4.2.2 GvHD Discussion

2.2.2.1 GvHD Use Case Description
2.2.2.2 GvHD Gating SOP
3.2.2    GvHD Use Case Results

One disappointment that we faced was needing to simplify the GvHD use case for our project. We have made a decision for our initial analysis of this data set to look only at the one population that appears to be predictive (CD3+/CD4+/CD8b+), rather than look at all potential populations for predicative ability. This decision is in line with the clear goal of having FlowDx provide an automated solution for clearly defined clinical flow assays with very specific target populations of interest. Using algorithms for exploratory flow cytometric research is not one of the aims of the current state of the project.

In Tree Star's initial iteration of this use case, we found a very high variability in the target populations selected by Tree Star scientists and the contributing lab. The average number of cells selected was 134, with a standard deviation of 171. This high standard deviation is most likely due to lack of a clear understanding of the reason for positioning of gates within the data analysis and very imprecise gating procedure instructions (SOP). This lack of precision will be overcome by confirming the gating strategy with Clay Smith, Ryan Brinkman, or Maura Gasparetto, who did the initial gating. We would like to have Clay Smith's group do the manual gating for consensus to make sure the variability within the consensus gating is acceptable to move forward with automated classifiers. Alternatively, we could use the analysis that was originally submitted by Ryan Brinkman and is analyzed by Maura as the gold standard and not use a consensus. The difficulty with these workspaces is that they are in the Apple Macintosh version of FlowJo and will require a utility to load and organize exported populations into the database, which represents 422 distinct populations for this use case.

Once Tree Star has streamlined the workflow and resolved the consensus gating issues, we will proceed to classify this data using all the algorithms (ANN, SVM, Magnetic, and Probability Bin Clustering). We will also be looking at the biological outcome (diagnostic decision) to see if our computational methods find the same answers as the manual gating. Functional Linear Discriminant Analysis (FLDA) was used to show the biological relevance of the CD3+/CD4+/CD8b+ population as described by Ryan Brinkman, et al. [45]. Tree Star will work with Ryan's group to apply this analysis.