Optimal CITRUS settings

[jamesaries]

CITRUS for CyTOF forum.pdf

(677.24 KiB) Downloaded 270 times

Dear fellow forum users,

I wonder if anyone has any great experience with CITRUS and the myriad of settings available. I’m new to using this so please, any suggestions would be gratefully received.
Attached are results of 3 runs for 2 different experiments using the PAMR. My aim, all be it with different endpoints for each experiment, was to search for any predictive abundance difference between the two experimental groups (X & Y in experiment A, and C & D in experiment B - see attached PDF). When I say CD+ lineage+ I mean I have excluded manually all those events that are CD45 positive but undefined by any other marker.

As you can see the results aren’t great! I got slightly excited by the first run in experiment A, given the cv.fdr.constrained model yielded a conceivable cell subset that I also found varied when I used Phenograph. However, on reflection, the CV.1se model revealed ‘0’ model features, so I’m not sure how valid a result it is. Further, runs 2 and 3 revealed no fdr.constrained models and the error rates are too high as well as being different from the first run!

I also have a different experiment (B) that reveals a better cross-validation curve, the existence of fdr.constrained models but once again no model features in the cv.1se model.

Clearly the experiments in A and B are different and I appreciate that comparison between the two is not appropriate. They are merely examples of different ‘unsatisfactory’ outcomes I am getting with CITRUS.

1. Does anyone have any comments on the settings I’ve used? Is it a case of just ‘playing around’ with them or can I be more directed and focused. I have checked out the Cytobank guide too…
2. Experiment A. My lowest frequency cell is invariant NKT cells but perhaps the cluster size should be adjusted for greater power with the compromise that any predictive subset featuring these cells will not be identified?
3. Experiment B. This is a ‘cruder’ analysis to see broadly if there was any difference between the two groups of samples. Perhaps there are too few samples as I know 8 per group is the suggested minimum (I don’t have 8 per group; I have more for one group but only used 3 in order to ‘balance’ the groups).

Of course it is possible there is no predictive difference to be found in either experiment!

Many thanks in advance and have a good weekend!

James