Analysis Methods

A) Demultiplexing Strategy

Each read contains a spatial barcode (SBC) that encodes the location on the substrate where the read originated. The pipeline uses the SBC to spatially locate each read on the substrate, then assigns the read to a sample based on the tissue masks from the ISIT OME-TIFFs. Reads whose spatial location falls outside of any tissue mask are not assigned to a sample.

B) UMI Processing

To avoid overcounting UMIs based on sequence errors, UMI error correction is performed among all reads with the same cell-barcode mapping to the same gene. UMI sequences that are likely errors of another UMI are not counted. After correction, reads that share the same SBC, UMI, and gene assignment are collapsed into a single molecule so that downstream counts reflect unique transcripts rather than read counts. The number of molecules after collapsing is reported as n_deduplicated_reads, and the fraction of reads removed as duplicates is reported as sequencing_saturation (see Alignment Stats).

C) Read Trimming

DRAGEN StrataMap performs a single combined trimming step on the cDNA payload read. It trims:

• Adapter sequences
• Low-quality bases (Q-score < 20)
• polyA tails
• polyG tails

Reads shorter than 22 nt after trimming are discarded. Per-category percentages are reported in the Trimming Stats file.

D) Tissue Mask Filtering

When a registered OME-TIFF is provided, the pipeline uses its tissue mask to restrict analysis to SBCs that fall within the tissue region. The unfiltered matrix (all SBCs on the substrate) is preserved for reference, but all downstream binning, clustering, and reporting use the tissue-mask-filtered matrix.

E) Sample Splitting and Subsamples

Because tissue samples can vary widely in size — from a few square millimeters up to the full capacity of the flow cell — DRAGEN StrataMap automatically splits very large samples into smaller subsamples for alignment, then merges the results back together. This keeps memory and runtime bounded for any size of input and lets subsamples be aligned in parallel.

1) How splitting works

1. During sample prep, the pipeline computes how many SBCs each sample covers based on its tissue mask.
2. If a sample has more SBCs than the per-subsample limit (configured by SBCs-per-subsample), its SBCs are partitioned into approximately-equal groups. Each group becomes a subsample named {Sample_ID}_x001, {Sample_ID}_x002, and so on.
3. Samples that fall under the limit are not split and have no _x00# suffix.

All subsamples of a sample share the same Sample_ID and are reassembled at the end of alignment.

2) What you see in the alignments folder

For a small sample that was not split, the alignments folder contains a single per-sample folder:

01_alignments/
└── SampleA/
    ├── SampleA_alignment_table_user_friendly.csv
    ├── SampleA.spatial.multimap_ties.tsv
    ├── Alignment_stats.csv
    ├── Fastqc_stats.csv
    ├── Quant_stats.csv
    └── Trimming_stats.csv

For a larger sample that was split into three subsamples, you'll see the parent folder plus one folder per subsample:

01_alignments/
├── SampleB/                    <-- merged, sample-level results
│   ├── SampleB_alignment_table_user_friendly.csv
│   ├── SampleB.spatial.multimap_ties.tsv
│   ├── Alignment_stats.csv
│   ├── Fastqc_stats.csv
│   ├── Quant_stats.csv
│   └── Trimming_stats.csv
├── SampleB_x001/               <-- per-subsample DRAGEN outputs
│   ├── SampleB_x001.mapping_metrics.csv
│   ├── SampleB_x001.quant_metrics.csv
│   ├── SampleB_x001.trimmer_metrics.csv
│   └── ...
├── SampleB_x002/
│   └── ...
└── SampleB_x003/
    └── ...

• The {Sample_ID}/ folder contains the merged, sample-level results that you should use for analysis.
• The {Sample_ID}_x00#/ folders contain the raw per-subsample DRAGEN outputs and are preserved for debugging and traceability. You generally do not need to interact with them directly.

3) Merging

After alignment, per-subsample count matrices and metrics are merged back into a single per-sample matrix and a single set of per-sample stats. All downstream steps (tissue mask filtering, binning, clustering, reporting) operate on the merged, sample-level matrix, so downstream outputs are always reported per-sample regardless of whether splitting occurred.

F) Read Alignment

1) Exonic and Intronic Alignments

By default, only exonic alignments are counted. Intronic alignments can be enabled through the UI. Exonic alignments have at least 50% overlap with annotated exons in the GTF file. Intronic reads require 20% overlap with an annotated gene but do not meet the exonic alignment criteria.

2) Multimapping

By default, the pipeline attempts to assign reads that exactly match multiple locations in the genome. Hits are rank-ordered, and each read is assigned to the top hit.

Rank Order:

1. Protein-coding gene
2. Non-protein coding gene
3. Ambiguous hit
4. Wrong strand
5. No feature

If a read maps equally well to two different protein-coding genes, the algorithm picks the first one listed in the GTF. These molecules are tracked and reported in the [SampleID]_multimap_ties.tsv and [SampleID]_multimap_ties_summary.csv files in the intermediate_results/alignment folder.

G) Cell Segmentation

What it does: Identifies individual nuclei masks from H&E-stained tissue images and expands the nuclei borders to estimate cell boundaries.

How it works: Uses a StarDist-based deep learning model trained from scratch on extensive, high-quality nuclei annotations from H&E images. For each image, the model performs instance segmentation and outputs nuclei masks, where positive integers (1, 2, 3, ...) denote individual nuclei instances (cell IDs) at 1 µm/px resolution. After nuclei segmentation, each nucleus mask is expanded outward by up to the user-defined distance (expansion_um, default 5 µm) to estimate the cell border. The expansion is capped if it runs into a neighboring cell's expansion: the boundary between two adjacent cells is placed at the midpoint between them, so a nucleus never expands further than half the distance to its nearest neighbor. On sides with no neighbor within the expansion distance, the boundary reaches the full expansion_um (illustration of cell border expansion below)

H) Cell and Grid Binning

Both methods consolidate individual sequencing barcodes (SBCs) into larger analytical units. This reduces data complexity while preserving spatial information.

1) Grid Binning

What it does: Grid binning divides the tissue sample into a regular, rectangular grid of uniform bins, then aggregates all transcripts within each bin.

How it works:

• You specify a bin size in micrometers, such as 10 µm or 20 µm.
• The pipeline creates a uniform grid overlay across the entire tissue sample.
• Each sequencing barcode is assigned to the grid bin containing its spatial coordinates.
• All transcripts from barcodes in the same bin are summed together.
• Each bin is identified by its center coordinates, such as bin10:150:200 for a 10 µm bin centered at x=150, y=200.

2) Cell Binning

What it does: Cell binning groups transcripts by parent cell. It uses a cell segmentation mask to identify individual cells and aggregate their transcripts.

How it works:

• It uses the cell segmentation image from the cell segmentation module, which is a labeled mask where each pixel contains a cell ID.
• The pipeline identifies the centroid of each cell.
• Each sequencing barcode is mapped to the cell containing its coordinates.
• All transcripts from barcodes within the same cell are summed together.
• Each cell is identified by its label and centroid coordinates, such as cell42:150:200 for cell 42 centered at x=150, y=200.

I) Sequencing Saturation Curves

Saturation curves estimate how many additional molecules or genes would be recovered at higher sequencing depths. They are produced for both cell-binned and grid-binned data.

How it works:

• The pipeline subsamples the aligned reads at a range of depths (expressed as fractions of the reads actually sequenced).
• At each depth, the reads are re-deduplicated using the same SBC/UMI/gene collapsing rule used in the main analysis. The number of unique molecules and unique genes per bin or cell is re-computed.
• A curve is fit to these subsampled points to extrapolate how recovery scales with depth.

Saturation is reported as 1 - (unique molecules / aligned reads). A value near 0 means nearly every additional read yields a new molecule (under-sequenced); a value near 1 means the library is saturated and additional reads will mostly be duplicates.

Output files and plot variants are described in 02 Sequencing Saturation.

J) Tertiary Analysis

After binning, each bin type (cell, grid, and optionally nuclei) is taken through the same tertiary analysis pipeline. Parameter names and defaults for each step are listed in the analysis config.

1) Cell and Bin Filtering

Low-quality cells and bins are removed before clustering. A cell or bin is dropped if any of the following apply:

• Its percent mitochondrial transcripts is above mt_threshold.
• Its total unique-molecule count is below umi_count_threshold and below the umi_percentile_threshold of the sample. Whichever threshold is stricter (higher) is applied.

Filtering is performed independently for each bin type, so a cell that is dropped from the cell-binned analysis does not affect the grid-binned analysis. Filtered-out cells are retained in the raw AnnData object but are not shown in the clustered output and appear as the "NA cluster" in downstream visualizations.

2) Normalization

Each remaining cell or bin is scaled so that its total UMI count equals normalize_target_sum (by default, the median total count of the sample prior to normalization). The normalized counts are then log1p-transformed. This makes UMI distributions comparable between cells of different library sizes.

3) Feature Selection

To reduce noise and computation, only the top top_n_genes most variable genes are retained for dimensionality reduction. Variability is measured on the log-normalized matrix using the standard Scanpy highly-variable-genes method.

4) Dimensionality Reduction

• PCA is run on the filtered, normalized, highly-variable gene matrix using n_pcs principal components.
• A nearest-neighbor graph is built in PCA space using n_neighbors neighbors.
• UMAP is computed on the neighbor graph using umap_min_dist as the minimum embedding distance. Smaller values preserve local structure; larger values emphasize global relationships.

5) Clustering

Cells and bins are clustered using the Leiden algorithm on the nearest-neighbor graph, controlled by leiden_resolution. Higher resolution values produce more, smaller clusters; lower values produce fewer, larger clusters. Cluster labels are written to the clustered.h5ad output and used to color the UMAP and spatial-cluster plots described in 05 Clustering.

6) Reproducibility

UMAP and Leiden are stochastic, but the pipeline uses fixed random seeds so that a rerun on the same inputs with the same parameters produces the same embedding and cluster labels.