et al.48https://github.com/satijalab/seuratBPCells (0.3.0)Parks et al.49https://github.com/bnprks/BPCellsscater (1.30.1)McCarthy et al.23https://bioconductor.org/packages/release/bioc/html/scater.htmlmiQC (1.10.0)Hippen et al.35https://github.com/greenelab/miQCddqcR (0.1.0)Subramanian et al.39https://github.com/ayshwaryas/ddqc_RscCDC (1.4)Wang et al.40https://github.com/ZJU-UoE-CCW-LAB/scCDCscDblFinder (1.16.0)Germain et al.33https://github.com/plger/scDblFinderscds (1.16.0)Bais et al.42https://github.com/kostkalab/scdsDoubletFinder (2.0.6)McGinnis et al.28https://github.com/chris-mcginnis-ucsf/DoubletFinderreactable (0.4.4)Linhttps://github.com/glin/reactable/DESeq2 (1.42.1)Love et al.59https://bioconductor.org/packages/release/bioc/html/DESeq2.htmlscRepertoire (2.0.0)Borcherding et al.43https://github.com/BorchLab/scRepertoireR (4.3.2)N/Ahttps://www.r-project.org/decontX (1.0.0)Yang et al.41https://github.com/campbio/decontXScTypeIanevski et al.51https://github.com/IanevskiAleksandr/sc-type#readmeHarmony (1.2.0)Korsunsky et al.32https://github.com/immunogenomics/harmonyCell Ranger (7.1.0)10× Genomicshttps://www.10xgenomics.com/support/software/cell-ranger/latestCOSG (0.9.0)Dai et al.60https://github.com/genecell/COSGvariancePartition (1.39.1)Hoffman et al.61https://github.com/GabrielHoffman/variancePartitiondreamlet (1.7.1)Hoffman et al.62https://github.com/GabrielHoffman/dreamletsingleCellTK (2.12.2)Hong et al.45https://github.com/compbiomed/singleCellTKGTEs (1.0.0)Zhou et al.53https://github.com/yzhou1999/GTEs/OtherCell-type proportion referenceThis studyhttps://github.com/WenZhong-Lab/SingleCellMQCQC html reportsThis studyhttps://doi.org/10.5281/zenodo.15120930 REAGENT or RESOURCESOURCEIDENTIFIERAntibodiesTotalSeq™-C Human Universal Cocktail V1.0BioLegendCat. No. 399905; RRID: AB_2876728Biological samplesHuman PBMCs, see Table S2This studyN/ADeposited dataRaw pbmc single-cell multi-omics sequencingThis studyGSA: HRA010489Breast single cell dataKumar et al.50GEO: GSE2353262,299 single cell dataLi et al.52DISCO (https://disco.bii.a-star.edu.sg/)Public PBMC single cell dataLuo et al.13GEO: GSE157007Software and algorithmsSingleCellMQC (0.7.1)This studyhttps://github.com/WenZhong-Lab/SingleCellMQCADT doublet detection methodThis studyhttps://github.com/WenZhong-Lab/SingleCellMQCSeurat (5.0.1 & 5.1.0)Hao et al.48https://github.com/satijalab/seuratBPCells (0.3.0)Parks et al.49https://github.com/bnprks/BPCellsscater (1.30.1)McCarthy et al.23https://bioconductor.org/packages/release/bioc/html/scater.htmlmiQC (1.10.0)Hippen et al.35https://github.com/greenelab/miQCddqcR (0.1.0)Subramanian et al.39https://github.com/ayshwaryas/ddqc_RscCDC (1.4)Wang et al.40https://github.com/ZJU-UoE-CCW-LAB/scCDCscDblFinder (1.16.0)Germain et al.33https://github.com/plger/scDblFinderscds (1.16.0)Bais et al.42https://github.com/kostkalab/scdsDoubletFinder (2.0.6)McGinnis et al.28https://github.com/chris-mcginnis-ucsf/DoubletFinderreactable (0.4.4)Linhttps://github.com/glin/reactable/DESeq2 (1.42.1)Love et al.59https://bioconductor.org/packages/release/bioc/html/DESeq2.htmlscRepertoire (2.0.0)Borcherding et al.43https://github.com/BorchLab/scRepertoireR (4.3.2)N/Ahttps://www.r-project.org/decontX (1.0.0)Yang et al.41https://github.com/campbio/decontXScTypeIanevski et al.51https://github.com/IanevskiAleksandr/sc-type#readmeHarmony (1.2.0)Korsunsky et al.32https://github.com/immunogenomics/harmonyCell Ranger (7.1.0)10× Genomicshttps://www.10xgenomics.com/support/software/cell-ranger/latestCOSG (0.9.0)Dai et al.60https://github.com/genecell/COSGvariancePartition (1.39.1)Hoffman et al.61https://github.com/GabrielHoffman/variancePartitiondreamlet (1.7.1)Hoffman et al.62https://github.com/GabrielHoffman/dreamletsingleCellTK (2.12.2)Hong et al.45https://github.com/compbiomed/singleCellTKGTEs (1.0.0)Zhou et al.53https://github.com/yzhou1999/GTEs/OtherCell-type proportion referenceThis studyhttps://github.com/WenZhong-Lab/SingleCellMQCQC html reportsThis studyhttps://doi.org/10.5281/zenodo.15120930 Experimental model and study participant details Human subjects Peripheral blood samples were collected from a healthy male donor (30 years old) at three time points with 2- or 3-day intervals: Day 0 (TP1), Day 3 (TP2), and Day 5 (TP3). This study was approved by the Ethics Committees of the Guangzhou National Laboratory (2023–0004). Written informed consent was obtained from the volunteer. The study was performed in accordance with the Declaration of Helsinki. Method details Sample collection and PBMC isolation We obtained three peripheral blood mononuclear cells (PBMCs) samples from a healthy donor at three-day intervals to allow for the assessment of natural biological variation within a single individual over time while also providing multiple replicates to evaluate the consistency and reliability of the QC framework. Due to technical issues, one of the samples (TP3) initially failed QC and was subsequently re-measured (TP3-rep) (Table S2). Each module of SingleCellMQC is described in detail below. At each time point, approximately 5 mL of venous blood was collected into ethylenediamine-tetraacetic acid (EDTA) blood collection tubes and processed immediately for PBMC isolation. Density gradient centrifugation was performed using 3.5 mL of Ficoll-Paque PLUS (Cat. No. 17144002) at 400 × g for 35 min (acceleration 5 and deceleration 0) at room temperature. The resulting PBMC-containing buffy coat was transferred to a 15 mL centrifuge tube and washed once with phosphate-buffered saline (PBS). The PBMC pellet was collected by centrifugation at 500 × g for 10 min, then resuspended in 2 mL of red blood cell lysis buffer (Cat. No. 00–4333–57, Invitrogen) and incubated at room temperature for 5 min. The lysed suspension was washed again with PBS and centrifuged at 500 × g for 5 min at 4°C. The final PBMC pellet was resuspended in a cryopreservation buffer containing 10% dimethyl sulfoxide (DMSO; Cat. No. C40050, Beyotime) and stored at −80°C for less than 6 months until single-cell multi-omics sequencing. Single-cell multi-omics sequencing All samples underwent scRNA-seq, surface protein profiling, and scTCR/BCR-seq on the 10× Genomics Chromium platform. The single-cell cDNA library amplification was performed with the 10× Genomics Chromium Next GEM Single Cell 5ʹ Reagent Kits (Cat. No. PN-1000244). The Chromium Single Cell 5ʹ Library Construction Kit (Cat. No. PN-1000190) was used to construct the DNA library. The libraries of V(D)J were amplified using a Chromium TCR amplification kit (Cat. No. PN-1000252) and BCR amplification kit (Cat. No. PN-1000253). Surface protein profiling used TotalSeq-C Human Universal Cocktail V1.0 (Cat. No. 399905) targeting 130 unique cell surface antigens and 7 isotype control antibodies. The constructed libraries were sequenced using PE150 on an Illumina Nova 6000 platform. Raw sequencing data were processed using 10× Genomics Cell Ranger (v7.1.0) with the command ‘cellranger multi’ and GRCh38 reference genome. The processed data comprised 28,498 cells across four samples (TP1, TP2, TP3, TP3-rep), each with comprehensive transcriptome, surface proteome, and immune repertoire profiles (details in Table S2). Quality control on PBMC multi-omics data was conducted with SingleCellMQC under default parameters for all QC modules. Multi-omics QC metrics calculation SingleCellMQC incorporated a comprehensive set of quality control (QC) metrics at sample-, cell- and feature-level, with comprehensive metric definitions and methodological descriptions available in Table S3. 10× metrics quality assessment SingleCellMQC processed Cell Ranger-generated ‘metrics_summary.csv’ files using the ‘Read10XMetrics’ function to standardize metric nomenclature, with specific prefixes: “RNA:” for RNA-seq, “VDJ B:”/“VDJ T:” for V(D)J, and “Antibody:” for ADT data. Samples outside predefined thresholds for 10× metrics are classified as warnings or errors. These quality thresholds were derived from 10× Genomics Cell Ranger requirements (https://github.com/10XGenomics/cellranger, Table S4). MAD-driven warning sample identification Median Absolute Deviation (MAD) analysis was employed in SingleCellMQC to establish statistical thresholds (median ±3×MAD) for each metric in flagging samples exhibiting deviations from group-wide distributions. The ‘FindSampleMetricsWarning’ function implemented this analysis, with optional cohort stratification via the split.by parameter (default for all samples). Default metrics included nCount_RNA, nFeature_RNA, percent.mt, nCount_ADT, nFeature_ADT and percent.isotype. Samples exceeding threshold boundaries were systematically flagged as warning samples. Single-cell clustering analysis pipeline Single-cell normalization RNA sequencing counts were normalized using the ‘LogNormalize’ method, while antibody-derived tag (ADT) counts were normalized via centered log-ratio (CLR) transformation, both performed through the ‘NormalizeData’ function in Seurat.48 Feature selection and scaling For RNA data represented in ‘dgCMatrix’ format, the pipeline identified 2,000 highly variable genes using the ‘FindVariableFeatures’ function (selection.method = “vst”) in Seurat.48 All detected antibody-derived tags (ADTs) were retained by default for analysis. The selected features underwent centering and scaling via ‘ScaleData’ function in Seurat. When processing BPCells RNA data, it employed the ‘matrix_stats’ function (BPCells package) to identify 2,000 highly variable genes, followed by scaling procedures.49 For ADT data, it proceeded directly to scaling without prior feature selection. Dimension reduction Principal component analysis (PCA): The Seurat workflow in SingleCellMQC computed principal components on the scale data using the ‘RunPCA’ function (npcs = 20) in Seurat package.48 The BPCells workflow performs dimensionality reduction through singular value decomposition on the scale data using the ‘svds’ function in BPCells package.49 Uniform manifold approximation and projection (UMAP): The UMAP embeddings were generated from the principal components derived from PCA, using function-specific implementations. The Seurat workflow in SingleCellMQC employed the ‘RunUMAP’ function in Seurat package, while the BPCells workflow utilizes the ‘umap’ function from the umap R package. Harmony batch effect correction Batch effect correction was performed on single-cell data by using Harmony integration method applied to precomputed PCA embeddings (either RNA or ADT-derived) through the ‘RunHarmony’ function in the harmony package.32 Clustering For single-omics clustering analysis, the Seurat framework in SingleCellMQC constructed a shared nearest neighbor (SNN) graph using the ‘FindNeighbors’ (dims = 1:20) function. Subsequent clustering was performed via the ‘FindClusters’ function in Seurat with a resolution parameter set to 1.48 Approximate nearest neighbors were identified in BPCells framework through the ‘BPCellsknn_hnsw’ function, followed by graph construction using ‘BPCellsknn_to_snn_graph’ and graph-based clustering via the ‘BPCellscluster_graph_louvain’ function.49 For CITE-seq clustering analysis, multi-omics integration was performed through weighted nearest neighbor analysis using ‘FindMultiModalNeighbors’ (knn.range = 300, prune.SNN = 1/20, dims.list = list(1:20, 1:10)) function in Seurat, with subsequent clustering via ‘FindClusters’ (algorithm = 3, resolution = 1).48 ScType-based cell type annotation SingleCellMQC performed automated cell type annotation for single-cell RNA sequencing data by ScType algorithm with three marker gene databases, including a curated in-house database (nine major cell types: T cells, B cells, NK cells, dendritic cells, monocytes/macrophages, endothelial cells, fibroblasts, granulocytes, and other cell types), the publicly available ScType database (https://sctype.app/), and the Cell Taxonomy database (https://ngdc.cncb.ac.cn/celltaxonomy/). The original ScType algorithm (https://github.com/IanevskiAleksandr/sc-type) had been incorporated into the ‘RunScType’ function in SingleCellMQC, where users can specify the desired marker database through the data_source parameter (“Main” for the in-house database, “ScType” for the ScType database, or “Cell_Taxonomy” for the Cell Taxonomy database).51,63 For outlier sample detection and batch QC modules, SingleCellMQC first performed clustering analysis on individual samples using the ‘RunPipeline’ function (with preprocess = “rna.umap”) following the single-cell clustering workflow to generate cell clusters, followed by cell type annotation using ‘RunScType’ with the in-house database (data_source = “Main”). Tissue-specific cell type composition reference Public single-cell RNA sequencing data were downloaded from the DISCO database using the DISCOtoolkit package, with sample selection performed through sequential application of the ‘FilterDiscoMetadata’ and ‘DownloadDiscoData’ functions.52 The sample inclusion criteria were as follows: (i) retention of unsorted samples (“cell_sorting” column); (ii) restriction to 10× Genomics 3′ and 5' platform; (iii) exclusion of nuclear transcriptome datasets; and (iv) inclusion of samples containing ≥2,000 cells. Blood tissue data only contained the Asian Immune Diversity Atlas (AIDA) dataset, and breast tissue data only contained GSE235326 dataset.50,64 All other tissues followed the default datasets. This process finally contained 2,299 single-cell samples encompassing 10 human tissue types (blood, bone, bone marrow, breast, intestine, kidney, liver, lung, pancreas, and skin). For each sample, clustering analysis was performed independently, followed by automated cell-type annotation using the ScType method51 (‘RunScType’ function with data_source = “Main”). Cell type proportions were then quantified per sample. Then, for every cell type within each tissue, potential outliers were identified and removed using the ‘isOutlier’ function from the scater package.23 From the retained values, summary statistics, including minimum, maximum, mean, standard deviation (SD), median, first quartile (Q1), third quartile (Q3), and interquartile range (IQR) were computed. The minimum and maximum values were defined as the reference range for each cell type within the respective tissue. Tissue-specific reference-based warning sample detection Cell type annotation for each sample was first performed using ScType-based method, followed by quantification of cell type proportions.51 For each sample, the proportion of each cell type was compared against the defined range (minimum to maximum) in the reference database. A sample was labeled as a ‘warning’, if outliers were detected in at least three distinct cell types. In addition to the built-in reference ranges, users can also provide custom expected cell type composition ranges through the ‘custom_ref_range’ parameter. PCA-based outlier sample detection analysis Before analysis, SingleCellMQC employed ScType-based cell type annotation and quantified cell type proportions for each sample. These cell type proportions as input for PCA using the ‘prcomp’ function (with scale. = T) from the stats package to generate a PCA transformation matrix. Then, the ‘fviz_contrib’ function (factoextra package) calculated the contribution of each cell type proportions to each principal component. For outlier detection, SingleCellMQC performed two approaches: (i) analysis via 95% confidence ellipses derived from the PCA projection, where samples falling outside these confidence boundaries were flagged as potential outliers, and (ii) density-based spatial clustering (DBSCAN algorithm via ‘dbscan’ function in dbscan package) to identify samples exhibiting anomalous clustering patterns in the PCA space. Biological sex prediction Sex-specific gene expression was utilized to predict biological sex, with XIST as the female marker and DDX3Y, UTY, and RPS4Y1 as male markers. A sample was classified as male if the fraction of cells expressing at least one Y-linked gene exceeded the fraction of cells expressing XIST; otherwise, female sex was determined. Empty droplet detection SingleCellMQC identified empty droplets by applying the ‘barcodeRanks’ (lower = 100) function (DropletUtils package) directly to unfiltered raw RNA count matrices.27 RNA-based doublet detection SingleCellMQC integrated five widely-used computational methods for RNA-based doublet detection: (i) DoubletFinder, which identified doublets using the ‘doubletFinder’ function in DoubletFinder package with parameters: PCs = 10, pN = 0.2528; (ii) scDblFinder, employing the ‘scDblFinder’ function in scDblFinder package33; (iii) cxds, utilizing the ‘cxds’ function from the scds package; (iv) bcds, implemented via the ‘bcds’ function (scds package); and (v) hybrid, employing the ‘cxds_bcds_hybrid’ function (scds package).42 Each detection method executed independently per sample. The default baseline doublet rate was set at 0.008 per 1,000 cells. For each sample, the expected doublet rate was computed as 0.08 multiplied by the total cell numbers, divided by 10,000. Benchmarking of RNA -based doublet detection methods For each method-sample combination, we first filtered out predicted doublets and then performed standard clustering using Seurat with default parameters across multiple resolutions (ranging from 1.0 to 1.5 in increments of 0.1).48 Cluster quality was quantified using the bluster, fpc, and clusterSim packages to compute ASW, CH, and DB metrics, while ARI and NMI metrics were calculated using the aricode package to measure the agreement between pre- and post-filtering cell type annotations. The overall rank for each method in each sample was determined through three steps. (i) Metric values were averaged across resolutions for each method-sample combination. (ii) These averaged values were ranked among all methods for each individual metric within each sample, with rank 1 representing optimal performance. (iii) Method performance was determined by averaging five metrics ranks within each sample, followed by a secondary ranking of these averages to obtain overall rank. ADT-based doublet detection For each sample, ADT-based doublet detection was performed through the following steps. (i) Clustering analysis was performed by using the ‘RunPipeline’ (preprocess = “adt.umap”) function on ADT count data. RNA-based clustering (preprocess = “rna.umap”) was implemented when fewer than 10 ADT features were available. (ii) Defining mutually exclusive protein pairs: selection of pairs with established lineage exclusivity based on biological knowledge (e.g., CD3+CD19−in T cells vs. CD3−CD19+ in B cells). (iii) Identification of protein-positive and negative cell clusters in ADT data: for each protein, the percentage of detected protein and mean expression level within each cluster were calculated, multiplying these values to generate a penalized mean score ranked in descending order. The protein-positive clusters were identified by the forward selection algorithm guided by the COSG scores.60 A forward selection algorithm iteratively expanded from the highest penalized mean cluster until COSG score improvement stagnates. Clusters retained after the selection process were classified as protein-positive clusters. (iv) Determining positive/negative CLR expression thresholds for mutually exclusive proteins: for each protein, optimal positive/negative classification threshold was calculated by maximizing the Area Under the Curve (via ‘roc’ function in cutoff package) or using Naive Bayes (via ‘glm’ function in stats package with family = binomial), based on pre-defined positive/negative labels. (v) Doublet identification: a cell was classified as doublet if its CLR-normalized expression values exceeded predetermined thresholds for both proteins in a mutually exclusive marker pair; otherwise was considered as singlet. All these steps can be implemented through the ‘RunDbt_ADT’ function. V(D)J-based doublet and low-quality cell detection SingleCellMQC calculated the number of TCR chains (TRA, TRB) and BCR chains (IGH, IGK, IGL) per cell from filtered contig annotation files. A cell was flagged as a doublet when: (i) it contained at least one BCR chain (IGH, IGK, or IGL) and at least one TCR chain (TRA or TRB); or (ii) it contained at least two immunoglobulin heavy chains (IGH); or (iii) it contained at least two light chains (IGK/IGL); or (iv) it contained at least two T cell receptor α chains (TRA); or (v) it contained at least two β chains (TRB). A cell was flagged as a low-quality cell when it only contained single chain. Cell-level fixed QC thresholds collection We collected 240 high-impact single-cell publications across 11 tissues. For each publication, we extracted the specific filtering thresholds applied to key QC metrics: nCount_RNA, nFeature_RNA, percent.mt, percent.hb (hemoglobin read percentage). To facilitate analysis, the range of these thresholds for each tissue can be conveniently retrieved using ‘GetMetricsRange’ function in SingleCellMQC. Cell-level fixed QC thresholds in three datasets For Dataset 1, cell-level fixed cutoffs were set as nFeature_RNA = 200, nCount_RNA = 500, and percent.mt = 10, as determined by the ‘GetMetricsRange’ function in SingleCellMQC. For Dataset 2, filtering thresholds of nFeature_RNA = 200 and percent.mt = 10 were directly adopted from its original publication. Similarly, for Dataset 3, filtering thresholds of nFeature_RNA = 200, nCount_RNA = 500, percent.mt = 10, and percent.rb = 50 were also directly adopted from its original publication. Data-driven low-quality cell detection (i) miQC, employing the ‘mixtureModel’ function (miQC package) with QC metrics (percent.mt and nFeature_RNA)35; (ii) ddqc, employing the ‘ddqc.metrics’ function (ddqcR package) with count matrix39; and (iii) MAD-based outlier detection (‘isOutlier’ function with nmads = 3 in scater package), applied to RNA metrics (nCount_RNA, nFeature_RNA, percent.mt) or ADT metrics (nCount_ADT, nFeature_ADT, and percent.isotype).23 Each detection method was executed independently per sample. Background noise correction (i) DecontX, employing the ‘decontX’ function in decontX package with RNA count data.41 (ii) scCDC, first identified potential contaminating features using ‘ContaminationDetection’ (restriction_factor = 0.1) function (scCDC) on RNA or ADT data, followed by systematic correction via the ‘ContaminationCorrection’ function (scCDC) with both count data and features (contamination features or user-defined features).40 Each method was executed independently per sample. Pseudobulk data generation Pseudobulk data was generated using the ‘RunPseudobulkData’ function by aggregating single-cell expression counts at the sample level (sample.by parameter), with optional further stratification by a user-specified metadata column (cluster.by parameter). When only ‘sample.by’ was provided, one pseudobulk profile was generated for each sample by summing counts across all cells from that sample. When ‘cluster.by’ was additionally specified, separate pseudobulk profiles were generated for each sample-group combination defined by the selected metadata labels (e.g., cell-type annotations, cluster assignments, or other user-defined categories). In this study, PCA was performed using pseudobulk data generated from all cells within each sample (cluster.by = NULL). For variance decomposition analysis, two types of pseudobulk data were used: one generated from all cells within each sample, and the other generated from cell type-specific subsets within each sample, using ScType-derived cell type annotation labels (‘RunScType’ function with data_source = “Main”) as the ‘cluster.by’ variable. PCA of pseudobulk data The pseudobulk data was processed following the standard Seurat workflow, including normalization (‘NormalizeData’ function), identification of variable features (‘FindVariableFeatures’ function), data scaling (‘ScaleData’ function), and principal component analysis (‘RunPCA’ function). Variance decomposition analysis For each principal component (PC) obtained from the PCA of pseudobulk data, SingleCellMQC fitted a linear mixed model for each PC with batch-associated covariates using the ‘lmer’ function from the lmerTest package. Then, the fraction of variance attributable to each batch-related factor was computed by ‘calcVarPart’ function from the variancePartition package.61 For each feature of pseudobulk data, voom-style normalization was performed using the ‘processOneAssay’ function from the dreamlet package.62 Next, the fraction of variance attributable to each batch-related variable for each feature was estimated by fitting linear mixed models with batch-associated covariates, using the ‘fitExtractVarPartModel’ function from the variancePartition package.61 GTE score calculation For each feature in the single-cell data, the GTE score was calculated using the ‘Run.GroupTechEffects’ function from the GTEs package.53 Interactive data visualization Interactive data visualization was implemented through reactable (interactive table), plotly (interactive plot), and rmarkdown (HTML report) packages. Public dataset Public single-cell RNA sequencing (scRNA-seq) data of 17 PBMC samples were downloaded from Gene Expression Omnibus (GEO): GSE157007.13 Public single-cell RNA sequencing (scRNA-seq) data of 167 breast samples from the primary site was downloaded from Gene Expression Omnibus (GEO): GSE235326.50,52 The count matrices were loaded and processed as BPCells matrix for memory-efficient computation.49 The QC task was performed using SingleCellMQC with default parameters. The comprehensive QC HTML report was generated by ‘RunReport’ functions under default parameters. Quantification and statistical analysis All statistical analyses and visualizations were conducted in R (version 4.3.2) (https://www.r-project.org/). Statistical analyses implemented in this study included MAD, GAM, and linear mixed model. Statistical analyses were performed using Dataset 1 (4 samples, 36,601 genes and 137 ADTs), Dataset 2 (17 samples, 33,689 genes), and Dataset 3 (167 samples, 32,320 genes), as appropriate for each analysis. The corresponding cell numbers, sample sizes (n), definitions of n, model parameters, and other figure-specific statistical details are provided in the corresponding method details and figure legends. No statistical significance testing (e.g., t-tests or wilcoxon rank-sum tests) was performed.