Introduction
The causes of many psychiatric disorders, including schizophrenia and the psychoses, are largely attributed to multiple gene–gene and gene–environment interactions (Rutter et al. Reference Rutter, Moffitt and Caspi2006) with many steps from genome to overt clinical disorder. Endophenotypes, or intermediate phenotypes, are hypothesized as intervening steps. They provide quantitative biomarkers that are abnormal in both patients and their first-degree relatives, who are themselves at greater than normal risk of developing the disorder (Gottesman & Shields, Reference Gottesman and Shields1973; Gottesman & Gould, Reference Gottesman and Gould2003). Endophenotypes thus represent a deconstruction of the clinical syndrome into underlying biological traits that are under more direct genetic control and may mediate genetic risk.
Decades of family and genetic studies have established that psychotic disorders are highly heritable (Cardno et al. Reference Cardno, Marshall, Coid, Macdonald, Ribchester, Davies, Venturi, Jones, Lewis, Sham, Gottesman, Farmer, McGuffin, Reveley and Murray1999), recently gaining momentum in the search for specific genetic causes (O'Donovan et al. Reference O'Donovan, Craddock and Owen2009). However, the familial co-aggregation of both disorders (Valles et al. Reference Valles, van Os, Guillamat, Gutierrez, Campillo, Gento and Fananas2000) and the considerable overlap of clinical phenotypes indicate that there may be shared risk factors. It may be inappropriate to regard the major psychoses as distinct disease entities (Craddock et al. Reference Craddock, O'Donovan and Owen2009).
Facial affect processing deficits have considerable implications for normal social functioning and are putatively associated with psychotic disorders (Mandal et al. Reference Mandal, Pandey and Prasad1998; Rocca et al. Reference Rocca, Heuvel, Caetano and Lafer2009). Several brain systems are involved, for example the occipitotemporal cortex, orbitofrontal cortex and limbic regions including the amygdala, basal ganglia and right parietal cortex (Adolphs, Reference Adolphs2002). Deficits in facial affect processing have been identified at the first episode of schizophrenia (FES) (Edwards et al. Reference Edwards, Pattison, Jackson and Wales2001) and in those at risk of developing psychosis (Addington et al. Reference Addington, Penn, Woods, Addington and Perkins2008). Therefore, its dysfunction seems to be meaningfully linked to the emergence of psychosis and may be genetically mediated. Endophenotype studies have probed the heritability of facial affect processing in psychotic disorders. One group identified FES patients as having general facial affect recognition deficits when compared with a control group. Their healthy siblings were significantly impaired in fear and disgust recognition (Bediou et al. Reference Bediou, Asri, Brunelin, Krolak-Salmon, D'Amato, Saoud and Tazi2007). Another study looked at a broader patient group of schizophrenia spectrum disorder patients and their healthy relatives, and identified a trend for facial affect recognition deficit in the relatives group when compared to a control group (Alfimova et al. Reference Alfimova, Abramova, Barhatova, Yumatova, Lyachenko and Golimbet2009). However, relatives have not always been found to exhibit impairment: one study looking at the families of autistic patients and schizophrenic patients in parallel showed no deficit in the relatives of schizophrenia patients in a generalized facial affect recognition test, or in the patients themselves (Bolte & Poustka, Reference Bolte and Poustka2003). To our knowledge, no previous study has investigated the candidacy of facial affect processing and related brain systems as an endophenotype for FEP.
Studies of healthy twins have demonstrated the heritability of magnetic resonance imaging (MRI) measurements of brain structure (Wright et al. Reference Wright, Sham, Murray, Weinberger and Bullmore2002). Structural MRI (sMRI) is therefore a viable tool for ascertaining endophenotypes of psychotic disorders. There are consistent effects of FEP versus chronic schizophrenia demonstrated by meta-analyses of the MRI literature (Ellison-Wright et al. Reference Ellison-Wright, Glahn, Laird, Thelen and Bullmore2008); both have similar profiles of abnormality, with more extensive volumetric decreases observed in chronic schizophrenia. However, MRI studies of the relatives of patients with schizophrenia or other psychotic disorders have so far produced inconsistent results; some studies have found an intermediate level of abnormality, most notably ventricular volume (Staal et al. Reference Staal, Hulshoff Pol, Schnack, Hoogendoorn, Jellema and Kahn2000) and cerebellar abnormalities (Marcelis et al. Reference Marcelis, Suckling, Woodruff, Hofman, Bullmore and van Os2003), but others have identified no significant difference in the relatives when compared to controls (Honea et al. Reference Honea, Meyer-Lindenberg, Hobbs, Pezawas, Mattay, Egan, Verchinski, Passingham, Weinberger and Callicott2008). Thus, MRI markers meet the classical endophenotype criteria of being heritable and abnormal in patients; it is less clear whether they are also abnormal in non-psychotic relatives of patients.
Most previous studies of psychosis-related familial effects on MRI markers have measured local structural differences using voxel-based morphometry (VBM; Fusar-Poli et al. Reference Fusar-Poli, Borgwardt, Crescini, Deste, Kempton, Lawrie, McGuire and Sacchetti2010). The conservative significance thresholds mandated by massively univariate analysis of all voxels in the image impact on statistical power of the small- to medium-sized samples in most primary reports. Psychotic disorders are increasingly characterized as dysconnectivity syndromes associated with abnormal integration or connectivity between multiple regions in large-scale anatomical or functional brain systems (Friston & Frith, Reference Friston and Frith1995; Stephan et al. Reference Stephan, Friston and Frith2009). Therefore, it may be more appropriate conceptually, and more powerful statistically, to use multivariate methods for mapping MRI data in psychosis studies, such as partial least squares (PLS; Krishnan et al. Reference Krishnan, Williams, McIntosh and Abdi2011).
Our strategy within this study was to identify endophenotypic signals at the behavioural level, then to characterize these signals at a brain structural level, moving hypothetically closer to the genome. We hypothesized that facial affect processing and related brain structure might provide an endophenotype of FEP. We therefore measured facial affect processing and grey matter (GM) density using sMRI in patients with an FEP, their first-degree relatives and healthy volunteers. PLS analysis was used to investigate the relationship between variation in behavioural data and brain structure. This patient group cuts across diagnostic boundaries and allowed us to probe broader psychosis endophenotypes (Weiser et al. Reference Weiser, van Os and Davidson2005).
Method
Participants and clinical assessments
The sample consisted of 30 FEP probands (one was later excluded after discovery of a pituitary microadenoma), 30 first-degree relatives and 31 unrelated healthy volunteers (Table 1). Probands were clinically followed at the time of recruitment from specialist clinical services in Cambridgeshire and surrounding regions. Eligibility criteria were: (1) age between 18 and 35 years at referral; (2) a first episode of psychosis within 1 year of testing; (3) no identifiable medical cause for psychotic symptoms; and (4) satisfied Melbourne criteria (McGorry et al. Reference McGorry, Edwards, Mihalopoulos, Harrigan and Jackson1996) for the presence of psychotic symptoms for at least 1 week.
Table 1. Demographic and behavioural data for FEP patients, their first-degree relatives and unrelated, healthy volunteers
FEP, First-episode psychosis; s.d., standard deviation; NART, National Adult Reading Test; BFRT, Benton Facial Recognition Test; RT, mean response time.
a Analyses covaried for age, gender and NART (e.g. ANCOVA).
b F 2,86 for behavioural data, one FEP patient was excluded.
Eligible probands gave permission to contact a first-degree relative; the relative could have no personal history of a psychiatric or neurological disorder. A sibling was recruited preferably, but parents (n = 8) were included when this was not feasible. Healthy volunteers were matched at a group-wise level with patient and relative groups for age, gender and pre-morbid intelligence. Volunteers with no family history of psychosis or personal psychiatric or neurological disorder were recruited from the GlaxoSmithKline (GSK) Healthy Volunteer Panel. Seventeen probands (57%) were receiving antipsychotic medication at the time of testing, for whom equivalent chlorpromazine doses were calculated, ranging from 75 to 800 mg/day (mean = 248 mg/day).
Participants were asked to avoid consumption of alcohol (8 h prior to visit) and cannabis (24 h). Any who tested positive for a psychostimulant drug (excluding cannabis) or had contraindications to MRI were excluded from the study. Assessments were made on a single visit to the GSK Clinical Unit, Addenbrooke's Hospital, Cambridge, UK. All MRI scanning was carried out in the Magnetic Resonance Imaging and Spectroscopy Unit, Addenbrooke's. The study was approved by the Cambridge Local Research Ethics Committee (REC reference number 06/Q0108/129). All participants provided written informed consent.
DSM diagnostic criteria were applied to clinical information collected over 6 months using the operational criteria (OPCRIT) checklist (McGuffin et al. Reference McGuffin, Farmer and Harvey1991) where possible. Twenty-two (73%) proband diagnoses were available, of which: schizophrenia = 9; bipolar disorder = 7; psychosis not otherwise specified = 2; depressive disorder with psychosis = 1; mania with psychosis = 2; delusional disorder = 1; and non-organic psychotic disorder = 1.
Behavioural testing: facial processing tasks
The Benton Facial Recognition Test (BFRT, short form; Benton et al. Reference Benton, Sivan, Hamsher, Varney and Spreen1994) assessed participants' ability to recognize facial identity. It requires participants to match a series of target faces from a six-stimulus array in different orientations and shadowing. The number of faces identified (recognition hits) was subsequently used as the measure of facial identity recognition accuracy.
We also used a facial affect intensity rating test that has previously identified group differences in a study of bipolar disorder (Lennox et al. Reference Lennox, Jacob, Calder, Lupson and Bullmore2004). The task was administered during functional MRI (fMRI). It uses faces taken from the ‘Facial Expressions of Emotion: Stimuli and Tests’ series of facial emotions (Calder et al. Reference Calder, Rowland, Young, Nimmo-Smith, Keane and Perrett2000), manipulated to produce four levels of intensity (0, 50, 100 and 125%) of six emotion types (anger, disgust, fear, happiness, sadness and surprise). Participants were asked to explicitly rate emotional intensity on a scale from low = 1 to high = 4. The faces were presented in a pseudo-randomized order, each face displayed for 3.7 s. Eight faces were shown at each intensity for each emotion, 192 faces in total. The paradigm was equivalent for all subjects and lasted 12 min. Subsequent analyses averaged across all intensities, thus using data from 24 trials for each emotion. Data from neutral faces were omitted from averages as they would not elicit normal facial affect processing
Each participant underwent a battery of psychological tests, of which only the aforementioned tests probed facial processing. As the hypothesis of this investigation was to investigate facial affect processing and related brain structure, the fMRI component of the facial affect intensity rating task is not reported here.
Data were analysed statistically using SPSS 17.0 (www.ibm.com/software/analytics/spss/). Paired t tests were used when testing differences between patient and relative groups, to account for their familial relatedness. A two-tailed general linear model with polynomial contrast was applied to test for a linear trend across data, with groups being ordered 0 < 1<2 (patients < relatives < controls). Thus, for a given variable, the null hypothesis is that group means will not represent proportionate change in this order (or its inverse).
sMRI data acquisition
Whole-brain MRI data were acquired using a GE Signa HDxt system (General Electric, USA) operating at 3 T. Inversion recovery (IR)-prepped fast three-dimensional (3D) gradient recalled echo (GRE) T1-weighted spoiled gradient recall (SPGR) images were acquired using the following parameters: repetition time (TR) = 9060 ms, echo time (TE) = 3880 ms, inversion time (TI) = 450 ms, number of slices = 120, slice thickness = 1.1 mm, field of view (FOV) = 240 × 240 mm, image matrix = 512 × 512, voxel dimensions = 0.4688 × 0.4688 mm.
Data were preprocessed using FSLVBM (www.fmrib.ox.ac.uk/fsl/fslvbm). Non-brain tissue was removed using the automated brain extraction procedure (Smith, Reference Smith2002). The T1-weighted images were segmented into probabilistic maps of GM, white matter and cerebrospinal fluid, by estimation of the partial volume for each voxel (Zhang et al. Reference Zhang, Brady and Smith2001). GM segmented tissue maps were registered in standard space by a linear affine transformation (Jenkinson & Smith, Reference Jenkinson and Smith2001; Jenkinson et al. Reference Jenkinson, Bannister, Brady and Smith2002).
sMRI data analysis: interaction between sMRI and facial affect processing speed
An ANCOVA model was applied to whole-brain GM maps to identify regions of GM density that differed significantly between the three groups. A pairwise comparison was applied to the patient group only, with the grouping variable defined as on (n = 16) or off (n = 13) antipsychotic medication. Both models covaried for intracranial volume and were implemented using Cambridge Brain Analysis software testing at the voxel cluster level (Bullmore et al. Reference Bullmore, Suckling, Overmeyer, Rabe-Hesketh, Taylor and Brammer1999).
To identify GM systems optimally correlated with facial affect intensity rating speed, we used the multivariate statistical technique partial least squares (PLS; McIntosh et al. Reference McIntosh, Bookstein, Haxby and Grady1996), implemented using PLSGUI software (www.rotman-baycrest.on.ca/pls/) in MatLab. Computationally, PLS carries out a singular value decomposition (SVD) of a cross-correlational matrix, that is a correlation matrix generated from two or more blocks of statistical variables (neuropsychological and MRI data). The decomposition produces a series of paired latent variables (LVs) for which PLS then computes the covariance between pairs of LVs using the statistic d, the first singular value of the cross-correlational matrix. Successive pairs of LVs have the maximum covariance under the constraint that each of the pairs is geometrically orthogonal to all LV pairs extracted previously. Hence, successive pairs of LVs will have smaller covariances (decreasing values of d). Each LV is paired so that one member of the pair is composed of saliences (or weights) corresponding to each measure in block 1 (e.g. MRI voxels) and the other is composed of saliences for each measure in block 2 (e.g. neuropsychological test scores). Each vector of weights (‘brain score’ or ‘behavioural score’) is exactly proportional to the profile of correlations of the variables of its block and is a summary measure of how well each individual (brain or behaviour respectively) fits these profiles (or LVs). GM maps (block 1) were smoothed with an isotropic Gaussian kernel (σ = 2) and each voxel was thresholded at 0.8. Across all participants, response times (RTs) for each emotion category (block 2) were correlated at each voxel. Brain scores were then extracted for the significant LV pair that explained the most covariance and these were investigated for group differences. The accuracy of BFRT performance was substituted into a parallel PLS analysis as block 2.
For statistical inference, the images were permuted among the group, with the behavioural measures remaining unchanged and the cross-correlation matrix and SVD performed again for 1000 permutations (minimum significance threshold of p = 0.001). A bootstrapping algorithm was performed to establish the stability of a particular voxel appearing in a particular LV image by repeating the SVD for the cross-correlation matrix 100 times (minimum p value = 0.01). In this way a Z score type statistic was computed as to the probability of that voxel appearing in that particular LV image.
Results
Demographic and clinical data
There was a trend towards significant group difference for age (F 2,88 = 2.94, p = 0.058), gender (F 2,88 = 2.48, p = 0.090) and pre-morbid intelligence (National Adult Reading Test, NART) (F 2,88 = 2.84, p = 0.064). Therefore, all were included as covariates in subsequent analyses.
Facial processing tasks
Accuracy in the BFRT showed a significant group difference (F 2,86 = 5.72, p < 0.001). Post-hoc t tests showed group differences to be significant between patients and relatives (df = 28, p = 0.009) and patients and controls (df = 58, p = 0.001) but not between relatives and controls. This deficit showed a significant linear trend across groups (t = 3.09, p = 0.003). However, mean correct recognition scores are almost indistinguishable between controls and relatives (Table 1).
Accuracy in the facial affect intensity rating test showed significant differences between groups for disgust (F 2,86 = 4.91, p = 0.001), sadness (F 2,86 = 2.94, p = 0.002) and surprise (F 2,86 = 3.10, p = 0.01). However, further post-hoc t tests did not identify any significant differences between groups. The accuracy of facial affect intensity recognition was significantly linear for fear only (t = − 2.1, p = 0.04); patients being more accurate than relatives who were more accurate than controls. The speed of response for this task showed significant differences between groups for fearful (F 2,86 = 3.73, p = 0.004), happy (F 2,86 = 2.75, p = 0.024) and surprised facial stimuli (F 2,86 = 3.69, p = 0.005). Post-hoc t test comparisons showed these significant differences in speed of facial affect processing to be between patient and controls for fear (df = 58, p < 0.001), happiness (df = 58, p = 0.001) and surprise (df = 58, p < 0.001); between relatives and controls for fear (df = 59, p = 0.010) and happiness (df = 59, p = 0.009); and between relatives and patients for surprise (df = 28, p = 0.017).
A linear trend was identified in speed of response across groups for each emotion: anger (t = − 2.15, p = 0.034), disgust (t = − 2.2, p = 0.031), fear (t = − 4.3, p < 0.001), happiness (t = − 3.5, p = 0.001), sadness (t = − 2.9, p = 0.005) and surprise (t = − 4.2, p < 0.001) (Fig. 1).
Fig. 1. Facial affect processing speed (ms) by group, for each emotion, showing significantly linear response times (RTs) for each emotion, such that patients are longest to respond, controls are quickest and relatives are intermediate. Error bars represent the standard error of the mean. * Significant at 5% level. ** Significant at 1% level.
The BFRT recognition hits did not correlate significantly with speed of response in the facial affect intensity rating task, for any emotion.
sMRI results
The ANCOVA model applied to whole-brain GM maps identified no significant clusters, nor did the whole-brain comparison between those patients on and off antipsychotic medication.
The first LV pair from the PLS analysis accounted for 28.71% of covariance between GM density and affect processing speeds (p = 0.047) (Fig. 2a). This LV image involves mainly temporal regions such as the right superior temporal gyrus, bilateral middle temporal gyrus, right rectal gyrus and limbic regions such as the right amygdala and bilateral parahippocampal gyrus. Frontal lobe, right rectal gyrus and left subcallosal gyrus also feature in this system along with bilateral cerebellum (see online Supplementary Table S1). A strong negative correlation between GM and RT (mean r = − 0.62, across all emotions) was displayed by controls but not by relatives (r = − 0.31) or probands (r = − 0.06) (Fig. 2b). There were no significant differences between the groups in extracted GM density.
Fig. 2. (a) Facial affect processing system, identified by partial least squares (PLS). This grey matter (GM) system represents where GM density was most strongly correlated with emotion intensity rating response time (RT), across all emotions and all groups. Red/yellow regions (positive saliences) indicate where increased GM density is correlated negatively with RT (i.e. with quicker RTs). Blue regions (negative saliences) indicate where increased GM density is correlated positively with RT (i.e. slower RTs). R and L markers indicate the right and left side of the brain respectively and numbers denote the z dimension of each slice in Montreal Neurological Institute (MNI) space. The image has been thresholded at Z > ± 1.96. (b) Scatterplots of GM density versus facial affect processing speed. GM density is from the positive regions of the latent variable (LV) image (a), plotted against facial intensity rating RTs (ms), separated by group and presented for each emotion. A Fisher's r to Z transform identified significant difference in the strength of correlation between patient and control groups for all emotions: anger (p = 0.035), disgust (p = 0.008), fear (p = 0.029), happiness (p = 0.017), sadness (p = 0.002) and surprise (p = 0.019). There were significant differences between relative and control groups for anger (p = 0.015), disgust (p = 0.019) and sadness (p = 0.013). There were no significant differences between patient and relative groups for any emotion except happiness (p = 0.001).
Accuracy in BFRT performance did not correlate significantly with brain scores across any group. The PLS analysis of BFRT accuracy and GM density did not produce any significant LVs.
Discussion
These data show that speed of emotional intensity rating exhibits a graded profile according to genetic proximity to psychosis: relatives demonstrated facial affect processing speeds intermediate to patients and healthy controls. This is evident across all six of the emotions tested rather than being restricted to just one. Normal variation in facial affect processing speed is strongly associated with anatomical variation in a cortical-limbic GM system for healthy controls. Normally, greater GM in emotionally salient brain regions is predictive of quicker facial affect processing; however, GM density in this system is a poor predictor for relatives, and even poorer for patients.
Therefore, there are two discoveries: (1) facial affect processing speed presents as a potential endophenotype of FEP, and (2) disruption to the anatomical and functional relationship between a cortical-limbic GM system and facial affect processing speed is associated with psychosis.
Affect processing speed and related brain systems
Although there were some significant differences in how accurately groups identified facial affect intensity (as shown by the main effect of group in an ANCOVA model), post-hoc comparisons did not reach significance (Table 1). Only accuracy of fear recognition was significantly linear, identifying patients to be more accurate than relatives who were more accurate than controls. In contrast, the speed with which participants responded within this task was significantly compatible with a ‘genetic dose–response’ relationship for each emotion (Fig. 1). In some cases, speed of response group differences did not reach significance in post-hoc testing. However, as the speed of response for each group showed a significant linear trend across all emotions, this may be attributable to a lack of power. Therefore, affect processing speed seems to be a candidate endophenotype of FEP, whereas affect processing accuracy does not. These variables have previously been shown to be differentially sensitive across several cognitive domains in a large family study of schizophrenia (Gur et al. Reference Gur, Nimgaonkar, Almasy, Calkins, Ragland, Pogue-Geile, Kanes, Blangero and Gur2007).
We have shown that facial affect intensity rating speed is significantly linked to variation in GM density in a distributed system that predominantly includes cortico-limbic regions involved in face perception and emotional processing, namely the visual areas (fusiform gyri, occipital gyri and lingual gyri), limbic areas (amygdala and parahippocampal gyri), subcortical regions (putamen), temporal areas and prefrontal areas (Fusar-Poli et al. Reference Fusar-Poli, Placentino, Carletti, Landi, Allen, Surguladze, Benedetti, Abbamonte, Gasparotti, Barale, Perez, McGuire and Politi2009). The system also includes regions of the cerebellum. Individuals from the control group who identified the intensity of emotion quickest tended to have greater GM density in these regions.
The superior temporal sulcus and lateral fusiform gyrus bilaterally were highly correlated with emotion intensity rating speed in the system. The superior temporal sulcus is key in the representation of variant aspects of faces (Engell & Haxby, Reference Engell and Haxby2007). The lateral fusiform gyrus is involved primarily in the invariant aspects of facial perception, that is facial identity (Hoffman & Haxby, Reference Hoffman and Haxby2000). The fusiform gyrus has also been implicated in facial affect processing deficits (Quintana et al. Reference Quintana, Wong, Ortiz-Portillo, Marder and Mazziotta2003), with further evoked response potential (ERP) and fMRI studies indicating its involvement at the early stages of any form of face processing, independent from emotional valence (Johnston et al. Reference Johnston, Stojanov, Devir and Schall2005; Fusar-Poli et al. Reference Fusar-Poli, Placentino, Carletti, Landi, Allen, Surguladze, Benedetti, Abbamonte, Gasparotti, Barale, Perez, McGuire and Politi2009). Therefore, involvement of the fusiform gyrus in the affect processing system reported here indicates that emotion intensity recognition is affected from the earliest stages of face perception. Furthermore, it suggests that this cortico-limbic system is not specific to any one emotion but is implicated in emotion processing more generally. This is further supported by the pattern of group differentiation in speed and correlating brain scores to be conserved across all emotions.
The inferior occipital gyri are considered to provide input into both the superior temporal sulcus and fusiform gyri because of their anatomical location (Haxby et al. Reference Haxby, Hoffman and Gobbini2000). The bilateral superior temporal gyrus, which is linked to word perception (Wise et al. Reference Wise, Chollet, Hadar, Friston, Hoffner and Frackowiak1991), has also been suggested to play a role in the ‘deciphering of emotional content from facial stimuli’ (Phillips et al. Reference Phillips, Young, Scott, Calder, Andrew, Giampietro, Williams, Bullmore, Brammer and Gray1998). The amygdala has putatively been shown to be involved in processing fearful faces (Adolphs et al. Reference Adolphs, Tranel, Damasio and Damasio1995) but has more recently been identified to play a general-purpose function in emotional processing (Fitzgerald et al. Reference Fitzgerald, Angstadt, Jelsone, Nathan and Phan2006). The anterior insula is predominantly associated with detecting and experiencing disgust (Phillips et al. Reference Phillips, Young, Senior, Brammer, Andrew, Calder, Bullmore, Perrett, Rowland, Williams, Gray and David1997).
It is to be expected that cerebellar regions would appear in an anatomical network associated with speed of response. However, variation in emotion intensity rating speed is predominantly correlated with anatomical regions implicated in facial affect processing. Thus, the widely distributed nature of this system indicates that group differentiation is not purely determined by a motor deficit.
We used the BFRT to explore facial identity processing per se. Patients were significantly impaired in their ability to recognize faces compared to relatives and controls. There was some evidence to suggest that relatives were intermediate in their ability to recognize faces, yet this was not statistically significant. It is possible that BFRT accuracy is a marker of the disorder rather than an endophenotype. Importantly, there was no significant correlation between accuracy in BFRT performance and speed of response measures or brain scores from the PLS analysis. This indicates that the anatomical network is primarily an affect processing system and supports the notion that there is considerable divergence between facial identity and affect processing (Bruce & Young, Reference Bruce and Young1986).
FEP endophenotypes
Antipsychotic medication has been associated with changes in brain structure (Smieskova et al. Reference Smieskova, Fusar-Poli, Allen, Bendfeldt, Stieglitz, Drewe, Radue, McGuire, Riecher-Rossler and Borgwardt2009). A meta-analysis of primary MRI studies of the neuroanatomy of antipsychotic-naïve patients with FEP or at high risk of psychosis identified that vulnerability to psychosis is associated with distributed GM decreases (Fusar-Poli et al. Reference Fusar-Poli, Radua, McGuire and Borgwardt2011). However, the present study identified no significant effect of antipsychotic use. GM deficits associated with FES have been characterized extensively (Ellison-Wright et al. Reference Ellison-Wright, Glahn, Laird, Thelen and Bullmore2008), but the present study identified no significant GM differences between groups. This result is difficult to interpret. It may reflect the heterogeneity of the sample: there are differences, and also similarities, between the structural profiles of schizophrenia and affective psychoses (McDonald et al. Reference McDonald, Bullmore, Sham, Chitnis, Wickham, Bramon and Murray2004; Kuroki et al. Reference Kuroki, Shenton, Salisbury, Hirayasu, Onitsuka, Ersner-Hershfield, Yurgelun-Todd, Kikinis, Jolesz and McCarley2006).
Several studies across various stages of psychosis have shown reduced GM in regions specifically included in this affect processing system. Goghari et al. (Reference Goghari, Macdonald and Sponheim2011) investigated GM volume across five a priori defined regions bilaterally (fusiform, superior temporal, middle temporal, amygdala and hippocampal regions), in patients with schizophrenia, their non-psychotic relatives and controls. Patients showed reduced hippocampal, middle temporal and fusiform volumes compared to controls. Additionally, their reduced temporal GM was associated with poor facial emotion recognition. Reduction in left fusiform gyri GM was also identified in the relatives group, indicating fusiform volume to be related to genetic liability to schizophrenia. A study observing FEP patients identified reductions in temporal gyri GM volumes (Kuroki et al. Reference Kuroki, Shenton, Salisbury, Hirayasu, Onitsuka, Ersner-Hershfield, Yurgelun-Todd, Kikinis, Jolesz and McCarley2006). A longitudinal study following people with high familial risk of psychosis identified significant differences in GM density in the left fusiform gyrus, left superior temporal gyrus, right amygdala and bilateral parahippocampal gyrus (Job et al. Reference Job, Whalley, Johnstone and Lawrie2005). Subsequently, left inferior temporal gyrus GM density reduction was suggested as a robust predictive measure of schizophrenia development (Job et al. Reference Job, Whalley, McIntosh, Owens, Johnstone and Lawrie2006). This was replicated when looking directly at GM volume reduction (Pantelis et al. Reference Pantelis, Velakoulis, McGorry, Wood, Suckling, Phillips, Yung, Bullmore, Brewer, Soulsby, Desmond and McGuire2003; Schaufelberger et al. Reference Schaufelberger, Duran, Lappin, Scazufca, Amaro, Leite, de Castro, Murray, McGuire, Menezes and Busatto2007).
Conversely, in the present study total GM density within the facial affect system showed no significant group differences. However, when GM density was observed in parallel with emotion intensity rating speed, the degree of positive correlation between GM density and RT showed group differentiation, such that relatives exhibited a level of correlation between those of patients and controls. This relationship between affect processing speed and a cortico-limbic network is therefore meaningfully linked to genetic risk of psychosis and informs potential mechanisms for the candidate endophenotype, facial affect processing speed.
Variation in the anatomical structure of brain systems can be explained by normal variation in genetic and environmental factors that act upon the developing brain. In the healthy brain, it is a viable hypothesis that neuroanatomical integrity is reflected functionally. Hence, for this GM system, an individual with greater GM density in emotionally salient regions will be better equipped to process emotional faces quicker than those with less GM in these regions. In this context, it is difficult to interpret our findings in patients as there is a disconnection from the normative association of structure and function. A key feature of normal neurodevelopment is axonal connectivity of anatomically distinct brain regions. Advances in diffusion tensor imaging (DTI) have highlighted the extent of axonal disorganization associated with schizophrenia (Kanaan et al. Reference Kanaan, Barker, Brammer, Giampietro, Shergill, Woolley, Picchioni, Toulopoulou and McGuire2009), bipolar disorder (Adler et al. Reference Adler, Holland, Schmithorst, Wilke, Weiss, Pan and Strakowski2004), FEP (Perez-Iglesias et al. Reference Perez-Iglesias, Tordesillas-Gutierrez, McGuire, Barker, Roiz-Santianez, Mata, de Lucas, Rodriguez-Sanchez, Ayesa-Arriola, Vazquez-Barquero and Crespo-Facorro2010), and in those at high risk of psychosis (Karlsgodt et al. Reference Karlsgodt, Niendam, Bearden and Cannon2009). With the weight of evidence for neuronal dysconnectivity underpinning the psychopathology of psychosis, slower facial affect processing speeds of the patient and relative groups could be interpreted as reflecting dysconnectivity of the emotional cortical-limbic system. Therefore, GM density in affect processing brain regions is within normal limits; however, processing speeds are significantly slower due to the compromised integrity of their connections. Alternatively, it could be that patients with psychosis process emotional faces in a distinct way altogether. Fakra et al. (Reference Fakra, Salgado-Pineda, Delaveau, Hariri and Blin2008) investigated two different conditions of facial affect processing with schizophrenic patients and matched controls using fMRI. They too identified no differences in accuracy in either condition across both groups, but identified a significant RT difference; schizophrenic patients took longer to respond than controls in the condition with least cognitive demand (i.e. the most emotionally intuitive condition). Patients also failed to elicit the same activation pattern in limbic regions, such as the amygdala, during this task condition. The authors attributed this slowness to patients adopting a more ‘cognitive approach’ when processing facial affect, regardless of the cognitive loading of task. In other words, schizophrenia patients possibly used a feature-based strategy of facial affect recognition that is slower (although accurate) compared to normative configural processing.
Methodological considerations
We must consider the trend towards significant age, gender and NART differences between groups as a limitation to the sample. A recent meta-analysis documents the impact of demographic factors on facial emotion perception results (Kohler et al. Reference Kohler, Walker, Martin, Healey and Moberg2010).
The strength of multivariate analyses is their sensitivity in identifying patterns between two independent data sets. However, they are not ideal for probing specific regions. Therefore, we cannot make direct parallels with univariate results. It should also be noted that GM density, not volume, is discussed. Using a linear registration, as we have, to produce density maps does not achieve such precise alignment of local anatomical structure as using a non-linear registration. However, it may be favourable when observing group-wise structural differences because volumetric maps have been criticized for being ‘overaligned’, potentially removing significant structural differences within the registration process (Toews et al. Reference Toews, Wells, Collins and Arbel2010).
Working with first-episode patients minimizes the effect of confounding variables such as chronic antipsychotic use, different treatment settings and a possible effect of chronicity, whereby the abnormal activity associated with the disorder itself may have secondary effects on structural parameters in the long term. Although we used an advantageous group for probing a broader psychosis spectrum, we may have diluted power to identify endophenotypes specific to a differentiated psychotic illness; despite the evidence for convergence of genetic risk factors, there must also be differentiating factors. We ascertained diagnosis at clinical follow-up using a structured assessment, but the missing follow-up patient diagnoses (27%) in the present study limits further investigation of our results in more homogeneous patient subgroups. Secondary studies are needed.
In conclusion, we found that accuracy of facial affect processing does not seem to be genetically linked to psychosis but speed of processing does. Therefore, speed of facial affect processing may be a candidate endophenotype of FEP. Greater GM density within a distributed facial affect processing system was predictive of quicker processing speeds in healthy controls, yet this coupling of structure and function was disrupted in patients and their unaffected relatives. More studies are needed to probe these neural underpinnings of facial affect processing speed as an endophenotype in more chronic, homogeneous patient samples.
Supplementary material
For supplementary material accompanying this paper visit http://dx.doi.org/10.1017/S0033291712001341.
Acknowledgements
This study was sponsored by Medicines Discovery and Development, GlaxoSmithKline R&D (GSK TMT106522). We thank the participants for their time and effort.
Declaration of Interest
A.M.D. holds an industrial Co-operative Award in Science and Engineering (CASE) studentship with the Biotechnology and Biological Sciences Research Council (BBSRC) and GlaxoSmithKline. Both P.J.N. and E.T.B. are employed by and hold shares in GlaxoSmithKline.
