Genome-Wide Pharmacogenomic Analysis of the Response to Interferon Beta Therapy in Multiple Sclerosis FREE

Multiple sclerosis (MS) is a leading cause of neurologic disability in young adults.¹ No curative therapy is available, and approximately 80% of individuals with MS are ultimately disabled. Recombinant interferon beta, a small protein with antiviral, antiproliferative, antiadhesion, and proapoptotic activity, is widely used as treatment to reduce clinical activity and possibly slow disease progression.^2- 4Despite interferon beta therapy, up to 50% of patients with MS continue to experience relapses and worsening disability. In addition, adverse effects, such as flulike symptoms and depression, are common, leading many patients to discontinue therapy.

The mechanism of action of interferon beta is incompletely understood, and there are no reliable clinical or biological markers that accurately forecast response to therapy. In this setting of variable responsiveness and clinical heterogeneity, pharmacogenomic research could uncover unexpected mechanistic processes and potentially achieve the elusive goal of personalized medicine in MS. So far, a few candidate-based studies have investigated germline variation in genes hypothesized to influence interferon beta therapy response, but findings await replication.^5- 9 Since MS is a complex genetic trait and interferon beta, a pleiotropic agent, it is likely that allelic variation at multiple genes contributes to the overall pharmacogenomic response. In contrast to candidate-based methods, the genome-wide pharmacogenomic approach described herein allows for the unbiased detection of DNA variants associated with interferon beta therapy response.

The screening group consisted of 206 subjects selected from a well-characterized and homogeneous cohort followed up at 4 collaborating centers in the Mediterranean Basin: Hospital Vall d’Hebron, Barcelona, Spain; University of Navarra, Pamplona, Spain; MS Centre, Toulouse, France; and Hospital Regional Universitario Carlos Haya, Malaga, Spain. We focused on southern Europeans to minimize differences in population structure.¹⁰ Patients were followed up prospectively for 2 or more years after the initiation of interferon beta therapy. Disability data were collected at 3-month intervals by neurologists experienced in Expanded Disability Status Score (EDSS) scoring, with low interrater and intrarater variability. Relapses were defined as a new symptom or worsening of a preexisting symptom attributable to MS activity, confirmed by examination within 3 days of onset. After the validation stage was complete, we identified 81 additional samples from the Pamplona, Barcelona, and Malaga centers that met our strict criteria for response. These additional subjects were added to the original study population to be included in the joint analysis.

Inclusion criteria were (1) clinically definite relapsing-remitting MS¹¹ treated with interferon beta (Betaseron [Bayer HealthCare Pharmaceuticals, Wayne, New Jersey], Avonex [Biogen Idec, Cambridge, Massachusetts], or Rebif [Pfizer, New York, NY]) for at least 2 years, (2) at least 2 documented relapses over the 2 years previous to treatment onset, and (3) 2 years of follow-up clinical data. We focused on extreme clinical phenotypes to maximize the ability to detect differences. Responders had no relapses and no increase in EDSS over the 2-year follow-up period; nonresponders had at least 2 relapses or an increase in EDSS of at least 1 point. The EDSS was required to have been stable over at least 3 consecutive visits.¹² Changes of 0.5 point were not considered significant. Magnetic resonance imaging was performed at the time of diagnosis but was not used to monitor treatment. Experimental protocols were approved by the committees on human research at each institution, and informed consent was obtained from all participants.

We used pooled DNA on single-nucleotide polymorphism (SNP) microarrays to assess population allele frequencies.^13- 14 DNA was extracted and samples were quantitated in duplicate at the University of California, San Francisco, using the PicoGreen dsDNA quantitation reagent (Molecular Probes, Inc, Eugene, Oregon). Samples were diluted to 20 ng/μL, requantitated, and pooled into groups of 20 subjects. Responders and nonresponders were always pooled separately. Each sample was included in 3 different pools on separate Affymetrix GeneChip 100K arrays (Affymetrix, Santa Clara, California) to provide technical replicates. Thirty-six microarrays were used for SNP detection using manufacturer-recommended procedures to process each chip (Oklahoma research facility). The Oklahoma research facility in-house proprietary software and algorithms were used to generate quantitative estimates of each pool's allele frequency from the raw data, and differential hybridization was corrected using individual allele profiles.

Promising candidate SNPs were selected on the basis of significance-based ranking or a clustering algorithm that considered genomic distance between SNPs. The clustering method assumes that significant SNPs are likely to segregate together in a haplotype. To find clusters of significant SNPs along each chromosome, we used a custom-designed algorithm to group SNPs with a Z²P value <.05 in each replicate. Single-nucleotide polymorphisms were considered part of a common cluster if they were positioned within 30 kilobases of each other. To be significant, a SNP was required to be a member of a cluster of 4 or more SNPs in all 3 replicates.

The top 35 candidate SNPs selected using the ranking methods were individually genotyped in each DNA sample from the original 206-subject data set using TaqMan assays (Applied Biosystems, Foster City, California). The DNA of 2 subjects was unavailable for individual genotyping. As an additional confirmatory measure, 5 nonsignificant SNPs were genotyped.

We used t tests to compare age at disease and treatment onset between responders and nonresponders and to compare origin, treatment duration, and number of relapses in the 2 years prior to treatment. Differences in allele frequency were tested using Z² binomial proportion tests. The Fisher exact test was used to compare genotype frequencies. Results were not adjusted for multiple comparisons. Logistic regression was used to determine odds ratios for response; analyses were adjusted for baseline EDSS and number of relapses prior to study entry. Statistical tests were completed using Intercooled Stata 7.0 (StataCorp, College Station, Texas). The MSSS test program was used for calculation of the Multiple Sclerosis Severity Score (http://www-gene.cimr.cam.ac.uk/MSgenetics).¹⁵

We used the Web-based GoStat program to assess which categories of SNPs were overrepresented or underrepresented in our list of pharmacogenomic candidates (http://gostat.wehi.edu.au).¹⁶ The GoStat program annotates genes with gene ontology terms; each gene can be associated with 1 or more gene ontology terms. We compared (1) genes associated with the top significant SNPs (average P < .05 over the 3 replicates) with (2) all genes detected by the microarray. Significance for the gene ontology analysis was set at .05; the Benjamini-Hochberg method was used to correct for multiple comparisons.¹⁷

This project was completed in 4 major stages: stage 1, DNA pooling on SNP microarrays for population allele frequencies (n = 206 individuals); stage 2, ranking of top candidate SNPs; stage 3, validation of association through individual genotyping of candidate SNPs (n = 204 individuals); and stage 4, joint analysis using new and original subjects (n = 285 individuals).

The genome-wide screening included 99 patients who met the criteria for positive response to interferon beta treatment and 107, for nonresponse. Responders and nonresponders were very similar in all respects, except with regard to baseline disability and relapses prior to onset of therapy (Table 1). On average, nonresponders had a baseline EDSS 0.4 point greater than responders, a pattern that has been observed previously.¹⁸ Responders also had a lower Multiple Sclerosis Severity Score than nonresponders, which adjusts the EDSS for duration (P = .01). Overall, there was a high degree of correlation between the 3 biological replicates (eTable 1).

Gene ontology analysis from the microarray data using GoStat showed that top-ranked SNPs that differed between responders and nonresponders were more likely to be related to ion channels and signal transduction pathways (Table 2). For example, γ-aminobutyric acid and glutamate receptor genes were overrepresented in the group of significant genes; 42.1% and 53.8% of the microarray's SNPs in this category were in the most significant subset, compared with their expected prevalence of 13.5%.

We used 3 different methods to rank top candidate SNPs:

In all SNPs selected for follow-up with individual genotyping, minor allele frequency differences between responders and nonresponders were significant (Z² binomial proportion tests, P < .05, data not shown). The average minor allele frequencies estimated using the triplicate microarrays were similar to the true frequencies obtained by individual genotyping (mean [SD] difference, 0.034 [0.03]). eTable 2 lists the frequencies from DNA pooling on SNP microarrays and individual genotyping. In addition, there was little difference in the performance of each of the biological replicates (vs average triplicate data) in estimating absolute allele frequencies (range of mean [SD] absolute difference was 0.034 [0.029] to 0.039 [0.036]). The absolute estimates generated with DNA pooling on SNP microarrays were robust enough that we did not detect a considerable improvement in the estimation of allele frequency differences over absolute frequencies.

Using individually genotyped data, we were able to determine genotypes for each of the subjects. Genotype differences between responders and nonresponders to interferon beta therapy using individual genotyping were found in 29 of the 35 candidate SNPs (Table 3). We also genotyped 5 SNPs that were not significant in the pooled microarray screening data; these controls remained nonsignificant in the validation stage. Altogether, population allele frequency differences detected on SNP microarrays accurately represented differences between responder and nonresponder subjects.

At this stage, we added an additional 81 subjects who were not available for the original screening to the analysis, recruited to the study using identical stringent inclusion criteria at the Barcelona, Pamplona, and Malaga centers (stage 1). This was possible because, overall, the clinical characteristics and genotypes of subjects were very similar between centers. Furthermore, southern Europeans from Spain, Italy, and southern France can be considered 1 genetic population.¹⁰ With the original group of 206, we had 44% power to detect an odds ratio effect size of 2 when the minor allele frequency was 0.1, whereas the power was 60% in the joint population. With larger minor allele frequencies, for example 0.2 and 0.3, the power increased to 85% and 99% in the original population and 95% and 99% in the joint population.

In the joint analysis,¹⁹ more than half of the SNPs remained significantly different between responders and nonresponders, and the significance of the effect increased in 5 SNPs (Table 3). Odds ratios for good interferon beta therapy response were calculated, comparing heterozygotes (Aa) or homozygote minors (aa) with homozygotes for the major allele (AA). The adjusted odds ratios were similar in effect size to unadjusted odds (data not shown). After adjustment for the baseline differences in relapse rate and baseline EDSS, candidate pharmacogenomic SNPs remained significant.

Two-thirds of the joint analysis group had information available on type of interferon beta therapy used. Among these subjects, there was no difference between responders and nonresponders in type of therapy (Avonex, Betaseron, or Rebif; P = .35).

We probed our screening data set to determine whether previously reported interferon beta candidate genes were significant in our population. We considered SNPs near or within 112 candidate genes investigated by others.^5- 9 We searched for differences in SNP allele frequencies within the (1) 112 candidate genes (intragenic) and (2) 75 kilobase pairs to each side of their transcriptional start site (neighboring area), which captured SNPs in haplotype and promoter regions of genes. For 100 genes, we detected both intragenic (n = 35) and neighboring (n = 65) SNPs; other genes had no SNP representation on the microarray. We found 15 significantly different SNPs within or near 11 genes (Table 4). Five interferon-receptor SNPs were detected (1 IFNAR1 and 4 IFNAR2 SNPs), none of which distinguished between responders and nonresponders. The IFNAR1 SNP rs1041429 has been studied previously; consistent with other studies, there is no strong pharmacogenomic relationship.^5,8 We did not detect differences in the frequency of SNPs in or near LMP7, CTSS, or MxA. Our microarray did not contain the same promoter SNPs studied by Cunningham and colleagues⁷ and Nicolae and colleagues²⁰ in the interferon-stimulated response elements of LMP7, CTSS, or MxA, so a direct comparison was not possible.

We completed a nonbiased genome-wide pharmacogenomic screen of interferon beta therapy response in relapsing-remitting MS, using a DNA pooling strategy on microarrays. This method was shown to be reliable, with outstanding correlation in frequency between our 3 replicates. In addition, allele frequencies were similar to the true population allele frequencies obtained using individual genotyping. Of 35 SNPs selected in stage 2, 18 SNPs maintained significance in a follow-up analysis with additional subjects: 10 of 15 of those significance ranked, 5 of 10 of those cluster ranked with a P < .05 cutoff, and 3 of 10 of those cluster ranked with a P < .01 threshold. Single-nucleotide polymorphisms selected using the significance-based ranking system were the most likely to replicate, probably because they were associated with the most extreme allele frequency differences. The magnitude and direction of the estimated effects are reflected in the odds ratios in Table 3. Their persistent significance after adjustment for baseline disability and relapse rate increases confidence that the findings reflect true differences between responders and nonresponders. Given how few microarrays are required for DNA pooling studies, future pooled SNP microarray pharmacogenomic and association studies completed in triplicate could result in cost savings compared with individual microarray genotyping.

Candidate SNPs that significantly differed between responders and nonresponders in the final joint analysis included 7 located within genes: glypican 5, collagen type XXV α1, hyaluronan proteoglycan link protein, calpastatin, TAFA1 (chemokinelike), neuronal PAS domain protein 3, and LOC442331 (similar to dynein). The remaining SNPs are located in intergenic regions.²¹ These may represent long linkage disequilibrium with annotated genes or polymorphisms in distant cis-regulatory regions. The results reflect the pleiotropic action of interferon beta and complex nature of MS. Results of the gene ontology classification—suggesting an enrichment of glutamate and γ-aminobutyric acid receptors in pharmacogenomic candidate genes—are provocative, implying an interaction between neuronal excitation and interferon beta therapy effect. This potential connection needs further exploration.

Glypican 5 polymorphisms arose several times in our candidate SNP lists. Glypicans are 1 of 2 main classes of heparan sulfate proteoglycans. They are implicated in synapse formation and axon regeneration and guidance and are found in dense networks in active MS plaques, where they may be involved in sequestering proinflammatory chemokines.^22- 23 In the peripheral nervous system, glypican 1 is required for Schwann cell myelination; glypican 5 is highly expressed in neurons.²⁴ Interferon beta may affect the expression of glypicans, as occurs with interferon gamma and glypican,²⁵ and the interaction of glypicans with growth factors, chemokines, and extracellular matrix proteins, including those detected in our pharmacogenomic analysis, may affect neuronal growth and repair.

Polymorphisms in extracellular matrix proteins, such as hyaluronan proteoglycans and collagen, were also significant in this study. These polymorphisms may affect the binding of matrix metalloproteinases, which are released by leukocytes and aid in their migration through basement membranes. In vitro, interferon beta inhibits production and secretion of these metalloproteinases,^26- 27 and polymorphisms in extracellular matrix proteins could further alter the efficacy of interferon beta therapy.

We also interrogated the microarrays for SNPs in genes previously considered pharmacogenomic candidates. Consistent with other studies, we did not find a pharmacogenomic relationship between the SNPs we detected in or near interferon receptors IFNAR1 and IFNAR2. Single-nucleotide polymorphisms in or near the interferon-stimulated genes LMP7, CTSS, or MxA were not significant in this analysis. We were, however, limited to the SNPs detected by the Affymetrix 100K microarray, which represents a fraction of common polymorphisms in the genome and oversamples intergenic areas.²⁰

We did not have the power to detect effects in rare alleles. Odds ratios for effects compare AA individuals with Aa and aa individuals, which may not reflect the true pattern of effect. While results are based on triplicate microarray data to reduce spurious findings, some of our findings are likely false positives, given the lack of stringent correction for multiple comparisons in the individual genotyping stage. In an attempt to address potential confounding by baseline differences between responders and nonresponders, we adjusted odds ratios for baseline EDSS and pretreatment relapse rate. Without a placebo group, however, we cannot definitively distinguish between the natural history of MS and response to treatment.

Neutralizing antibodies (Nabs) may confound the relationship between SNPs and pharmacogenomic response, since Nabs are associated with interferon beta and persistently high titers of Nabs may be associated with decreased effectiveness of interferon beta therapy.²⁸ Some of the significant SNPs identified herein may be related to Nab status, something we could not examine because Nabs were not measured in all subjects. Finding SNPs related to Nabs would be of great interest, nevertheless, to better understand the mechanism associated with Nab production.

The beneficial outcomes of interferon beta therapy for patients in the relapsing-remitting phase of MS have been clearly shown. On the other hand, the effect of this treatment is partial, and a substantial amount of patients are not responders. Hence, in the absence of prognostic clinical, neuroradiological, and/or immunological markers of response, the question remains who and when to treat when adverse effects, inconvenience, and cost of the drug are significant. The identification of pharmacogenetic polymorphisms provides important new insights into the mechanism of interferon beta action, bringing the paradigms of rational drug design and personalized medicine one step further. These results, however, require replication in a larger, prospective data set and confirmation in functional assays to directly assess the relationship between genotypes, control of cell division, and response to immunotherapy. Because of the heterogeneous and multifactorial nature of MS and the complex role of interferon beta in the immune response, multianalytical approaches that incorporate and integrate genomic, laboratory, and clinical data will be necessary to predict therapeutic outcomes based on molecular evidence.

Correspondence: Jorge R. Oksenberg, PhD, Department of Neurology, University of California, San Francisco, 513 Parnassus Ave, Room S-256, San Francisco, CA 94143-0435 (jorge.oksenberg@ucsf.edu).

Accepted for Publication: October 26, 2007.

Published Online: January 14, 2008 (doi:10.1001/archneurol.2008.47).

Author Contributions:Study concept and design: Byun, Montalban, Villoslada, Barcellos, Baranzini, and Oksenberg. Acquisition of data: Byun, Caillier, Villoslada, and Fernández. Analysis and interpretation of data: Byun, Montalban, Fernández, Brassat, Comabella, Wang, Barcellos, Baranzini, and Oksenberg. Drafting of the manuscript: Byun and Oksenberg. Critical revision of the manuscript for important intellectual content: Byun, Caillier, Montalban, Villoslada, Fernández, Brassat, Comabella, Wang, Barcellos, Baranzini, and Oksenberg. Statistical analysis: Byun, Villoslada, Wang, Barcellos, Baranzini, and Oksenberg. Obtained funding: Fernández and Oksenberg. Administrative, technical, and material support: Caillier, Montalban, Fernández, Brassat, and Oksenberg. Study supervision: Montalban, Fernández, and Oksenberg.

Financial Disclosure: Dr Fernández has performed clinical trials with the following promoters: Bayer Schering Pharma, Merck Serono, Biogen Elan, Sanofi-Aventis, and Teva Pharmaceuticals.

Funding/Support: This work was funded by National Institutes of Health grant 1RO1 AI42911. Dr Byun was supported in part by the University of California, San Francisco, School of Medicine Genentech Foundation Fellowship.

Additional Contributions: Ru-Fang Yeh, PhD, provided invaluable advice on bioinformatics and Sonia Hurani, BSc, interrogated the data set. We are grateful to the patients with MS who participated in this study.