Genome Scan Meta-Analysis of Schizophrenia and Bipolar Disorder, Part II: Schizophrenia

Introduction

Schizophrenia (SCZ; locus SCZD [MIM #181500]) is a common disorder with a heritability of 0.70–0.85 and a 10-fold increase in risk to siblings of probands (Levinson and Mowry 2000). A large number of genome scan projects have been completed (see below). Genomewide significant evidence for linkage has been reported in three chromosomal regions, each with some degree of support in other data sets. In a study of 265 Irish pedigrees, Straub et al. (1995) reported a LOD score of 3.51 on chromosome 6p24-22, a region where evidence for linkage was also reported by Moises et al. (1995), Maziade et al. (2001), Lindholm et al. (2001), Schwab et al. (2000), and a multicenter analysis (Schizophrenia Linkage Collaborative Group for Chromosomes 3, 6 and 8 [hereafter referred to as “SLCG”] 1996). Straub et al. (2002a) recently presented evidence supporting dysbindin (DTNBP1; 6p22.3) as an SCZ candidate gene in this region. In 54 European-ancestry pedigrees, Blouin et al. (1998) reported a nonparametric linkage (NPL) score (Z_all) of 4.18 (P<.00002) on chromosome 13q32, as well as positive evidence for linkage in 51 additional pedigrees, with a second report of significant evidence for linkage (Z_max=4.42) in 22 Canadian pedigrees (Brzustowicz et al. 1999). In the same Canadian data set, Brzustowicz et al. (2000) reported a Z_max of 6.50 on chromosome 1q21-22 (D1S1679), and Gurling et al. (2001) reported supportive evidence in the region in 13 U.K. and Icelandic pedigrees. In two other regions, several groups have produced highly suggestive evidence for linkage. On chromosome 8p21-22, Blouin at al. (1998) reported a Z_all of 3.64, with support from Brzustowicz et al. (2000), Gurling et al. (2001), Kendler et al. (1996), a multicenter analysis (SLCG 1996), and Stefansson et al. (2002). The latter group also presented evidence for NRG1 (8p21-p12; neuregulin 1) as an SCZ candidate gene, although it is located 9–13 cM centromeric to other maximal linkage results. Finally, Cao et al. (1997) reported suggestive evidence for linkage on chromosome 6q21-22.3 (P=.00024), a finding supported by data from the National Institute of Mental Health (NIMH) Genetics Initiative data set (Martinez et al. 1999), a multicenter analysis (Levinson et al. 2000), and Lindholm et al. (2001).

None of these regions has produced strong support for linkage in the majority of genome scan projects. There are many possible explanations for this inconsistency of results, such as genotyping and diagnostic errors, differences in statistical strategies, or true differences in genetic effects of these and other loci across ethnic groups or types of pedigrees. Given the large number of projects and analyses that have been undertaken, spurious findings could also be reported several times (Lander and Kruglyak 1995). Another possibility is that each of these loci has a small populationwide effect on susceptibility and is thus difficult to detect consistently without very large samples. Suarez et al. (1994) demonstrated this problem in a simulation study of polygenic disease models, and Göring et al. (2001) recently showed that, for small genetic effect sizes, genome scans of pedigree samples with low power tend to overestimate the genetic effects at loci with the highest scores in the scan—that is, genetic parameters have effectively been maximized. This would be true whether a locus had a small effect on risk in each individual carrier (polygene model) or a larger effect on risk in a small proportion of cases (locus heterogeneity model). Thus, although optimistic observers might conclude that strong evidence has been emerging for linkage to SCZ in a number of chromosomal regions, skeptics have been more impressed by the many failures to replicate SCZ linkage findings. A major difficulty here is that SCZ genome scan sample sizes have been small—the largest scan included 294 small pedigrees (DeLisi et al. 2002b), and the typical sample has been 20–100, whereas samples of (for example) 600–1,000 affected sibling pairs would be required to reliably demonstrate locus-specific genetic effects causing a 27%–30% populationwide increase in risk to siblings (Hauser et al. 1996), and there are multiplicative gene effects that are even more difficult to detect (Rybicki and Elston 2000).

To address this problem, we applied the rank-based genome scan meta-analysis (GSMA) method (Wise et al. 1999) to data from the 20 complete genome scans of SCZ described in table 1. The first article in this series (Levinson et al. 2003 [in this issue]) described the GSMA method in greater detail and presented a simulation study of the power of the method to detect linkage in data sets resembling the SCZ and bipolar disorder scans that we have studied. In brief, in GSMA, the autosomes are divided into 30-cM bins, the evidence for linkage in each study is rank-ordered across bins with and without weights for sample size, and the average ranks across studies are evaluated for statistically significant evidence for linkage in several ways, as discussed in the first article in this series (Levinson et al. 2003 [in this issue]) and summarized below.

Badner and Gershon (2002a) reported a meta-analysis of published SCZ genome scans using their multiple scan probability (MSP) method (Badner and Gershon 2002b), which combines P values across scans in regions with clusters of positive scores after adjusting for the size of the region. Differences between these methods have been discussed in the first article in this series (Levinson et al. 2003 [in this issue]). The GSMA presented here included studies that were not available to Badner and Gershon (2002b) and excluded some of the studies they analyzed (see the “Discussion” section for details); our analysis also included data provided by the investigators for every marker in each scan, whereas the MSP used only published data (from scans with data in all nominally significant regions) and substituted P values of .5 or 1.0 for missing data points. By including only complete whole-genome data from each scan, the present analysis may be expected to avoid two problems inherent in publication bias: (1) investigators tend to present their most positive results, as well as weakly positive results that seem to confirm others’ findings, which could inflate meta-analysis results in those regions; and (2) conversely, GSMA can detect significant cross-study results for regions that are weakly positive in many studies but not sufficiently positive to have been presented as “one of the best results” in any study.

The GSMA for SCZ has been performed under a single narrow diagnostic model (with a few exceptions, as described below), which includes SCZ (psychotic symptoms such as delusions and hallucinations and deficits in emotional and social functioning, usually with a chronic course) and schizoaffective disorder (mixed schizophrenic and manic or depressive disorders with persistence of psychosis when mood symptoms remit), which coaggregates with SCZ in families (Levinson and Mowry 2000). Although family study data suggest that a broader spectrum of disorders coaggregate with SCZ in families (Levinson and Mowry 2000), the “core” diagnoses involving overt psychosis and chronic course are made most reliably across research groups (Faraone et al. 1996) and have been the focus of most linkage studies.

If at least weak linkage is present in many or most samples, larger samples would be expected to contribute more information. Therefore, we consider our primary analysis to be one in which the ranks of bins for each scan are weighted to account for sample size, as described below. Unweighted analyses are also reported for comparison. This method of meta-analysis should be most useful in detecting loci that increase disease susceptibility in a substantial proportion of the samples under study, although the observed effect in each sample may be both small and variable because of stochastic or other effects. One would not expect the method to be useful in cases where a locus is having an effect only in a few unique samples, such as those with specific pedigree structures or ethnic or geographical backgrounds. However, we have included samples collected in more isolated populations, because the similar clinical features and prevalence of SCZ in different cultures suggest that at least some of the susceptibility loci are common to many populations (Risch 1990). We followed the same rationale in including studies of one or a few extended pedigrees and of many smaller pedigrees.

Material and Methods

Results

Figure 1 shows the markers at the boundaries of each bin and their location on the Marshfield map, the average rank for each bin across all scans weighted by the square root of the number of genotyped affected cases, and a condensed representation of each bin’s within-study ranks (“1” denotes R_study=1–5, etc., as shown in the legend). A complete list of weighted and unweighted R_study and R_Avg values is available at the URL listed in the “Electronic-Database Information” section. In the row labeled “bin” for each half of the figure, the bins with P_AvgRnk<.01 are shaded in black, and those with P_AvgRnk<.05 are shaded in gray. This figure illustrates which studies contributed the lowest ranks to the average rank for each bin. For example, bin 2.5, which achieved the best R_Avg, had R_study values between 1 and 10 in 8 of 20 scans, whereas bin 3.2, with the third-best R_Avg, was ranked between 1 and 20 in only 4 scans, and its low R_Avg was due to lower R_study values than the average in most scans. The scans are sorted in descending size so that the relative contributions of larger and smaller data sets can be appreciated.

Figure 2 is a scatterplot of the weighted and unweighted average ranks for each bin, with solid diamonds representing the primary, weighted analysis, and open diamonds representing the unweighted analysis for comparison. The 99%, 95%, and 90% thresholds of significance are shown; the Y-axes for the two analyses have been adjusted so that the lines marking the thresholds are the same.

Table 2 lists P_AvgRnk and P_ord values for weighted and unweighted analyses for the 19 bins with either value <.05 by either analysis. Results for the primary weighted analysis will be discussed here and are illustrated in figure 3. Only bin 2.5 (2p11.1-q21.1) met the criterion for genomewide significance of P_AvgRnk<.000417. The nominally significant P_AvgRnk of the adjacent bin, 2.6, provides additional evidence for linkage here, because, in simulated data, bins adjacent to those containing disease loci often also achieve nominal significance. This bin had one of the five lowest ranks in six scans (U.S./International, MCV/Ireland [c], U.S./Australia, NIMH-Eur-Am, Kiel, and Palau), and ranks between 6 and 10 in two others (Cardiff and Utah).

Three aggregate criteria for genomewide significant linkage were met:

• 1.

  Twelve bins had P_AvgRnk<.05 (shown in boldface italic type in table 2). This exceeds the threshold of ⩾11 such bins observed in <5% of simulated whole-genome replicates of similar data sets with no linkage present.

• 2.

  Each of these 12 bins also had P_ord<.05, indicated by asterisks in table 2 and illustrated in figure 3. This exceeds the 5% empirical threshold of ⩾4 bins with P_AvgRnk and P_ord values <.05 for GSMA replicates of 14 or 20 studies (the threshold was 5 with only 9 studies); 12 such bins were never observed in 1,000 unlinked replicates of 20 studies resembling those in the SCZ GSMA data set.

• 3.

  The 19 bins with the lowest average ranks had P_ord values <.05 (see fig. 3). The 5% threshold in unlinked simulated data was ⩾4 such values among the 10 best bins, and a cluster of 19 was never observed in 1,000 unlinked replicates of 20 studies.

In simulated data with linkage present, all three of these thresholds were exceeded with increasing frequency as the number of linked bins and their genetic effect increased.

Finally, to determine whether the placement of bin boundaries was having a critical effect on results, adjacent bins were combined into 60-cM bins, for two possible combinations (e.g., for chromosome 1, bins 1+2, 3+4, etc., and then 2+3, 4+5, etc.), for the weighted analysis. Results were unchanged: the same regions had nominally significant P_AvgRnk P values, as are shown in table 2 for 30-cM bins, with similar levels of significance, and no new regions of interest were identified. Thus, it appears that the placement of bin boundaries did not, in this case, bias the results against any region.

Discussion

Genomewide evidence for linkage was observed in this meta-analysis of 20 SCZ genome scans. The average rank of bin 2.5 (chromosome 2p12-q22.1) was associated with a genomewide significant P value after correction for multiple testing. In addition, all three empirical aggregate criteria for linkage were met: more bins than expected by chance achieved nominally significant values of P_AvgRnk (pointwise P value of the average rank), of P_ord (pointwise P value of the average rank taking into account its place in the order of average ranks), and of both P_AvgRnk and P_ord. The number of bins meeting each of the latter two criteria was highly significant—that is, it was greater than was observed in any of 1,000 GSMAs of simulated whole-genome replicates of 20 similar studies (genomewide P<.001).

The aggregate criteria suggest that these 20 scans support linkage of multiple loci to SCZ. However, there is no straightforward way to determine which of the implicated bins represent true positive results. In simulations of genomes containing 5–10 linked loci, each of small effect, bins with nominal significance for P_AvgRnk and P_ord are most likely to contain a linked locus or to be immediately adjacent to a bin that does; when clusters of bins meet this criterion, few, if any, are false positives. Therefore, the 10 chromosomal regions represented by the first 12 bins in table 2 (marked with asterisks) should be considered strong SCZ candidate regions. In simulated data, both the linked bin and an adjacent bin often achieve nominally significant P_AvgRnk and P_ord, and it is possible that this phenomenon is being observed in the pairs of significant bins on chromosomes 2q (2.5 and 2.6) and 6p (6.1 and 6.2) and perhaps also on 1q (1.6 and 1.7), where bin 1.7 achieved a nominally significant P_ord.

Six additional chromosomal regions were included in the cluster of 19 consecutive bins with nominally significant values of P_ord (table 2 ). In simulated data (see table 7 in the first article in this series [Levinson et al. 2003 {in this issue}]), when there are 5–10 weakly linked loci in the genome, there are still more unlinked than linked bins with P_ord<.05, but these tend to be distributed throughout the overall order of average ranks rather than clustered among the best average ranks. Thus, it is possible that some or all of these regions are also linked to SCZ, but the support for these regions must be considered weaker than for the cluster with nominally significant P_AvgRnk and P_ord.

The simulation studies reported in the previous article (Levinson et al. 2003 [in this issue]) did not consider a sufficiently broad range of situations to interpret these data more definitively. For example, we did not study the effects on GSMA of placing two or more very weakly linked loci in the same bin or in adjacent bins or of varying the genetic effect of multiple linked loci. However, the single-bin and aggregate criteria for genomewide significance discussed above were validated by data from simulations of genomes with no linkage present, so that these conclusions do not depend on the comprehensiveness of the simulations that assumed linkage.

Data for the unweighted analysis, with each study weighted equally, are shown in table 2 for comparison. Genomewide significance was still observed for bin 2.5. Only 9 bins had nominally significant P_AvgRnk values, fewer than the aggregate significance threshold of 11, but 6 of the best 10 bins had both P_AvgRnk and P_ord values <.05. In discussions among the authors and with colleagues, it has been debated whether small studies with larger numbers of cases per pedigree should be given relatively more weight, because of the possible advantages of such pedigrees for linkage analysis. The unweighted analysis can be considered the most extreme possible alternative, in that the smallest study is weighted equally with the largest study, but it produced less evidence for linkage in most regions, including those that have been supported by studies of larger pedigrees.

These data do not exclude the possibility that other chromosomal regions harbor SCZ susceptibility loci that could not be detected by the present methods or that substantially influence risk in only one or a few populations. GSMA does not currently consider X or Y chromosome data, so no conclusions can be reached concerning possible linkage on those chromosomes (DeLisi et al. 2000). There are many other limitations to our analysis—for example, we made decisions about multiple tests (we maximized over them), subdivided data sets (we considered them separate scans), used only the first stage of two-stage scans, accepted variable map density both within and across studies, as well as variable numbers of incomplete families (missing parents), and, in some cases, used data with denser maps of certain candidate regions. It is possible that other decisions regarding these or other issues would have produced different and perhaps more accurate results.

Regarding variable marker density, the mean number of markers genotyped in the 12 bins with P_AvgRnk<.05 was 5.21, versus 3.1 for all other bins (t=3.39; P=.006). Wise (2000) simulated an unlinked 30-cM region in 120 sib pairs (10 markers, 500 replicates). Multipoint sib pair analysis using 3, 5, 7, or 10 markers yielded mean maximum MLS values (within the bin) of 0.272, 0.333, 0.373, and 0.385, respectively, because of increasing SDs. A GSMA could thus be biased by extreme variation in density, especially if there was biased selection of regions (e.g., increased density in candidate regions only where positive scores had been observed by chance). In the present data, the densest typing was in a few studies that first reported evidence for linkage in a region; however, because the bin had one of the best ranks before the extra markers were typed, the biasing effect should be small. For example, in bin 6.1, mean coverage was 8.65 markers, but without the Bonn and three MCV/Ireland scans (18–27 markers) the mean was 4.85. For bin 22.1, the mean was 6.85, but without Johns Hopkins, U.S./International, and Bonn (13–16 markers) it was 4.82. However, this type of bias was not systematically studied here, and an effect cannot be ruled out. Future investigators should be aware that it is advantageous to perform an initial analysis of a screening map with uniform spacing prior to adding additional markers in candidate regions.

One previous meta-analysis of SCZ genome scan data (Badner and Gershon 2002a) applied a method for combining P values from many of these studies and found significant results on chromosomes 8p and 13q. They analyzed published results, whereas we obtained linkage data for all markers from each investigator, to permit detection of evidence for linkage in regions that had not come to attention because scores were not strikingly positive. The differences in results could also have been due primarily to the use of different sets of scan analyses. Badner and Gershon (2002a) included nine data sets that were used in this analysis (Bonn, Canada, Johns Hopkins, Kiel, the two NIMH scans, U.K./Iceland, U.S./Australia, and Utah), a different data set for one scan (the second-stage analysis of candidate regions in the Cardiff scan, using an enlarged sample), incomplete data for four scans (U.S./International, the two Finnish scans, and Sweden), no data for six scans (the three MCV/Ireland subsets, and the Costa Rica, Texas, and Palau scans), and data from four scans that we excluded because of small sample size (Bailer et al. 2002), incomplete genome coverage (Barr et al. 1994), selection of probands for periodic catatonia rather than SCZ (Stober et al. 2000), or possible overlap with another sample (Rees et al. 1999) leading us to include the larger sample (Williams et al. 1999). The same region of chromosome 8p was significant in both analyses, and it seems understandable that we identified additional significant regions with a larger data set and more complete data. It is more difficult to explain why the 13q region was highly significant in the Badner and Gershon (2002a) analysis but did not approach significance here. Evidence for a susceptibility gene in this region has recently been published (Chumakov et al. 2002) and awaits confirmation.

In the present analysis, bin 2.5 (2p11.1-q21.1) had the best average rank (31.8), with a genomewide significant P_AvgRnk=.0004. This bin had ranks between 1 and 5 in six scans (U.S./International, MCV/Ireland [c], U.S./Australia, NIMH-Eur-Am, Kiel, and Palau) and had ranks between 6 and 10 in two scans (Cardiff and Utah). DeLisi et al. (2002b) observed slightly more positive results in bin 2.4, with the peak MLS in the initial genome scan at the boundary marker for bins 2.4 and 2.5 (D2S139), whereas the other scans with high ranks in this region maximized within bin 2.5 or the more distal bin 2.6.

Among the other bins listed in table 2, previous findings on chromosomes 6p and 8p have been discussed in the Introduction. For 6p, maximal evidence for linkage was in bin 6.1 (Lindholm et al. 2001; Straub et al. 2002b) or 6.2 (Schwab et al. 2000), whereas all of the highly positive linkage reports on 8p have been in bin 8.2 (Kendler et al. 1996; SLCG 1996; Blouin et al. 1998; Brzustowicz et al. 1999), except for the linkage peak and candidate gene locus (NRG1) reported recently by Stefansson et al. (2002), which are in bin 8.3. Positive findings in bin 5.5 have previously been presented by Straub et al. (1997) and by Schwab et al. (1997). Pulver et al. (1995) reported a modestly positive LOD score on chromosome 3p in bin 3.1, whereas the current finding is in bin 3.2; a multicenter study that found no evidence for linkage on 3p (SLCG 1996) did not include all of bin 3.2. Gurling et al. (2001) reported evidence for linkage in bin 11.5. On chromosome 1q, the maximum evidence for linkage reported by Brzustowicz et al. (2000) was in bin 1.6, and the nearby peak reported by Gurling et al. (2001) was in bin 1.7. Bin 22.1 on chromosome 22q contains the deletion region for velo-cardio-facial syndrome, which is associated with schizophrenic symptoms (Pulver et al. 1994b; Karayiorgou et al. 1995; Murphy et al. 1998) and which came to the attention of the field after Pulver et al. (1994a) reported modestly positive evidence for linkage in the region. Chiu et al. (2002) reported an interaction between linkage at markers in bins 8.2 and 14.1 in the Johns Hopkins data set.

Among the bins in the cluster of nominally significant P_ord values, bin 10.1 had one of the highest ranks in four scans (U.S./International, MCV/Ireland [a], Bonn, and NIMH-Eur-Am) and had been proposed as a candidate region by Faraone et al. (1998), Schwab et al. (1998), and Straub et al. (1998), with a multicenter study providing only weak support (Levinson et al. 2000). Bin 6.4 contains the region of chromosome 6q discussed above. No previous reports have called attention to chromosomes 16p-q, 15q, or 17q as SCZ candidate regions. Bin 15.3 is considerably distal to the location of the α-7-nicotinic receptor in bin 15.1 (Freedman et al. 1997), where positive evidence for linkage to SCZ has been reported (Freedman et al. 2001a, 2001b; Gejman et al. 2001; Tsuang et al. 2001).

It is of interest that, in several cases, the GSMA provided support for linkages that had been observed previously in small samples of extended pedigrees from more homogeneous or isolated populations (e.g., regions of chromosomes 1q, 2q, 6q, and 11q, as discussed above), despite the fact that the weighting procedure greatly reduced the contributions of these smaller samples to the results. This supports the hypothesis that there are at least some loci that contribute to SCZ susceptibility in many or most populations and that loci whose effects are detected particularly strongly in a unique population may also prove relevant in many other populations.

In conclusion, we can suggest with considerable confidence that SCZ susceptibility loci are likely to exist in some or all of the chromosomal regions noted above. A meta-analysis is one step toward defining genomic regions that harbor SCZ susceptibility loci. Investigators might wish to consider performing collaborative linkage analyses using raw genotypes and integrated marker maps or regenotyping available pedigrees with a common, dense map. Investigators of other complex disorders might be further stimulated to collaborate in various ways, to study larger data sets. Additional details of the present analysis, including ranks for each study, are available at the University of Pennsylvania Web site listed in the Electronic-Database Information section. These results are cause for considerable optimism about the prospects for identifying SCZ susceptibility genes.

We plan to update this analysis when several new SCZ genome scans are completed, and we invite investigators with new genome scan data to contact the present authors.