The algorithms described on this page are used to generate the cell-feature matrix file annotations and secondary analysis results, differential gene expression files, and associated metrics in the output bundle:
cell_feature_matrix.zarr.zipcontains all of the secondary analysis results as an AnnData objectdiffexp_genes/contains differential expression results in CSV format, which could be read by e.g., LLMs for a quick check of annotating clustersanalysis_summary.html: Cell Annotation and Analysis tabs
The transcripts Zarr file contains all decoded transcripts. Secondary analysis uses the cell-feature matrix and is performed on:
- Transcripts that pass the default quality value (Q-Score) threshold of Q20 and are assigned to cells
- Cells with more than five transcripts per cell
The Atera Onboard Analysis secondary analysis pipeline produces UMAP, graph-based clusters, and differential gene expression results using RAPIDS-singlecell tools (details below).
Cell type annotation clusters are also generated for human samples. The results provide cell types that are well-studied and documented within the Pan-human Azimuth model (details below). However, there may be expression profiles that are distinct even within the annotated cell types. The graph-based clustering can be useful for further exploration of these cell subtypes, as well as cell types that are not covered by the Pan-human Azimuth model.
Normalization, dimensionality reduction, and clustering can be memory intensive for large datasets. To improve memory usage for samples with more than 1 million cells, the Atera Onboard Analysis generates a "sketch" dataset using uniform subsampling. Sketching is a technique that generates a subset of the dataset that captures the biological complexity of the entire dataset, runs the computationally intensive parts of the algorithm on this small subset, and then projects results to the entire dataset.
The secondary analysis workflow includes these tasks:
- Data filtering: Filter low-quality cells or low-signal genes from the cell-feature matrix. They are unlikely to be biologically meaningful and can add noise to downstream steps.
- Normalization and scaling: Normalize expression values across cells, select informative genes based on expression data, and scale values to support stable PCA and UMAP generation.
- Principal Components Analysis (PCA): Compress the normalized expression matrix into fewer dimensions that capture biological variation and enable efficient downstream analysis.
- Neighborhood graph and Uniform Manifold Approximation and Projection (UMAP): Following the data dimensionality reduction via PCA, the pipeline computes a neighborhood graph by measuring Euclidean distances between neighboring cells in this PCA space. The UMAP algorithm uses the graph to project cells in 2D, allowing you to visually explore cell groups and outliers. The cell assignments in the UMAP projection can be visualized by unsupervised cluster results or by annotated cell type.
- Clustering and differential gene expression: Cells are clustered based on their neighborhood graph, so related cell populations can be identified and compared. These clusters are used for the differential expression task, which identifies genes that distinguish one group of cells from another. When viewing graph-based cell clusters in 10x Explorer, the cells with fewer than five transcripts are labeled "Unassigned".
- Cell type annotation: Cells are clustered based on their cell type annotation (details below). When viewing annotated cell clusters in 10x Explorer, the cells with fewer than five transcripts are labeled "Unannotated", while the cells that the Pan-human Azimuth model could not assign (e.g., due to low transcripts or ambient RNA) are labeled "Unassigned".
The following parameter values and methods are set as the defaults in the Atera Onboard Analysis pipeline using the GPU-accelerated rapids-singlecell package (v0.16.1). Consult the rapids-singlecell documentation for implementation details and the preprint for benchmarking and algorithm information (Dicks et al., 2026).
| Task | Parameter or method |
|---|---|
| Global parameter | The pipeline sets a random seed. Fixing the seed makes UMAP and clustering results more reproducible across runs, which is important for downstream steps (i.e., differential gene expression), publishing results, and analysis reproducibility. |
| Filtering | Minimum cell counts per gene: Remove genes that are expressed in fewer than 3 cells. Minimum transcript counts per cell: Remove cells that have fewer than 10 transcripts per cell. Minimum transcript counts per gene: Remove genes with fewer than 10 total transcripts. |
| Normalization and Scaling | HVG method: The default Highly Variable Gene (HVG) method, Seurat V3, ranks genes using normalized dispersion (variance/mean, Z-scored within expression bins) calculated from log-normalized counts. Number of HVGs: 2,000 Normalization method: Log Norm Scaling: True. The pipeline scales gene expression values prior to PCA to a comparable range, so PCA is not dominated by a few high-variance genes. Target sum: Each cell is normalized so its total transcript count sums to 1,000. This puts normalized expression values on a common scale across samples, making them directly comparable. |
| Principal Components Analysis (PCA) | Number of PCs: The default is 10, consistent with 10x Genomics Cell Ranger and Xenium Onboard Analysis pipelines. |
| Uniform Manifold Approximation and Projection (UMAP) | Number of neighbors: 15 Minimum distance: 0.5 Spread: 1.0 |
| Clustering | Clustering method: Leiden Resolution: 1.0 |
| Differential Expression analysis | Differential expression method: The default method, Wilcoxon, is the standard, robust choice for single cell workflows and is recommended for most analyses. |
The pipeline performs cell type annotation for human Atera WTA samples using the Pan-human Azimuth model, which was developed by the Satija lab as part of The Human BioMolecular Atlas Program (HuBMAP) (Sarkar et al., 2026). This model is only compatible with human samples.
The first iteration of this model was trained on scRNA-seq and snRNA-seq data from 23 different tissues. Cancer datasets were excluded from the training process. Cell types are organized into a unified cell ontology. For model details and updates, see the Pan-human Azimuth page from the Satija lab.
- Dicks S, et al. GPU-accelerated single-cell analysis at scale with rapids-singlecell. arXiv. 2026.
- Sarkar S, et al. Organism-scale annotation with Pan-human Azimuth. bioRxiv. 2026.