10x
Support
Search Support
Contact us
10x Genomics Support/In Situ Gene Expression/Documentation/Onboard Analysis/

Overview of Xenium Algorithms

This page describes the algorithms underlying the processing and analysis steps that occur on the Xenium instrument.

Table of contents

Important
See what's new in the Xenium Onboard Analysis software pipeline. Click here to read the release notes.

Background: workflow overview

The Xenium workflow begins with sample preparation. Fresh frozen (FF) or formalin-fixed paraffin-embedded (FFPE) samples are mounted on Xenium slides. The samples are fixed and permeabilized (FF samples) or deparaffinized and decrosslinked (FFPE samples). Then, probe hybridization, ligation, and rolling circle amplification (RCA) are performed.

Once the sample has been prepared, imaging is performed in cycles on the Xenium Analyzer. During each cycle, fluorescently labeled probes for detecting RNA target sequences and other reagents (see Controls), are automatically cycled in, imaged, and removed. The internal image sensor captures image data across multiple Z-planes (with a 0.75 μm step size across the entire tissue thickness) for every field of view (FOV) in the user-selected region (see Region Selection Guidelines in the Xenium Analyzer instrument user guide). Image data are captured for multiple fluorescence channels every cycle, and are processed and stitched to build a spatial map of the transcripts across the tissue section.

Thus, over the course of a run, the Xenium Analyzer’s internal image sensor collects 3D volumes across: 1) multiple FOVs, 2) multiple fluorescence channels, and 3) multiple cycles of chemistry and imaging. This produces terabytes of internal sensor data that are processed efficiently and analyzed across all cycles to decode transcripts. Once transcripts are decoded, downstream analysis and visualization of Xenium raw data output (the spatial map of transcripts) can proceed (see Xenium data).

Onboard analysis pipeline

The primary steps performed by the pipeline across cycles are:

  • DAPI image processing: DAPI (Xenium Nuclei Staining Buffer) is a blue fluorescent DNA stain for visualizing nuclear DNA in fresh and fixed cells. In the Xenium workflow, DAPI staining is used to locate nuclei, inform cell segmentation, and produce a 3D tissue morphology image. DAPI images are captured once across all FOVs in the first cycle.
  • RCA product image processing: For each cycle, images are captured in multiple color channels across all FOVs. Punctate fluorescent signals (blobs) are detected and filtered, and image distortion is corrected.

Cell segmentation occurs between cycles. DAPI images are used to infer cell boundaries using a machine learning algorithm.

After all cycles complete, the final steps of the pipeline include image registration, decoding, deduplication, and generation of Xenium output data. Quality scores (Q-Scores) are estimated using controls for calibration.

DAPI image processing

This process produces a complete 3D morphology image (the morphology.ome.tif output file) for each of the stained regions and determines the reference image volume that cells are segmented from and decoded transcripts are assigned to.

First, the lens distortion in internal sensor data is corrected. This is done computationally based on instrument calibration data, which are collected in order to characterize the optical system and are saved on-instrument. Next, the Z-stacks from internal sensor data are further subsampled to a 3 μm step size. This subsampling step size was determined empirically to be a useful resolution for cell segmentation quality.

Image features are then extracted from the regions where FOVs overlap. Feature matching is performed to estimate the offsets between adjoining FOVs. The offsets are used to ensure consistent alignment across the image (global alignment). Finally, the 3D DAPI image volumes (Z-stacks) generated across FOVs are blended together to construct a stitched volume.

RCA product image processing

The goal of RCA product image processing is to detect and filter blobs and correct distortion. Performed for every channel and cycle, the 3D image volumes (Z-stacks) obtained for each FOV are processed to detect the blobs in 3D space that correspond to labeled RCA products. Images are currently captured in four color channels and 15 cycles (subject to change to accommodate platform growth). The raw image is scanned for blob signals that stand out from the local background. The XYZ coordinates of each blob are refined by examining local brightness. The signal intensity of the blob is determined based on a fitted shape.

Next, the pipeline filters out blobs that are unlikely to be true transcripts (non-punctate or low quality signals). Similar to DAPI images, curvature distortion is corrected.

Cell segmentation

The goal of cell segmentation is to approximate boundaries between cells so that transcripts can be assigned to cells. Downstream, these results will be used to produce a cell-feature matrix, similar to those output by existing single cell and spatial technologies.

The first step is to detect the locations of nuclei using the DAPI images and a custom neural network for nucleus segmentation. The neural network is trained on thousands of manually labeled image patches covering multiple tissue types.

Once the locations of nuclei in the sample have been identified by the model, a heuristic cell boundary expansion step is performed. The nuclei boundaries are expanded by 15 µm or until they encounter another cell boundary in X-Y. If cell boundaries overlap during expansion, they are resolved using an algorithm that is conceptually similar to Voronoi tessellation.

Xenium cell segmentation takes into account the 3D output from the DAPI image processing step for all Z-slices for better accuracy, but ultimately produces a flattened 2D segmentation mask for ease of use. The nuclear boundaries are consolidated to form non-overlapping 2D objects when projected in X-Y. Since the segmentation mask is 2D, transcripts are assigned to 2D shapes based on their X and Y coordinates.

It is possible to use third-party segmentation tools with the same morphology image that Xenium uses. Analysis Guides on this topic can be found on the 10x website. Xenium’s cell segmentation algorithm will evolve to accommodate platform growth - stay tuned!

Image registration

For each FOV, blobs across channels and cycles must be aligned to reduce differences in image offset, rotation, and magnification. This is important for accurate transcript decoding. The localized blobs from each channel and cycle are registered so that blobs corresponding to the same original RNA molecule are aligned tightly in 3D. Nonlinear transformations are fitted such that all the blobs are aligned to the reference morphology image.

Decoding, quality scores, and controls

In order to proceed from blobs to transcripts, decoding must be performed. The Xenium codebook contains a collection of codewords that are assigned to genes in a gene panel. The pipeline uses the gene_panel.json to specify a given gene name to an indexed codeword. Each codeword is defined by a pattern of fluorescent signals (or absence of signal) recorded across channels and cycles (see diagram below). Some codewords are reserved for negative controls.

The fluorescent signals from all channels and cycles are compared to the codebook using a global (across all FOVs) maximum likelihood approach based on probabilistic modeling. This approach considers attributes such as blob locations, their color and cycle of detections, and signal intensities.

Quality scores and controls

A Phred-style calibrated quality score is assigned to each decoded transcript to signify the confidence in the decoded transcript identity. The quality score is derived from the likelihood of the maximum likelihood codeword (i.e., the codeword that best explains the observed data), compared to the likelihood of other sub-optimal codewords. This yields a raw Q-Score. Codewords are then mapped to targets using the gene panel information.

Final Q-Scores are obtained first by putting the full range of raw Q-Scores into bins. Then, the raw Q-Scores in each bin are calibrated by the proportion of “Negative Control Codewords” in the bin. A final Q-Score is assigned to each raw Q-Score bin, to ensure that Q-Scores in each bin are correctly calibrated. Control probes are built into the process to ensure that the final Q-Scores are accurately calibrated.

  • Negative control codewords are codewords in the codebook that do not have any probes matching that code. They are chosen to meet the same requirements as regular codewords and can be used to assess the specificity of the decoding algorithm.
  • Negative control probes are probes that exist in the panels but target non-biological sequences. They can be used to assess the specificity of the assay.
  • Unassigned codewords are unused codewords. There is no probe in this particular gene panel that will generate the codeword.

The cell-feature matrix and Xenium Analyzer’s secondary analyses only include transcripts with a Q-Score ≥ 20. Final Q-Scores are reported in the transcript data output files.

Deduplication

After decoding, the results are combined from all FOVs. Duplicate decoded transcripts in overlapping regions between adjoining FOVs are reconciled into one transcript, based on a nearest neighbor analysis that considers transcript identity and Q-Score. The deduplicated transcripts are assigned global coordinates based on the reference morphology image obtained from the DAPI image processing step (morphology.ome.tif).

Xenium output data

The Xenium raw output data consists of decoded transcript counts and morphology images. These data reduce low level internal sensor data as described above, preserving details needed to assess decoded transcript quality. Raw output and other standard output files derived from them are included in the Xenium output directory for each region. Because they represent raw data from the Xenium platform, decoded transcript counts and morphology images should be archived for possible off-instrument reprocessing and reproducibility.

For a complete table of output files, see At a Glance: Xenium Output Files and for output file descriptions, see Understanding Xenium Outputs.

Cell-feature matrix

Decoded and deduplicated transcripts in 3D are assigned to directly overlapping segmented cells to produce a cell-feature matrix. This matrix can be analyzed using conventional and novel single cell and spatial analysis approaches, facilitating integration with existing single cell and spatial datasets.

Secondary analysis

Xenium’s onboard analysis uses the same algorithms as Cell Ranger for single cell gene expression for performing secondary analysis on the cell-feature matrix (PCA, UMAP, graph-based clustering, differential expression analysis). t-SNE projection is not supported in Xenium onboard analysis.

Next steps

Document Type
Software

Last Modified
March 13, 2023