Gene Expression in HIV-Associated Neurocognitive Disorders

A Meta-Analysis

Abstract

Objectives: 

To identify differentially expressed (DE) genes in HIV-associated neurocognitive disorders (HAND) patients in comparison with HIV-infected patients without HAND and controls.

Design: 

A meta-analysis of publicly available gene expression data from HIV postmortem brain tissue studies.

Methods: 

We selected studies using clearly defined inclusion and exclusion criteria. Within study data preprocessing and individual analyses were performed for each brain region. The following meta-analytic methods were applied: combining P values, combining effect sizes with and without a permutation method. The DE genes were defined with a false discovery rate less than 5% using Benjamini–Hochberg method.

Results: 

Our meta-analysis on 3 studies encompasses analyses of over 48 postmortem brains [25 HAND, 7 HIV encephalitis (HIVE), 8 HIV-infected patients, and 8 controls]. Overall, 411 genes in white matter were DE in HAND with HIVE patients when comparing with controls. Of these, 94 genes were significantly expressed in all statistical methods. These 94 genes participate in significant pathways such as immune system, interferon response, or antigen presentation. Sixty-six of the 94 genes were significantly upregulated with log₂ intensities greater than 2-fold. Strong examples of the highly upregulated genes were PSMB8-AS1, APOL6, TRIM69, PSME1, CTSB, HLA-E, GPNMB, UBE2L6, PSME2, NET1, CAPG, B2M, RPL38, GBP1, and PLSCR1. Only BTN3A2 was expressed in HAND with HIVE patients as compared with HAND patients without HIVE.

Conclusion: 

A number of genes were DE in our meta-analysis that were not identified in the individual analyses. The meta-analytic approach has increased statistical power for identifying DE genes in HAND.

INTRODUCTION

HIV-associated neurocognitive disorders (HAND) are a common neurological complication of HIV infection. HIV penetrates the central nervous system (CNS) shortly after initial viral exposure,¹ whereas HIV RNA can be detected in the cerebrospinal fluid (CSF) as early as 8 days after infection.² Similar to all retroviruses, HIV RNA is transformed to proviral DNA by reverse transcriptase enzyme to incorporate into the host cell genome. HIV manifests pathogenic effects both in the immune system such as infecting T lymphocytes, microglia and macrophages, and the nervous system. Both systems generally use chemical mediators, such as cytokines and chemokines to regulate the complex interactions between cells required for normal functioning.^3,43,4 Nevertheless, the infected glial cells release toxic factors, such as viral products, excitotoxins, and cytokines or chemokines in turning damage neurons and affecting neuronal functions in vulnerable CNS areas, including white matter, frontal cortex, basal ganglia, and hippocampus brain regions.^{4–84–84–84–84–8} As a result of HIV disruption, the processes induce neuronal apoptosis in both the immune and CNS systems, causing neurodegenerative diseases in 50% or more of those infected.^{2–42–42–4}

HAND remains prevalent despite the widespread use of antiretroviral (ARV) treatment. A 20-year survival rate in HAND patients was approximately 70% compared with about 90% in HIV-infected patients without HAND with a hazard ratio of about 3-fold.⁹ In addition to the increased mortality, HAND diagnosis also remains unclear. Most patients present with mild conditions and no identified symptoms, and that dissemble symptoms make HAND difficult to identify. Comparing HAND severity between the pre- and post-antiretroviral therapy eras indicates that HIV-associated dementia has decreased from 18% to less than 5%, whereas mild neurocognitive disorder and asymptomatic neurocognitive impairment have increased from 12% to 17% and from 20% to 28%, respectively.^10,1110,11 The asymptomatic HAND also had a shorter time conversion to symptomatic HAND than normal neurocognitive patients.¹²

Access to formal neuropsychological testing for detecting HAND is limited. Various neuropsychological testing instruments that screen for HAND perform well in assessing severe HAND but perform poorly when detecting milder HAND conditions. In a meta-analysis on screening tools, the HIV dementia scale has poor pooled sensitivity (0.48), and the international HIV dementia scale has moderate pooled sensitivity (0.62).¹³ Moreover, the accuracy of HAND detection and differentiation is highly dependent on availabilities of controls used in standardizing neuropsychological testing scores. In other words, HAND diagnosis requires comparison of cognitive performance specifically to local normative data, which are different by clinical research settings, indicating that appropriate normative data are substantially required in the neuropsychological testing in HAND research.^{10,11,1410,11,1410,11,14}

Furthermore, neither systemic nor plasma biomarkers can clearly detect HAND. Both CD4 count and viral load (VL) in plasma and CSF are not associated with HAND in post-cART era.^{2,14,152,14,152,14,15} Neurological abnormalities can develop even with viral suppression.¹⁶ Also nadir CD4 count has been recently indicated as a predictor of HAND but none of the nadir CD4 thresholds could be used to identify HAND.^{17–1917–1917–19} Meanwhile, HIV detections in plasma and CSF can be discordant, detectable in CSF but undetectable in plasma, uncommon but appear 10% in HIV-infected patients.^2,172,17

Because access to formal neuropsychological testing is limited, gene expression studies have been important for HAND investigations. Several gene expression studies in HAND have been conducted. However, genetic markers identified in the gene expression studies may differ due to data quality issues, application of different microarray platforms, sample characteristics, or insufficient controls. Also due to the limited availability of postmortem brain samples, small data sets are still a major obstacle to ensuring adequate statistical power in microarray experiments. In this research, we conducted a meta-analysis of publicly available gene expression data from HIV postmortem brain tissue studies to increase statistical power for identifying significant genes in HAND and HIV encephalitis (HIVE).

METHODS

Data Acquisition

Microarray gene expression data were gathered from searches using the public data repository in Gene Expression Omnibus (GEO) from the National Center for Biotechnology Information (NCBI) (http://www.ncbi.nlm.nih.gov/geo/) using the keywords “HIV-associated neurocognitive” or “HIV neurocognitive” or “HIV cognitive” or “HIV dementia” or “HIVE” or “HIV brain” or “HIV frontal cortex” or “HIV white matter” or “HIV ganglia” or “HIV cortex” or “HIV hippocampus,” and filtering for series (entry type) and “homo sapiens (organism)” and “expression profiling by array (data type).”^20,2120,21 These keywords were also searched in ArrayExpress database, the European Bioinformatics Institute (EBI) (http://www.ebi.ac.uk/microarray-as/ae/), and Stanford Microarray database (http://smd.stanford.edu/).^22,2322,23 Twenty-two series of RNA expression profiling in the GEO database were selected for first review. Most of the studies could also be accessible in the ArrayExpress database, whereas no studies were found in the Stanford Microarray database. The eligible criteria for data acquisition were as follows: (1) the data sets were publicly accessible, (2) the samples were from human brain regions, (3) the samples contain any of these controls: HIV seropositive without HAND, HIV seronegative, or HIVE, (4) the data sets contain phenotypic data and published articles describing the data, (5) the raw or normalized intensity data were defined as gene expression levels, (6) the data sets with or without original gene expression level files were included. For each study, we reviewed the minimum information about a microarray experiment (MIAME) in the GEO website, including research methods and results described in the articles, and data summary of the phenotypic data. We then excluded the studies without brain samples, without controls, without a supporting article, with redundant sub–data sets, and experiments in other species. We also reviewed a previous meta-analysis of HAND studies,²⁴ and found that 7 studies were conducted on postmortem brain samples^{25–3025–3025–3025–3025–3025–30} (see Table S1, Supplemental Digital Content, https://links.lww.com/QAI/A730). Of the 7 studies, RNA expression data sets in only 3 were accessible in the GEO database. Based on the inclusion and exclusion criteria, a total of 3 studies were finally included in this study.^{26,29,3026,29,3026,29,30} The microarray data in the selected studies differ by sample characteristics, brain regions, microarray platforms, and number of sample replications. Patient and sample characteristics in the data sets are described in Table 1.

TABLE 1: Patient and Sample Characteristics

Data Retrieval

The NCBI GEO database contains 4 basic entity types: samples, platforms, series, and complied and curated data sets. GEO series contain complete, original, submitter-supplied records form the basis of microarray data set.^20,2120,21 As data quality may influence gene expression results and the selected studies gather original data, we primarily evaluated data quality of the samples.^31,3231,32 The selected microarray data sets were performed using Affymetrix Human Genome U95 Version 2 Array (HG-U95v2) and Affymetrix Human Genome U133 Plus 2.0 Array (HG-U133 Plus 2.0), consisting of 12,625 and 54,675 probe sets,³³ respectively. Accessible cell intensity (CEL) files were also retrieved for data quality control (QC).^29,3029,30 The CEL files contain all the intensity-related information of the probes on the array such as intensity itself and other meta information.³⁴ One of the 3 studies includes only normalized expression and phenotypic data.²⁶

Data QC

To assess quality of hybridizations, we evaluated the consistency of the percent present calls among all arrays hybridized.³⁵ We expected the percent of probe sets called present to roughly fall in the range from 30% to 60%, and this was consistent across all the arrays in the experiments.³⁶ We also evaluated the 3′/5′ ratio of 2 control genes, beta-actin and glyceraldehyde-3-phosphate dehydrogenase (GAPDH). The ratios represent RNA integrity and efficiency of the first strand cDNA synthesis. Affymetrix also suggests scaling factors used for measuring arrays on a similar scale should not differ by more than about 3-fold. Large differences in scaling factors may indicate either poor-quality RNA or situation where the normalization assumption may not hold.³⁶ The QC on the samples in the selected studies was also presented (see Figure S1–S4, Supplemental Digital Content, https://links.lww.com/QAI/A730). In this study, the QC criteria indicative of poor-quality samples were having a 3′/5′ GAPDH ratio greater than 3 and/or percent of present calls less than 30%.³⁷ We therefore excluded 12, 1, and 8 samples leaving 16, 34, and 64 samples in the GSE3489, GSE28160, and GSE35864 studies for our meta-analysis, respectively.

Image plots for array artifacts, histograms, and box plots of the probe intensities by arrays for technical variation were evaluated. MA plots were constructed to evaluate technical replicates of reproducibility of the assay in the 2 studies.^26,2926,29 Spearman rank correlation coefficients of the replicated samples present high correlations of the intensities (see Figure S5–S6, Supplemental Digital Content, https://links.lww.com/QAI/A730). For the studies with available CEL files,^29,3029,30 the Robust Multi-Array Average method, which includes quantile normalization to standardize intensity distributions to be identical across arrays, was used to obtain probe set expression summaries.^{38–4038–4038–40} Meanwhile, the study without CEL files provided background corrected and normalized probe set summaries using Microarray Suite 5.0 (MAS5.0).²⁶ The MAS5.0 normalizes each array independently and sequentially, whereas the Robust Multi-Array Average method uses a multichip model.^{38–4038–4038–40} The preprocessing steps were performed using the affy and simpleaffy Bioconductor R packages.^{34,36,4134,36,4134,36,41} The QC plots of the preprocessed data are presented in Figure S7A–S7D (see Supplemental Digital Content, https://links.lww.com/QAI/A730).

Data Analysis

The microarray data from HG-U133 Plus 2.0 platform of the 2 individual studies were directly combined.^29,3029,30 This platform comprised 54,675 probe sets that cover more than 38,500 human genes. However, the data from different platforms, HG-U95v2 and HG-U133 Plus 2.0 platforms with corresponding 12,625 and 54,675 probe sets, were matched at the gene level before combining the meta-analysis data sets, which contains 5,322 of the same probe sets extracted from the cross platforms.^{42–4442–4442–44} Their raw intensity data remained consistent between 2 platforms (see Figure S8, Supplemental Digital Content, https://links.lww.com/QAI/A730). After data preprocessing, we performed individual analysis and then conducted meta-analysis based on the objectives for each part of the brain as described in Table 2.

TABLE 2: Number of Differentially Expressed Genes in Individual and Meta-Analysis Results

Individual Analysis

Statistical analyses were performed on the log₂ normalized intensities. The expression values of replicate samples were combined by averaging the values into single values. A linear model was fit for each gene with sample group as the independent variable. Empirical Bayes method was applied to obtain robust estimators due to the small number of arrays.⁴⁵ Benjamini–Hochberg (BH) method was used to control the false discovery rate (FDR) due to multiple hypothesis testing of differential expression. An overall F test with an FDR of 5% and nested F-tests within contrasts were used. The moderated t-statistics in pairwise comparisons were computed for each gene in each study. The DE genes were defined as those with FDR less than 5%. Unsupervised hierarchical clustering using Ward's method and one minus Pearson's correlation coefficients for measures of similarities by a heatmap plot were used to present graphically the DE genes in the individual analysis. These analyses were performed using limma package in the R programming environment.^45,4645,46

Meta-Analysis

The small sample sizes in these studies of postmortem brain samples may result in inadequate statistical power for identifying differential gene expression. The genetic markers identified in individual analyses may also differ due to data quality issues, different microarray platforms, sample characteristics, or insufficient controls.^{33,47,4833,47,4833,47,48} We conducted a meta-analysis to increase the statistical power in this study.

Several statistical methods have been used in meta-analysis for gene expression studies. Two common statistical methods were applied. First we combined P values by an inverse normal method, weighted by the number of samples from an individual analysis. Then BH method was applied to control the FDR.⁴⁹ We also applied the effect size combination method from two level hierarchical models, derived from the true differences of mean expressions among groups. A modified BH method was used for controlling the FDR in multiple testing. We generated the z-statistics from empirical distributions by a thousand random permutations to obtain the modified FDR by the order statistics of the actual and permuted z-statistics.^{49–5149–5149–51} These 2 combination methods were implemented using MAMA package in the R programming environment⁵² and modified programs from the package.

Pathway Analysis

To generate biologically meaningful results, we applied gene set enrichment analysis (GSEA) to assess whether a gene set or a pathway was statistically significant. We created gene sets based on metabolic pathways annotated in the Kyoto Encyclopedia of Genes and Genomes (KEGG) database^53,5453,54 and Gene Ontology annotations of Biological Processes (GOBP) terms.⁵⁵ A hypergeometric test was used to test association of genes and biological process categories. Pathway enrichments were defined at nominal P value less than 0.05 and FDR less than 10%. These analyses were performed using GSEABase package in R programming environment.^56,5756,57 In addition, we also performed gene network analysis using weights assigned from DE genes in GeneMANIA, a publicly available web interface, to identify gene–gene interactions and predict additional genes that may be involved in HAND.⁵⁸

RESULTS

After data preprocessing and individual analysis, 183 genes in white matter were differentially expressed in HAND patients with HIVE in the GSE28160 study, whereas only 2 genes were differentially expressed in the GSE35864 study when both comparing with controls. Similarly, 101 genes in white matter were differentially expressed in all HAND patients in the GSE28160 study, and no genes were differentially expressed in the GSE35864 study. Some variability of individual analysis results was found (Table 2). We displayed an example of sample classification based on the significant DE genes of all HAND patients and controls using hierarchical clustering (see Figure S9, Supplemental Digital Content, https://links.lww.com/QAI/A730), and also Venn diagrams presenting number of DE genes with an FDR of 5% in white matter and basal ganglia (see Figure S10, Supplemental Digital Content, https://links.lww.com/QAI/A730).

Our meta-analysis on 3 studies encompasses analyses of more than 48 postmortem brains (25 HAND, 7 HIVE, 8 HIV-infected patients, and 8 controls) from white matter and frontal cortex, regions. In white matter, a number of genes were differentially expressed by both combining methods, P values and effect sizes, whereas these expression differences were not detected in the individual analysis (Table 2). In total, 411 genes were differentially expressed in HAND patients with HIVE, when comparing with controls. Of the 411 genes, 94 were significantly expressed in all statistical meta-analysis methods. These 94 genes can be used to effectively classify HAND patients with HIVE and control samples (see Figure S11, Supplemental Digital Content, https://links.lww.com/QAI/A730). The expression pattern of these genes was apparently similar in both individual studies (Fig. 1). The list of gene symbols, log₂ normalized intensities, and fold changes values of the expressions are presented in Supplemental File 2 (see Supplemental Digital Content, https://links.lww.com/QAI/A730).

FIGURE 1: Heatmaps of expression patterns of 94 differentially expressed genes in white matter in HANDHIVE and control samples using meta-analysis. HANDHIVE: HIV-associated neurocognitive disorder with HIV encephalitis.

Sixty-six of the 94 genes were significantly upregulated with log₂ normalized intensities greater than 2-fold in HAND patients with HIVE when comparing with controls. The top 10 upregulated genes were proteasome activator subunit 1 PSMB8 antisense RNA 1 (PSMB8-AS1), apolipoprotein L 6 (APOL6), tripartite motif containing 69 (TRIM69), proteasome activator subunit 1 (PSME1), cathepsin B (CTSB), signal transducer and activator of transcription 1 (STAT1), major histocompatibility complex, class I, E (HLA-E), glycoprotein nmb (GPNMB), ubiquitin-conjugating enzyme E2L 6 (UBE2L6), and proteasome activator subunit 2 (PSME2). After a brief summary of these genes, PSMB8-AS1 is a long noncoding RNA containing 6 spliced transcript variants. APOL6 belongs to an apolipoprotein L gene family, which might influence the movement of lipids in the cytoplasm or binding of lipids to organelles. TRIM69 is a protein-coding gene involving ubiquitin-protein ligase activity and may contribute to apoptosis. UBE2L6 is a protein-coding gene involving an important mechanism for targeting abnormal or short-lived proteins for degradation. PSME1 has been implicated in immunoproteasome assembly and required efficient antigen processing. The protein encoded by CTSB is lysosomal cysteine proteinase, which is produced from a single protein precursor known as amyloid precursor protein (APP) secretase and involved in the proteolytic processing of APP. Incomplete proteolytic processing of APP has been believed to cause dementia in Alzheimer disease. STAT1 responds to cytokines and growth factors, provides signal transducer and transcription activator those involve cellular responses to interferons. HLA-E is peptide antigen binding and receptor binding and a presence of foreign antigens to the immune system. GPNMB is related to integrin binding and heparin binding and involved in the regulation of sensory transmission, thermoregulation, tissue development, proliferation, and differentiation. PSME2 is implicated in immunoproteasome assembly and required for efficient antigen processing.^{53,54,5953,54,5953,54,59}

Seven of the 94 genes were significantly downregulated in HAND patients with HIVE: ribosomal protein L37a (RPL37A), DEAD (Asp-Glu-Ala-Asp) box polypeptide 58 (DDX58), NLR family, Caspase recruitment domain-containing 5 (NLRC5), poly (ADP-ribose) polymerase family (PARP9), butyrophilin, subfamily 3, member A3 (BTN3A3), sialic acid binding Ig-like lectin 1, sialoadhesin (SIGLEC1), and STAT1. The RPL37A is a gene encoding ribosomal proteins belonging to the L37AE family. The DDX58 is involved in viral double-stranded RNA recognition and the regulation of immune response. The NLRC5 plays a role in cytokine response and antiviral immunity through its inhibition of NF-kappa-B activation and negative regulation of type I interferon signaling pathways. The BTN3A3 is a group of major histocompatibility complex-associated genes that encode type I membrane proteins with 2 extracellular immunoglobulin domains. The SIGLEC1 encodes a member of the immunoglobulin superfamily, expresses only by a subpopulation of macrophages, and involves in mediating cell–cell interactions. The STAT1 respond to cytokines and growth factors is important for cell viability in response to different cell stimuli and pathogens.^{53,54,5953,54,5953,54,59}

In addition, 35 genes in white matter were differentially expressed in all HAND patients when comparing with controls (see Figure S12, Supplemental Digital Content, https://links.lww.com/QAI/A730). A few of these genes were significantly differentially expressed by 2 statistical methods. Twenty-eight genes in white matter were differentially expressed in HAND patients with HIVE when comparing with HAND without HIVE (see Figure S13, Supplemental Digital Content, https://links.lww.com/QAI/A730). One of the 28 genes was significantly differentially expressed for all statistical meta-analysis methods, butyrophilin, subfamily 3, member A2 (BTN3A2). The BTN3A2 is a protein-coding gene in which the encoded protein plays a role in T-cell responses in the adaptive immune response and inhibits the release of interferon gamma from activated T cells.^{53,54,5953,54,5953,54,59} However, no genes were differentially expressed in overall HAND patients or HAND without HIVE (results not shown) when comparing with controls. Also, no genes in frontal cortex were significantly differentially expressed in HIVE patients in comparison with HIV without encephalitis patients (Table 2).

Using GSEA with the 94 DE genes, we identified 13 significant pathways enriched among 48 pathways tested using KEGG analysis, and 473 pathways enriched among 1,513 pathways tested using gene ontology enrichment analysis. The 13 significant KEGG pathways included the immune system, antigen processing and presentation, transport, and catabolism in cellular process, folding, sorting, and degradation in genetic processing, signaling molecules and interaction, endocrine and metabolic diseases, and cardiovascular diseases. As with the KEGG pathway analysis, the gene ontology list also includes pathways important in the host response to infectious agents even at the very top. The first-10 were interferon related, immune response, defense response, immune system process, antigen processing, cytokine signaling, and immune effector process (Table 3; see Supplemental File 3, https://links.lww.com/QAI/A730). In addition, our gene expression network consists of 71 recognized DE genes, 20 additional genes that strongly connect to the query genes, 10 consolidated pathways, and 3,030 total links. GeneMANIA retrieved these genes with known coexpression (68.42%), consolidated pathways (25.08%), physical interactions (2.71%), colocalization (2.01%), predicted interactions (1.09%), shared protein domain (0.58%), and genetic interactions (0.07%) (Fig. 2). The major functional networks and the genes sharing the same protein domains can be found in Figure S14A–S14D (see Supplemental Digital Content, https://links.lww.com/QAI/A730).

Table 3: List of significant pathways correlated with 94 differentially expressed genes

FIGURE 2: Gene networks produced using GeneMANIA (www.genemania.org). The network consists of recognized differential expressed genes or query genes (black circles), 20 additional genes (small gray circles) that strongly connect to the query genes, 10 consolidated pathways (gray diamonds), and 3,030 total links.

DISCUSSION

Our meta-analysis shared some significant pathways based on neuroimmunological activation in HAND with HIVE with previous studies, such as immune response, defense response, cytokine and interferon response, and antigen presentation.^24,3024,30 These pathway expressions show increased inflammation and neuroimmunity associated with HIVE.³⁰ At a single gene level, our results support a previous study that these DE genes were upregulated in HAND patients with HIVE: ISG15, GBP1, IFIT3, IFI44L, BTN3A3, IFIT3, STAT1, MX1, IFIH1, PSMB8, and HLA-B.³⁰ These are interferon response genes and interferon regulatory factors expressed with high fold-change values (see Supplemental File 2, https://links.lww.com/QAI/A730).

Our results additionally confirm that no interferon response genes were significant expressed in HAND patients without HIVE when comparing with controls (results not shown). However, in this study, we were able to detect more genes with expression values greater than 4-fold from individual studies: GBP1, TRIM69, TRIM14, APOL6, STAT1, PSMB8-AS1, CTSB, GPNMB, CAPG, RARRES3, RPL38, HLA-E, PSME2, and FUCA1.

Two HAND and 2 control samples from the GSE28160 study were in common with the GSE35864 study. Another HIVE sample from the GSE35864 study was the HIVE patient in the GSE3489 study.³⁰ Due to small data sets, data quality issues, and no selection criteria for duplicated samples among settings, we included the samples as they were in the original studies and also provided additional results when the duplicated samples were excluded from one of the studies. We found 81 significant DE genes in HAND and HIVE vs control samples in all meta-analysis methods when the 2 overlapping cases were excluded from the GSE28160 study, all of which were genes among the 94 significant DE genes as previously presented. However, the number of DE genes was substantially reduced when the 2 overlapped controls were excluded (see Table S2, and Figure S15–S16, Supplemental Digital Content, https://links.lww.com/QAI/A730). Another duplicated case in the GSE3489 and GSE35864 studies were previously excluded from the GSE3489 study due to poor sample quality.

The meta-analytic approach combines results from independent studies to increase not only statistical power but generalizability of individual analysis. As of today, microarray technologies are the most effective way to obtain comprehensive gene expression profiles, and a number of gene expression studies published try to address similar biological questions. However, conducting gene expression studies in HIV postmortem brain samples is substantially difficult as there is naturally time delay between death and sample collection. Combining data from multiple existing studies can benefit both by statistical power increases as sample size increases and result generalization across sample characteristics in HAND gene expression studies.^32,3332,33

In the cART era, the pathophysiology concept of inflammation, encephalitis, and neurodegeneration has been changed.⁶⁰ The prevalence of HIV-associated dementia and HIVE has declined, whereas milder HAND conditions are difficult to identify.^{24,30,6024,30,6024,30,60} Because the actual mechanism of HAND is unknown,⁶⁰ the meta-analysis approach combining published studies over time should be undertaken with caution and performed along with individual analysis. We presented a meta-analysis result of the combined studies across the pre-cART and post-cART eras as there was a strong correlation of neocortical gene expression in HIVE vs HIV in the GSE3489 and GSE35864 studies.³⁰ However, with small data sets, we could not identify DE genes in HIVE vs HIV in our analyses.

HAND was nasologically renamed and its categories were formalized in 2007.⁶¹ Many of the subjects were diagnosed before the diagnostic scheme. The HIVE was diagnosed likely as end-stage AIDS illness^30,6230,62 in which the HIVE or HAND with HIVE were more likely dementia with high prevalence. After cART, the HIVE diagnoses were decreased as neuropathological traces such as astroglial scars or lasting residua in developed HIVE disappeared at autopsy. HAND without HIVE may thus be more likely minor neurocognitive conditions.^30,6030,60 This is a possible reason that our meta-analysis identified mostly DE genes in the HAND with HIVE category.

Several biological differences among subjects can influence gene expression patterns. In a previous study, antiretroviral therapy received at the time of death improved nervous system functions and reduced dysregulated expression in HAND, no matter HIVE occurrence. Higher dysregulation of immune responses and endogenous antigen presentation pathways was also presented in untreated patients with HIVE than those without HIVE.²⁴ Types and duration of ARV regimens would be helpful to sort out clinicopathological data in HAND. Meanwhile, this evaluation was limited in our meta-analysis because the ARV regimens at death were available from only one study.²⁴ Additionally, the expression pattern can be variable in different brain regions. In HIVE, many upregulated genes were found in neostriatum, and few found in white matter. However, in frontal neocortex, several genes were dysregulated,^26,3026,30 whereas some genes were upregulated.²⁶ Our meta-analysis was restricted to only the samples from white matter and frontal cortex regions. We found many upregulated genes in white matter, but no genes were expressed in frontal cortex.

The success of cART has recently changed HIV infection to a chronic disease.⁶³ HAND with HIVE becomes less prevalent, whereas mild cognitive impairments still persist. The cliniconeuropathological relevancy to HAND remains suspicious.⁶⁰ Various clinical research in the field mainly focuses on therapeutic intervention targeting on viral suppression.⁶⁴ In addition, some evidence suggested that HAND has neuropathological changes similar to other neurodegenerative diseases such as Alzheimer's or Parkinson diseases^{30,60,6530,60,6530,60,65}; however, only a handful of studies investigated disease mechanisms of HAND and the neurological conditions. In particular, only one underpowered study identified common genes and pathways associated with the cognitive impairments between HAND and Alzheimer's disease in the current era.⁶⁶ Future investigation for neurocognitive impairments of the disease in a larger study or metadata would be useful to better understanding HAND neuropathogenesis.

Data qualities have a strong impact on results of gene expression. We found a diversity of gene expression patterns in both individual analysis and meta-analysis, when the microarray data contained either one or more samples with poor qualities, particularly in the individual analysis with small data sets (Table 2; see Figure S10, Supplemental Digital Content, https://links.lww.com/QAI/A730). We suggest sample exclusion of poor-quality hybridizations by 3′/5′ GAPDH ratio greater than 3 and/or percent present calls lower than 30%, in addition to data visualizations. This will remove artifacts, noise, noninformative, and problematic arrays, resulting in decreased data heterogeneity and improved conditions for accurate statistical testing.^{33,48,6733,48,6733,48,67}

Although a brain biopsy could not be recommended for genetic investigations in HIV-infected patients still alive, our study maybe benefit HIV research on the CNS, providing more information for understanding common disease mechanisms of HAND and help researchers to extend future work for HAND gene expression studies. The importance of our study will also benefit other gene expression studies using postmortem brain samples about data quality evaluation, as the data quality issues have a huge impact on results of gene expression studies using postmortem brain samples.

ACKNOWLEDGMENTS

The authors thank the reviewers and the editor for their helpful comments and suggestions, and the National NeruoAIDS Tissue Consortium (NNTC) data coordinating centre to identify sample duplications.