Statistical Bioinformatics Webinar: Eling -- Informed Region Selection and Analysis of Imaging Mass Cytometry Data

Presented by Dr. Nils Eling, University of Zurich 

The development of highly multiplexed imaging technologies has led to deeper insights
into the spatial organization of healthy and diseased tissues. In particular, the
structure of the immune compartment within the tumour microenvironment is a predictor
for immuno-oncology treatment success. The IMMUcan (Integrated immunoprofiling of large
adaptive cancer patient cohorts) project studies immune-tumour interactions within the
tumour microenvironment and its impact on therapeutic interventions. As part of this
initiative, the team acquires multiplexed immunofluorescence and imaging mass cytometry
data of thousands of patient samples from five cancer types. While multiplexed
immunofluorescence captures the expression of about seven proteins across all cells of
the tumour section, imaging mass cytometry focuses on measuring smaller regions (about 1
mm squared) with higher content (about 40 proteins). To study detailed immune-tumour
interactions using imaging mass cytometry, it is therefore crucial to perform informed
selection of regions of interest that capture the cell types of interest. We have now
developed a set of computational tools that guide and facilitate the selection of
regions for imaging mass cytometry based on multiplexed immunofluorescence
measurements. First, a custom-made, user-guided workflow robustly identifies the same
cell-types across different patients and cancer indications. Via an online tool, we are
able to select three to four regions containing about 50% of tumour cells and a
diversity of immune cells of interest. The Python package napping was developed to
transfer the selected regions’ coordinates onto brightfield images of consecutive
sections to be measured by imaging mass cytometry. Upon data acquisition and quality
control, we use the cytomapper Bioconductor package to manually label cell types on
imaging mass cytometry images and train a random forest classifier for cell type
classification. Based on this workflow, we were able to identify broad (e.g. T cells)
and rare (e.g. exhausted CD8+ T cells) cell types across all acquired samples with good
agreement to matched multiplexed immunofluorescence data.