gms | German Medical Science

Biological Background: Several studies have demonstrated a protective effect of farming environments on asthma development during childhood. However, the underlying cellular processes remains unclear [1].

In the context of childhood asthma, we aimed to analyze the effect of dust stimulation on cellular level. Therefore, a subset of peripheral blood cells of 20 children (10 healthy controls and 10 asthmatics) with and without ex vivo farm dust stimulation were characterized using mass cytometry by time of flight (CyTOF).

Table 1 [Tab. 1]

Data Background: CyTOF is a single-cell technique and can measure up to 50 markers for each cell. Those markers can capture four different dimensions: Allocation to a sample ID, recognition of intact cells, definition of cell types (lineage marker), and quantification of protein expression.

Our data set had 11.5 million cells with 48 marker (22 lineage marker) expressions per cell.

When analyzing a mass cytometry data set, the first and second dimension are straightforward to handle. They assign each cell to its initial sample and remove non-intact cells from the data set.

The next step is defining cell populations based lineage marker expression. In a high dimensional space (n=22) traditional motivated biaxial gating approaches can be both unreliable and inefficient. Thus a variety of machine learning methods based on clustering was developed. Both approaches share the basic idea of finding similar cells and assigning those to a group. However, the declaration of a group is done at different steps: Biaxial gating a priori vs machine learning retrospectively [2], [3].

The next step is analyzing the data according to the research question. Examples are comparing cellular composition across samples or looking for cellular groups with high protein expression.

Aims: This project aims to define cell populations and utilizes both biaxial gating and unsupervised learning. Our research questions were:

• How to achieve a cluster solution based on flowSOM [4]?
• What is the effect of hyperparameters on our solution?
• Can we use biaxial gating as guidance for hyperparameter tuning?

As our data comprises roughly 11.5 million cells we dealt with common problems of high dimensional clustering throughout the project.

Results: When running an automatized cell identification algorithm we had to set a number of hyperparameters that all affect the clustering solution. An objective criterion for determining the best suitable choice of hyperparameters remained unclear. When comparing our two approaches (biaxial gating and machine learning), we found concordance for major cell populations (84–92%).

Conclusion: As an automated cell identification approach relies on a set of hyperparameters, we believe that combining it with a biaxial gating strategy is a beneficial step to achieve a meaningful cluster solution. It not only eases biological interpretation but also provides some guidance for feature selection.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.