Rs and non-responders. This scenario calls for a “supervised” algorithm, in which we know the answer (response, non-response) and are looking for a set of analytes that help us arrive at that answer. A decision tree is one such supervised approach. Alternatively, if one is looking for a variety of patterns in the data that help us to better Thonzonium (bromide) cancer understand the relationships between patient characteristics and analytes, then an “unsupervised” approach, in which there is not a specific answer is appropriate. Hierarchical clustering and association rule mining are examples of unsupervised approaches. Ideally, the analytical approaches will provide both quantitative and visual results. Another consideration is whether the analytical techniques are magnitudeinsensitive, that is, able to easily support data from assays yielding wildly different numeric ranges. Furthermore, the results suggested by any analysis should be vetted for biological relevance and replicated in independent data sets or PubMed ID:https://www.ncbi.nlm.nih.gov/pubmed/28607003 studies. The following five techniques, detailed below, can provide insight into the systemic host response and are applicable to het sets: regression modeling, network of cross-compartment correlations, penalized regression, decision tress, and association rule mining. Regression modeling supports both simple models (such as response 1 x analyte) and more complex models (such as response 1 x analyte + 2 x treatment + 3 x sex + 4 x age). In both simple and complex models, the terms are the estimated coefficients or contributions of the predictor variables to the outcome variable. Complex multivariable models can be longitudinal models or time-to-event (survival) models and account for variables like treatment type, sex, and age. Longitudinal models may be particularly appropriate for characterizing immune response over time and can account for patient-specific trends. Response can be categorical (responder versus non-responder) or continuous (progression-free survival). A strategy that is common in gene expression analysis is to build such a model for all genes and focus on a handful with the smallest p-values on the coefficient of interest. While it is fast and easily understood, this approach does not provide a comprehensive picture that accounts for systemic responses or for correlations amongst analytes.Analyte CD4+ Treg CD4+ Treg IL2 IL2 IL2 IL2 Readout 3.2 5.1 3.8 2.7 2.5 10.1 Units of parent of parent pg/ml pg/ml pg/ml log normalized expressionTable 1 Sample extract from a representative integrated heterogeneous data set (het set)Stroncek et al. Journal for ImmunoTherapy of Cancer (2017) 5:Page 14 ofOne approach to building a systemic network of crosscompartment correlations is to start with a regression model in which one analyte is the outcome and another is the predictor, e.g., assayA.analyte1 1 x assayB.analyte2 + 2 x response. As with multivariable regression, a variety of other predictors can be included in the model. Once the model results for all possible pairs of analytes are obtained, the results can be filtered to pairs of analytes from different assays or tissues and have reasonably small p-values on effects of interest, such as both the correlation between the analytes, and the effect of the response. Given 50 to 100 of such correlations, the relationships across the analytes can be tallied and the networks of correlations can be visualized. For example, Whiting et al. identified a network of 61 highly correlated analytes spannin.
