Statistical analysis of CyTOF data

We used a mixed effects logistic regression model to identify markers predictive of infection with influenza. Cell infection status (i.e. mock, Cal/09 or Vic/11) constituted the response variables, while antibodies were considered explanatory variables. Raw counts were transformed using an arcsinh transformation with a standard cofactor of 5. Estimated coefficients are in units of log-odds, which are logarithm of ratios of two probabilities: the probability that a cell is in condition B (for example, exposed to Cal/09) (encoded as 1) over the probability that a cell is in condition A (for example, exposed to Vic/11) (encoded as 0). The logistic regression model for the ith cell is a linear combination of weighted (denoted byβ) marker intensities (denoted byx);

To account for between donor variability, we added a random effects term to the logistic regression (denoted by z):

Markers with positive coefficients increase the log-odds, whereas markers with negative coefficients decrease the log-odds. In the case of zero-inflated marker expressions, the estimated marker coefficients are shrunk towards zero and provide more conservative estimates. Multimodal distributional markers do not violate the logit model assumptions. We estimated fixed and random effects using the R packages mbest (61).