Ethics Statement

The recruitment and diagnosis of patients and their family members are described elsewhere [13]. In brief, this clinical cohort was previously examined and published using the behavioral and genetic data utilized in this study in a genome scan of schizophrenia [13]. As described previously [13] subjects were recruited from five geographic centers beginning in 1985: 213 families from the United States (based at Stony Brook, N.Y.), 50 from the United Kingdom (Oxford), 33 from Italy (Milan), 11 from Chile (Santiago), and two from Belgium (Leuven). Recruitment included catchment area screening, recruitment by health professionals at hospital and outpatient facilities, and advertisement through organizations which support families of mentally ill individuals. Written consent was obtained from all participants in the study after receiving an explanation of the study procedures and their implications. Consent was obtained using the same procedures in all five countries, and each center was granted approval with Single Project Assurance status by the Office of Protection From Research Risks of the U.S. Department of Health and Human Services. This study was conducted according to the principles expressed in the Declaration of Helsinki. Individual Institutional Review Board approval from the institutions to which the authors are affiliated were not obtained as this was a reanalysis of a previously published de-identified dataset.

Samples and Subjects

Simulated Data (GAW14).

The GAW14 data was designed to model the genetic influences on a complex psychiatric disorder [14]. The made-up disorder termed Kofendred Personality Disorder (KPD) could be subcategorized into three distinct latent phenotypes denoted as P1, P2, and P3. These latent phenotypes were in turn defined by 12 behavioral traits labeled a-to-l. The genetic architecture of this disorder involves four disease gene loci D1, D2, D3, and D4 (Figure 1A). The genes interact in an epistatic fashion together with two modifier genes, D5 and D6 (Figure 1B). For the purposes of evaluating our methodology we assumed that the latent phenotypes were unknown, instead all analyses were based directly on the 12 behavioral traits.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Genetic model and analysis of the GAW14 dataset.

(A) Graphical representation of the genetic model used in the GAW14 simulation. D1-to-D4 are the major disease-causing loci while D5 and D6 are modifier loci that influence disease penetrance if the disease genotype is present. P1-to-P3 are latent phenotypes and a-to-l are sub-clinical phenotypes both caused by the disease loci, as seen by the connecting lines. The letters “A” and “B” associated with the latent phenotype reflect identical phenotypes, which are caused by different underlying loci or acted upon by modifier loci. (B) Graphical representation of the network showing relationships of behavioral phenotypes in the GAW14 dataset using a majority rule consensus tree. There are three key groupings referred to as clades 0_21, 0_4, and 0_1, which define the phenotypic groups P1, P2 and P3, and the unaffecteds respectively. The group shaded in blue, labeled 0_21 (P1), denotes the clade containing individuals who have the latent phenotype P1. The group in red labeled 0_4 (P2+P3) is defined in a similar manner. Finally, the group in green labeled 0_1 represents the unaffected individuals. The high level ratchet based tree search employed in the network generation procedure will not always resolve subtle differences within clades. Given that half of the individuals in clade 0_4 share certain traits, additional analyses within this clade were conducted to further refine the latent phenotypes. The clade 0_4 depicts the majority rule consensus tree resulting from a tree search within this clade. The expansion of clade 0_4 demonstrates the improved resolution of this clade as well as the separation of the latent phenotypes by the presence of trait b in P3 and the absence of trait b in P2. Specifically, the clade labeled in yellow is defined by the absence of trait b defining phenotype P2 and the clade labeled in blue have symptom b present defining phenotype P3. (C) A histogram plot for all sub-clinical traits present in the primary clades defined in part B. The x-axis represents the traits, each with three bars for each clade, and the y-axis represents the proportions of individuals with the presence of the trait. Within these clades, the proportions of individuals with each of the behavioral traits were examined and found to concur with the simulated models relating the behavioral traits with the latent phenotypes. In clade 0_21, 100% of individuals in this group have the traits b, e, f, and h, which define the P1 latent phenotype. The clade 0_4 is not as clearly resolved, where 100% of individuals share the behavioral traits c-to-h, but only fewer than 50% possess the traits b and k. This is because clade 0_4 defines the latent phenotypes P2 and P3 that are separated by only one behavioral trait (b) and share 6 traits in common (c-to-h). Finally, clade 0_1 contains most of the unaffected individuals not in clades 0_4 or 0_21 and is not defined by any one specific or subset of behavioral trait(s).

https://doi.org/10.1371/journal.pone.0009714.g001

In each population the families were ascertained based on the following criterion assuming a family contains at least one latent phenotype in order to be considered for inclusion in the study: for the Aipotu population at least two family members must have either P1, P2 or P3; for the Karangar population at least two family members must have P2 or P3; for the Danacaa population at least two family members must have P1. In this way, 100 families were ascertained from each population, constituting a single replicate (Replicate #1 was used in this study). A total of 300 families were analyzed, consisting of 2,077 individuals of which 781 were affected with KPD and 1,296 were unaffected. All genotype data for the 10 simulated chromosomes were examined, with 416 microsatellite markers approximately 7.5 centimorgans (cM) apart.

Algorithm for Behavioral and Genetic Network Analysis

The hierarchical behavioral network representing phenotypic relatedness is obtained through the use of a character-based analysis. This approach optimizes the change of behavioral traits, where a change is defined by the loss (1→0) or gain (0→1) of the presence of a behavioral trait as the network is traversed from one individual to another[16]. This scoring system allows for the separation of clinical symptoms within diagnostic categories and for overlap between diagnostic categories. The structure of this network is then utilized to create increasingly more inclusive or nested sets of individuals based on phenotypic similarity each of which is tested via traditional linkage and/or association analyses. As such, the nesting of individuals based on the structure of the behavioral network allows for the identification of genetic loci that have both local and global influences on clusters of symptoms. Each group of individuals, defined by the nesting group derived from the behavioral network structure, is then considered as a candidate behavioral endophenotype (or a candidate phenotypic model) that is adopted for subsequent genetic linkage or association testing. If significant evidence for linkage or association is detected for a particular behavioral endophenotype, the symptom profile of that group can then be examined to identify those symptoms shared by the majority of individuals. Since the behavioral network algorithm maximizes the traits that individuals close to one another on the network share, the majority of individuals in a group with significant linkage or association are likely to all share one or more behavioral traits in common. In this way, relationships between specific endophenotypes and genotypes may be established.

The character-based algorithm described here, which generates a behavioral network based on individual symptoms, accounts for phenotypic heterogeneity. The goal is to identify genetically meaningful phenotypic groups. Since an optimal arrangement of shared behavioral symptoms was used to build the phenotypic network, the symptoms defining a phenotypic group of interest can easily be determined. As phenotypic groups are analyzed at multiple, increasingly inclusive levels, genetic loci that show linkage or association within small or large groups of individuals with particular symptoms can be detected (denoted as local or global affects respectively). This provides a direct means of detecting genetic loci that can potentially modulate behavioral symptoms.

Specifically, our algorithm consists of four distinct components including (1) estimating a behavioral network, (2) nesting of the behavioral network into inclusive groups/clades[17], (3) performing genetic analysis, and (4) evaluating statistical significance of behavioral symptoms (see Figure 2 for an outline of the algorithm, and Figure S1 for an example of network estimation). The statistical evaluation of behavioral symptoms involves comparing symptom distributions of individuals belonging to a given nested group to the symptom distribution within the diagnostic category to which they were assigned. To this end, we adopt a likelihood-based approach to test for significant differences in the symptom distributions in these groups. The likelihood ratio test is formulated as follows,

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Pipeline of Algorithm for Behavioral and Genetic Analysis.

^a Templeton, A. R., Boerwinkle, E., Sing, C. F. 1987. A cladistic analysis of phenotypic associations with haplotypes inferred from restriction endonuclease mapping. I. Basic theory and an analysis of alcohol dehydrogenase activity in Drosophila. Genetics 117(2):343–351. ^b Lipscomb DL. 1992. Parsimony, homology and the analysis of multistate characters. Cladistics (8):45–65.

https://doi.org/10.1371/journal.pone.0009714.g002

The likelihoods for the null and alternative hypotheses (H₀ and H_A respectively) are straightforward, where assuming p₂ = p₃ = p₄ with corresponding maximum likelihood estimates of ; . Similarly, assuming p₃ = p₄ with corresponding maximum likelihood estimates of ; ; . Note, x_i is the number of individuals with symptoms in group i and n_i is the total number of individuals in group i. Here p is the probability of the presence of a behavioral trait for any of four possible groupings, such that p₂ is the probability of observing a given trait in the clinical diagnostic group that does not overlap with the clade, and p₃ is that trait's probability in the clade of interest that does not overlap with the major diagnostic group. p₄ is the probability of trait overlap between p₂ and p₃, and p₁ is the probability of the trait in individuals that are not in any of the previously noted groups. The null hypothesis (H₀) assumes that the trait probability distribution for the individuals in the clade and the clinically defined diagnostic group (as well as the overlap) are the same (i.e., p₂ = p₃ = p₄). Whereas, the alternative hypothesis (H_A) assumes that the probability of the trait is not the same (i.e., p₂≠p₃ = p₄). For the boundary conditions where one group is contained within the other, for H_A p₂ is compared to p₄, while for H₀ they are equal. The situation is similar in the case of no overlap between categories where p₂ and p₃ would be the key parameters. The likelihoods thus formulated are evaluated for each trait of interest and assessed via the likelihood ratio test at the parameter's maximum likelihood estimates. Furthermore, for assessment of statistical significance, empirical p-values were estimated using 10,000 permutations of the data.

Data Analysis

Generating a network from the behavioral symptoms of individuals in the dataset involves examination of the entire “tree space”. This requires that all possible network configurations, which group individuals with respect to their shared symptom profiles must be enumerated and evaluated. This is not computationally feasible. Instead we used a heuristic ratchet search algorithm that has been developed for applications of exceptionally large datasets [18]. The ratchet search algorithm assumes that there are islands of local optimality in the search space for the shortest network [19], that is, a network with an arrangement of individuals and their shared symptoms that minimize the number of character state changes. However, such heuristic algorithms typically produce multiple equally parsimonious trees. A solution to this problem is to adopt a majority rule consensus tree, where a resolution among individual nodes on the tree is arrived at if it is supported by the majority of trees in the tree space. In this application, 51% is the minimum allowable support. For this analysis, all parsimony based ratchet searches were conducted using TNT v 1.0 [20]. All trees were visualized and all symptoms were mapped onto trees using MACCLADE v 4.08 [21]. To improve our tree resolution for a specific clade, we performed a tree search using the SPR branch swapping in PAUP v 4.0b10 [22].

The behavioral networks thus generated were examined and nested groups/clades were identified for genetic analysis. Since the datasets considered consist of family data, we performed genetic linkage analysis using the program MILINK from the LINKAGE software package [23], [24]. Heterogeneity LOD scores (HLOD) were computed based on an admixture likelihood model that jointly tests for linkage and heterogeneity (i.e., the maximum HLOD is compared with the log-likelihood under a null hypothesis of “no linkage and no heterogeneity” rather than with the maximum homogeneity LOD) [25], [26]. Linkage analyses were performed on those clades that were comprised of individuals from at least 40 families. This clade constraint was applied, a priori, before conducting the linkage analyses. This limited the number of tests performed for small samples with potentially little power to detect linkage. Of all nested groups considered, 6 clades met this criteria in the simulated dataset and 39 clades in the schizophrenia dataset. For each nested clade, LOD scores maximized over multiple models were calculated with corresponding estimated empirical p-values adjusting for the testing of multiple phenotypic and genotypic models as well as multiple marker loci. LOD scores were calculated under models of both homogeneity (LOD) and heterogeneity (HLOD)[26]. Four genetic models were considered including (1) a fully penetrant dominant model (Dom-1) and disease allele frequency of 0.01, (2) fully penetrant recessive model (Rec-1) and disease allele frequency of 0.09, (3) a dominant model with 55% penetrance (Dom-2) and phenocopy rate of 0.0005, and disease allele frequency of 0.01, and (4) a recessive model with 55% penetrance (Rec-2), phenocopy rate of 0.0005, and disease allele frequency of 0.09. The first two models (Dom-1 and Rec-1), though the simplest and perhaps overly optimistic, were chosen because they provide the greatest power to detect linkage [27]. The latter two models (Dom-2 and Rec-2) were based on a previously published analysis of this schizophrenia dataset[28]. In the absence of any knowledge regarding the mode of inheritance of the disease, as is typically the case for complex diseases, application of a limited set of simple genetic models have been shown to work well when testing for linkage [29].

To account for multiple testing, we estimated empirical p values, allowing for multiple phenotypic and genetic models, as well as multiple marker loci tested. To this end, we randomly assigned genotypes to the founders within each family while conditioning on the observed allele frequency and intermarker distances in the data and then mating individuals within families according to the pedigree structure using the program SIMULATE [30]. In this way, 100 randomized replicates of the data were generated. These replicates were used to estimate both the model-based and the global empirical p-values. The model-based p value (p_M) is evaluated by recording the number of times a maximum LOD score from a replicate exceeds the maximum LOD score from that of the observed data, across all genetic models within each clade (phenotypic model) and therefore corrects for the testing of multiple genetic models and marker loci tested for a particular phenotypic model. The global p value (p_G) is calculated by making comparisons to the maximum LOD scores from randomized replicates across all genotypic and phenotypic models (clades) and marker loci tested. This approach accounts for multiple testing of all combinations of genetic-phenotypic models as well as genetic markers tested.

Simulated Data (GAW14)

Behavioral traits from all family members were analyzed to generate a behavioral network. The network correctly grouped the three latent phenotypes, as shown in Figure 1C. Within these groups, referred to as clades, the proportions of individuals with each of the behavioral traits were examined and found to concur with the simulated models relating the behavioral traits with the latent phenotypes (Figure 1C). Furthermore, genetic linkage analysis of these key clades identified the major disease loci contributing to KPD. In contrast to the findings observed from conducting a traditional genome scan (where KPD is used as the major diagnosis for defining affectedness), the observed linkage signals from this analysis were more significant (with the exception of one locus), and combinations of loci contributing to specific latent phenotypes were identified (Table S2). This analysis indicates that our algorithm is able to identify the genetic factors underlying each of the latent phenotypes. Indeed, this degree of resolution in detecting the contribution of genetic loci to specific sub-phenotypes in a complex trait model is not attainable using existing approaches.

Real Data (Schizophrenia)

Analysis of schizophrenia samples identified four distinct behavioral groups (clades) with corresponding evidence for genetic linkage. The salient clinical features of these groups were similar to major depression and schizophrenia with two groups resembling “depression” (clades 6_4 and 4_28) and two “schizophrenia” (clades 6_6 and 6_1). These clades can be represented in a behavioral network (Figure 3). In this network representation, the most inclusive nesting level contains two clades (8_0 and 8_1) that encompass the four clades noted above, where these clades become smaller and less inclusive as the network is traversed from left to right (Figure 3). In this way, individuals with shared behavioral symptoms are grouped through successive nesting levels, in that specific symptom(s) are shown to define specific clades (Figure 3). The significance of these specific symptom profiles was further evaluated (Figure 4). Those behavioral symptoms that were observed in >70% of the individuals within each clade were identified, and the relative proportion of these symptoms were compared to the assigned diagnostic categories. Specifically, those symptoms that occur in a significantly greater number of individuals in the clade under consideration were compared to the diagnostic category (of e.g., depression or schizophrenia) to which the majority of the individuals in the clade were assigned. In this way, statistically significant symptom profiles were identified that characterize each clade (Figure 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Nested structure of the network showing only the five clades that exhibit significant findings from the schizophrenia dataset.

From left to right are the most to least inclusive nesting levels. The most inclusive nesting level, 8, divides the tree into three parts, clades 8_0, 8_1 and 8_2. The five clades found to have suggestive to significant linkage (4_28, 5_15, 6_1, 6_4, 6_6) are shaded with colors that are consistently used to represent them in the text. More inclusive nesting levels that encompass these five clades are also shown; however, nesting levels below level 8 that have no significant linkage findings are not shown. Clades with individuals who have a greater proportion of affective symptoms are grouped more closely together (4_28, 6_4). The clade with a greater proportion of positive symptoms typical of schizophrenia is grouped by itself (6_6), while the clade with both affective and positive symptoms is grouped between the two (6_1), and the clade with individuals with minimal symptoms is grouped by itself (5_15).

https://doi.org/10.1371/journal.pone.0009714.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. The spectrum of behavioral symptoms defining the depression and schizophrenia clades.

Symptoms present in greater than 70% of individuals with the diagnosis of depression, schizophrenia, or schizoaffective disorder are depicted as filled-in circles with depression (in blue), schizophrenia (in purple) and schizoaffective (in gray). Significant differences in symptom distributions within clades compared to the diagnostic categories of depression (Dep), schizophrenia (Schiz), and schizoaffective (Schizaf) are depicted with a “*” (with statistical significance reported at the 0.05 level). Unfilled circles with a “*” imply statistically significant differences between the clade and diagnostic category but that less than 70% of individuals in that clade have that symptom. Boxes which encompass the circles in the three columns of depression, schizophrenia and schizoaffective disorder are shaded when that symptom appears in greater than 70% of the individuals in that diagnostic category. Five symptoms distinguish the depression^4_28 clade from the diagnostic category of depression, which include trouble getting to sleep, trouble staying asleep, early morning awakening, severe impairment in both childhood and adolescent adaptation to school. The depression^6_4 clade is distinguished by one symptom, mainly difficulty concentrating. These depression clades are mutually exclusive. Further, language disorder symptoms distinguish the schizophrenia^6_6 clade from the diagnostic category schizophrenia (i.e., circumstantial speech and excess details) as well as an increased incidence of violence and aggression. The schizophrenia^6_1 clade is distinguished from both diagnostic categories schizophrenia and schizoaffective disorder by three symptoms, including thoughts inserted into one's head that are not one's own, thoughts stopping or withdrawn, and thoughts being read by others. However, this clade shares behavioral symptoms characteristic of both of these diagnostic categories.

https://doi.org/10.1371/journal.pone.0009714.g004

The depression group depression^6_4 is characterized by the symptom “difficulty concentrating” (Figure 4), since the majority of individuals belonging to this clade share this symptom. In contrast, the second clade, depression^4_28 is characterized by a spectrum of symptoms, that include “trouble getting to sleep”, “trouble staying asleep” and “early morning wakening”, and “severely poor adaptation to school with poor scholastic performance”. Such symptoms appear in significantly greater proportions of individuals in these clades than those with the assigned diagnosis of major depression (Figure 4). Further, these individuals belong to mutually exclusive groups accounting for 83 of the 139 patients with major depression in the total dataset (Table 1), identifying two depression sub-phenotypes with distinct symptom profiles.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Diagnostic and LOD score statistics for clades.

https://doi.org/10.1371/journal.pone.0009714.t001

Linkage analysis of these depression clades provided significant evidence for linkage on chromosome 2 (Table 1). Individuals with behavioral symptoms in each of the depression groups were designated as affected and their family members who did not share these defining symptoms were designated as unaffected for consideration in the linkage analyses. We observed significant linkage signal for the depression^6_4 clade with loss of concentration as the defining symptom on chromosome 2q35 (heterogeneity LOD score HLOD = 4.07 at marker D2S2248). We also observed suggestive evidence for linkage on 2q13 for the second depression^4_28 clade (HLOD of 2.86), where the dominating symptoms were those of sleep disturbances. Comparatively, analysis of families in the total dataset using the diagnosis of major depression to define affectedness yielded lower linkage signals. In fact, the maximal linkage signal overlapped with the same chromosomal region identified by the depression^4_28 clade, with a lower score (HLOD = 2.6 with 'alpha = 1 at marker D2S160; Figure S2). The next highest linkage signal was 5 centiMorgans (cM) away from our strongest finding with the depression^6_4 clade with a substantially lower LOD score (HLOD = 2.1 and alpha = 0.45 at marker D2S126; Figure S2). Taken together, these results may indicate the presence of two distinct candidate loci influencing the symptoms defining these two groups, since the clades contain two mutually exclusive groups of individuals with potentially stronger, more homogeneous signal resulting from such refinement of phenotypes.

The schizophrenia groups also revealed intriguing symptom profiles. The schizophrenia^6_6 clade was characterized by language disorder symptoms such as circumstantial speech and excess details, and aggressive and violent tendencies (Figure 4). The schizophrenia^6_1 clade was composed mainly of individuals with the major diagnoses of schizophrenia or schizoaffective disorder with symptoms including loss of concentration, pressure of speech, low mood, suicidal ideation and psychomotor retardation (Figure 4). Individuals belonging to this clade have symptoms more characteristic of those with the diagnosis of schizoaffective disorder even though the majority of these individuals had the diagnosis of schizophrenia (56% schizophrenia vs. 32% schizoaffective). This group is enriched with language disorder and hallucination symptoms resembling schizophrenia and depressive symptoms resembling schizoaffective disorder. This is a clear demonstration of the overlap between and difficulty in separating these two so called clinically distinct diagnostic categories.

Linkage analysis of these schizophrenia clades provided suggestive evidence for linkage (Table 1). We observed a novel linkage finding with schizophrenia^6_6 clade on chromosome 2p25 (HLOD of 2.19 and alpha  = 0.66 proximal to marker D2S168). This linkage peak is approximately 66 cM from the peak LOD score of 2.99 reported previously [13] from linkage analysis of the same samples with the diagnosis of schizophrenia or schizoaffective disorder. Linkage analysis with the schizophrenia only diagnosis resulted in substantial improvement of the LOD score rising to 5.13 in the same region when only those with schizophrenia were considered as affected [31]. The analysis of the schizophrenia^6_1 clade yielded maximal linkage signal on chromosome 2p21 (HLOD of 2.42 and alpha  = 1 at marker D2S2298; no evidence for linkage was observed at this locus in the comparative analysis with diagnosis of schizophrenia or schizoaffective disorder for affectedness (Table 1). These findings are unique to each schizophrenia clade with a relative distance of 19 cM between the most significant loci from each clade, underscoring our ability to detect potential linkage to distinct genetic regions due to the refinement in the phenotypic definitions.

Finally and most importantly, the present study also illustrates that behavioral symptoms follow a continuum of psychosis and extend into normal personality traits. This is best illustrated by clade 5_15, which has no defining symptom profile and encompasses a collection of individuals from many diagnostic categories with the largest represented group being the unaffecteds. Many of these individuals have very few symptoms, which is why they cluster more closely with the unaffected individuals. However, closer examination of their symptom distribution revealed that approximately 57% of all individuals carry at least one symptom of a personality disorder (data not shown), including paranoid, schizoid, schizotypal and borderline traits (with significant evidence for linkage, HLOD of 3.4 proximal to D2S391 approximately 4 cM from the peak LOD in schizophrenia^6_1 clade see Table S1). This result captures the notion of a continuum of psychosis. As expected, unaffected family members of the mentally ill often share some symptoms with their affected family members.