Genetic Epidemiology 14:549!561 (1997)

Genetic Analysis of Bipolar Disorder:Summary of GAW10John Rice

Department of Psychiatry, School of Medicine, Washington University, St. Louis,Missouri

Participants in the Bipolar Disorder component of Genetic Analysis Workshop 10had access to five distributed data sets containing chromosome 18 marker data andfive data sets containing chromosome 5 data. A total of 25 groups participated inanalyses and applied a myriad of methodologically innovative approaches to thesedata. Contributors focused on how to: (1) best define the phenotype from thespectrum of affective diagnoses; (2) test for a parent-of-origin effect in thetransmission of bipolar illness and assess whether sharing in affected sib pairsdepends on the sex of the transmitting parent; (3) evaluate the effects ofmisspecification of marker allele frequencies; (4) examine the putative candidateloci provided; (5) investigate the mode of inheritance; and (6) perform ameta-analysis to combine multiple data sets in a single analysis. Taken as awhole, the results would appear suggestive, but not definitive for linkage to abipolar susceptibility locus on chromosome 18. The evidence for linkageappeared to increase as the diagnostic definition of the phenotype was broadened.Multipoint analyses seem to provide less evidence. It is possible that, becauseadjacent markers may be present in different data sets, the multipoint methods arecombining marker data from different studies in a more comprehensive way thansingle marker analyses. Evidence on chromosome 5 and evidence for candidateloci were minimal. A discussion of problems inherent in combined analyses isgiven. © 1997 Wiley-Liss, Inc.

Key words: bipolar disorder, chromosome 18, chromosome 5, linkage analyses, parent-of-origineffect

INTRODUCTION

The affective disorders refer to a group of psychiatric conditions in whichdisturbances of mood (i.e., affect) predominate. However, depression and euphoria are

Address reprint requests to Dr. John Rice, Department of Psychiatry, Box 8134, Washington UniversityMedical School, St. Louis, MO 63110.

© 1997 Wiley-Liss, Inc.

550 Rice

not the only symptoms: insomnia, anorexia, suicidal ideation, and feelings of worthlessnessare also associated with low moods; hyperactivity and flight of ideas are associated withelation. There is considerable debate concerning the many subclassifications proposed tosubdivide the affective disorders. One major separation that is generally accepted is thatof unipolar/bipolar disorder. Unipolar affective disorder refers to individuals who havedepression only, whereas bipolar affective disorder refers to individuals who have episodesof both mania and depression, or episodes of mania only. GAW10 focuses on bipolarillness; however, relatives under study may have a range of affective diagnoses, so thatphenotypic definition is of major importance.

There is convincing evidence that susceptibility towards bipolar disorder istransmitted through genetic factors [Craddock and McGuffin, 1993] but the basic questionsof mode of inheritance and the identification of specific susceptibility loci remainunanswered.

Linkage to chromosome 18 was initially reported to markers on 18p by Berrettini etal. [1994], with additional evidence reported by Stine et al. [1995]. The latter report foundevidence on 18p and 18q, with evidence for linkage in families with paternal transmissionon 18q. The data sets which led to these reports, along with three others were availableto GAW10 participants [Goldin et al., this issue].

Kelsoe et al. [1996] reported suggestive evidence for linkage on chromosome 5 nearthe dopamine transporter gene (DAT), and data were also available in this region [Goldinet al., this issue].

DIAGNOSES

The following is a description of commonly used diagnoses in studies of the affectivedisorders:

Bipolar I Disorder (BPI)

The essential feature of bipolar affective disorder is the occurrence of manicepisodes. These consist of distinct periods of elevated, expansive mood associated withincreased activity, push of speech, flight of ideas, inflated self-esteem, decreased need forsleep, and distractibility. Most people who experience manic episodes eventually haveepisodes of depression, characterized by intense, pervasive, and persistent symptoms thatinterfere with their normal functioning. The disorder is commonly called bipolar I disorderor manic depressive disorder and has a population frequency of approximately 0.3 - 0.5%.

Schizoaffective Disorder, Manic Subtype (SA/M)

An episode of illness characterized by a manic syndrome may be diagnosed asschizoaffective, manic subtype if, in addition to the bipolar features described above,specific types of delusions, hallucinations, or marked formal thought disorder are present.There is some consensus that this syndrome be included with the diagnosis of bipolar Iillness. A depressive episode with such features is referred to as schizoaffective disorder,

Bipolar Disorder: Summary of GAW10 551

depressed subtype (SA/D). There is some consensus that this syndrome should not beincluded with bipolar disorder.

Bipolar II Disorder (BPII)

Some (but not all) classification schemes include a diagnosis of hypomania, amanic-like episode that is not of sufficient intensity to meet the full criteria for a manicsyndrome. Operationally, the hypomania in these individuals is distinguished from maniain that the intensity of the hypomania does not lead to major impairment or hospitalization.Individuals with both hypomania and depressive episodes are often referred to as havingbipolar II disorder.

Major Depressive Disorder (MDD) or Unipolar Depression (UP)

The diagnosis of major depression requires a distinct period of depressive or irritablemood (lasting from 2 to 4 weeks depending on the diagnostic system used), some degreeof impairment, and the presence of symptoms such as weight loss (or weight gain), sleepdifficulties, loss of interest or pleasure in usual activities, feelings of guilt orself-worthlessness, difficulty concentrating, or suicidal ideation. There is no clearconsensus on establishing a threshold on a gradient between normal mood and clinicalstate. This uncertainty is reflected in variable population rates that depend on the criteriaused and on the particular diagnostic instrument used to codify the criteria. Some analysesdistinguish individuals with more than one episode (recurrent major depression , UPR)from those with a single episode (UPS).

Overview of Individual Contributions

A total of 25 groups analyzed the bipolar linkage data. One initial challenge facedby most investigators was how to combine/pool the multiple data sets. The various studiesincluded somewhat different diagnostic schema and different sets of markers. Mostcombined analyses used the program CRIMAP to construct a genetic map; however, theestimation of allele frequencies for the pooled map was complicated by the problem of howto match alleles across studies. One novel approach used by Daly et al. was to avoidmatching entirely. They created a copy of a marker for each data set in which it waspresent and then computed allele frequencies within that data set. They then placed eachset of copied markers in the same position on the pooled map. Greenwood and Bull useddifferent allele sets to produce study-specific allele frequency estimates in their combinedanalysis.

The Hopkins data contained markers in which the allele nomenclature was notconsistent across families, and most groups deleted these markers before analysis. Thiscould contribute to differences between the GAW analyses and the analyses of Stine et al.[1995].

The number of groups which analyzed various data sets is given in Table I. OnlyDavis et al. and Simonsen et al. analyzed both chromosome 18 and chromosome 5 data.

The methods used by the contributors varied greatly and included parametric andnonparametric approaches. Most groups used multiple methods. Table II indicates thevarious computer programs cited and gives a feel for the spectrum of analyses used.

552 Rice

TABLE I: Data Sets Used by GAW10 Participants

Data sets Number of groupsa

Chromosome 18

All 5 data sets 10

4 data sets 2

3 data sets 1

2 data sets 2

Hopkins only 5

NIMH only 3

Chromosome 5 4

Number sums to 27 since two groups analyzed both chromosome 5 and 18 data.a

TABLE II. Genetic Programs Used in Analysis

Program Number of users Reference

SIBPAL 6 S.A.G.E. [1994]

GENEHUNTER 4 Kruglyak et al. [1996]

MAPMARKER/SIBS 3 Kruglyak and Lander [1995]

FASTLINK 3 Cottingham et al. [1993]

LINKAGE 2 Lathrop et al. [1994]

HOMOG 1 Ott [1991]

LODLINK (S.A.G.E.) 1 S.A.G.E. [1996]

GAS 2 Young [1995]

SimIBD 1 Davis et al. [1996]

SIMWALK2 1 Sobel and Lange [1996]

ASPEX 1 Hinds and Risch [1996]

MIM 1 Goldgar and Lewis [1996]

SIB-PAIR 1 Duffy [1996]

SIBPAIR 1 Terwilliger [1996]

APM 2 Weeks and Lange [1988]

LIPED 1 Ott [1974]

Thus, individual contributors made choices as to which data sets and how many,which diagnoses, and which analytic methods to use. Given these choices, it is notsurprising that different conclusions are reached and that results are not directlycomparable from one paper to another. We summarize some of these approaches.

Diagnostic Definition

Three groups focused on how to use multiple liability classes to model the multiplediagnostic categories (BPI, BPII, MDD, etc.) and one group used a permutation test toevaluate diagnostic models. Van Eerdewegh et al. considered (1) BPI and SA/M, (2) BPII,(3) subsyndromal mania, and (4) nonpsychotic major depression. They computed a set ofweights which were used to compute penetrances for each liability class. They

Bipolar Disorder: Summary of GAW10 553

also used GENEHUNTER for a nonparametric approach with weights incorporated intothe scoring functions. Using this approach, they found a lod score of 2.2 at D18S57 underan intermediate model. A model without diagnostic weights yields much lower lod scores.The weights appeared to have less of an effect on the NPL score.

Levinson also modeled a diagnostic spectrum by reducing the penetrance in theliability classes corresponding to broader diagnoses. He used the NIMH data with BPI andSA/M as a narrow diagnosis and BPII, unipolar disorder and SA/D as a broad category.He conducted: (1) NPL analyses using GENEHUNTER on the narrow, and then narrowplus broad diagnoses, and (2) a weighted additive parametric analysis allowing forheterogeneity. The maximum lod2 score (where lod2 is the heterogeneity lod scoreassuming that a proportion of families are linked) was 1.81 for " = 0.45 under a broaddiagnostic model and 1.53 under a narrow model. The two maxima were 25 cM apart.The p-values associated with the NPL statistic were less than those of the lod2 statistic.

Turecki and colleagues investigated three phenotypic approaches: (1) affection statusas defined in the original data, (2) a liability class model and (3) a pseudo-quantitative traitmodel. They found no single phenotypic model to provide a consistent advantage overothers, with the maximum lod score occurring at different markers. The maximum lodscore was 2.5 at D18S56 in the Columbia data; the maximal lod scores were under 2 in theHopkins, NIMH and Bonn data.

Simonsen et al. used all data sets on chromosomes 18 and 5 to explore narrow versusbroad diagnoses. Under the assumption that individuals with genetically related disorderswould be similar at marker loci linked to susceptibility genes, they develop three teststatistics to test the null hypothesis that the grouping of diagnoses within groups is arandom partition. This leads to a permutation test of the diagnostic partitioning. They didnot reject the BPI, BPII, and UP model in four of the five chromosome 18 data sets(Hopkins being the exception). However, they found evidence in the Quebec and UCSDchromosome 5 data sets that single episode unipolar illness should not be counted asaffected.

Wyszynski et al. analyzed the NIMH and Hopkins data using several analyticapproaches. They applied SIBPAL, GAS and SIB-PAIR, the last being an implementationof the APM method of Weeks and Lange [1988]. They used narrow (BPI, SA/M, BPIIwith MDD), intermediate (also including MDD) and broad (also including several otherdiagnoses) phenotypic definitions in the NIMH data. The results were consistent with theoriginal Berrettini et al. [1994] and Stine et al. [1995] reports. While all of the linkagemethods found some evidence of linkage to the chromosome 18 markers, they tended tohave the strongest evidence at different markers. They noted that the evidence tended toincrease as the classification of affected individuals was broadened.

Maternal/Paternal Transmission

Motivated by the Stine et al. [1995], the McMahon et al. [1995] and the Gershon etal. [1996] reports of excess maternal transmission and differences in identical-by-descent(IBD) sharing depending on whether the family showed maternal or paternal transmission,several groups used the chromosome 18 data sets to explore parent-of-origin effects.

Lin and Bale considered all data sets and defined the transmitting parent as theparent who was affected himself/herself or who had an affected first degree relative otherthan the offspring. Transmissions were classified as maternal, paternal, or unknown. The

554 Rice

latter included bilineal transmission families. They restricted analyses to D18S37, andobserved an excess of maternal transmissions (204 versus 107); however, they did notobserve IBD score differences among all possible pairs of transmission patterns. Theexcess maternal transmission could not be explained by the higher proportion of affectedfemales, and there was a significant increase in allele sharing at D18S37 in affected sibpairs regardless of the sex of the transmitting parent.

Greenwood and Bull also analyzed all chromosome 18 data sets to examineparent-of-origin effects. They adapted the EM algorithm of Kruglyak and Lander [1995],allowing the IBD sharing proportions to depend on covariates. This extension also leadsto subgroup-specific lod scores, and the development of a test for heterogeneity. Theyanalyzed 648 sib pairs in 260 nuclear families where plausible parent-of-origin could bedefined under a broad diagnostic model. They included parent-of-origin, an indicatorvariable comparing BPI to other diagnoses and four data set indicators for the five setsanalyzed. Their results suggested (but did not confirm) the 18p region of Berrettini andStine linked to BPI with some possibility of parent-of-origin effects on allele sharing (peaklod of 3.0). They did not see evidence for 18q in the paternal subgroup and did not detectstudy heterogeneity. This is contrary to the original report of Stine et al. [1995] in theHopkins data.

Dorr et al. also performed a logistic regression analysis allowing IBD to depend onparent-of-origin, broad versus narrow diagnosis, and study. They used ASPEX to estimateIBD and then performed the logistic analysis. In agreement with Greenwood and Bull,there was no significant heterogeneity among studies, although a paternal parent-of-origineffect was found in the D18S41-D18S64 region on the q-arm in agreement with Stine etal. Their maximal lod score was 2.99 at D18S40 under a broad phenotype.

Collins and Go examined parental transmissions in the 22 families of the NIMH dataset, considering male to male, male to female and maternal transmission patterns. Theyperformed analyses in the three transmission groups separately, using three differentaffection status classifications. Their results concurred with Berrettini et al. [1994] in thatno significant lod scores were found with the program LINKAGE. Using SIBPAL, theyfound the greatest evidence for linkage in the male-to-male transmission group, consistentwith Stine et al. [1995], suggesting strongest evidence for linkage in paternal pedigrees.The strongest evidence was at D18S40.

Donald et al. classified families according to the sex of the transmitting parent andintroduced a mixed category in which transmission was from both parents. They used amore strict criterion in that to be classified as maternal (or paternal), all transmissions inthe family had to occur exclusively in females (or males). Linkage analyses were carriedout separately for each type of family. They found positive lod scores in the paternal, butnot maternal, families at D18S41-D18S38 on the q-arm, but using only data from Hopkinsfor these markers. No evidence was found on the p-arm using this division of families.Moreover, they tested for excess maternal transmission and found that the excess could beexplained by the excess of affected females.

The majority of analyses in GAW10 assumed equal recombination in men andwomen. If recombination occurred less often in men, there may be greater sharing inaffected sib pairs in paternal pedigrees at a marker near a susceptibility gene due to thesmaller recombination rate. That is, is there confounding between the parent-of-origineffect and male/female differences in recombination? Durner and Abreu restricted theiranalyses to D18S37 and D18S40 and utilized all data sets. They used two phenotypic

Bipolar Disorder: Summary of GAW10 555

models and assumed either a dominant or recessive mode of inheritance under severalassumed sets of penetrances and with gene frequencies set to correspond to fixed populationprevalences. The lod score was maximized over the set of penetrances and separaterecombination fractions for males and females. There was little evidence for a difference inthe recombination fractions in males and females, suggesting that the parent-of-origin effectdoes not result from misspecification of male/female recombination differences.

There are two distinct phenomena being studied—increased maternal transmissionand greater evidence for allele sharing in affected sib pairs in paternal transmissionfamilies. The above results are not totally consistent, although different approaches areused in different data sets. Most of the markers on the q-arm were in fact from the Hopkinsdata set, so that the combined analyses were disproportionately influenced by this singledata set. There does appear to be a paternal effect on this area of the q-arm, with apossible weaker effect on the p-arm. The two analyses that tested for study heterogeneity[Greenwood and Bull; Dorr et al.] did not detect heterogeneity across studies.

Analyses of Hopkins Data

Several groups limited their analyses to the data contributed by Hopkins. Thisconsisted of 28 pedigrees. Stine et al. [1995] identified a subset of 11 pedigrees withpaternal transmission and reported a lod score of 3.51 in this subset analyzed under adominant model. This lod score occurred at D18S41 on the q-arm of 18.

Cleves and colleagues limited their analysis to D18S41 in the Hopkins data. Theytested the null hypotheses 2 + 2 = 1 versus the alternative 2 + 2 < 1. This is moref m f m

general then the usual test of 2 = 2 = ½, and leads to a chi-square test with one degreef m

of freedom corresponding to the single linear constraint. They tested both dominant andrecessive models using LODLINK and did not find evidence for linkage in the entire setof 28 pedigrees. They analyzed the subset of 11 paternal transmission pedigrees and founda lod score of 1.87. This contrasts with the lod score of 3.51 reported by Stine et al. [1995]in the same 11 families under a dominant model. Schnell et al. attribute differences toeither assumptions made in their age-dependent penetrance function or in their estimatesof allele frequencies. In the Discussion, they note that Stine et al. [1995] modeled agedependency with a step function corresponding to three liability classes. Results similarto those reported were obtained when age-specific penetrance was estimated from the data.However, the lod score was maximized at the boundary 2 = 1 and 2 = 0. They discussm f

the difficulties in interpreting these estimates.Zhu et al. also analyzed D18S41 in the Hopkins data with an emphasis on

age-of-onset methods. They used the Haseman and Elston [1972] approach asimplemented in SIBPAL and performed age adjustment as well as analysis of the residualsfrom a parametric survival analysis. They found both the age-of-onset and survival modelresiduals analyses to increase the evidence for linkage of bipolar disorder to markerD18S41. Consistent with other analyses, linkage was detected only in the paternaltransmission pedigrees.

Cordell and Olson considered two methods for the estimation of confidence intervalsfor relative risk using markers D18S41 and D18S37 in the Hopkins data. A total of 28families were typed at D18S41, and 17 families were typed for D18S37. They found theconfidence intervals to be quite wide, and predict that at least 100, and preferably 200,

556 Rice

fully informative affected sib pairs would be required to yield reasonably sized confidencelimits.

Li and Schaid applied GENEHUNTER to the Hopkins data and proposed a newscore statistic to incorporate unaffecteds. Their maximum Z-all score was in the paternaltransmission pedigrees, but the associated p-value of 0.013 at D18S61 was not striking.There was no evidence in the maternal pedigrees. They also describe the new scorestatistic to utilize information on unaffected individuals, but the approach has yet to beimplemented and was not applied to the GAW10 data.

Misspecification of Marker Allele Frequencies

Two groups examined the impact of marker allele frequencies. Gordon et al.examined the (false) assumption of equal allele frequencies on the p-values computedusing SIBPAL. They used the candidate locus ACTHR112 from the NIMH data set.Approximately 50% of the affected sib pairs had both parents typed. Gordon et al.randomly deleted about half the parental genotype information so that approximately 20%of affected sib pairs had typed parents, and then deleted all parental information so that 8%were informative (this is not 0 since some parental genotypes could be inferred from thechildren). They found the test for proportion of alleles shared IBD among concordantlyaffected sib pairs to show a greater percentage of significant p-values with decreasingparental genotypic information, whereas the Haseman-Elston test, based on the regressionof squared trait differences using all types of sib pairs, produced significant p-valuescomparatively less frequently.

Margaritte-Jeannin et al. first selected a total of 82 sib pairs where both members hada diagnosis of BPI. Due to heterogeneity in allele frequencies, they required pairs fromHopkins and Bonn to have both parents genotyped. They found the maximum lod score tobe low (0.59-0.81) in this sample of sib pairs. They then considered the Amish pedigree(family 9000) and used the affected pedigree method (APM) with two sets offrequencies—one set estimated from the entire NIMH data set, and one from the Amishfamily alone. The p-values from the APM were 0.0002 versus 0.16 using these two sets ofmarker allele frequencies. This underscores the potential impact of incorrect specificationof allele frequencies. However, to avoid this problem in the pooled data, they limitedanalyses to 82 sib pairs. They discuss the potential loss of power versus protection againsttype I error when attempting a meta-analysis as in the chromosome 18 data sets.

Association Studies

Three groups used versions of the transmission/disequilibrium test (TDT).Bickeböller et al. examined ACTHR and Golf (on 18p), two candidate loci provided onchromosome 18. They considered three versions of the TDT: (1) a biallelic version, witha Bonferroni correction applied when multiple tests are done for markers with more thantwo alleles, (2) a test of symmetry in the m × m table of parental transmitted andnontransmitted alleles and (3) a multiallelic TDT using a logistic regression model.Evidence for linkage and association was found for ACTHR for paternal transmissionpedigrees only under a broad diagnosis (at approximately the .025 significance level).

Waldman et al. used data from two studies (UCSD/UBC and Cardiff) onchromosome 5. They found evidence for linkage disequilibrium between the PCR marker

Bipolar Disorder: Summary of GAW10 557

of DAT1 and bipolar disorder in the Cardiff sample, but not in the UCSD/UBC sample.In the Cardiff sample, stronger evidence was obtained using a more narrow diagnosis.Overall, the evidence for association is marginal given the number of statistical testsperformed.

Badner et al. employed two multiallelic versions of the TDT: (1) a test based on thecontingency table comparing marginal frequencies of parental transmission, and (2) thelikelihood ratio test of Terwilliger [1995]. They cite simulations which show the latter ismore powerful when a single allele is positively associated, whereas the former is morepowerful when multiple alleles are weakly associated. They find eight markers nominallysignificant in at least one data set, and two (D18S53 and D18S37) significant in more thanone data set. However, they note that none would be significant if corrected for multipletesting.

Mode of Inheritance

Although the emphasis of GAW10 was on linkage analysis, two groups consideredevidence for transmission of bipolar disorder in these families. Schnell et al. restrictedanalyses to D18S41 in the Hopkins data. They considered a class D regressive model[Bonney, 1984] in which the likelihood depends on the ordering of the affected siblingswithin the sibship. They included the proportion of siblings in a sibship who are affectedas a common sibship factor to create an exchangeable sibling effect, and used the programsREGTL and LODLINK of S.A.G.E. [1994] for analyses. They considered twono-transmission models as well as dominant and recessive models. They did not detectlinkage in the full set of pedigrees, but did have a lod score of 3.44 in the paternalpedigrees under a dominant model. However, they found little evidence for genetictransmission in the full set of families and when regressive coefficients were included inthe model, the lod score was no longer statistically significant. They interpret the lack ofconsistent evidence for transmission of bipolar illness in these families as an indicationthat the linkage results may reflect type I error.

Homer et al. examined 52 families in the chromosome 5 data sets and the twoflanking markers for DAT (D5S392 and D5S406). They computed two-point lod scoresunder dominant and recessive modes of inheritance and compared the absolute value of thedifference between the maximum lod scores calculated under the two models. The largestlod score was 1.25 for D5S406 under a dominant model for bipolar disorder, so that theevidence for linkage was not sufficient to discriminate between the dominant and recessivemodels.

Meta-analysis of Chromosome 18 Data

As noted in Table I, fewer than half the contributions used data from four or five ofthe GAW10 chromosome 18 data sets. We describe here those which attempted acombined/pooled analysis. Some of the contributions [Simonsen et al., Badner et al.,Turecki et al., and Donald et al.] performed parallel analyses in all data sets and have beenpreviously described. Margaritte-Jeannin et al. analyzed a combined data set of 82affected BPI sib pairs, as described above. Their evidence for linkage was minimal, butthis may be due to the relatively small sample.

558 Rice

Daly et al. used the nonparametric features of GENEHUNTER and performedseparate analyses of narrow and broad diagnostic models. They used 115 pedigrees withtwo or more BPI affecteds. There were two peaks 40 cM apart at D18S39/D18S35 andD18S53 with associated p-values of 0.05 and 0.06, respectively. When the phenotype wasbroadened to include diagnoses of BPII, schizoaffective disorder and MDD, a p-value of0.005 was seen at D18S39/D18S35, with a secondary peak at D18S53. When the data setswere analyzed separately, only the NIMH data gave a significant result (p = 0.007), withthe strongest evidence for linkage coming from the Amish pedigree.

Davis et al. used a simulation-based APM (SimIBD) and a Markov chain Monte Carlotechnique to examine marker allele clustering among all affected relatives (SIMWALK2).These methods overcome a difficulty of earlier APM methods which were very sensitive toallele frequencies and only used identical-by-state (IBS) information. They found threemarkers on chromosome 18 in the NIMH data to be suggestive of linkage (p < 0.01) usinga broad definition of the phenotype: D18S40 (p = 0.008), D18S35 (p = 0.003), and D18S54(p = 0.008). No markers had p < 0.01 using the narrow phenotype, and no other data setsgave suggestive linkage. The three suggestive markers span more than 60 cM, so that linkageto a single susceptibility disease locus would seem doubtful.

Haghighi et al. considered only individuals diagnosed as BPI to be affected. Theyanalyzed the NIMH, Hopkins, Bonn, and Columbia data separately, and then did a combinedanalysis. They found suggestive evidence (p < 0.001) for D18S45 in the Columbia data set.The multipoint results were not significant for any of the four data sets. They restricted thecombined analysis to only four markers in the NIMH, Columbia, and Bonn data since allelesizes (in base pairs) were not available for all markers. D18S53 gave suggestive evidence.A formal heterogeneity test did not demonstrate significant heterogeneity among the threestudies. The multipoint analysis using MAPMAKER/SIBS was not significant.

Greenwood and Bull also used MAPMAKER/SIBS to obtain IBD estimates for use in acombined analysis. Unlike others, they used study-specific allele frequency estimates, butcalculated the multipoint IBD estimates while using the same map in each data set. Thecombined analysis (no covariates) then showed a maximum lod score of just under 1.0 on 18p.

Lin and Bale combined all data and analyzed only marker D18S37. They analyzed3,394 individuals in 185 families using the program MIM under a broad diagnosis. Theyestimated 58% IBD sharing using 382 affected sib pairs (p < 0.0000091). As notedearlier, they found evidence in both paternal and maternal families.

Dorr et al. analyzed an intermediate phenotype (BPI and BPII and SA/M) and abroad phenotype (with UPR included). They separated analyses which used IBD only (asimplemented in sib_ibd of ASPEX) and those which use all data and depend on markerallele frequencies. The percent sharing was consistent in these two approaches. Thehighest lod score was 2.99 at D18S40 under the broad diagnosis using 551 affected sibpairs (it was 1.32 under the intermediate diagnosis with 351 pairs). The lod score was 1.86at D18S37. Thus, their strongest evidence for linkage was in a region near D18S40; theyalso detected excess paternal sharing on the q-arm near marker D18S64. They did notperform multipoint analysis.

SUMMARY AND CONCLUSIONS

Taken as a whole, the above results would appear suggestive, but not definitive. Itmust be emphasized that one reason that chromosome 18 was chosen for GAW10 is the

Bipolar Disorder: Summary of GAW10 559

publication of the positive reports by Stine et al. [1995] and Berrettini et al. [1994], so thatthe GAW10 analyses should not be viewed as an independent attempt at replication. Stineet al. [1995] reported a lod score of 3.51 in their 11 paternal transmission pedigrees; nocombined analyses, using up to 185 families, found a lod score of this magnitude for anyregion of 18. One group using a different age correction in the same 11 families had a lodscore of 1.87 at the same marker. It is not surprising that different approaches makingdifferent assumptions can lead to quite disparate statistical evidence.

Overall, the evidence for linkage appeared to increase as the diagnostic definition ofthe phenotype was broadened. This is consistent with contributions which formally analyzedthe phenotypes themselves, and with the contributions which used phenotypic severity (i.e.,BPI versus a broader diagnosis) as a covariate and did not find it significant. There wererelatively modest results when the definition of who is affected is restricted to BPI alone.

Overall, there appears to be evidence for a parent-of-origin effect on the q-arm of18, with increased sharing in affected sib pairs when bipolar illness is transmitted throughthe father. It should be noted that, for example, D18S64 was available only in the Hopkinsand NIMH data, and D18S38 was available only in the Hopkins and Bonn data, so thatcombined analyses on the q-arm were leveraged by the Hopkins data.

One of the most significant findings (i.e., smallest p-value) was reported by Lin andBale for D18S37, the marker positive in both the Berrettini et al. [1994] and the Stine etal. [1995] reports. Stine et al. had evidence in all these families (not just the paternal ones)in this region. Interestingly, the evidence in the NIMH data of the Berrettini reports isminimal when parametric lod-score analysis is done under a single locus model as inLINKAGE or FASTLINK—the original evidence was striking with an APM method. Thechoice of method of analysis might contribute to differences among results in GAW10.

The multipoint analyses using GENEHUNTER or MAPMAKER/SIBS seem toprovide less evidence than analyses using a single marker at a time (an exception to thisis in the NIMH data analyzed by Van Eerdewegh et al.). Investigators who commented onmap length did not report excessive length which might have been diagnostic of typingerrors or incorrect order. Moreover, multipoint methods would seem to be especiallyinformative when parents are untyped and help reduce dependency on allele frequencyestimates. It is possible that, since adjacent markers may be present in different data sets,the multipoint methods are combining marker data from different studies in a morecomprehensive way than single marker analyses.

The evidence for association with DAT on chromosome 5 was minimal, as was theevidence for ACTHR and GOLF on chromosome 18.

As a group, the methodologic questions and approaches utilized in GAW10 are ofconsiderable interest. Comparison of methods when applied to a common set of data areinstructive. This aspect of the 25 contributions cannot easily be summarized; rather itrequires a close comparison of individual contributions.

It is interesting to contrast the GAW10 data sets and analyses to those of the 1987GAW5, which also had bipolar data available [Rice and Risch, 1989]. In these data,families to use for segregation analysis, X-linked marker data (color blindness and G6PDdeficiency), and HLA haplotypes for association were contributed. The molecular andanalytic approaches to complex diseases have clearly made great strides in the interveningdecade.

One hidden message in many of the current contributions is that data collectionefforts should endeavor to collect and code their linkage data in ways to facilitate

560 Rice

combined analyses. If marker alleles were coded as base pair sizes with a commonstandard (such as a CEPH individual), then calibration of alleles across studies would besimplified. Given the sensitivity of analytic methods to marker allele frequencies, as wellas the labor involved in calibrating alleles across studies in the GAW data, a commonprotocol would seem warranted. Related to this are concerns of ethnic admixture in whichsubgroups of families may have different marker allele frequencies. Race was notprovided as a potential covariate with the GAW10 data sets, so that average frequencieswere used.

Several papers commented on the difficulty in calibrating diagnoses across studiesand on deciding which ones to use. Psychiatric genetics has a tradition of performingparallel analyses under multiple schema—e.g., narrow, intermediate and broad diagnoses.These lead to multiple, nonindependent tests and problems associated with interpretingresults.

One approach used by several groups was to allow allele sharing in affected sib pairsto depend on several covariates such as parent-of-origin, severity of diagnosis andcontributing data set. This allows for a formal statistical test of heterogeneity as well asthe test of interaction effects. The techniques for doing this are not yet fully developed,but already offer an alternative to many parallel analyses using small subsets of the data.The detection of susceptibility loci for complex traits such as bipolar disorder may requirea meta-analysis of many studies to achieve appropriate statistical power.

ACKNOWLEDGMENTS

Supported in part by NIH grants MH37685, MH31302 and MH46280. We expressour gratitude to all workshop participants for their thoughtful comments.

REFERENCES

Berrettini WH, Ferraro TN, Goldin LR, Weeks D, Detera-Wadleigh S, Nurnberger JI, Jr, Gershon ES(1994). Pericentromeric chromosome 18 DNA markers and manic-depressive illness: evidence fora susceptibility gene. Proc Natl Acad Sci USA 91:5918-5921.

Bonney GE (1984): On the statistical determination of major gene mechanisms in continuous human traits:regressive models. Am J Med Genet 18:731-749.

Cottingham RW Jr., Idury RM, Schaffer AA (1993): Faster sequential genetic linkage computations. AmJ Hum Genet 53:252.

Craddock N, McGuffin P (1993): Approaches to the genetics of bipolar disorders. Ann Med 25:317-322.Davis S, Schroeder M, Goldin LR, Weeks DE (1996): Nonparametric simulations-based statistics for

detecting linkage in general pedigrees. Am J Hum Genet 58:867-880.Duffy D (1996): SIB-PAIR. A program for elementary genetic analyses. Johns Hopkins University,

Baltimore.Gershon ES, Badner JA, Detera-Wadleigh SD, Ferraro TN, Berrettini WH (1996): Maternal inheritance

and chromosome 18 allele sharing in unilineal bipolar illness pedigrees. Am J Med Genet(Neuropsychiatr Genet) 67:202-207.

Goldgar D, Lewis CM (1996): Multipoint Identical-By-Descent Method, Release 2.1. Computer packageavailable from the Department of Genetic Epidemiology, University of Utah, Salt Lake City.

Haseman JK, Elston RC (1972): The investigation of linkage between a quantitative trait and a markerlocus. Behav Genet 1:3-19.

Hinds D, Risch N (1996): The ASPEX package: affected sib pair exclusion mapping (computerdocumentation).

Bipolar Disorder: Summary of GAW10 561

Kelsoe JR, Sadovnick AD, Kristbjarnarson H, Bergesch P, Mroczkowski-Parker Z, Drennan, Rapaport MH,Flodman P, Spence MA, Remick RA (1996). Possible locus for bipolar disorder near the dopaminetransporter on chromosome 5. Am J Med Genet 67:533-540.

Kruglyak L, Lander ES (1995): Complete multipoint sib-pair analysis of qualitative and quantitative traits.Am J Hum Genet 57:439-454.

Kruglyak L, Daly MJ, Reeve-Daly MP, Lander ES (1996): Parametric and nonparametric linkage analysis:A unified multipoint approach. Am J Hum Genet 58:1347-1363.

Lathrop GM, Lalouel JM, Julier C, and Ott J (1994): Strategies for multilocus linkage analysis in humans.Proc Natl Acad Sci USA 81:3443-3446.

McMahon FJ, Stine AOC, Meyers DA, Simpson SG, and DePaulo JR (1995): Patterns of maternaltransmission in bipolar affective disorder. Am J Hum Genet 56:1277-1286.

Ott J (1974): Estimation of the recombination fraction in human pedigrees: Efficient computation of thelikelihood for human linkage studies. Am J Hum Genet 26:588-597.

Ott J (1991): “Analysis of Human Genetic Linkage.” Baltimore: Johns Hopkins University Press.Rice J, Risch N (1989): Genetic analysis of the affective disorders: Summary of GAW5. Genet Epidemiol

6(1):161-177.S.A.G.E. (1994): Statistical analysis for genetic epidemiology, Release 2.2. Computer program package

available from the Department of Epidemiology and Biostatistics. Case Western Reserve University,Cleveland.

S.A.G.E. (1996): Statistical analysis for genetic epidemiology, Release 3.0. Computer program packageavailable from the Department of Epidemiology and Biostatistics. Case Western Reserve University,Cleveland.

Sobel E, Lange K (1996): Descent graphs in pedigree analysis: applications to haplotyping, locationsscores, and marker-sharing statistics. Am J Hum Genet 58:1323-1337.

Stine OC, Xu J, Koskela R, McMahon FJ, Gschwend M, Friddle C, Clark CD, McInnis MG, Simpson SG,Breschel TS, Vishio E, Riskin K, Feilotter H, Chen E, Shen S, Folstein S, Meyers DA, Botstein D,Marr TG, DePaulo RJ (1995). Evidence for linkage of bipolar disorder to chromosome 18 withparent-of-origin effect. Am J Hum Genet 57:1384-1394.

Terwilliger JD (1995): Powerful likelihood method for the analysis of linkage disequilibrium between traitloci and one or more polymorphic marker loci. Am J Hum Genet 56:777-787.

Terwilliger JD (1996): ANALYZE computer package, version 2.1. Weeks DE, Lange K (1988): The affected-pedigree-member method of linkage analysis. Am J Hum Genet

42:315-326.Young A (1995): Genetic Analysis System, Version 2.0 (Computer Program). Oxford University, Oxford.