Daniel Harari 1, *, Irit Orr 2, Ron Rotkopf 2, Sergio E. Baranzini 3 and Gideon Schreiber 1, * Abstract. Introduction ORIGINAL ARTICLE

Transcription

1 Human Molecular Genetics, 2015, Vol. 24, No doi: /hmg/ddv071 Advance Access Publication Date: 26 February 2015 Original Article ORIGINAL ARTICLE A robust type I interferon gene signature from blood RNA defines quantitative but not qualitative differences between three major IFNβ drugs in the treatment of multiple sclerosis Daniel Harari 1, *, Irit Orr 2, Ron Rotkopf 2, Sergio E. Baranzini 3 and Gideon Schreiber 1, * 1 Department Biological Chemistry and 2 Bioinformatics and Biological Computing Unit, Department of Biological Services, The Weizmann Institute of Science, Rehovot 76100, Israel and 3 Department of Neurology, University of California at San Francisco, San Francisco, CA 94158, USA *To whom correspondence should be addressed. Tel: ; Fax: , daniel.harari@weizmann.ac.il (D.H.); Tel: ; Fax: , gideon.schreiber@weizmann.ac.il (G.S.) Abstract We analysed gene expression microarray data from whole blood samples from 228 multiple sclerosis (MS) patients either untreated or treated with one of three alternative commonly used interferon beta (IFNβ) disease modifying drugs: Avonex ( 1 weekly), Betaseron (every second day) or Rebif ( 3 weekly). Patient injections were not timed to coordinate sample collections, thus providing a global transcriptomic profile for each population of patients studied. Three hundred and fifty one genes were significantly differentially expressed by at least one of the IFNβ drugs. Despite the different drug sources with distinct injection and dosage protocols, a striking similarity was found in the identity and functional classes of the differentially expressed genes induced. Using the 25 most-upregulated genes, we defined a robust IFNβ gene expression signature that quantifies the IFN activation state per blood sample collected irrespective of the type of IFNβ therapy. This 25-gene signature also defined basal IFN activation states among untreated MS patients, which differed among individuals but remained relatively constant per patient with time. The maximum drug-induced IFN-activation state was similar for all three drugs despite a fold diminished average effect for Avonex. This and a more erratic effect of Avonex per patient across longitudinal measurements is likely a result of its reduced injection frequency. In summary, we have defined a robust blood-derived type I IFN gene signature from MS patients. This signature could potentially serve to generically quantify the systemic Type I IFN activation status for any other clinical manifestation, inclusive of other autoimmune diseases. Introduction IFNβ has been a preferred first-line therapeutic option for the treatment of the relapse remitting form of multiple sclerosis (MS) for over 20 years. Meta-analyses of clinical trials show that IFNβ therapy reduces the frequency of MS relapses and delays disease progression. The safety profile of IFNβ therapy for long-term use is good, which is an important consideration for treatment of this chronic disease (1 3). In patients with the milder relapse remitting forms of the disease in particular, IFNβ has proved to provide long-term benefits not only in terms of measurements in clinical response to disease but also in the overall survival rate (4). Therefore, despite the increase in the Received: December 14, Revised and Accepted: February 16, 2015 The Author Published by Oxford University Press. All rights reserved. For Permissions, please journals.permissions@oup.com 3192

2 Human Molecular Genetics, 2015, Vol. 24, No number of disease modifying drugs approved for the treatment of MS, especially due to its good safety record, IFNβ is continuing to be used as a first-line therapy and is anticipated to remain a common therapeutic choice for the foreseeable future (5,6). IFNβ is marketed by a number of pharmaceutical companies for clinical use. Each company produces its own version of the drug generated by alternative recombinant protein expression systems. Additionally each drug has its own recommended protocol of administered dosage, injection route and prescribed frequency of injection. The drug Betaseron (in some parts of the world marketed as Betaferon or Extavia ) is produced by bacterial fermentation, and has hence been given the name IFNβ-1b. To facilitate solubility of the non-glycosylated bacterially generated IFNβ-1b form, cysteine 17 was mutated to serine. IFNβ-1b is administered to patients by sub-cutaneous (SC) injection of a 250 μg dosage once every 2 days. Two other IFNβ drugs, Avonex and Rebif are generated by mammalian expression systems and are naturally glycosylated, in common with that of native IFNβ. Despite being produced by independent companies, they have both been given the name IFNβ-1a (3). Rebif is administered SC three times a week and can be injected using either lower (22 μg) or higher (44 μg) dosage options. Alternatively, Avonex is administered by single weekly intramuscular injection at 30 μg/dose. The different modes of IFNβ production, recommended drug dosage, routes and frequencies of drug injection, has led to some confusion as to how similar/dissimilar these different drug alternatives are to each other, and how they may alternatively benefit MS patients. Numerous clinical trials have assessed the therapeutic benefit of IFNβ in the treatment of MS, which has been extensively reviewed previously (3). Many such trials are single agent comparing the relative efficacy of a single IFNβ drug therapy against MS untreated controls. Such single agent trials are not extremely informative in providing quantitative cross-correlations for effectiveness between different IFNβ drugs, particularly due to considerable variation in baseline disease progression that often can be demonstrated to take place across trials. Rather, head-to-head clinical trials comparing the efficacy of one MS drug in relation to another within the same study is a more reliable approach to cross-compare effectiveness of different IFNβ drugs (3): The INCOMIN trial compared injections of Avonex (30 μg IFNβ-1A IM) to Betaseron (250 μg IFNβ-1b SC). It involved 188 patients and took place over a 2-year period. The respective annualized average relapse rate was found to be 0.7 versus 0.5 (P = 0.03), with the percentage of relapse-free patients being 36 and 51% (P = 0.03) for Avonex and Betaseron, respectively (7). The EVIDENCE trial included 677 patients examined over a 16-month period for patients receiving either Avonex (30 μg IFNβ-1A IM) or higher dose Rebif (44 μg; IFNβ-1A SC). Here Avonex was outperformed by Rebif, the latter with decreased measured annualized relapse rate (0.65 versus 0.54 P =0.033) and with a larger percentage of relapse-free patients (48 versus 56%, P = 0.023) (8). In a Danish trial with 301 MS patients comparing the lower dose Rebif (22 μg IFNβ-1A SC) to Betaseron (250 μg IFNβ-1b SC) therapy over a 2-year period, no statistical difference in treatment outcome between the two therapies was found (9). In yet another, but smaller study (n = 90, 2-year follow-up) comparing all three therapies (using 44 μg/dose Rebif), the relapsefree percentage was 20, 43 and 57% for Avonex, Betaseron and Rebif injections, respectively, albeit because of the smaller sample size, the differences found between Rebif and Betaseron were not deemed significant (P = 0.3)(10). Other measures comparing clinical disease outcome such as use of the Expanded Disability Status Scale (EDSS), requirement for steroid use, or more recently magnetic resonance imaging provide alternative methodologies to follow MS disease status and progression. As yet there remains a lack of consensus as to which of these measures best serves to quantitatively characterize this complex disease. Nevertheless a common interpretation of these clinical studies provide a general consensus indicating that Rebif and Betaseron provide a very similar degree of therapeutic efficacy in the treatment of MS, whereas the beneficial effects of Avonex are more modest. A new PEGylated IFNβ variant (Plegridy ) has recently been endorsed in the USA and in Europe for the treatment of MS. Protein PEGylation is a method that has been successfully used to increase the pharmacodynamic half-life of proteins in serum, providing patients with an opportunity to inject a drug at lower frequency (11). PEGylated forms of low affinity IFNα have been used in the clinic for a number of years to provide improved patient benefit for the treatment of Hepatitis C (12,13). Likewise, the PEGylated IFNβ variant serves the advantage of providing a sustained plasma drug half-life, where patients were only injected either once every second week or once monthly. Patients administered with Plegridy demonstrated improved clinical response over a 48-week period in relation to non-treated MS controls, exhibiting an adjusted annualised relapse rate of in the placebo group versus in the every 2 weeks group (P = ) and in the every 4 weeks group (P = c.f. placebo) (14). As this drug was not tested in comparison to a pre-existing IFNβ alternative, it remains to be established if and how well Plegridy provides improved clinical benefit to MS patients in relation to the established IFNβ drug counterparts (14). Another long-life IFN drug variant has been developed in our laboratory and was tested in a preclinical mouse model. We generated an engineered human IFN superagonist with biological activity superseding IFNβ and also with a 10-fold increased pharmacological lifespan in mice. This superagonist outperformed IFNβ in the ability to decrease clinical symptoms in a pre-clinical experimental autoimmune encephalomyelitis model using transgenic mice harbouring humanized IFN receptors. This took place even though we injected 1/16 the overall drug dosage and with a 4-fold decreased frequency of injections required (15,16). Another strategy to compare the relative effects of different IFNβ therapies in MS is to assess the transcriptional landscape of patients receiving treatment. By use of high-density geneexpression microarray technology, we now have the ability to assess the relative effects of these three major IFNβ drug alternatives at a transcriptome-wide level. Here, we describe the reanalysis of a pre-published transcriptomic profiling of whole blood collected IFN-treated MS patients compared with that of non-treated MS controls (17). Whereas in the initially reported study, the effects of the three IFNβ treatments Avonex, Betaseron and Rebif were pooled (17), here we dissect the differential transcriptomic effects between drug therapies. We found a quantitative difference in the degree of activation of the IFNresponse genes by these three different therapies, but with the same transcriptional program activated by all therapies. A considerable degree of variability in IFN activation status was found to take place between individual patients, which was particularly pronounced for Avonex-treated patients. In the course of this study, a robust IFN gene expression signature was uncovered in whole blood samples, which can be exploited to test IFN activity status per patient by analysis of their blood-derived RNA, irrespective of the IFNβ therapy administered. This gene signature was even of sufficiently high resolution to determine different levels of basal IFN activation states in non-treated MS patients.

3 3194 Human Molecular Genetics, 2015, Vol. 24, No. 11 Results Different drugs same genes different activation levels To test the effects of the different IFNβ therapies given to MS patients, we compared gene expression values for Avonex, Betaseron and (44 μg dosage) Rebif-treated groups over a 2-year longitudinal study, comparing changes in patient gene expression relative to non-treated MS patients. As per recommended protocol, Avonex, Betaseron and Rebif were prescribed to be injected once weekly, every second day or three times a week, respectively. For those that completed this study, three blood draws were taken; one at the beginning of this study (T =0),the second at T =12 and the final at T = 24 months. The time of blood draw for each patient relative to the time of last injection was not timed and thus considered as random. Patients had already commenced interferon therapy prior to their first blood draw thus T = 0 represents a bona fide response to IFN therapy and not as pre-injection baseline measurements. The data used in this study comprised 501 microarrays, each measuring an independent patient sample (Table 1). A majority of the data (327 arrays) describe the analysis of 109 patients receiving the same therapeutic option and hence with three sample measurements taken over the course of this 2-year study. The remainder of the expression data represent patients who either switched therapeutic option during the study, or for patients whose sample data were not provided for all of the three time-points tested (Table 1). Initial findings from this study have been published elsewhere (17) and focused on the pooled effects of the different IFNβ therapies, but not their differential effects. For clarity, unless otherwise indicated, the data shown here represents the global analysis of all 501 microarrays, inclusive of multiple measurements taken from the same patients over three different time-points. The number of statistically confirmed (P < 0.05) differentially expressed genes was very similar for Betaseron- and Rebif-treated patients but was substantially smaller for the Avonex-treated group (Fig. 1A). Using an arbitrary cutoff of ± 1.4-fold gene expression for example, 102 statistically verifiably differentially regulated genes were identifies for Avonex, 300 with Betaseron and 275 with Rebif. Genes were both up- and downregulated using all treatment options, although strong bias for up-regulation was evident (Fig. 1B). The 25 most-highly differentially expressed genes for Avonextreated patients in relation to untreated MS controls were compared with the other two IFNβ therapy groups to assess how the expression of these genes might change as a function of drug therapy choice. All these 25 genes were upregulated and thus can be considered as interferon stimulatory genes (ISGs). This same list also defined the 25 most-upregulated ISGs for the Betaseron and Rebif groups as well (Fig 2). Plotting the average fold-change of gene expression as a function of IFNβ drug type shows that the amplitude of average ISG activation is higher for the Betaseron- and Rebif-treated versus Avonex-treated patients (Fig 2A). From visual examination of this list of 25 ISGs, it was evident that per gene, the degree of upregulation for each gene was more or less conserved as a function of IFNβ therapy option. To examine this further, we compiled a list all the genes in this study with at least ± 1.4-fold change in gene expression in at least one of the IFNβ therapies (relative to non-treated MS patients). This amounted to 351 non-redundant genes (P < 0.05). The average changes in gene expressions were plotted as to compare the different drug treatments. This is represented from a panel of pairwise dot-plots, with each data-point demonstrating expression levels for a different gene as a function of drug treatment (Fig. 2B). A curve of best fit revealed an extremely close linear relationship between averaged expression of genes and choice of drug therapy (R 0.98). However, the slopes of the curves substantially differ, with Rebif versus Betaseron data producing a slope of 0.83 while Betaseron or Rebif versus Avonex producing a slope of 2. These results show that all three interferon therapies are activating the same genes, however, their degree of activation (averaged over all patients) is drug dependent. With the exception of a single outlier (IFI27), the behaviour of each gene is extremely predictable and at least when averaged for a number of patients taking the same therapy, the degree of differential activation of each gene is proportional to the drug therapy chosen (Fig 2B). Overall, the three treatments produce an extremely similar gene activation profile, but the average degree of activation by Betaseron and Rebif is about double that measured for Avonex. Different IFNβ therapies activate the same functional classes of genes The higher degree of gene activation by Betaseron and Rebif in relation to Avonex results in a higher number of genes significantly regulated (±1.4 or more) by the two more potent IFNβ therapies, which may result in the activation of different classes of genes. To test this formally, the independent gene lists generated from each IFNβ treatment were examined for enrichment to different functional classes of genes as defined by the Gene Ontology (GO biological process subset) database. All three therapies activate similar classes of genes, all of which are typical for functions characterized as Type I interferon response (Table 2). This includes the innate immune response, inflammatory response and response to virus. The number of genes found in each category (and hence the relative enrichment and ensuing adjusted P-values) shows a much higher significance for both Betaseron and Rebif treatment (which are almost identical) to that for Table 1. Patient numbers: years of consecutive IFNβ therapy Years of consecutive therapy Avonex Betaseron Rebif Untreated Patients Samples Once Twice Three times Sample totals MS patients in this longitudinal observational study received one of four therapeutic options: Avonex, Betaseron, Rebif or no treatment. Samples were collected at three different time-points: at the start of this study (T = 0), at 12 months (T = 12) or after 24 months (T = 24). Patients had already commenced their IFNβ therapies prior to sample collection, so T = 0 does not represent baseline pre-injection measurements for those taking an IFNβ drug. Blood was collected from patients in the mornings but was not coordinated with the time of last IFNβ injection. About a half of the patients in this study (109/228 = 48%) received the same therapeutic option without change over this observational time-course. The remaining patients changed their therapeutic choice over the course of this study.

4 Human Molecular Genetics, 2015, Vol. 24, No Figure 1. Gene expression profiles of MS patients on alternative IFNβ therapies. (A) Average number of significantly differentially expressed genes (absolute ± values, P < 0.05) for patients on one of three IFNβ therapies in relation to untreated MS controls. A total of 351 non-overlapping genes are differentially expressed by over 1.4- fold at least one of the three IFNβ treatments relative to untreated MS patients. (B) The same data are represented by volcano plot format. The relative distribution of upregulated and downregulated genes is presented. Genes with ± 1.4-fold change in expression were converted to log2 scale and are plotted in gray. Avonex treatment, consistent with the lower number of genes activated by Avonex. Supplementary Material, Table S1 lists additional GO terms found along with the genes identified in each enrichment. We next tested for the presence of enriched transcription binding motifs in the promoter regions of each of the IFN-regulated genes for each therapy examined. Typical modulators of type I IFN response were identified for all three treatments, including the interferon stimulatory response element (ISRE) as well as the interferon regulatory factors (IRFs) 1, 2, 7 and 8 (Table 2). Once again, similar amplitude of enrichment of transcription factor binding sites was observed for the Betaseronand Rebif-treated groups with a more modest enrichment found for Avonex-treated patients. These data confirm that the three differential drug therapies provide altered phenotypic response due to a difference in the quantity and degree of amplitude of differential expression of their activated genes, but not due to the activation of alternate signalling pathways. A tight cluster of co-expressing IFN response genes defines IFN activation status independent of IFNβ therapy option On the basis of our comparisons of averaged gene activation for different IFNβ therapies (Fig 2B), it seemed evident that there is a close inter-relationship of expression of different ISGs irrespective of treatment option. In order to assess this formerly, we analysed the relationship of co-expression of ISGs after pooling the whole 501-microarray dataset. This combines patients taking one of the alternative IFNβ therapy options and samples taken from untreated patients with MS (Table 1). In an initial assessment, we compared the pattern of gene expression for the prototypical IFN-response gene MX1, against two of the highest upregulated genes in this study, RSAD2 and IFI44L. Both of these genes demonstrated an extremely tight linear correlation of gene co-expression (Fig. 3A and B). As negative control, we compared co-expression of MX1 to the gene HPRT1, the latter commonly used as a reference control and which has been previously shown not to be perturbed by type I IFN signalling (16). As expected no correlation was found between MX1 and HPRT1 expression (Fig. 3C). Thus at least for the genes MX1, RSAD2 and IFI44L, a strong correlation of gene co-expression is evident, independent of IFN therapy option. To perform a more general analysis, we took the 25 most-upregulated ISGs, and tested for correlation of gene co-expression for all gene pair combination by Pearson s analysis. The findings are plotted as a matrix, with correlation values (R) shown (Fig. 3D). Most genes demonstrated a good fit of gene co-expression (R 0.8), and are consistent with the findings of the pairwise plots (Fig. 3A and B). A small cluster of genes exhibited an exceptional degree of co-expression, including, for example, the genes: RSAD2, OAS3, IFIT1, MX1 and CMPK2 (R 0.95). This extraordinary high level of gene co-expression confirms the high precision of measurements generated from this microarray data study. Interestingly, the same correlative pattern of gene co-expression

5 3196 Human Molecular Genetics, 2015, Vol. 24, No. 11 Figure 2. Fold-change gene expression induced by the different IFNβ therapies. (A) Average fold-change in gene expression for the 25 most-differentially expressed genes (all which are upregulated) as a function of type of IFNβ therapy. (B) Dot plot of the relative levels of the 351 differentially expressed genes (showing the averaged absolute values of the same genes to those genes described in Fig. 1B). Linear correlation fits are shown. The dotted line indicates the hypothetical plot gradient should two treatments give the same degree of gene activation. The high linear correlations (R) of >0.98 demonstrate that the same genes are upregulated proportionally by the different treatments, albeit the slopes that differ from 1 suggest different levels of induction. Table 2. Gene enrichment analyses: Avonex, Betaseron and Rebif induce the same subclasses of genes but to different quantitative extents (A) Gene ontology a (biological process category) Gene ontology ID. Avonex Betaseron Rebif Number of genes (Probability FDR adjusted) Immune response (10 24 ) 82 (10 31 ) 80 (10 32 ) Innate immune response (10 25 ) 61 (10 30 ) 59 (10 32 ) Innate system process (10 17 ) 97 (10 25 ) 88 (10 23 ) Regulation of innate immune response (10 7 ) 20 (10 8 ) 21 (10 9 ) Immune effector process (10 22 ) 46 (10 20 ) 46 (10 22 ) Response to stress (10 15 ) 106 (10 14 ) 96 (10 13 ) Response to virus (10 30 ) 44 (10 29 ) 43 (10 30 ) Negative regulation of viral replication (10 08 ) 14 (10 15 ) 13 (10 14 ) Cytokine-mediated signalling pathway (10 18 ) 35 (10 16 ) 33 (10 15 ) Response to type I interferon (10 21 ) 24 (10 22 ) 23 (10 22 ) Response to interferon gamma (10 07 ) 21 (10 15 ) 19 (10 13 ) (B) Transcription factor binding sites a Database ID. Avonex Betaseron Rebif Number of genes (Probability FDR adjusted) ISRE DB_ID: (10 13 ) 28 (10 22 ) 29 (10 25 ) IRF DB_ID: (10 14 ) 23 (10 17 ) 24 (10 19 ) IRF1 DB_ID: (10 2 ) 13 (10 6 ) 14 (10 8 ) IRF2 DB_ID: (10 2 ) 9 (10 5 ) 9 (10 6 ) IRF7 DB_ID: (10 09 ) 25 (10 19 ) 25 (10 20 ) IRF8 DB_ID: (10 08 ) 26 (10 20 ) 23 (10 18 ) a This analysis was performed using Web Geshtalt (2014 version). This server used data provided by: (A) Gene Ontology (version 1.2, 11/11/2012). (B) Regulatory modules [Motif gene sets: MsigDB (11/11/2012)]. b Probability values were adjusted for multiple tests using the Benjamini-Hochberg procedure.

6 Human Molecular Genetics, 2015, Vol. 24, No Figure 3. Interferon activation is robustand conserved independent of drug therapyoption. (A and B) Pairwise gene expression comparisons were generated fora sample of the highly upregulated interferon stimulated genes RSDA2 versus MX1 (A), IFI44L versus MX1 (B). (C) Plotting expression of MX1 against the reference gene HPRT1 as non-induced control. These plots were generated using all 501 samples from this study. This includes not only patients taking one of the three IFNβ therapies, but also non-treated MS patients (189 samples). Lines of best fit demonstrate the high robustness of gene co-expression for IFN-response genes. (D) Cross-comparison of gene co-expression for the 25 most-upregulated genes from the pool of 501 samples in this study (Pearson correlation). Correlation coefficients (R) approaching a hypothetical perfect score of 1.0 demonstrate a remarkable degree of similarity in co-expression for most of these genes, with IFI27 and CXCL10 being clear outliers. for these genes was also detected amongst MS patients who were not treated with an IFN therapy (Supplementary Material, Fig. S1), indicating that these genes can even be used to determine a range of basal Type I signaling states in blood samples from non-treated controls. One of the most highly upregulated genes in this study IFI27, has already been identified as a potential biomarker for MS by others (18 20). This gene stands rather unique in that it shares a weaker degree of co-expression with other IFN-response genes in this study (Fig. 1D) despite its strong upregulation by IFNβ. Another gene with relative weak coexpression in relation to other genes shown here is CXCL10 (Fig. 3D). Thus although clearly defined as ISGs, both IFI27 and CXCL10 exhibit exceptional properties altering their mrna expression patterns, possibly because they are under the control of unique transcriptional regulators or otherwise with alternative mrna stabilities.

7 3198 Human Molecular Genetics, 2015, Vol. 24, No. 11 Significant IFN signalling heterogeneity is observed between individual MS patients The data presented so far showed the averaged expression profile among a population of patients taking the same therapy, but provides no information regarding variation in ISG activation between individual patients or individual samples (same patient, different time points). To this end, we generated an interferon gene signature by determining the averaged expression of the 25 most-upregulated ISGs (as shown in Fig. 2A). This provides a quantitative measure which reflects the type I IFN activity status per sample tested. The distribution of IFN gene signature status normalized to the average signal of non-treated MS patients was then determined for the different populations of patients in this study either untreated or treated with Avonex, Betaseron or Rebif and is shown in Figure 4A. For the untreated MS group, gene expression was not elevated for most individuals. Nevertheless, a tail of higher IFN-activation states extending to 10-fold upregulated signal [IE: log(2) transformed = 3.32] relative to the untreated average was detected for a small proportion of untreated samples (Fig. 4A). This contrasts to both Betaseron and Rebif treatments where the IFN-activation state for the majority of patient samples peaks at 10-fold, and with a small tail of samples demonstrating low IFN activation states. For example, only about 15% of the Betaseron- and Rebif-treated samples demonstrated an IFN-activation state <3.0-fold. In contrast, samples from Avonex-treated patients demonstrated a much wider distribution in IFN-activation. Whereas a reduced proportion of samples shared approximately the same maximum activation state found for Betaseron and Rebif (IE: 10-fold), a relatively large tail of decreasing IFN-gene signature values was detected (Fig. 4A). Due to the large variation in IFN signature between patient samples, we next tested if this signature remained stable for the same patient over the time course of this study. In order to do this, we selected only the subset of patients who provided samples for gene expression data over all three longitudinal time-points (T = 0, 12, 24 months) of this study while receiving the same type of therapy. Data from individual untreated MS patients were also studied. This selection criterion resulted in narrowing the data to 327 arrays, IE: three samples taken from each of the 109 patients (Table 1). As the sample size of Betaseron- and Rebif-treated patients was considerably reduced when considering this limitation (11 and 14 patients respectively, versus 42 patients for the Avonex-treated group), and since Betaseron and Rebif were shown to provide similar transcriptional profiles (Figs 2 and 4A), we pooled their data. The alteration in gene expression at the level of individual patients was plotted in the form of a heat map displaying the log2 of average expression for the IFN signature genes (Fig. 4B) with each row representing a single patient. The untreated MS controls show only minor variations in gene levels between patients along the 2-year study. The occasional spikes in the IFN signature may arise from activation of Type I IFNs as a natural course of response to bacterial or viral infections (21). Interestingly, in three of the 42 untreated patients, a sustained elevation in IFN signature was observed, even though these patients did not receive IFNβ therapy. For the Avonex-treated group, the variation both between patients and between treatments was significantly larger (Fig. 4B), while for the Betaseron/Rebif group, the gene expression of the individual patients along the time axis clustered together. In order to quantify this variation, we determined both the average IFN-signature value per patient (IE: an average of three measurements/patient) as well as the standard deviations between the time-points (Fig. 4C). A wider distribution of both the average signal of individual patients as well as the standard deviation of the signal was evident for the Avonex-treated group [log2 (1.89 ± 0.94)], in comparison to the Betaseron/Rebif-treated group [log2 (2.32 ± 0.71)] and the untreated group [log2 (0 ± 0.24)]. As a control for data quality, we generated a signature of reference genes that are not IFN-activated. This signature comprised the average value of expression of four genes- Actb, Hprt1, Polr2a and RPS11, all which are used as reference probes to quantify qpcr studies. As can be seen in the lower panel of Figure 4C, the average expression of the reference gene signature does not alter as a consequence of treatment option. Furthermore, the standard deviation of the reference gene signals for the three independent samples (0, 12, 24 months) are low and tightly clustered, further supporting the high quality of the data. Thus, the variations seen in the IFN-treated groups are real and not due to experimental artefacts. We next tested if IFNβ-induced transcriptomic signal alters in patients over time. ANOVA over repeated measurements revealed that there was no significant change in amplitude of IFN gene signature for patients taking the same therapeutic option over the three different time-points of this study (Pr = 0.41). Thus for at least the limited duration of this observational study, we found no clear evidence of dampening of IFN-activation for the patients over time. Reduced averaged transcriptional activation of IFNresponse genes by Avonex is related to injection frequency While some of the Avonex patients exhibited high activation of IFN-induced genes (Fig. 4A), the average value for all Avonex patients is fold lower than that for Rebif and Betaseron, respectively (Fig 2B). Moreover, the variation in gene expression betweenyears0,1and2forthesamepatientismuchhigher for Avonex-treated patients than for the Betaseron/Rebif group (Figs 4B and C). Avonex is injected once weekly in comparison to Rebif and Betaseron, which are injected three times weekly or every second day, respectively. To model the effect of the different injection regimes on the pharmacodynamics of gene induction, we used previously published data. The serum half-life of the two glycosylated formulations of IFNβ (IE: Avonex and Rebif) were shown to be extremely similar after single-dose injection despite their different regimes of injection (22), allowing us to model their pharmacodynamics together. We next turned to a microarray gene expression study by Fernald et al.(23) measuring the transcriptomic response of two individuals receiving single-pulse injection of Avonex after which multiple blood samples were taken up to 156 h post drug injection for timed data collection. We re-analysed the Fernald microarray, in which expression of 20 of the 25 IFN-induced genes used in this current study were measured and averaged. Using this trimmed 20-gene IFN signature, we quantified the IFN mrna-activation status as a function of time after single dose IFN injection (gray line; Fig. 5A). This is compared with the serum levels of IFNβ after single-dose Avonex injection (black line, data extracted from) (22). For both sets of data, a double exponential curve was generated according to the Bateman function, i.e. one exponential representing increase of the signal during the distribution phase and the second exponential representing decay during elimination. From these curves, we extrapolate that the pharmacokinetic half-life for IFNβ is 15 h (IE: measuring serum IFNβ levels), compared with a more robust half-life of the IFN-induced RNA transcripts measuring 40 h. With these measured parameters now in hand, we

8 Human Molecular Genetics, 2015, Vol. 24, No Figure 4. Drug-dependent patient-to-patient heterogeneity in IFN gene expression. A sample-specific type I Interferon signature was generated by averaging the expression of the 25 most-upregulated genes and expressed this as fold-change in relation to the averaged 25-gene-expression values of untreated controls. (A) Plotting the population fraction of IFN-induced gene signal for all treated groups in this study. Data was taken from all 501 gene chip measurements in this study. Each plot was normalized to account for its population size for each therapeutic option. (B) Here the 25-gene IFN signature is plotted per-patient as a function of time in a heat map format. In this case, we have selected only data from patients who received the same therapy option for the duration of the study (228 samples from 109 patients, Table 1). Due to the lower number of patients receiving continued Betaseron (11 patients) or Rebif (14 patients), their data were pooled. (C) Average IFN gene activation per patient (averaged over all three time-points for 25 genes) and their corresponding standard deviations. As a control, a reference gene signature (the average expression values per sample for the genes Actb, Hprt1, Polr2a and RPS11) was generated per sample and their averaged expression values per patient and corresponding standard deviation are shown. then extrapolated a model to describe the activation state of IFNinduced genes for three times weekly rather than the prescribed once weekly dose (Fig. 5B, gray and black lines). While, to the best of our knowledge, there is no publicly available high throughput transcriptomic data following the behaviour of IFN-induced genes after three weekly injections, the results shown in Figure 5B

9 3200 Human Molecular Genetics, 2015, Vol. 24, No. 11 own study. This model provides supportive evidence to implicate that differential frequency of IFNβ injections is the major determinant that separates the transcriptomic drug activation status for Avonex to that of Betaseron or Rebif treatments. Highly variable clinical responses masks possible associations with ISG expression We next wished to assess if our gene expression data correlated with changes in MS clinical disease status. Both the expanded disease status scale (EDSS) and the multiple sclerosis severity score (MSSS) were available for all patients (25,26). Most of the patients entered the trial with moderate disease state (averaged EDSS and MSSS score of 1.6 and 2.3, respectively) with the Betaseron/Rebif-treated patients having a slightly higher baseline of disease severity than the non-treated or Avonex-treated MS groups, but with a large variation in disease severity (Table 3A). During the course of this 2-year study, however, the overall advance of clinical disease inclusive of non-treated controls was small (0.51 and 0.52 for EDSS and MSSS, respectively), yet the degree of variation within all treated population groups was substantially larger than the change in disease score (Table 3B). Although the averaged advance in clinical disease state was slightly lower for the Betaseron/Rebif-treated patients, no correlation between gene expression and disease status was found when testing the 25-gene IFN signature, nor for any other IFN-induced gene in this study. A graph showing the changes in MSSS in relation to the 25-gene IFN signature is shown (Fig. 6) reflecting the high degree of variation in the clinical data. This is also representative of what was found with comparisons of changes in EDSS gene expression (data not shown). Thus in contrast to the very stable IFN gene signatures that can be used to accurately quantify the IFN status per patient per longitudinal measurement (at least for Rebif and Betaseron treatments), the clinical data at least presented here is extremely noisy, and precludes proper statistical analysis to compare gene expression with clinical disease status. Discussion Figure 5. Half-life of IFNβ and IFNβ induced genes. (A) Experimental data for IFNβ serum concentrations (black line) were taken from Munafo et al. (22) withno significant difference between Avonex and Rebif found despite their different routes of injection. Gene induction data (gray line) were averaged from 20 of the 25 IFN signature genes taken from Fernald et al. (23). The data were fitted to a double exponential according to the Bateman function. The fitted pharmacokinetic half-life for IFNβ is 15 h, while for IFN-induced genes is 40 h. (B) The area under the curve (integral) of gene product for once weekly (black) and three times weekly injections (gray). The three times weekly injection data are extrapolated from the once weekly results, as no measured data are available. Overall, three times weekly injections result in about 2-fold increase in interferon induced gene products. are reminiscent to those presented by Williams and Witt (24) for MX2 protein levels comparing Avonex with Betaseron. The area under the curve demonstrates that once weekly injection provides on average an 2-fold decrease in activation of IFN-induced genes. Moreover, the simulation also explains the larger variation seen for Avonex in individual patient gene expression, as blood draws were more or less randomly timed in relation to the last injection. Thus, the simulation provides an explanation to the differences noted between Avonex and Rebif/Betaseron found in our We have cross-examined the transcriptome-wide effects of samples taken directly from MS patients actively prescribed with one of three major IFNβ treatments. In doing so, we have identified a robust gene signature that can quantify type I IFN activation on a per-patient basis regardless of the IFNβ treatment administered. This study included a relatively large cohort of MS patients, most who were followed with repeated measurements over a 2-year period. At the level of transcriptional profiling, two of the three IFNβ therapies [Betaseron (IFNb-1β) injected subcutaneously once every second day and Rebif (IFNβ-1a) injected SC, three times a week] were both found to activate a similar profile of genes to a similar amplitude and for a similar proportion of patients. This comparable transcriptomic signature of IFNβ response between these two drugs was found despite Betaseron and Rebif being generated as recombinant proteins from different sources (bacterial versus mammalian) and Betaseron having a slightly modified protein sequence to facilitate recombinant expression in bacteria (3). Avonex (IFNβ-1a) injected intramuscularly once weekly is activating the same genes but on average more weakly than the other two drugs. Furthermore, the effects of Betaseron/Rebif treatment per patient remained relatively constant over the three measurements over a period of 2-years, whereas the IFN activation state induced by Avonex per patient

10 Human Molecular Genetics, 2015, Vol. 24, No Table 3. Clinical severity (A) and delta clinical severity (B) of MS patients in this study (average value ± SEM in relation to year 0) (A) EDSS average ± SEM Untreated Avonex Betaseron /Rebif MSSS average ± SEM Untreated Avonex Betaseron /Rebif Year ± ± ± 2.0 Year ± ± ± 2.8 Year ± ± ± 1.9 Year ± ± ± 2.8 Year ± ± ± 1.8 Year ± ± ± 2.4 (B) ΔEDSS average ± SEM Year ± ± ± 1.1 Year ± ± ± 2.2 Year ± ± ± 0.73 Year ± ± ± 1.56 Figure 6. Highly variable clinical score measurements, coupled with a very modest progression in overall disease severity masks possible associations between expression of IFN-response genes and clinical responsiveness (MSSS). was more erratic. Our study suggests that the smaller variance in transcriptional response of Betaseron and Rebif in comparison to Avonex therapy is related to the different timing of administration of these drugs. The pharmacodynamic effect of a drug is dependent on its pharmacological half-life. Single injections of Avonex demonstrated a pharmacological activity that returned to baseline after 3 4 days of injection (16,23,24,27,28) and a pharmacodynamic transcriptomic half-life of 40 h (Fig. 5). Using our simulated data, it is indeed expected that patients receiving Avonex will show a lower ( 2-fold) average activation of IFN-response genes compared with that of Betaseron and Rebif treatment with larger variations between individual samples, as shown in Figure 4A. This would further explain the erratic change in IFN-signature genes measured for different blood draws for Avonex patients (Fig. 4A and B), as the timing of last IFN injection prior to blood collection was not coordinated. Transcriptome-wide profiling of one IFNβ treatment against another for MS patients has been reported in other studies, but has been limited to ex-vivo examination of drug effects (20,29,30), or having a much smaller number of patients or genes sampled (19,31 34). In the original publication studying this MS patient group, the different IFNβ therapies were pooled and considered as a single therapy. Using the pooled data, 265 transcripts were found to be differentially expressed between MS controls and IFNβ-treated MS patients (17), in comparison to the 351 genes found in this study where we stratified the therapy groups. Once stratified, we noticed that on average, the same genes are being activated to different degrees by different IFN therapies, with a very tight correlation found between the different IFN-induced genes (Fig. 2B). This suggested that perhaps all the IFN-response genes are being activated concordantly, but the degree of average activation relates to different injection regimes of the drugs. This was confirmed by combining all the patient data inclusive of all IFNβ therapygroupsaswellasnontreated MS controls where an extremely tight co-correlation of expression of IFN-response genes was uncovered (Fig. 3). That such strong correlations of ISG co-expression (e.g. IFIT1 and OAS1 with a Pearson s correlation of 0.95) was uncovered was a great surprise for us, considering that the dataset used here is from a diverse population of MS patients receiving differential IFNβ therapies. The robustness of this signalling module has allowed us to generate a reproducible 25-gene-signature that defines type I IFN activation state from blood RNA samples. For all therapies, a wide range of IFN-activation states was noted. Even in untreated MS patients, some variation in gene activation was observed between patients, which tended to remain constant for the same patient over the course of this study (Fig. 4). IFNs are naturally expressed in healthy mammals, albeit at low basal levels, which in turn results in basal activation states of IFN-signalling (reviewed in 35). Differential basal IFN activation states among MS patients and healthy individuals have been previously reported from studying different tissues from MS patients but where baseline measurements were taken before patients began to receive IFNβ therapy. In these studies, a common theme has emerged that high basal IFN-signalling state correlates with reduced or no activation of IFN signalling after commencement of IFN therapy (19,36 38). For at least one of these studies, an association between upregulated basal IFNβ signalling to failed responsiveness to IFNβ therapy was found (36). Furthermore there have been reports that attenuated basal type-i interferon signalling may in itself characterize the MS phenotype, perhaps more evident with patients with the chronic form of MS (39 41). If substantiated, these findings would indicate decreased basal type I IFN signalling in MS may be a component of the disease state in itself, and if so, IFNβ therapy would provide a means to correct this deficit. What might be the causation of this high variation in baseline IFN activation status for some individuals is currently not known. An interplay between certain commensal bacterial populations and their ability to trigger interferon signalling in the host organisms in which they reside has recently been reported (42 44) and just might hold a key to this intriguing phenomenon. From another perspective, it is

11 3202 Human Molecular Genetics, 2015, Vol. 24, No. 11 perhaps worth considering that Type I IFN signalling is subject to stringent down-regulation mechanisms (45 49), as is typical for other homeostatic biological systems. Such mediators of Type I IFN down-regulation can in some circumstances be transactivated by alternative signalling pathways such as that by inflammatory cytokines (50,51). The understanding of how this IFN-regulatory machinery might relate an individual MS patient s response to IFNβ therapy, to our opinion, is a subject worthy of close investigation. This observational clinical study did not provide transcriptomic data regarding the IFN activation state before commencement of IFNβ therapy. Nor did it address an important issue of the development of drug-neutralizing antibodies for a proportion of MS patients as a consequence of prolonged therapy. The presence of such antibodies can severely curtail the activation of IFN-response genes and hence the therapeutic potential of IFNβ therapy in the treatment of MS, this has been extensively studied for the different IFNβ drugs elsewhere (52 54). Although other studies have provided recommended approaches to measure IFN activity status in MS patients (e.g. by following expression levels of the prototypical type-i IFN response gene MX1), we suggest that averaging the robust 25-gene panel described in this study may provide a more accurate means to measure IFN activity levels before and during the time-course of long-term therapy for individual patients, and can help medical practitioners in making informed decisions if a patient is responding to therapy by the activation of IFN-response genes. Indeed, our study shows that even for the Betaseron/Rebif treated group, in 15 20% of patients IFN gene upregulation is <3-fold in all measurements along the time axis, which might be due to development of neutralizing antibodies. Interestingly, no systematic decrease in interferon responsiveness was observed over time, suggesting that patients who tolerate interferon without an immunological reaction do so for extended periods. Do we really need all 25 genes to generate a robust signature to measure IFN-activation from blood RNA samples? In this study, we have derived average gene expression from extremely high quality Affymetrix exon-microarrays. That multiple exons were typically measured per gene before deriving an average expression value means that in practice expression levels for most genes was derived from a large number of measurements. This we assume has contributed to the robustness of our data and has allowed us to derive extremely concordant gene co-expression for 25 of the highest expressed genes derived from this study (Fig. 3 and Supplementary Material, Fig. S1). The very high Pearson correlation for these 25 genes indeed indicates that theoretically, only a subset of these genes might be required to generate a robust IFN-gene signature from blood mrna, with the required prerequisite that the data are also of high quality. Often, however, expression data is of much lower quality in some cases, and in other cases, such as the emerging technology of high throughput RNA sequencing but with a limited number of sequence reads, having a larger list of genes in which to derive an IFN-signature has its distinct advantage. In our study, the low progression of clinical disease response markers (EDSS and MSSS) over this 2-year study and the very high noise in these data precluded our ability to accurately assess if and how IFNβ gene expression markers correlate with protection from disease progression. Longer-term studies may resolve this issue and provide biomarkers to predict which patients gain from IFNβ treatment. This is of particular interest as we found that the bulk of IFNβ-induced genes behave similarly, except for IFI27 and CXCL10 (Fig. 3) and thus it would be interesting to see which of those is a better predictor. Nevertheless, the data here suggest that RNA profiling alone can be used to test if a patient is responding to therapy, to what degree, and if this activation is maintained over time. This would provide medical professionals with valuable information to help assess if the injected IFNβ is bioactive in the patient. The robust gene signature defined here can also be useful to quantify IFN activation status when using whole blood/pbmcs for other disease indications besides that of MS. Type I IFNs play a major role particularly in the innate immune response but are also mis-regulated in a number of autoimmune diseases (55). We caution, however, that definition of which genes are activated by interferons can be very much tissue specific (16) and thus for alternative cell types (e.g. purified monocytes), a different set of genes may better define type I IFN signalling in the context of the cells being studied (36). Furthermore, if the quest of a clinical investigation is to relate gene expression to clinical response, then the cell type most relevant to the disease condition in question may provide the richest transcriptomic profiles relating to the disease state being studied. For example, we have previously shown that profiling CD4 regulatory T cells specifically provided important information concerning the relationship between CD274 (PD-L1) and disease progression in an IFN-treated mouse model of MS (16). The same gene was suggested in another study to be overexpressed in a novel regulatory T-cell population, playing a key role in IFN-treatment for MS (56). But for another disease condition such as chronic tuberculosis infection where type I IFN signalling plays an important clinical role, the IFN-signature found relates to infected macrophages (57). In this context, a more targeted approach would be beneficial. In summary, in this study, we have demonstrated from transcriptomic profiling that the biopharmacological effects of Betaseron and Rebif therapy are on average very similar. The effects of Avonex are on average more reduced and more erratic, apparently due to the random regime of blood draws relative to last IFN injection (with Avonex being injected only once weekly). We have identified a robust 25-gene IFN gene signature which can provide accurate quantification of IFN signalling status from whole blood RNA for all IFNβ treatments and is furthermore sufficiently sensitive to determine different basal IFN activation states among non-ifn injected individuals. This gene signature has the potential to be developed into a clinical standard to help tailor a personalized medicinal approach to assess baseline systemic type I IFN activation status in MS as well as for other clinical indications where a functional role of Type I IFN signalling has been implicated. Materials and Methods Data analysis The data used here were generated for a MS transcriptomic study conducted by the UCSF School of Medicine measuring gene expression from cdna from whole blood samples from MS patients that were either untreated or treated with one of three IFNβ drugs, Avonex,Betaseron or Rebif,theprimary data which has already been published (17). This was an observational study and not blinded nor randomized. Patients were treated by the recommended prescribed dosages and injection protocol for each respective therapy. A higher dose of Rebif (44 μg) was used in this study. All blood draws were performed in the mornings, but without coordination or timing of the last IFNβ administration for each patient. RNA was extracted for microarray analysis using Affymetrix Human Exon 1.0 ST Arrays