Skip to content

Pairwise alignment

In the previous step, you processed and mapped transcriptomic reads into spatial coordinates with spacemake. Before aggregating transcriptomic information per cell, we need to align this transcriptomic data into tissue space, leveraging imaging data.

openst can automatically (or manually) perform pairwise alignment of transcriptomic and imaging modalities, via advanced computer vision algorithms. Alignment is performed in two steps:

  1. Coarse (low-resolution) of H&E images to pseudoimages of ST data, to roughly align the transcriptome to the tissue.
  2. Fine (high-resolution), using fiducial marks detected at both modalities for sub-micron accurate alignment.

Data required

Two data modalities are used for pairwise alignment:

  1. A single Open-ST h5 object (from spacemake) containing all the spatially-resolved locations of a sample.
  2. A single, high-resolution image of the capture area and tissue

Tissue transcriptome

If you ran spacemake with a puck with a coordinate_system, and a run_mode without meshing, e.g.:

pucks:
  openst:
    coordinate_system: puck_data/openst_coordinate_system.csv
    spot_diameter_um: 0.6
    width_um: 1200
...
run_modes:
  openst:
    clean_dge: false
    count_intronic_reads: true
    count_mm_reads: true
    detect_tissue: false
    mesh_data: false
    mesh_type: 'hexagon'
    n_beads: 100000
    polyA_adapter_trimming: true
    spatial_barcode_min_matches: 0.1

Then you are all set to proceed with the Tissue image preprocessing (unless you had the issues below).

My run_mode and/or puck do not look like that

It is likely that you didn't specify a coordinate_system or you only have mesh(ed) Open-ST h5 objects, therefore you need to generate a puck_collection Open-ST h5 object, manually:

openst from_spacemake \
    --project-id openst_demo_project \
    --sample-id openst_demo_sample \
    spatial_stitch \
    --tile-coordinates fc_1_coordinate_system.csv

Please replace --tile-coordinates with the adequate coordinate_system for your barcoded flow cell. This one is available under the examples section

If you want to select only specific tiles for a sample, e.g., to exclude some spurious tiles, you can puck_collection files fully manually, without the from_spacemake command:

openst spatial_stitch \
    --tiles dge.fc_1_tile_2267.h5ad dge.fc_1_tile_2268.h5ad \
    --tile-id fc_1_tile_2267 fc_1_tile_2268 \
    --tile-coordinates fc_1_coordinate_system.csv \
    --h5-out spatial_stitched_spots.h5ad

Tissue image

The tissue transcriptome data you just preprocessed will be aligned into a tissue (hi-res) image. For this, openst expects a single, high-resolution tiff image where the tissue is visible (and preferrably at least a small part of the capture area, too).

Expand one of the cases below to decide how to proceed with image data:

My microscope gives a single, hi-res image

Fantastic! Just copy it to the relevant spacemake subfolder for the specific project and sample, for instance:

spacemake_folder
`-- projects
    `-- <project_id>
        `-- processed_data
            `-- <sample_id>
                `-- imaging # you might need to create
                    `-- Image_Stitched_Composite.tif # copy with this name

Make sure to copy the image to that location, with the name Image_Stitched_Composite.tif.

projects/openst_demo_project/processed_data/openst_demo_sample/imaging/Image_Stitched_Composite.tif

My microscope doesn't give a single image, but many tile files

When performing imaging, i.e. as tile-scans, you might need to assemble a single, large image from a collection of smaller tiles (fields-of-view, FOVs).

We provide tools to perform tile stitching from raw data with the microscopy setup showcased in our implementation (a Keyence BZ-X710 inverted fluorescence phase contrast microscope). If you use a different microscope setup, please refer to the documentation of your microscope for how to stitch tile-scans into a single image.

For this keyence microscope, copy the raw data (including *.tiff and .bcf files) into the spacemake folder specific for that project and sample:

spacemake_folder
`-- projects
    `-- <project_id>
        `-- raw_data
            `-- imaging # you might need to create it
                `-- <sample_id>
                    |-- Image.bcf
                    `-- *.tif

Then, run:

openst image_stitch \
    --microscope='keyence' \
    --imagej-bin=<path_to_fiji_or_imagej> \
    --image-indir=projects/<project_id>/raw_data/imaging/<sample_id> \
    --image-out=projects/<project_id>/processed_data/<sample_id>/imaging/Image_Stitched_Composite.tif

This leverages the Grid/Collection stitching plugin included in Fiji to create a single, composite tile-scan image. Replace <path_to_fiji_or_imagej> with the path where the Fiji executable is.

Question

If you don't know how to specify the <path_to_fiji_or_imagej>, please follow the official instructions provided for Running Headless. For example, under linux you can download it using

cd ~
wget "https://downloads.imagej.net/fiji/latest/fiji-linux64.zip"
unzip fiji-linux64.zip
# then, <path_to_fiji_or_imagej> will be "~/Fiji.app/ImageJ-linux64"

Note

If your imaging setup consisted of a Z-stack, each tile image will likely be a FOV and a specific Z plane (e.g., file names like Image_00001_Z001.tif, Image_00001_Z002.tif,...). In this case, you can specify a regular expression (regex) to parse the Z plane from the file names, and a single file per FOV will be created before stitching.

This can be specified in openst image_stitch with the argument --join-zstack-regex (empty, by default)

Either way, openst expects a single .tif file at projects/<project_id>/processed_data/<sample_id>/imaging/Image_Stitched_Composite.tif.

Unless you had the issue below...

The quality of images is suboptimal (e.g., for segmentation)

Most of the times, large tile-scans may have uneven illumination, focus or noise. This can be challenging for downstream processing, like segmentation, feature extraction or quantification (i.e., of fluorescence images). There is a plethora of methods to address these issues (e.g., Flatfield Correction from BigStitcher, or CARE, to name some). This might be highly dependent on your microscope, imaging settings, sample type, sample width... Always look at your images so you can take an informed decision.

In our publication, we leveraged a CUT model that allowed to homogeneize the style of the whole tile-scan - that is, reduce possible biases in illumination, noise and focus across the entire tile-scan. You can run this by running the following command on the stitched image.

CUT restoration showcased in the paper

openst image_preprocess \
    --image-in Image_Stitched_Composite.tif \
    --image-out Image_Stitched_Composite_Restored.tif
    # --device cuda # in case you have a CUDA-compatible GPU

If you ran openst merge_modalities, then imaging data will be contained inside the Open-ST h5 object, and the command can be adapted:

openst image_preprocess \
    --h5-in multimodal/spots_stitched.h5ad # just a placeholder, adapt
    # --device cuda # in case you have a CUDA-compatible GPU

By default, the image will be loaded from the key uns/spatial/staining_image, and the CUT-restored image will be saved to uns/spatial/staining_image_restored. You can preview the image restoration results using:

openst from_spacemake \
    --project-id openst_demo_project \
    --sample-id openst_demo_sample \
    --image-key uns/spatial/staining_image uns/spatial/staining_image_restored

This will load the two images and visualize it using napari. Later, you can run segmentation and pairwise alignment using either the default merged image (uns/spatial/staining_image), or the restored image (uns/spatial/staining_image_restored). Always assess these preprocessing choices (quantitatively and qualitatively) to decide whether these make sense for your data.

Once the single image has been generated, make sure that the image is the full-resolution and not a downsampled version. As a rule of thumb, at least a few fiducial markers must be clearly distinguishable.

... you are all set to proceed with pairwise alignment.

Merging data modalities

Once the ST and imaging data have been stitched, you can create a single object containing both modalities:

openst from_spacemake \
     --project-id openst_demo_project \
     --sample-id openst_demo_sample \
     merge_modalities

Pairwise alignment

From the merged object you created above, you can do pairwise alignment of the transcriptomic data into the tissue image automatically, semiautomatically, or fully manually:

For aligning transcriptomic data into H&E-stained tissue sections, we recommend leaving the rest of arguments with default parameters - you can get a full list of configurable parameters by running openst pairwise_aligner --help.

openst from_spacemake \
      --project-id openst_demo_project \
      --sample-id openst_demo_sample \
      pairwise_aligner
      # --metadata alignment.json # to create a visual report
      # --only-coarse for semiautomatic

When running automatic alignment, the spatial keys will be: 1. obsm/spatial: the original (unaligned) coordinates 2. obsm/spatial_pairwise_aligned_coarse: after low-res alignment 3. obsm/spatial_pairwise_aligned_fine: after hi-res refinement

Similarly, the aligned image will be saved under uns/spatial_pairwise_aligned/staining_image_transformed.

Therefore, if you want to visualize with the GUI or perform manual fine-tuning of the alignment, you need to specify the correct --spatial-key and --image-key:

# for visualization and/or manual fine-tuning
openst from_spacemake \
      --project-id openst_demo_project \
      --sample-id openst_demo_sample \
      manual_pairwise_aligner \
      --spatial-key obsm/spatial_pairwise_aligned_fine \
      --image-key uns/spatial_pairwise_aligned/staining_image_transformed

# if you want to perform semiautomatic alignment,
# continue with openst manual_pairwise_aligner
# see "Manual alignment" tab

Also, in some of the following steps (e.g., segmentation), make sure to populate the --image-key and --spatial-key arguments with the correct values.

We provide a Graphical User Interface (GUI) for selecting keypoints between imaging & ST modalities, for full manual alignment or refinement of automatic results. This GUI requires a single Open-ST h5 object (after spatial stitching). There are two kinds of workflow:

openst from_spacemake \
      --project-id openst_demo_project \
      --sample-id openst_demo_sample \
      manual_pairwise_aligner \
      --spatial-key obsm/spatial \
      --image-key uns/spatial/staining_image

In the GUI, you first perform coarse alignment (not tile by tile), and then click 'Apply to data'. Save the transformed coordinates as obsm/spatial_manual_coarse. Then, re-render with with spatial key, select at least 3 corresponding keypoints per tile, and save a keypoints file, e.g., as keypoints.json.

Then you can apply the transformation, per tile, separately:

openst from_spacemake \
      --project-id openst_demo_project \
      --sample-id openst_demo_sample \
      apply_transform \
      --keypoints-in keypoints.json \
      --spatial-key-in obsm/spatial_manual_coarse \
      --spatial-key-out obsm/spatial_manual_fine \
      --per-tile

This will save the refined (similar to automatic hi-res alignment) coordinates into obsm/spatial_manual_fine. The image data will still be (by default) under uns/spatial/staining_image. Make sure to populate --image-key and --spatial-key arguments in later steps with the correct values.

We provide a video tutorial showcasing the GUI, with an illustrative example of refinement from (only coarse) automatic alignment.


Walkthrough of the GUI for manual alignment by @danilexn – 5m – Learn how to visualize and align STS and imaging data in a step-by-step guide.


Now you are all set to proceed with the visual (qualitative) assessment of the alignment (unless you had the issues below).

The coarse (or fine) alignment didn't work

There are several parameters that you can change.

The most important ones are: - --rescale-factor-coarse or --rescale-factor-fine: the highest, the lower resolution the image; thus, more global features are used for registration - --threshold-counts-coarse or --threshold-counts-fine: the highest, the less spots there will be on the image

These allow to increase the number of matches and possibly the number of RANSAC inliers. There are other RANSAC-specific parameters that can be changed, such as --ransac-coarse-residual-threshold, --ransac-coarse-max-trials and --ransac-fine-min-samples. For more parameters, check openst pairwise_aligner --help.

I get an error: 'This application failed to start [...]'

In some environments, the following error might happen:

qt.qpa.plugin: Could not load the Qt platform plugin "xcb" in "/home/user/miniconda3/envs/openst/lib/python3.11/site-packages/cv2/qt/plugins" even though it was found.

This application failed to start because no Qt platform plugin could be initialized. Reinstalling the application may fix this problem.

Available platform plugins are: xcb, eglfs, linuxfb, minimal, minimalegl, offscreen, vnc, wayland-egl, wayland, wayland-xcomposite-egl, wayland-xcomposite-glx, webgl.

This can be solved by removing the files inside /home/user/miniconda3/envs/openst/lib/python3.11/site-packages/cv2/qt/plugins, e.g. see this StackOverflow thread

Visual assessment of alignment

With HTML report

With automatic alignment, if you specified the --metadata argument, you can generate an HTML report showing a qualitative summary of the alignment (images, parameters...)

openst report --metadata=alignment.json --html-out=alignment_report.html

With interactive GUI

Alternatively, you can visualize the images & ST data interactively using the integrated interactive GUI.

openst manual_pairwise_aligner

We provide a Graphical User Interface (GUI) for selecting keypoints between imaging & ST modalities, for visualization and refinement of automatic results. This GUI requires a single Open-ST h5 object, the output of openst pairwise_aligner.

Expected output

After automatic or manual alignment, you will have a single h5ad file, containing the transformed spatial coordinates. This will be used in the following step to aggregate the transcripts by a spatially-corresponding cell, in order to get a cell-by-gene matrix that can be used in later downstream analysis.