Gene set enrichment analysis is ideally suited to identifying small but coordinated changes in gene expression in sets of biologically related genes [13, 21]. It has been used to check details identify biological processes such as metabolic changes [21] and signaling flux [22] that are evident across

networks of genes but subtle at the level of individual gene expression. The ability to build predictive models from small but coordinated changes in transcriptional programs is particularly important for clinical applications such as the detection of a vaccine response in which the transcriptional signal in responders compared to nonresponders is small. We therefore anticipate that this approach to gene expression predictor development will be generally useful in clinical Midostaurin molecular weight situations in which the difference in gene expression between outcome classes is limited. Future studies will be able

to use this approach to test whether analogous enrichment of B cell and proliferation signatures are characteristic of vaccine response in different vaccines. Alternatively, analysis of different vaccines and in larger cohorts may be able to identify different gene sets representing other biological processes that underlie vaccine response. An advantage of gene set based predictors is that their biological meaning is more transparent. While predictive features based on individual genes may contain important, novel information about the vaccine response, their mechanistic basis is not always much obvious without additional experimental inquiry [4, 16]. Instead, we developed our predictive model from a library of well-annotated signatures derived from previously published microarray experiments and expert curation. Together with a novel analysis and visualization method—the constellation plot (Figs. 1 and 2)—this allowed the predominant biological themes that correlated with vaccination response to be readily identified. We also anticipate that in addition to vaccine response, this approach may also be useful for identifying subtle features that vary across a group

of responders, allowing the heterogeneity that is part of all human studies to be better interrogated. Moreover, the use of gene set-based classifiers may also prove useful in features predictive of adverse effects to vaccines. A theoretical concern with our method is that the biological processes involved in the vaccine response may not be represented in the compendium of signatures currently used in the analysis. However, our results suggest that at least some of the biological signatures that predict vaccine response — such as proliferation — are already present in the database of signatures used for this study. Moreover, because the method we used can draw on any collection of annotated gene sets, it can easily be extended to additional collections of gene sets.