1 of 24

Canadian Bioinformatics Workshops

The Brain Single Cell Initiative

Panoramics - A Vision

www.bioinformatics.ca

bioinformaticsdotca.github.io

1

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

2 of 24

2

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

3 of 24

Module #5: Visualizing your gene expression in tissue context

Introductory Spatial ‘Omics Analysis

July 9-10, 2024

Introductory Spatial ‘Omics Analysis

February 20-21, 2025

Alyona Ivanova

PhD Candidate

Brain Tumour Research Centre,

The Hospital for Sick Children

Institute of Medical Science, University of Toronto

3

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

4 of 24

Learning Objectives

By the end of this lecture, you will

understand the value of tissue segmentation in spatial transcriptomics analysis

and learn how to:

use annotated H&E images and integrate them with the previous gene expression derived spatial domains
use your image to confirm, fine-tune or validate spatial clusters
complement image-defined regions where spatial clustering fails

4

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

5 of 24

Lesson plan

Key steps involved in image analysis

Softwares available

Practical with QuPath

Applications

5

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

6 of 24

Pathology images play crucial role in clinical practice

Broad class of histological structures found in breast tissue images

(Left) two normal ducts outlined in magenta and black; ductal carcinoma in situ outlined in green; a large nest of invasive carcinoma outlined in blue; adipose tissue (fat cells) outlined in different colors; (Right) a large tumor nest with infiltrating lymphocyte outlined.

Nguyen et al. 2018

6

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

Pathology images of surgically removed or biopsied tumor tissue sections play a crucial role in modern clinical practice aiding in cancer diagnoses, subtype classification and treatment planning. In the clinics, these tasks are predominantly performed via manual microscopic inspection of H&E stained images by expert pathologists for morphologic evaluation of tumors. Visual inspection is often complemented with molecular characterization of the tissue, which further enhances our understanding of tumor progression and helps identify important biomarkers essential for precision oncology.

Histological structure also helps understand spatial tumor biology and inform pathological basis of cancer. For example, in the H&E stained tissue image shown in this figure (left), normal ducts (top left corner) have two layers of nuclei, epithelial (inner) and myoepithelial (outer), surrounding a cavity (lumen). The structure of the normal ducts is disturbed when ducts develop into carcinoma in situ (outlined in green), in which the epithelial nuclei proliferate in close proximity and invade the lumen cavity. The structure is further perturbed as the proliferating nuclei become invasive carcinoma (outlined in blue), destroying the duct confinement, freely infiltrating through the breast stroma, and heading toward a blood vessel (outlined in teal), indicating an increased risk of metastasis. Host response to invasive carcinoma can be seen on the right panel, where the tumor nest is infiltrated with lymphocyte nuclei (small, dark purple).

7 of 24

Histological structure helps understand spatial tumour biology and inform pathological basis of cancer��Accurate segmentation of histological structures can thus help build a spatial interaction map to serve as an exploratory tool for pathologists��Segmentation can also facilitate precision medicine studies which perform microdissection for deep molecular profiling��

7

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

8 of 24

Key steps involved

Image preprocessing

Feature extraction from H&E image

Cell segmentation
Morphological features
Histological structures

Data integration

8

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

We have already extracted spatial clusters in module 3 How can we now visualize this data and integrate it with the histopathology image?

One of the outputs of Visium HD run is the high-resolution H&E image of your tissue. First step is to align the H&E image to the spatial transcriptomics grid, ensuring each spot on the grid corresponds to the correct location in the H&E image.

Next, we need to extract features from the H&E image. Cell segmentation will help to identify individual cells or nuclei. Morphological features like cell size, shape, and intensity of staining for each cell can be used to infer specific information about different cell types present in the tissue. Identification of histological structures may be more informative for spatial clustering, as it helps reveal how anatomy directly relates to tissue patterning. Beyond individual cells, spatial patterns and interactions of larger cellular neighborhoods or communities play pivotal roles in determining cellular behavior, cooperation and overall tissue dynamics. Decoding the intricate web of cell–cell communications, be it through direct physical contact or the release and reception of signaling molecules, offers insights into the collective behavior of cellular communities.

This is especially relevant in the context of cancer and anatomical structures.

Data integration refers to combining the extracted features from the H&E image with the corresponding gene expression data from the spatial transcriptomics data based on their spatial coordinates

In this module, we explore how we can visualize the output of spatial clustering and interpret these findingds�

9 of 24

Neuroanatomical Structure of Mouse Brain

Annotated Mouse Coronal Section, Allen Brain Atlas

9

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

10 of 24

Software options for histology annotation

Augmentiqs - combines microscope annotation with histopathology imaging. It allows pathologists to create inline annotations, morphometric measurements, and share images in real-time

QuPath - allows users to annotate, view, markup, and analyze pathology images. It provides tools for tumor identification, biomarker evaluation, and batch processing

CellProfiler - allows users to analyze and batch-process cells in biological images

3D Slicer - medical image processing and visualization; can be used for 3D image segmentation and registration

10

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

11 of 24

Mouse Brain Hippocampus annotated in QuPath

11

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

12 of 24

QuPath Demo

How to use QuPath for annotation of images
How to download your annotations
How to import the spatial clusters
How to simultaneously view annotated image and spatial clusters

12

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

13 of 24

Scale factor

scalefactors_json.json: The purpose of the file is to record the relative scales of the user-supplied image, the images in the spatial outs, and the Visium array:

tissue_hires_scalef: A scaling factor that converts pixel positions in the original, full-resolution image to pixel positions in tissue_hires_image.png

µm

pixels

Scaling factor

13

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

14 of 24

Visium HD slide

2/8/16 µm

Defines the grid size

14

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

15 of 24

Output from Module 3 Spatial Clustering

1. Each cluster is bit (binary) == channel

2. Assign each grid with a value that is either 0 or 255

3. Create an empty numpy array that is equivalent to the grid size of your bins

4. Scale to fit H&E resolution

5. QuPath

15

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

16 of 24

Reflection questions

What is the value of spatial clustering? Does spatial clustering explain differences between anatomical regions?
Can you identify any areas within one anatomical region that contain several spatial clusters?

16

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

17 of 24

MBP (Myelin basic protein): protein that helps form the myelin sheath around nerve fibers

SYP (synaptophysin): involved in synapse formation and function in the hippocampus

HPCA (hippocalcin):expressed in the hippocampus and is involved in calcium signaling

GFAP (Glial fibrillary acidic protein): protein that's expressed in the hippocampus and is a marker of glial cells

PROX1 (Prospero-related homeobox 1): regulates cell differentiation and proliferation, and is expressed in the dentate gyrus (DG) of the hippocampus

LCT: expressed in the dorsal hippocampus, and it's a marker for the dorsal division of the dentate gyrus

One step further…

17

bioinformatics.ca

18 of 24

Applications

Automating morphologic evaluation tasks through computational pathology

FastGlioma workflow. From Foundation models for fast, label-free detection of glioma infiltration.

Kondepudi et al., 2024

18

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

With the digitization of histopathology glass slides into Whole-Slide Images (WSIs), unprecedented opportunities arise for cost-efficient analyses of tumor properties. Notably, WSIs are available without additional cost as they are obtained in routine clinical practice for diagnostics. Despite providing only morphological information, WSIs can also reflect the molecular traits of tumors. Over the past decade, machine and deep-learning methods have been developed to extract hidden morphological features from WSIs that are associated with molecular properties.

The emergence of machine learning is revolutionizing the field of pathology by automating morphologic evaluation tasks through computational pathology.

Moreover, ML algorithms have demonstrated the capability to predict various molecular features such as microsatellite instability, gene mutationsand chromosomal instability. However, most existing algorithms focus on end-to-end predictions at the patient level, lacking insight into the spatial architecture of tumors.

Growing evidence highlights the critical role of spatial context and cellular composition in the tumor microenvironment on cancer prognosis and treatment outcomes. To better understand how cellular organization affects tumor progression and treatment response, there is a need for more interpretable models that not only make accurate predictions but also provide clear insights into the reasoning behind those predictions (e.g., using attention-based learning). Incorporating spatial information into ML algorithms can enhance both explainability and predictive accuracy

For example, scientists have recently developed a machine learning model that allows to detect the degree of tumour infiltration within seconds in fresh, unprocessed surgical tissue.

19 of 24

Applications

Computational staining

Human-interpretable image feature extraction workflow.From Human-interpretable image features derived from densely mapped cancer pathology slides predict diverse molecular phenotypes

Diao et al., 2021

19

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

20 of 24

Applications

Subtype prediction and identification of spatially-resolved biomarkers

20

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

21 of 24

Applications

AI tool 'sees' cancer gene signatures in biopsy images

A new AI program, SEQUOIA, can analyze a microscopy image from a tumor biopsy (left, purple) and rapidly determine what genes are likely turned on and off in the cells it contains (gene expression shown in shades of red and blue on right).�Images courtesy of Olivier Gevaert�Photo illustration by Emily Moskal

21

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

22 of 24

Limitations

High Heterogeneity

High Cost

Representativeness of the sample

22

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

Risk stratification in patients with cancer is usually based on cancer staging, grading, clinical characteristics, and molecular profiling.Cancer molecular subtyping has become a useful tool to further stratify the disease risk and to refine the prediction of patient prognosis and treatment effects. Pathologic subtypes on the basis of manual assessments of pathologists have been proposed for multiple tumor types.However, high interobserver variability has been a major challenge.

Drawing parallels with the early days of the genomics era, when sequencing a whole genome cost millions of dollars, the trajectory of progress indicates that technological advances can lead to cost reduction and enhanced accessibility. If history has taught us anything, it is that technology evolves rapidly, prices come down swiftly and hence we must continue to guide the direction of progress. The same trajectory of advances and cost reductions might apply to spatial omic technologies, enabling them to be seamlessly integrated into clinical practice without the need for extensive downsizing. If these technologies can reliably inform clinical outcomes, they will find their way into the clinic and be further fueled by the integration of AI algorithms to decode spatial features that provide valuable insights for disease management. As a result, the clinical application of spatial omic technologies holds immense promise and has the potential to reshape the landscape of precision medicine.

23 of 24

Questions?

Email me alyona.ivanova@panoramics-a-vision.com

23

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

24 of 24

We are on a Coffee Break & Networking Session

Workshop Sponsors:

24

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision