Improvements in technology (new assays, multiplex panels) and large repositories of clinical samples accelerate investigations of numerous individual biomarkers with potential for increased risk of type 1 error. Approaches to combine biomarkers can gain biological understanding and improve statistical efficiency to assess risk of clinical events; however, the best high-dimensional analytic methods are not clear.
Seventeen biomarkers measured at year 1 of suppressive ART were analyzed, from an ACTG case-control study of predictors of subsequent non-AIDS clinical events (141 cases and 310 matched controls). A targeted subset of 10 biomarkers of monocyte and macrophage activation was pre-specified, which we hypothesized to be most predictive of cardiovascular events. Three statistical approaches were considered for data reduction: 1) simple, equally weighted average of all cytokines, 2) optimally weighted combination of targeted cytokines from confirmatory factor analysis, 3) optimally weighted combination of all cytokines from principal components analysis (PCA). Cytokines were log-transformed and normalized (mean 0, SD 1) prior to analysis. Scores from each method were analyzed in conditional logistic regression models to evaluate risk of subsequent non-AIDS events.
The best analytical approach was PCA, as principle component (PC) scores had reduced variability when included in regression models compared to other approaches. Based on scree plots and other diagnostics, selecting 2 PCs described the data best, compared to 1 or 3 PCs. The 2 PCs were primarily composed of distinct subsets of the cytokines (Table). IL6, IP10, KT Ratio, sCD14, sCD163, sTNFrI, sTNFrII, and suPAR primarily contributed to PC1 (consistent with the a priori list); CMV IgG, EBV IgG, and EBV DNA primarily contributed to PC2. Higher levels of PC1 were associated with increased risk of non-AIDS events and the subset of MI/Strokes; effect sizes were similar to those reported individually for sTNFrI and II and suPAR (Tenorio 2014, Hoenigl 2018).
PCA is a useful high-dimensional analytic approach to reduce the number of statistical comparisons, to characterize unique, meaningful cytokine profiles, and to identify specific biomarkers to consider for endpoints in future interventional trials. In our analysis we confirmed that sTNFrI and II and suPAR were the strongest predictors of non-AIDS events, supporting sTNFrII as the primary endpoint of ACTG studies of letermovir and anti-CMV vaccine.