Blood Gene Expression Profiling of Neurologic Diseases: A Pilot Microarray Study

Background  Tissue gene expression profiling with arrays measures the transcription of thousands of genes. However, this approach cannot be readily used to guide clinical neurologic practice.

Objectives  To determine whether clinical neurologic diseases are associated with unique patterns of up- and down-regulated genes in whole blood and to explore the possibility of using peripheral blood as a surrogate tissue in these diseases.

Design  Case-control study.

Setting  University-based pediatric and adult neurology clinics.

Participants  Patients with neurofibromatosis type 1, epilepsy, or Tourette syndrome diagnosed using traditional clinical criteria; controls without disease; and controls with neurologic disease.

Main Outcome Measure  Blood gene expression levels of greater than 12 000 genes, measured using U95A arrays.

Results  Neurofibromatosis type 1 and childhood epilepsy treated with carbamazepine or valproic acid are associated with distinct patterns of blood gene expression. Patients with valproic acid–responsive vs valproic acid–refractory epilepsy formed distinct subclusters. Tourette syndrome was characterized by several gene expression clusters. In 1 cluster, 6 genes—all associated with immune cell function—were overexpressed.

Conclusion  Blood gene expression profiling can provide surrogate markers for neurologic diseases without obvious blood phenotypes.

Global gene expression profiling with DNA microarrays measures the transcriptional activity of thousands of genes. Among other uses, gene arrays can produce a molecular fingerprint for disease diagnosis, classification, and prognosis. Cancer expression profiling has distinguished tumor subtypes, evaluated sensitivity to chemotherapy, and predicted clinical outcomes.¹ This technology has also been applied to examining the brain genomic changes of several neurologic diseases.^2,3 However, findings from brain genomic profiling cannot be readily used to guide clinical practice owing to the difficulty of obtaining brain tissue. This made us consider whether genomic profiling of peripheral blood could provide meaningful surrogate markers for brain diseases. Peripheral blood cells inherit the same genetic information as brain cells and are equipped with abundant signaling pathways that respond to a myriad of pathologic changes, so blood genomic profiling may reflect changes in the brain.

To test the hypothesis that gene expression in blood could reflect brain abnormalities, our group⁴ previously examined the gene expression profiles in the blood of rats subjected to a variety of acute neurologic insults. We found that microarrays identified specific changes in gene expression in blood 24 hours after stroke, seizures, and hypoglycemia.

Consequently, we studied neurologic diseases in humans. We postulated that single gene abnormalities should produce blood transcriptional changes even in the absence of obvious blood phenotypes. We tested this hypothesis in neurofibromatosis type 1 (NF1), an autosomal dominant disease caused by mutations of the NF1 gene (chromosome band 17q11.2).⁵ In addition, we tested whether gene expression patterns in blood could identify markers of medication responses in pediatric epilepsy.⁶ Because many genes may contribute to an inherited basis of drug response,⁷ we reasoned that a high-throughput approach might yield important insights.

Our final aim was to explore whether blood gene expression patterns distinguish complex, heritable neurologic diseases for which no causative gene has been identified, which we tested in Tourette syndrome (TS). This syndrome seems to follow an autosomal dominant inheritance pattern, but linkage analyses have been unsuccessful, and multiple genetic or environmental factors may contribute to the phenotype.

A total of 129 individuals participated in the study: control subjects without neurologic disease (n = 13; 7 males and 6 females; mean ± SD age, 27.7 ± 6.9 years), children with headache (n = 15; 7 boys and 8 girls; mean ± SD age, 12.7 ± 3.0 years), children with epilepsy (n = 24; 12 boys and 12 girls; mean ± SD age, 10.6 ± 3.6 years), patients with NF1 (n = 12; 7 males and 5 females; mean ± SD age, 28.0 ± 19.4 years), patients with bipolar disorder or schizophrenia (n = 17; 15 men and 2 women; mean ± SD age, 54.4 ± 8.4 years), adults with epilepsy (n = 20; 7 men and 13 women; mean ± SD age, 35.2 ± 10.5 years), adults with Parkinson disease or progressive supranuclear palsy (n = 12; 5 men and 7 women; mean ± SD age, 65.6 ± 12.6 years), and patients with TS (n = 16; 12 males and 4 females; mean ± SD age, 17.4 ± 13.9 years).

Diagnoses of NF1 were based on established clinical criteria.⁸ Children with epilepsy were treated with valproic acid (n = 11) or carbamazepine (n = 6) or were newly diagnosed and untreated (n = 7). Patients with TS had to meet Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition,⁹ criteria for TS and have at least 1 affected first-degree family member. Protocols were approved by the institutional review boards at the Cincinnati Children’s Hospital Medical Center and the University of Cincinnati, and all participants or their guardians provided informed consent.

Ten- to 15-mL blood samples were collected in tubes containing EDTA and then were mixed with Trizol LS reagent (Invitrogen, Carlsbad, Calif) within 15 minutes. Total RNA was isolated and purified using the RNeasy mini kit (Qiagen Inc, Valencia, Calif). The quality of total RNA was assessed using a bioanalyzer (Agilent 2100; Agilent Technologies Inc, Palo Alto, Calif) and was quantified by spectrophotometry.

Sample labeling, hybridization to U95A arrays (>12 000 sequences), and image scanning were performed as described in the Affymetrix Expression Analysis Technical Manual.¹⁰ Arrays were normalized using the invariant set normalization method. Expression values were calculated using the perfect match–only model-based expression index and data analysis software (dChip version 1.2; Harvard University, Cambridge, Mass).¹¹

The variation of gene expression in human blood has been related to many factors,¹² which would tend to obscure important correlations (type II error) or create spurious correlations (type I error) when comparing disease cases with controls. We therefore analyzed 266 highly variably expressed genes in 14 samples taken on separate days from 7 healthy donors using expression data analysis software (GeneSpring version 6.0; Silicon Genetics, Redwood City, Calif), which performs “unsupervised cluster analysis,” pairing samples by the highest correlation of gene expression.¹³ If the intraindividual gene expression patterns are stable, reflecting the genetic and environmental factors unique to each individual, the program clusters each individual’s samples.

We hypothesized that blood gene expression patterns in patients with neurologic diseases of interest were different than those of controls with and without disease. For the NF1 study, we compared 12 NF1 blood samples with 3 independent sets of 12 healthy and patient controls matched for age and sex. For the study of treatment response in pediatric epilepsy, we compared valproic acid, carbamazepine, and no treatment. For the TS study, we compared 16 TS blood samples with those of 8 healthy and disease control groups to identify genes with the highest possible specificity for TS. We anticipated that TS might not be a single class associated with 1 distinct gene expression profile.

As an internal means of validation, we performed permutation analysis (BRB ArrayTools; developed by Richard Simon, DSc, and Amy Peng Lam), which determines the number of genes differentially expressed at an appropriate significance, then performs random permutations of the class labels (ie, which samples correspond to which classes) and computes the proportion of the random permutations that gave as many genes significant as were actually observed. This proportion provides a global test of whether the observed expression profiles differed by chance. P < .05 is sufficient to establish that class-associated differences in gene expression exceed what would be expected by chance.

Unsupervised cluster analysis of the 266 genes with the highest variation paired individuals’ samples (data not shown), indicating that variation in blood gene expression related to the temporal and technical reasons is smaller than the interindividual variation.

For the comparisons between patients with NF1 and the 3 sets of controls, 24, 55, and 58 genes were differentially expressed in NF1 blood compared with that of controls, and the probability of generating the same number of genes by chance was 2.3%, 0.2%, and 0.7%, respectively. Hierarchical cluster analysis using these gene lists demonstrated that the NF1 samples clustered separately from each set of controls (Figure 1). In contrast, the 3 comparisons between controls produced 1, 3, and 0 genes, and the probability of generating the same number of genes by chance was 70%, 29%, and 100%, respectively. Of pathophysiologic interest, many dysregulated genes in NF1 blood are related to tissue remodeling and tumorigenesis.⁵

The global permutation-based test showed that the expression patterns in blood caused by valproic acid and carbamazepine are significant (P = .005 for valproic acid vs no treatment and P = .02 for carbamazepine vs no treatment). Cluster analysis segregated the valproic acid samples into 2 subclusters. One subcluster included all 3 valproic acid–resistant patients, and the other included all 8 valproic acid–responsive patients (Figure 2). Many mitochondrial genes are overexpressed in valproic acid responders, pointing to the possible involvement of mitochondria in the determination of valproic acid efficacy.⁶

Permutation analysis showed that the blood gene expression pattern associated with TS has a P = .20. In other words, 20% of random permutations of the class label generated the same number of up- and down-regulated genes. Thus, we did not find evidence that the clinical diagnosis of TS is associated with a single, unique gene expression profile in whole blood.

Subgroup analysis showed that there were 6 up-regulated genes and 1 down-regulated gene in TS (P < .05 for each of 8 comparisons) (Figure 3). Granzyme B is involved in the target-killing process of cytotoxic T cells or natural killer cells.¹⁴ NKG2E encodes a lectin-like receptor, which plays a role in the recognition of the major histocompatibility complex molecules by natural killer cells and some cytotoxic T cells. CD94 is also preferentially expressed by natural killer cells and forms heterodimers with NKG subunits.¹⁵ NK-p46 participates in natural killer cell–mediated lysis of cells infected with intracellular bacteria.¹⁶ IMPA2, expressed in caudate, a region shown in neuroimaging studies to be involved in TS and obsessive-compulsive disorder, was down-regulated.

Post hoc, K-means cluster analysis for the 6 cytotoxic T-cell/natural killer cell genes yielded cluster A (low expressers) and cluster B (high expressers). The proportion of patients with TS in the higher expression group was significantly greater than the proportions in the control groups (P < .05 by χ² test). Within TS, permutation-based analysis showed that less than 4% of random permutations generated the same number of differentially expressed genes as those in clusters A and B (Figure 4).

Global gene expression profiling holds potential for classifying diseases and predicting clinical outcomes based on molecular criteria. Because blood is the most accessible tissue, blood gene expression profiling has been used to explore hematologic malignancies¹ and autoimmune¹⁷ and infectious¹⁸ disorders. Our studies suggest that this approach can be extended to brain diseases.

Brain tissue is not accessible in vivo for most neurologic and psychiatric diseases. Our findings in NF1 support the notion that monogenetic neurologic or multiorgan disorders can be identified by distinct blood gene expression profiles. The results of our pediatric epilepsy and TS studies suggest that this approach may be extended to complex disorders. Blood genomic profiling might provide an accessible platform to categorize a polygenic condition into meaningful molecular subtypes. As exemplified in the anticonvulsant study, part of the interindividual variation in valproic acid efficacy may be related to an identifiable, efficacy-related drug effect at the blood transcription level.

We did not identify a unique, important blood expression pattern in TS. However, because TS may not be a monogenetic disorder, our findings may be consistent with the hypothesis that TS is the result of multiple genetic or environmental factors, creating multiple phenocopies. The most useful application of unbiased, high-throughput technologies like gene expression profiling might be to identify profiles that distinguish molecular subtypes of complex disorders, from which the association of genomic pattern and various clinical features can be probed. Our findings relating to the functioning of immune cells is particularly intriguing, given studies¹⁹ that suggest that autoimmune mechanisms triggered by infection with group A β-hemolytic streptococci are involved in the pathogenesis of some patients with TS.

The results of these and future blood gene expression studies in neurologic diseases must be interpreted cautiously. Most genes regulated in blood have low fold changes and high variability. The low fold change is not unexpected considering the absence of obvious blood phenotypes. The variation in gene expression patterns in peripheral blood come from multiple non–disease-related sources. RNA quality is also critical.

Finally, the internal statistical validation techniques, for example, permutation analysis, used in this pilot study have limitations. Among genes that seem to be up- and down-regulated, there are disease-related and artifactual changes. When internal validation methods demonstrate that the number of abnormally expressed genes exceeds what would be expected by chance, this does not confirm that an individual gene’s expression level is disease related. Selectively using quantitative reverse transcription polymerase chain reaction (data not shown) confirms the gene expression levels measured by the array but does not discriminate between spurious statistical differences that result from multiple comparisons vs real differences associated with disease. The predictive value of any set of dysregulated genes ultimately requires external validation, for example, cosegregation of gene expression profiles with differential survival or treatment response, comorbidity, or inheritance patterns, or abnormal expression of genes in disease-causing pathways. A final test is to apply the identified genes as a prediction rule in a second cohort to determine whether the genomic profile accurately distinguishes diseased individuals from controls.

Correspondence: Donald L. Gilbert, MD, MS, Division of Neurology, Cincinnati Children's Hospital Medical Center, ML2015, 3333 Burnet Ave, Cincinnati, OH 45229-3039 (d.gilbert@cchmc.org).

Accepted for Publication: July 1, 2004.

Author Contributions:Study concept and design: Tang, Gilbert, Glauser, and Sharp. Acquisition of data: Tang, Gilbert, Glauser, and Hershey. Analysis and interpretation of data: Tang and Gilbert. Drafting of the manuscript: Tang, Gilbert, and Sharp. Critical revision of the manuscript for important intellectual content: Tang, Gilbert, Glauser, and Hershey. Statistical analysis: Tang. Obtained funding: Sharp. Administrative, technical, and material support: Gilbert, Glauser, and Hershey. Study supervision: Sharp.

Funding/Support: This study was supported by grants NS41920 (Dr Gilbert); NS28167, AG19561, NS38084, NS42774, and NS43252 and an American Heart Association Bugher Award (Dr Sharp); NS040261 and NS044956 (Dr Glauser); and NS045752 (Dr Hershey) from the National Institutes of Health, Bethesda, Md.

Previous Presentation: This study was presented in part at the 32nd Annual Child Neurology Society Meeting; October 2, 2003; Miami, Fla.