Systemic and intrathecal immune activation in association with cerebral and cognitive outcomes in paediatric HIV

Despite treatment, immune activation is thought to contribute to cerebral injury in children perinatally infected with human immunodeficiency virus (HIV). We aimed to characterize immune activation in relation to neuroimaging and cognitive outcomes. We therefore measured immunological, coagulation, and neuronal biomarkers in plasma and cerebrospinal fluid (CSF) samples of 34 perinatally HIV-infected children aged 8–18 years, and in plasma samples of 37 controls of comparable age, sex, ethnicity, and socio-economic status. We then compared plasma biomarker levels between groups, and explored associations between plasma/CSF biomarkers and neuroimaging and cognitive outcomes using network analysis. HIV-infected children showed higher plasma levels of C-reactive protein, interferon-gamma, interferon-gamma-inducible protein-10, and monocyte chemoattractant protein-1 than controls. In HIV-infected participants, plasma soluble CD14 was positively associated with microstructural white matter (WM) damage, and plasma D-dimer was negatively associated with WM blood flow. In CSF, IL-6 was negatively associated with WM volume, and neurofilament heavy-chain (NFH) was negatively associated with intelligence quotient and working memory. These markers of ongoing inflammation, immune activation, coagulation, and neuronal damage could be used to further evaluate the pathophysiology and clinical course of cerebral and cognitive deficits in perinatally acquired HIV.

This study includes data concerning 36 perinatally human immunodeficiency virus (HIV)-infected children and 37 uninfected controls. Interest lies with exploring inflammation and immune activation as measured through plasma and cerebrospinal fluid (CSF) markers. It was previously found that HIV-infected children display neuroimaging abnormalities and impaired cognitive functioning in comparison to controls. Four questions directed the analyses: (1) Are the plasma markers differentially expressed between HIV-infected children and controls?
(2) Are the plasma markers concordant with their corresponding CSF markers?
(3) Can we characterize, for HIV-infected children, associations between HIV-related measurements and plasma/CSF markers?
(4) Can we characterize, for HIV-infected children, associations between plasma/CSF markers and neuroimaging abnormalities/cognitive functioning measurements?

Data preprocessing
Plasma markers were measured in all study participants. Prior to analysis, we removed two cases due to >10% missing values (due to unavailable plasma samples for those measurements). For further analyses, we had plasmamarker data available from 34 HIV-infected children and 37 controls. CSF of 25 HIV-infected participants was available for soluble cluster of differentiation (sCD)14 and sCD163 measurements, and of 23 participants for the remaining biomarker measurements.
We then identified plasma and CSF measurements that were listed as below the lower limit of quantification (LLOQ).
Biomarkers were excluded from analysis if >30% of values were below the LLOQ. Remaining values below the LLOQ were imputed by assigning the value of the LLOQ as indicated by the manufacturer (Supplemental Table 1). One missing value of serum NFL was attributable to a value exceeding the upper LOQ, and was thus assigned the value of the upper LOQ. All markers were then log-transformed using the natural logarithm.
Approach to question 1.
Question 1 asked if there were (subsets of) plasma markers that were differentially expressed between HIV-infected children and controls. A two-tier approach was considered: Blokhuis et al. Systemic and intrathecal immune activation in association with cerebral and cognitive outcomes in paediatric HIV Most plasma features were termed approximately normally-distributed on the ln-scale. As the features displayed some non-normality and as the group sizes differed (implying unequal group variances) a nonparametric alternative to the independent samples t-test was used: the Mann-Whitney U test. This test was employed to each feature (i.e., plasma marker) for the juxtaposition of interest. (When possible) Exact p-values were calculated for each test instance. Multiplicity correction (p-value adjustment) was performed based on the false discovery rate (FDR), which was controlled at .05.
In addition, global differential expression testing was performed. Specifically, a Global Test was employed. 1 This test may be seen as a method that looks for differentially expressed feature-sets. It can be used for testing the overall expression profile to calculate whether it is notably different between conditions. This test is suitable when there may be insufficient or low power to detect individual plasma markers. This test may then indicate/detect if the overall plasma-marker expression profile differs markedly between the HIV and control conditions. Moreover, this test may take possible confounders into account. The Global Test provides a single p-value for a chosen feature set. The chosen feature set is the set of all (retained) plasma markers (33 features).
These two approaches were considered for complementarity. The first approach was a standard approach. The second approach is lesser known, but could have advantages in situations of low-power. In addition, consistency in findings over these two approaches (in terms of the differential expression signature) strengthen the findings. The possible confounders that were considered in the Global Testing exercise were age at inclusion and gender.

Approach to question 2.
Question 2 asked if the plasma markers were concordant with their corresponding CSF markers. This question thus concerned markers that, within the same person, were measured in both plasma and CSF. This resembles a testretest setting. Hence, the notion of concordance was appropriate to assess the consonance between corresponding plasma and CSF markers. Here, concordance was operationalized with Kendall's W. 2 We constructed 95% Bootstrap confidence intervals around the concordance values. The number of Bootstrap iterations was set to 10,000.

Blokhuis et al.
Systemic and intrathecal immune activation in association with cerebral and cognitive outcomes in paediatric HIV

Approach to questions 3 and 4
To explore associations between soluble biomarkers, clinical characteristics, neuroimaging, and cognitive functioning, we assessed conditional associations as measured through partial correlations. 3 The non-zero partial correlations between these variables form a network that can be interpreted as a conditional independence graph.
Linkage in such a network implies that variables, conditional on the effects of all other variables, are associated.
Hence, linkage means that the association between two linked variables cannot be explained by conditioning on the other variables.
The variables selected for the network analyses included the following: 1) Biomarkers representing immune activation, inflammation, endothelial function, and neuronal damage in relation to HIV. [4][5][6][7] Available markers were slightly different per compartment (see also Supplemental Table   1 2) Clinical variables: age, as important factor concerning brain and cognitive development; HIV viral load (VL) at study inclusion; CD4 + T-cell count Z-score at nadir, which represents the standard deviation from the ageappropriate mean at its lowest point, before or shortly after cART was initiated; and the age at which cART was initiated. Further technical details of network modeling are explained below.

Graphs: a language for networks
Networks are represented by graphs. We consider graphs G = (V, E ) consisting of a finite set V of vertices (or nodes) and set of edges E . The vertices of the graph correspond to a collection of random variables with a multivariate probability distribution. Edges connect pairs of vertices. We thus focus on Gaussian graphical modeling.
The support of a Gaussian precision matrix (i.e., the inverse of the covariance matrix) represents a Markov random field. This means that conditional independence between a pair of variables corresponds to zero entries in the precision matrix. Now, let Ω denote a generic estimate of the precision matrix. When the jj ' th entry of the precision matrix is zero, this implies that the variables Y j and Y j ' are independent given the remaining variables, which implies that Y j and Y j' are unconnected in the graph. Hence, model selection efforts in Gaussian graphical models focus on determining the support of the precision matrix.
To determine the support, we first need an estimate of the precision matrix. We employ a regularized estimate as the data can be high-dimensional in the sense of containing more variables than observations (p > n). Regularization will also stabilize estimates when the data are not or almost high-dimensional. The penalized Maximum Likelihood ridge estimator 3 is employed, given by: with S denoting the sample covariance (or correlation) matrix, where T denotes a symmetric positive definite target matrix, and where λ ∈ (0, ∞) denotes a penalty parameter. The target matrix is taken to be the (p × p)-dimensional identity matrix I p . The optimal penalty parameter was determined by leave-one-out cross-validation of the negative log-likelihood score. Support determination was performed based on partial-correlation thresholding (the partial correlation matrix is a scaled version of the precision matrix). The 30 strongest partial correlations were retained.
The sparsified matrix then represents the network. All employed machinery is available through the package rags2ridges in the statistical language R. 12