Psychophysiology, 51 (2014), 1321–1322. Wiley Periodicals, Inc. Printed in the USA. Copyright © 2014 Society for Psychophysiological Research DOI: 10.1111/psyp.12351
COMMENTARY
Do our “big data” in genetic analysis need to get bigger?
LAURA A. BAKER Psychology Department, University of Southern California, Los Angeles, California, USA
Abstract Individual papers in this special issue might seem disappointing in their lack of discovery of specific genes of potential relevance to mental disorders. Yet, collectively, they yield information that could not be gleaned otherwise. Combining genome-wide complex trait analysis and classic approaches to estimate heritability in the same sample, and supplementing genome-wide association studies of common variants with exome and sequencing analyses, provides an unprecedented opportunity to examine major issues encountered in genetic research of complex traits, in ways not easily done with a series of unrelated studies using different samples, measures, and analytical approaches. Extending molecular genetic approaches to fully multivariate analyses will be an important future direction. These will require bigger analyses of even bigger big data, but will be essential in efforts to redefine psychopathology in the Research Domain Criteria (RDoC) approach promoted in the NIMH strategic plan. Descriptors: Multivariate GWAS, RDoC, Missing heritability
Coupling GWAS SNP heritabilities (which focus on common variants) with exome chip analyses (focusing on rare variants) and sequencing analyses (covering both rare and common variants) further helps explain the missing heritability, by addressing the issue of DNA chips used in GWAS being composed of common variants. Rare variants could still account for the heritability not explained by GCTA, since the exome chip used covers only a small portion (2%) of such effects (Vrieze et al., 2014). But the combined analyses for these 17 indicators using the same sample at least provides a more comprehensive consideration, and future analyses with additional DNA chips in this sample will shed more light on the rare variant hypothesis. Another explanation of the missing heritability in these papers that cannot be easily resolved even with their comprehensive, multipronged approach, however, is that different genes may be important at different ages. This is suggested by systematically higher correlations for DZ (dizygotic) twins compared to parents and offspring (rPO < rDZ). Developmental changes in relative sizes of genetic variance for skin conductance measures (e.g., orienting responses) do appear across childhood and adolescence (Tuvblad et al., 2012), although it is yet unknown (a) how stable genetic influences are between onset of adulthood (i.e., age 17 for twins’ assessments) and midlife (i.e., age of Minnesota Twin Family Study [MTFS] parents) and, importantly, (b) whether the same genes are important across ages. It could be advantageous to conduct separate within-generation analyses, although this would necessarily compromise power due to smaller Ns in the MTFS samples. There is a clear need for much larger GWAS and GTCA studies of narrower age ranges in single generations (e.g., twins and other siblings). Notably, these papers used only univariate regression models, while multivariate structural models may be required to optimize
This special issue provides an impressive set of papers describing various levels of analyses of an enormous amount of “big data” to understand the genetic basis of 17 psychophysiology measures, each having demonstrated significance in one or more psychiatric disorders. To consider these as true endophenotypes, and whether they are actually “closer to the genes” than the disorders themselves, we must understand not only their heritability, but the exact nature of their genetic influences. That is, after all, the primary objective for identifying and studying endophenotypes in the first place. Any of the genome-wide complex trait analysis (GTCA)-based “SNP (single nucleotide polymorphism) heritabilities” of the 17 psychophysiology variables in the five genome-wide association studies (GWAS) papers, if considered in isolation, might simply appear as further illustrations of the “missing heritability” problem found for other complex human traits (Visscher, Brown, McCarthy, & Yang, 2012). However, using both GCTA and classic biometrical (twin-family) approaches to estimate heritability in the same sample helps argue against common explanations for discrepancies between the two approaches, that is, nonadditive genetic effects and gene by environment (G×E) interactions (Visscher et al., 2012). The idea that such effects inflate heritability estimates from twin correlations can be reasonably ruled out given the median rMZ is approximately equal to 2*rDZ. Of course, nonadditive effects could still be important, and there is indeed variation in the magnitude of rMZ–2*rDZ across measures. But nonadditive effects cannot be used across the board to explain the majority of “missing heritability” in these 17 measures.
bs_bs_banner
Address correspondence to: Laura A Baker, Psychology Department (SGM 501), University of Southern California, Los Angeles CA 900891061, USA. E-mail: [email protected] 1321
1322
L.A. Baker
information gained from these already big data. Multivariate models could be highly informative in molecular genetic analyses, in several ways in the search for specific genes or SNPs. First, combining structural equation modeling (SEM) with GWAS would allow “true score” factor models for endophenotypes (i.e., using multiple indicators), reducing measurement error and potentially increasing power for finding genetic associations. Second, the same approach could incorporate multiple endophenotypes into one model, to identify common (pleiotropic) SNPs or candidate genes across measures. Finally, joint analysis of the disorder with purported endophenotypes could help identify common genetic mechanisms—the primary objective of the endophenotype approach. The extent to which the same candidate genes or SNPs, or clusters of them, influence multiple measures (e.g., Medland & Neale, 2009; Visscher et al., 2014) could be fruitful in discerning the genetic basis of endophenotypes and their relationships to psychopathology. Such multivariate extensions—although computationally demanding—will undoubtedly prove useful in considerations of these psychophysiology measures, in combination with one another and with other biological indicators (e.g., HPA axis functioning)—in the wider realm of Research Domain Criteria (RDoC) being promoted through the NIMH strategic plan for redefining psychopathology. Finally, the utility of GWAS and other molecular analyses of endophenotypes for discovering genes for psychopathology must be considered, in light especially of small correlations typically
found between psychophysiology measures and behavior. Even if both the disorder and endophenotype have moderate to strong heritabilities, and even with genetic correlations (rG) of 1.0 between endophenotype and disorder, the genetic influences on each will be largely specific rather than common to the two. For example, while there is a robust inverse correlation between resting heart rate and antisocial behavior, which is entirely due to correlated genetic effects (i.e., rG = −1.0), the genetic effects in heart rate (h2 = 0.49) explain < 1% of heritability of antisocial behavior (Baker et al., 2009). Thus, finding SNPs or candidate genes associated with an endophenotype is no guarantee that these will in turn be important to the disorder. For example, there are likely genes involved in auditory or visual systems that influence heritability of measures such as P300 or startle eye blink, which have no impact on correlated disorders. This highlights the need for multivariate extensions of GWAS approaches to enable joint analysis of both endophenotype and behavior in order to find common genetic mechanisms. Our models thus need to get bigger in many ways (e.g., multivariate extensions of GWAS; much larger Ns), but perhaps smaller in others. Examining narrower groups of individuals (e.g., based on age, generation, sex), might be required to increase power for finding specific genetic associations. This will, of course, require even larger Ns, which speaks to the importance of multisite collaborations and combining samples. Increasing speed and sophistication of computing power will undoubtedly yield opportunities for bigger analyses of bigger data.
References Baker, L. A., Tuvblad, C., Reynolds, C., Zheng, M., Lozano, D. I., & Raine, A. (2009). Resting heart rate and the development of antisocial behavior from age 9 to 14: Genetic and environmental influences. Development and Psychopathology, 21, 939–960. Medland, S. E., & Neale, M. C. (2009). An integrated phenomic approach to multivariate allelic association. European Journal of Human Genetics, 18, 233–239. Tuvblad, C., Gao, Y., Isen, J., Botwick, T., Raine, A., & Baker, L. A. (2012). The heritability of the skin conductance orienting response: A longitudinal twin study. Biological Psychology, 89, 47–53. doi: 10.1016/ j.biopsycho.2011.09.003
Visscher, P. M., Brown, M. A., McCarthy, M. I., & Yang, J. (2012). Five years of GWAS discovery. American Journal of Human Genetics, 90, 7–24. doi: 10.1016/j.ajhg.2011.11.029 Visscher, P. M., Hemani, G., Vinkhuyzen, A. A. E., Chen, G.-B., Lee, S. H., Wray, N. R., . . . Yang, J. (2014). Statistical power to detect genetic (co) variance of complex traits using SNP data in unrelated samples. PLoS Genetics, 10, e1004269. Vrieze, S. I., Malone, S. M., Pankratz, N., Vaidyanathan, U., Miller, M. B., Kang, H. M., . . . Iacono, W. G. (2014). Genetic associations of nonsynonymous exonic variants with psychophysiological endophenotypes. Psychophysiology, 51, 1300–1308.
COMMENTARY
Do our “big data” in genetic analysis need to get bigger?
LAURA A. BAKER Psychology Department, University of Southern California, Los Angeles, California, USA
Abstract Individual papers in this special issue might seem disappointing in their lack of discovery of specific genes of potential relevance to mental disorders. Yet, collectively, they yield information that could not be gleaned otherwise. Combining genome-wide complex trait analysis and classic approaches to estimate heritability in the same sample, and supplementing genome-wide association studies of common variants with exome and sequencing analyses, provides an unprecedented opportunity to examine major issues encountered in genetic research of complex traits, in ways not easily done with a series of unrelated studies using different samples, measures, and analytical approaches. Extending molecular genetic approaches to fully multivariate analyses will be an important future direction. These will require bigger analyses of even bigger big data, but will be essential in efforts to redefine psychopathology in the Research Domain Criteria (RDoC) approach promoted in the NIMH strategic plan. Descriptors: Multivariate GWAS, RDoC, Missing heritability
Coupling GWAS SNP heritabilities (which focus on common variants) with exome chip analyses (focusing on rare variants) and sequencing analyses (covering both rare and common variants) further helps explain the missing heritability, by addressing the issue of DNA chips used in GWAS being composed of common variants. Rare variants could still account for the heritability not explained by GCTA, since the exome chip used covers only a small portion (2%) of such effects (Vrieze et al., 2014). But the combined analyses for these 17 indicators using the same sample at least provides a more comprehensive consideration, and future analyses with additional DNA chips in this sample will shed more light on the rare variant hypothesis. Another explanation of the missing heritability in these papers that cannot be easily resolved even with their comprehensive, multipronged approach, however, is that different genes may be important at different ages. This is suggested by systematically higher correlations for DZ (dizygotic) twins compared to parents and offspring (rPO < rDZ). Developmental changes in relative sizes of genetic variance for skin conductance measures (e.g., orienting responses) do appear across childhood and adolescence (Tuvblad et al., 2012), although it is yet unknown (a) how stable genetic influences are between onset of adulthood (i.e., age 17 for twins’ assessments) and midlife (i.e., age of Minnesota Twin Family Study [MTFS] parents) and, importantly, (b) whether the same genes are important across ages. It could be advantageous to conduct separate within-generation analyses, although this would necessarily compromise power due to smaller Ns in the MTFS samples. There is a clear need for much larger GWAS and GTCA studies of narrower age ranges in single generations (e.g., twins and other siblings). Notably, these papers used only univariate regression models, while multivariate structural models may be required to optimize
This special issue provides an impressive set of papers describing various levels of analyses of an enormous amount of “big data” to understand the genetic basis of 17 psychophysiology measures, each having demonstrated significance in one or more psychiatric disorders. To consider these as true endophenotypes, and whether they are actually “closer to the genes” than the disorders themselves, we must understand not only their heritability, but the exact nature of their genetic influences. That is, after all, the primary objective for identifying and studying endophenotypes in the first place. Any of the genome-wide complex trait analysis (GTCA)-based “SNP (single nucleotide polymorphism) heritabilities” of the 17 psychophysiology variables in the five genome-wide association studies (GWAS) papers, if considered in isolation, might simply appear as further illustrations of the “missing heritability” problem found for other complex human traits (Visscher, Brown, McCarthy, & Yang, 2012). However, using both GCTA and classic biometrical (twin-family) approaches to estimate heritability in the same sample helps argue against common explanations for discrepancies between the two approaches, that is, nonadditive genetic effects and gene by environment (G×E) interactions (Visscher et al., 2012). The idea that such effects inflate heritability estimates from twin correlations can be reasonably ruled out given the median rMZ is approximately equal to 2*rDZ. Of course, nonadditive effects could still be important, and there is indeed variation in the magnitude of rMZ–2*rDZ across measures. But nonadditive effects cannot be used across the board to explain the majority of “missing heritability” in these 17 measures.
bs_bs_banner
Address correspondence to: Laura A Baker, Psychology Department (SGM 501), University of Southern California, Los Angeles CA 900891061, USA. E-mail: [email protected] 1321
1322
L.A. Baker
information gained from these already big data. Multivariate models could be highly informative in molecular genetic analyses, in several ways in the search for specific genes or SNPs. First, combining structural equation modeling (SEM) with GWAS would allow “true score” factor models for endophenotypes (i.e., using multiple indicators), reducing measurement error and potentially increasing power for finding genetic associations. Second, the same approach could incorporate multiple endophenotypes into one model, to identify common (pleiotropic) SNPs or candidate genes across measures. Finally, joint analysis of the disorder with purported endophenotypes could help identify common genetic mechanisms—the primary objective of the endophenotype approach. The extent to which the same candidate genes or SNPs, or clusters of them, influence multiple measures (e.g., Medland & Neale, 2009; Visscher et al., 2014) could be fruitful in discerning the genetic basis of endophenotypes and their relationships to psychopathology. Such multivariate extensions—although computationally demanding—will undoubtedly prove useful in considerations of these psychophysiology measures, in combination with one another and with other biological indicators (e.g., HPA axis functioning)—in the wider realm of Research Domain Criteria (RDoC) being promoted through the NIMH strategic plan for redefining psychopathology. Finally, the utility of GWAS and other molecular analyses of endophenotypes for discovering genes for psychopathology must be considered, in light especially of small correlations typically
found between psychophysiology measures and behavior. Even if both the disorder and endophenotype have moderate to strong heritabilities, and even with genetic correlations (rG) of 1.0 between endophenotype and disorder, the genetic influences on each will be largely specific rather than common to the two. For example, while there is a robust inverse correlation between resting heart rate and antisocial behavior, which is entirely due to correlated genetic effects (i.e., rG = −1.0), the genetic effects in heart rate (h2 = 0.49) explain < 1% of heritability of antisocial behavior (Baker et al., 2009). Thus, finding SNPs or candidate genes associated with an endophenotype is no guarantee that these will in turn be important to the disorder. For example, there are likely genes involved in auditory or visual systems that influence heritability of measures such as P300 or startle eye blink, which have no impact on correlated disorders. This highlights the need for multivariate extensions of GWAS approaches to enable joint analysis of both endophenotype and behavior in order to find common genetic mechanisms. Our models thus need to get bigger in many ways (e.g., multivariate extensions of GWAS; much larger Ns), but perhaps smaller in others. Examining narrower groups of individuals (e.g., based on age, generation, sex), might be required to increase power for finding specific genetic associations. This will, of course, require even larger Ns, which speaks to the importance of multisite collaborations and combining samples. Increasing speed and sophistication of computing power will undoubtedly yield opportunities for bigger analyses of bigger data.
References Baker, L. A., Tuvblad, C., Reynolds, C., Zheng, M., Lozano, D. I., & Raine, A. (2009). Resting heart rate and the development of antisocial behavior from age 9 to 14: Genetic and environmental influences. Development and Psychopathology, 21, 939–960. Medland, S. E., & Neale, M. C. (2009). An integrated phenomic approach to multivariate allelic association. European Journal of Human Genetics, 18, 233–239. Tuvblad, C., Gao, Y., Isen, J., Botwick, T., Raine, A., & Baker, L. A. (2012). The heritability of the skin conductance orienting response: A longitudinal twin study. Biological Psychology, 89, 47–53. doi: 10.1016/ j.biopsycho.2011.09.003
Visscher, P. M., Brown, M. A., McCarthy, M. I., & Yang, J. (2012). Five years of GWAS discovery. American Journal of Human Genetics, 90, 7–24. doi: 10.1016/j.ajhg.2011.11.029 Visscher, P. M., Hemani, G., Vinkhuyzen, A. A. E., Chen, G.-B., Lee, S. H., Wray, N. R., . . . Yang, J. (2014). Statistical power to detect genetic (co) variance of complex traits using SNP data in unrelated samples. PLoS Genetics, 10, e1004269. Vrieze, S. I., Malone, S. M., Pankratz, N., Vaidyanathan, U., Miller, M. B., Kang, H. M., . . . Iacono, W. G. (2014). Genetic associations of nonsynonymous exonic variants with psychophysiological endophenotypes. Psychophysiology, 51, 1300–1308.
