ENIGMA and the individual: Predicting factors that affect the brain in 35 countries worldwide

1Here we provide an update on the progress of the ENIGMA consortium, a global alliance of over 500 scientists from over 200 institutions in 35 countries to study brain imaging data worldwide, discovering factors that modulate brain structure, integrity, connectivity, and patterns of brain differences in major brain diseases. Founded in 2009, ENIGMA's initial aims were to perform genome-wide analyses to identify common variants in the genome that are reliably associated with normal variability in brain structure. Since the initial effort discovered consistent effects worldwide of genetic variants that explained less than 1% of the variance in brain measures (; Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,, ), over 500 scientists have joined ENIGMA. ENIGMA is now (as of October 2015) a worldwide consortium, organized into over 30 working groups, studying major brain diseases (detailed at http://enigma.ini.usc.edu). The work in ENIGMA is divided into projects on (1) genetics, screening genomic data for predictors of individual variations in brain structure, function, and connectivity; (2) disease, screening brain measures to identify patterns of differences in the major brain diseases and factors that affect them; and (3) methods development. New “Big Data” methods are being developed and implemented around the world to perform genetic analysis of high-dimensional features that arise in neuroimaging — such as brain networks or “connectomes”, 3D or 4D maps of brain changes over time, and more complex imaging data from functional MRI and EEG/MEG.

2For this issue of NeuroImage we review the work ENIGMA has done, and how it relates to making individual predictions to support the emerging discipline of precision medicine — where personalized medical decisions are made considering an individual's genetic make-up, other risk factors, and the large body of scientific knowledge detailing genotype-phenotype relationships. ENIGMA's genetic and disease-related studies are discovering new factors that affect the brain throughout life, how the diseased brain differs from the healthy brain, and how patterns of brain measures differ from one disease to another. The potential to use machine learning methods in this context is vast, and we point to future opportunities and challenges, and what we have learned already about how individual genetic variants and diseases affect the brain.

3One major thrust of ENIGMA's work is genomics, so we first review studies that discovered individual loci in the genome that are linked to variations in brain structure (Stein et al., 2012; Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,, ). The effect of these common genetic variants tends to be small, but the aggregate effect of thousands of them accounts for a substantial proportion of the variance in brain measures. The relevant genes can be difficult to discover in individual cohorts, but they can be detected by meta-analyzing data across multiple sites. We discuss multivariate and machine learning methods needed to combine some of these predictors in more powerful models that can make valuable predictions about individuals, such as predicting deviations from normal lifetime aging, risk for mental illness, or recovery from trauma.

4There have been numerous recent surprises regarding the nature of gene effects on the brain, including surprisingly poor reproducibility of candidate gene effects on imaging measures and risk for mental illness, and the very large sample sizes needed to reliably detect any genetic associations at all. There have also been dramatic claims of poor reproducibility of findings in genetics, neuroimaging, and neuroscience studies in general. Meta-analyses, such as those conducted by ENIGMA, have been proposed as a way to screen for false positive findings. If claims of “significance chasing” and “fishing” in neuroscience studies are true, then predictive models based on them should fail more often than models based on meta-analyzed studies of large numbers of independent cohorts, analyzed in a harmonized way. ENIGMA is dedicated to replication, and a number of initiatives are underway to develop methods to replicate imaging genomics findings.

5We discuss factors that affect reproducibility of models that predict specific gene effects on the brain, including technical factors of image acquisition and analysis. Low effect sizes for individual predictors make genetic effects hard to detect, so meta-analysis is valuable in demonstrating effects that no single cohort can detect on its own. Clearly, if we build a model to classify a person into a certain diagnostic group, based on a set of predictors, we also need to know how to decide if we have measured the predictors well enough, or if the context where the model was fitted is similar enough to the current situation for the prediction to make sense and be accurate. Apart from the choice of predictive model and predictors, there are many other reasons why imaging or genetic models of diagnosis or prognosis may generalize poorly or not at all, depending on the context. Factors that affect model prediction will include age and environment, and the demographic history of the populations sampled; these may affect whether or not a predictor is relevant to a new cohort or an individual. In the ENIGMA studies below, we point to examples in which predictors in the genome and image would be valuable in making individual predictions about brain volume or about a person's diagnosis, but only in certain contexts, such as in certain parts of the lifespan, or only after considering certain confounds or variables that are known to drive brain differences (duration of medication and duration of illness are often confounded, and modeling each effect independently may produce paradoxical conclusions, e.g., that medication is bad for the brain). Individual predictive models are likely to become increasingly nuanced, as we find out more about how predictors interact and contexts where different models work best.

6In the course of ENIGMA's efforts, a vast quantity of normative data has been gathered and analyzed from different countries and continents of the world, allowing us to make some inferences about the normal trajectory of brain development and aging (ENIGMA-Lifespan; ). We discuss the challenges and opportunities in using models based on these data to make assertions about individual and group deviations from normal, or to generate cohort, or national norms, if they exist and if their value outweighs the costs of generating them.

7We also discuss several concepts that have increased the power of ENIGMA to find factors with very small effects on the brain, including how we assess their generality and extensibility to new cohorts.

8By December 2009, many researchers worldwide had collected genome-wide genotyping data from cohorts of subjects for whom brain imaging information such as anatomical MRI was available.

9It had long been presumed that genetic and environmental factors, and the complex interactions among them, play a role in shaping brain structure. Decades of work in behavioral and medical genetics had convincingly shown that many of the major brain diseases – from Alzheimer's and Parkinson's disease to psychiatric illnesses such as schizophrenia and major depression – had a strong additive genetic component. Similar genetic risks exist for neurodevelopmental disorders such as autism. Even so, studies of identical twins who share the same genome show that genetic factors do not fully account for disease risk, and discordant twin pairs provide valuable information about the impact of environmental and epigenetic factors on disease. Furthermore, many common disorders are likely to reflect a constellation of modest gene differences acting in concert, which smaller individual studies are unlikely to find. Instead, larger studies that capture heterogeneity have begun to unravel the influence of multiple ‘low level’ minor but important gene differences on disease expression.

10As high-throughput genotyping methods became available, genome-wide association studies (GWASs) began to reveal specific sources of risk in the genome for several major brain diseases (Fig. 1). To fully appreciate this kind of study, we need to understand that much of the genome is invariant between humans. Many kinds of individual genetic variations – common or rare – can occur, including polymorphisms, insertions and deletions of genetic material, loss or retention of homozygosity (LOH/ROH), or copy number variations (CNVs) — where the number of copies of pieces of genomic material differs from the normal two alleles in some individuals but not others. Polymorphisms are a common marker of individual differences, where a single nucleotide polymorphism (SNP) is essentially a “single-letter” change in the genome: a change in a single base pair between individuals.

11Some genomic changes interfere with the viability of the organism, leading to very low frequencies in the population. Others remain and some have a moderate or severe impact on a person's health, or their risk for disease. For example, a common variant (present in 1 in 100 in the general population) in the HFE gene impairs a person's ability to metabolize iron. Excessive iron levels can then accumulate in bodily organs, which can cause liver and kidney failure. Multiple deletions in the 22q region of the genome provide another example. Individuals with these deletions have a characteristic neurodevelopmental profile associated with mild to severe abnormalities in the face, brain, and heart, and are at heightened risk for schizophrenia and autism. 22q deletions occur frequently de novo, so they do not really remain in the population; rather 22q is a vulnerable spot in the genome for mutation. Even so, 22q deletion syndrome – and other neurogenetic disorders such as Fragile X, Williams syndrome, and Turner syndrome – have often been studied to help identify potential mechanisms that may contribute to more prevalent psychiatric conditions. ENIGMA's 22q working group has been set up to understand brain differences associated with deletions at this locus, and how they relate to those found using the same analysis protocols in ENIGMA-Schizophrenia and ENIGMA-Autism.

12Genetic risk for many major psychiatric illnesses is thought to be mediated in part by common genetic variants that have persisted in human populations for thousands of years. In many cases, the adverse effects of disease risk genes – such as the Alzheimer's risk gene, APOE – are not apparent until later in life (Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,, ). Because of this, the variants tend to be preserved in the gene pool and continue to drive disease risk worldwide.

13Geneticists continue to debate the relative contribution of common versus rare genetic variants to risk for various diseases, but a recent large-scale screen of schizophrenia patient cohorts worldwide implicated over 100 genetic loci in risk for the disease (; Fig. 1). This highly successful study pointed to several genes in the dopamine neurotransmission pathway that had long been implicated in schizophrenia and its treatment — for example, a functional polymorphism in the DRD2 promoter region, which modulates levels of gene expression, and affects antipsychotic drug efficacy. This same genomic screen pointed to other unexpected genetic variants in immune system pathways that offer tantalizing new leads about disease mechanisms, and the role of modifiable factors in eventually treating or averting the illness. Similar efforts in bipolar illness, major depression, and ADHD uncovered genes driving risk for these disorders that overlapped to some extent with those for schizophrenia and with each other. Members of the ENIGMA Consortium have recently demonstrated the usefulness of polygenic risk scores for schizophrenia (based on the 108 loci shown in Fig. 1A) in revealing an association between early cannabis use and brain maturation during adolescence — replicated in three samples.

14Many successful genomic screens involve over 100,000 individuals. For example, the most recent GWAS of height, educational attainment, and body mass index (BMI) identified 56 novel BMI-associated loci in a sample of up to 339,224 individuals. Similarly, the Psychiatric Genomics Consortium's discovery of genetic loci implicated in schizophrenia risk took a ‘quantum leap’ once the sample sizes exceeded 75,000, after less successful searches in smaller samples. Several factors may contribute towards this need for large sample sizes in genome-wide association. First, there are biological variation and ascertainment differences among cohorts. A person diagnosed with a specific illness may have other co-morbid illnesses, and diagnostic criteria may vary somewhat worldwide in terms of who is included in the groups of patients and controls.

15However, the main reason GWAS needs large samples is power: a genome-wide association analysis comprises approximately a million independent tests, so a threshold of p < 5 × 10–8 is employed to minimize false positives. Early GWAS estimated their required sample sizes based on published effect sizes of candidate genes that have since been shown to be greatly overestimated. Although the genetic architecture of each trait is unique, for most complex traits the effect sizes of individual SNPs are typically less than half a percent. Thus, it follows from power analyses that GWAS and GWAS meta-analyses typically require data from tens of thousands of individuals.

16In the imaging field, initial studies also attempted genome-wide screens of brain imaging measures, such as brain size, the volume of the temporal lobes on MRI, in cohorts of around 800 subjects (see, for a review). This type of analysis became feasible as large cohort studies, such as the Alzheimer's Disease Neuroimaging Initiative, started to put their images and genomic data online. In line with accepted practice in genetics, it is customary to require replication of such genetic effects in independent cohorts.

17While some effects appeared to replicate, most did not as the studies were underpowered, and it was unclear whether cohort factors, biological differences, or technical factors were to blame.

18As the field of imaging genetics grew, some researchers hoped that imaging might offer a more efficient approach to discover genes involved in mental illness. The reason for this optimism was based on the observation that many brain measures are consistently reported as affected in psychiatric cohort studies (see later, under ENIGMA Disease Studies), so they could maybe serve as quantitative traits, or markers, correlated with the illness.

19There was also some hope that the biological signals in images – measures of neurotransmitters, receptors or metabolite levels, blood flow, the volume of specialized brain areas such as the hippocampus, or its chemical content – might be influenced by genetic variants because of their proximity to primary gene action. Likewise, it was argued that brain-derived measures may have a simpler genetic architecture – perhaps with fewer individual genes or pathways influencing them – compared to the multitude of factors driving a person's overall risk for developing a disease. Brain measures may also offer a more precise or reproducible diagnostic scale. noted that GWAS can be more efficient when researchers analyze continuous measures (such as brain volumes) rather than binary traits, such as diagnosis, which may also disguise complexities such as co-morbidity, etc.

20This endophenotype theory2 led to confidence that genome-wide screening of brain measures would yield “hits” – genetic loci consistently associated with brain measures – relatively efficiently and, some believed, in much smaller samples. Several countervailing arguments should also be considered. The genetics of brain traits may reveal common pathways involved in a number of mental illnesses, but one loses some specificity when moving from a psychiatric disorder to brain measures — different disorders may have very similar brain abnormalities. For this reason, ENIGMA's Disease Working groups have analyzed tens of thousands of brain scans to see which measures best distinguish patients from controls, across a range of 12 diseases, with a view to understanding similarities and differences. Collecting brain imaging data is more expensive than diagnostic testing. Also, genes that affect brain measures may be of less interest to a patient or physician unless they are also connected to disease risk or prognosis. In ENIGMA, however, the costs of collecting the imaging data had already been incurred, making the feasibility of a large-scale analysis the main consideration. Others voiced a muted optimism: noted that effect sizes for gene effects on neuroimaging data were not likely to be any greater than for any other trait, but the value in studying them came from the ability of brain measures to help understand mechanisms that might underlie associations between genes and more conventional traits (see also ). Yet, the potential to find genetic factors that jointly influence risk for mental illness and a neuroimaging trait could dramatically improve statistical power and provide an important link between the genome and the behavioral symptoms used to diagnose psychiatric and neurological illnesses.

21In ENIGMA's first paper in Nature Genetics, Stein and 158 authors (2012), including 4 existing consortia (SYS, EPIGEN, ADNI, and IMAGEN3), meta-analyzed GWAS data from cohorts worldwide and found genetic loci consistently associated with the size of the human hippocampus and total intracranial volume. Notably, in a partnership with another consortium, CHARGE, the top “hits” – the genetic variants with greatest effect sizes – were anonymously exchanged and found to be the same, supporting the replicability of the findings in completely independently designed efforts.

22In a follow-up study in a larger sample (N = 21,151 individuals; Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication; called “ENIGMA2”), eight genetic loci were discovered that were reliably associated with the size (volume) of several subcortical structures, including the putamen, caudate, and pallidum. With the increased sample size, earlier findings regarding the hippocampus and intracranial volume were replicated and reinforced; new genetic loci were also discovered. Several of the SNPs implicated lie within or close to genes involved in cell migration, axon guidance, or apoptosis — all cellular processes likely to lead to observable differences in the size of cellular nuclei in the brain. Parallel work in mice by the Williams lab in Memphis began to study mouse homologs of these variants; recent data suggest that variation of the top putamen gene, KTN1, can predict putamen volume and cell counts in outbred mice (R. Williams, pers. commun.).

23Several lessons were learned from the first two ENIGMA genetic studies, in addition to a third pair of papers currently in submission, involving an even larger sample (N > 31,000; Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,,; Adams and the CHARGE and ENIGMA2 Consortia, submitted for publication). First, through meta-analyses, it was possible to detect factors (here, SNPs) that accounted for less than 1% of the variance in brain measures. This was despite the fact that the participating studies were designed with different goals in mind, and many used scanners of different field strengths, processed by researchers who had not all met, and communicated through email and teleconference calls.

24Much of the consistency in brain measures capitalized on the ongoing refinement of standardized protocols for analyzing images and genomes; in turn, those protocols relied on decades of work by developers of widely used and extensively tested analysis packages such as FreeSurfer, and FSL. The supplement of the first ENIGMA paper (Stein et al., 2012) contained 104 pages of ancillary tests supporting the validity and reliability of the data, including tests comparing different imaging software for brain volume quantification.

25On the genomic side, the ability to compare genomic data in a common reference frame depended on the availability of the HapMap3 (The International HapMap3 Consortium, 2010) and later the 1000 Genomes reference datasets. These reference panels are continually updated and refined, and allow genotyping data collected with one kind of genotyping array (“chip”) to be imputed to match data collected using others, and pooled in the same overall study.

26A second issue is whether these findings could have been detected more efficiently using only some of the samples. In a sense, this is a “meta-question” — how might the study have been designed more efficiently after seeing the results?

27As in any meta-analysis, the weight assigned to each cohort in the final statistics can be made to depend on its total sample size, or on the standard error of the regression coefficients (which is in fact what ENIGMA does). As such, it is not vital for every cohort to reject the null hypothesis on its own. In fact, any cohort study, however small, can partner with other sites to contribute to the discovery of effects that it cannot detect alone. In ENIGMA1 (Stein et al., 2012), only 5 of the 21 cohort studies were able to detect the effect of the SNPs on the brain in their cohort alone, at the nominal significance level of p = 0.05. By the time of ENIGMA2, 20 of the 38 Caucasian European (CEU) cohort studies could detect the effects of the top SNP. Even so, the aggregate support of the discovery and replication samples was crucial to making sure the effects were credible and unlikely to be false positives.

28The quest to identify genetic variants associated with brain measures is partly motivated by finding variants that affect our individual risk for disease. Any modulators of health outcomes in populations may have a vast impact on society, even if they are not the main factors explaining risk for any one individual. As well as affecting risk for disease, genetic differences may also affect symptom severity, treatment response, and prognosis.

29As such, several clinical trials for Alzheimer's disease drugs already stratify their cohorts by APOE genotype — a major risk gene for AD that may have a bearing on treatment response as well as disease risk (see Riedel et al., submitted for publication, for a review of APOE effects, which are remarkably complex). At the time of writing, several manuscripts are under review addressing the overlap between ENIGMA's genomic findings and accepted or emerging markers of disease risk (Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,,; Adams and the CHARGE and ENIGMA2 Consortia, submitted for publication; ). Here we simply review their overall design. Some initial reports have appeared in abstract form, relating brain-related SNPs to risk for Parkinson's disease (Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,, ), obsessive compulsive disorder (Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,, ), schizophrenia, and multiple sclerosis (Rinker et al., submitted for publication). An initial negative report has appeared for epilepsy. Even so, given the low fraction of heritability explained by the SNPs discovered, the studies so far are widely accepted as underpowered.

30One method to assess an individual's relative risk for disease, based on genome-wide genotyping data, involves computing a polygenic risk score (PRS) for each individual. In Alzheimer's disease, for example, carrying one copy of the APOE4 genotype boosts lifetime risk for AD by a factor of 3, and carrying two copies may boost risk by 15 times. These odds ratios are not constant across human populations and even vary by ethnicity, or circumstances, so some caution is needed when extrapolating them to new data; but as AD GWAS data accumulate, over 20 common genetic variants have been found to affect AD risk — 3 of them, in the genes CLU, PICALM, and CR1, appear to be associated with a difference in disease risk of over 10% per allele. If an individual's genotype is known for these loci, it is possible to create a polygenic risk score in a number of different ways, depending on whether the goal is to predict diagnosis, outcome, or brain measures. The simplest approach is to count risk loci, although that clearly ignores the vastly different odds ratios from each locus. It is more common to weight the loci based on their odds ratio for disease, or by their regression coefficients. APOE4, for example, is just a single genotype that might contribute to calculation of a polygenic risk score together with other risk loci. As shown by the PGC analyses, the predictive accuracy of PRS scores increases as the number of variants included increases. Calculation of these scores does not need to be restricted to genome-wide significant loci.

31Recent efforts to predict disease status based on polygenic risk scores have had varied success, but the reasons are quite well understood. First, for the most prevalent neurological or psychiatric diseases, we do not yet have a set of common variants that account for more than a small fraction of disease risk (except for APOE4, where a single copy may triple a person's risk for AD, other factors being equal). In AD, there are rare mutations in genes related to AD pathology – such as presenilin and APP – that invariably produce early-onset AD. Carriers of these genetic variants are the targets of major neuroimaging initiatives. A very important aspect of this – relevant to the field of personalized medicine – is that the person's genotype in conjunction with amyloid imaging can accurately predict the age of onset for the disease and the symptoms.

32Another cause for optimism is the efforts of the Psychiatric Genomics Consortium (PGC). When the PGC Schizophrenia Working Group increased their sample size to 36,989 cases and 113,075 controls, they discovered over 100 loci associated with risk for schizophrenia, suggesting that other GWAS may experience similar boosts, depending on where they are in the arc of discovery. The rate of success of these efforts, and yield on the efforts invested, also depends on the polygenicity of each disease, and the distribution of risk loci across the genome. Holland et al. (submitted for publication) used recent data from the ENIGMA study and the PGC to estimate what sample sizes are needed for a GWAS to discover enough SNPs to account for, say 50% or 80% of the chip-based heritability, i.e., the amount of the population variance predictable from genotyped SNPs. They argued that some traits are more polygenic than others, and that, relative to some brain measures, GWAS studies of schizophrenia and major depressive disorder may require much larger sample sizes to discover enough SNPs to account for high levels of the chip-based heritability. If that is true, then imaging genetics may be well on the way to a significantly higher rate of discovery, and a more complete understanding of common variants driving individual differences in brain measures.

33In recent years, a number of powerful methods emerged to estimate what fraction of the population variance in a trait could be predicted, in principle, from all the SNPs on the genotyping chip, even if the exact genes and SNPs were not yet known.4 Predictions can be made from the full set of association statistics: models (linear or Gaussian) are first fitted to the observed effect sizes of all the SNPs, even if most SNP effects fail to reach the accepted standard for genome-wide significance. In much the same way as FDR (the false discovery rate method) is used in imaging to confirm evidence for a distributed signal — spread out across the brain, the overall effect of genome-wide SNPs on a trait can be estimated without having to pinpoint which exact regions — of the image or the genome — contribute unequivocally to the effect.

34used genome-wide summary statistics to estimate heritability and found that common variants across the genome explained around 19% of the variance in hippocampal volume, which is comparable to SNP-based estimates of heritability for many psychiatric disorders and other biological traits. More recently, introduced a similar method based on linkage disequilibrium5 (LD) scores that is also able to recover heritability from summary statistics. The LD score method assigns an LD score to each SNP — the sum of its squared correlations (r2) with all other SNPs in a 1 centimorgan window. One then regresses the chi-squared statistics from a GWAS against the LD score for each SNP. The slope of the resulting regression line depends on the sample size and the SNP-heritability — the proportion of trait variance accounted for by all the genotyped SNPs (see,, for derivations).

35A related method, GCTA (genome-wide complex trait analysis; ) suggested that a still higher proportion of population variance in brain volumetric measures may be accounted for based on all genotyped SNPs, even in cases where we do not know which SNPs help as predictors of the trait. Members of the ENIGMA Consortium have applied this method to estimate SNP-based heritability for structural and functional brain measures. A working group in ENIGMA, ENIGMA-GCTA, is now comparing the GCTA and LD score methods to better estimate how much brain variation is explainable by genotyped SNPs, at least for the brain measures that are most readily computed from MRI. SNP-based heritability estimates of cortical surface area for different cortical subdivisions calculated by GCTA were recently published. These cortical subdivisions were defined by a genetically based cortical parcellation scheme.

36The reason ENIGMA and other GWAS researchers are interested in measuring heritability – and ideally the fraction of heritability explained by common genetic variants – is that it should be possible to prioritize brain measures for deeper genetic analysis based on their heritability, reliability, polygenicity, and relevance to disease. Such rankings or “Bayesian priors” would help in prioritizing research, making studies more efficient and better powered (; Becker et al., submitted for publication; Holland et al., submitted for publication; Wang et al., submitted for publication). Even so, there is no evidence that phenotypes with higher heritability show stronger associations with SNPs. One such example is white matter hyperintensities — a brain measure with high heritability, for which specific genomic risk factors have been hard to find. The main benefit of focusing on highly heritable phenotypes comes from the fact that measurement error is typically lower, and prioritizing brain measures is important as there are so many ways to quantify brain structure and function.

37A recurring caveat in this work is that the SNP effects are not expected to be constant in all cohorts. They may depend on a person's age, environment, or other circumstances. We now know from ENIGMA2 that the top 8 loci associated with the volumes of subcortical structures were detectable consistently worldwide, even though each one accounts for < 1% of the variance. A later screen for age × SNP effects suggested that some genes have a greater effect on brain measures later in life (Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,, ), perhaps because they interact adversely with other biological processes or environmental stressors. In other words, although ENIGMA primarily uses meta-analysis to assess evidence, we do not assume that the effect size is always the same. Heterogeneity of effects is also assessed – a SNP effect important late in life may not be replicated in younger samples. Conversely, since most psychiatric disorders occur at a young age, one may expect to find associations that link genetic vulnerability, brain structure and disease at a younger age, with effects that may diminish later. Moreover, for certain disorders such as addiction, the psychological, neurobiological and genetic factors most relevant at one age (e.g., impulsivity or sensation-seeking in adolescents experimenting with drugs) may be quite different from the factors when dependent (e.g., compulsivity or habit-based behavior) or when recovering (e.g., stress regulation or cognitive control). Even so, ENIGMA's genomic screens so far are only well-powered to detect SNP effects that are consistent — there may also be SNP effects, so far undetected, that depend on the demographics of the cohort assessed, or disease status, or other circumstantial factors.

38This is a reminder that predictive models work best in cohorts similar to those where discoveries were made. Because of this concern, which to some extent affects all brain imaging studies — and all human studies — ENIGMA has diversified to over 33 countries. Recently, ENIGMA partnered with other consortia such as the Japanese consortium, COCORO; encouragingly, effects of psychiatric illness on brain structural measures were replicated in Western and Eastern populations, not just in the structures affected the most, but in their rank order, showing congruence between independent studies.

39After the initial success of the genetic analyses (; Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,, ), ENIGMA investigators had analyzed brain MRI data from well over 30,000 individuals — around a third of the data came from patients with a range of psychiatric conditions. In the primary GWAS studies, analyses were run with and without patients, and excluding patients did not affect the main findings; of course the possibility remains that some SNP effects may be easier to detect in some patient cohorts, but ENIGMA's overall results were not driven by the presence of patients.

40In 2012, ENIGMA formed working groups on schizophrenia, bipolar disorder (Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,, ), major depression, and ADHD; groups meta-analyzing data on 8 additional disorders have been formed since, with current sample sizes detailed in Table 1; a map of participating sites is shown in Fig. 2. In the summer of 2015, additional working groups were formed on anorexia nervosa, recovery after stroke, and Parkinson's disease — the current “roadmap” showing relationships between ENIGMA's working groups is shown in Fig. 3 (also see http://enigma.ini.usc.edu for the latest status). The diseases surveyed include many where controversy exists on the nature and scope of disease effects on the brain. Given this controversy, the main benefit of meta-analysis is to discover which effects are strongest or most reliably found, and which depend on known or unknown factors of the cohorts assessed.

41The initial goal of ENIGMA's Disease working groups has been to meta-analyze effects of these disorders on the subcortical brain measures studied in the GWAS study. As scans had already been analyzed with a harmonized protocol, and subtle genomic effects had been discovered, there was some interest in ranking brain measures in terms of disease effects (i.e., differences between patients and controls).

42A secondary goal was to find factors that might moderate how these diseases impact the brain, such as a person's age, the duration or severity of illness, comorbidities, or treatment-related effects, such as which medications the patients had been treated with, and for how long. Clearly, treatment effects on the disease or the brain depend on many factors. ENIGMA's multiple cohorts, in some cases, offered the opportunity to gauge their generality or consistency. At the same time, many groups joined ENIGMA and provided only brain measures as their initial case–control analyses did not require genome-wide genotyping data on their cohorts. As such, truly vast samples began to be analyzed (N = 8,927, in the published ENIGMA-Depression study; N = 10,194 in the ENIGMA-Lifespan study; see Table 1).

43At the time of writing, ENIGMA's first studies of schizophrenia and major depression have been published; results are compared in Fig. 4. Some caveats are needed in showing these data side by side: the schizophrenia and major depression patients were not ascertained at the same sites, so site or geographic effects may be present.

44Among the subcortical structures so far assessed, the hippocampus shows the greatest differences in each disorder in terms of statistical effect sizes — but in major depression, it is the only structure showing differences, of those assessed so far. Many other structures show volume deficits or even hypertrophy in schizophrenia; basal ganglia enlargement has been widely noted in prior studies of patients taking second-generation antipsychotics. In people with schizophrenia, abnormal ventricular enlargement has long been reported (as far back as ), but the natural variations in ventricular size make the effect size smaller for this structure, even though the absolute volume difference, on average, is greater than for other structures assessed. In major depression, the hippocampal volume difference was greater in patients who experienced more depressive episodes, and in those diagnosed before the age of 21 years, which were at least partly independent effects. This is in line with many prior reports of greater brain differences in those with an earlier onset of the disease. Studies of cortical measures are now underway across all ENIGMA disease working groups; many cortical regions are commonly implicated in psychiatric illness, so these analyses may offer a more complete picture relating brain structural differences to clinical measures, medications, and outcomes. At the same time, diffusion imaging studies are also underway; initial reports reveal consistent deficits in fractional anisotropy – a measure of white matter microstructure – for major white matter tracts in schizophrenia; an interesting question is whether antipsychotic medications affect white matter and brain connectivity in a way that fits with their known effects on structural anatomy.

45Because of the worldwide scope of the ENIGMA studies, only the brain measures that were most readily measured have so far been examined. Clearly, there are measures that may be more relevant to each disease or closer to the action of disease-causing genes, but if they are difficult to harmonize and measure in a standard way, the available sample sizes will lag behind those available for the simpler measures. Because of decades of work on shape analysis of anatomy, several of the ENIGMA disease groups have begun to analyze and meta-analyze subcortical shape, to map the profile of volumetric effects with more spatial precision. These efforts will also determine whether shape metrics offer additional predictive value over and above standard metrics, and in which situations.

46The ENIGMA-Laterality group is studying global trends in the profile of left–right differences in brain structure, and whether they relate to handedness, sex, and disease status, in over 15,000 people (, submitted for publication). Reduced or abnormal brain asymmetry has been reported in many brain disorders, but the scope and generality of these differences is not yet understood. Also, many important aspects of human brain function show lateralization in terms of the underlying processing networks, but the biology of this specialization is poorly understood, as are factors that influence it. Whether brain asymmetry measures add value as diagnostic predictors, will be testable across ENIGMA.

47ENIGMA-EEG is studying the influence of genetic variants on brain functional activity measured with scalp recorded electrical signals, in a combined dataset from 10,155 individuals, ranging from 5 to 74 years of age. EEG metrics of brain function mature rapidly with age, and relate to aspects of cognition such as the brain's processing efficiency; they also show abnormalities across many neurodevelopmental and psychiatric disorders. Combining data from several large twin and family datasets, the ENIGMA-EEG working group is performing a genome-wide association analysis of brain oscillatory power – a highly heritable trait – before proceeding to in-depth analyses of lateralized activity, brain connectivity, and network properties.

48Voxel-based mapping methods are complementary to approaches that measure the volumes of specific regions of the brain, and they allow comprehensive and unbiased searches for effects of disease or genetic variations across the brain. “Brain-wide” genome-wide searches, or “voxelwise GWAS” can involve over a trillion statistical tests. However, once we account for the covariance within the image and genomic data, the number of independent tests being conducted drops to less than 15,000 × 1,000,000. Given the extremely low p-values of some genetic associations in ENIGMA (p ~ 10–23 in Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,, ), several effects can still survive a “double” Bonferroni correction for multiple testing across both the image and the genome.

49As a result, several recent approaches have been developed to perform brain-wide genome-wide association studies to identify “spatial” features associated with genetic variants, such as specific WM pathways and their components, patterns of cortical thickness, or even activation patterns, rather than “global” measures such as brain or subcortical structure volumes. These approaches may be broadly divided into (1) “brute force” methods, that use mass-univariate testing to test every SNP for associations at each voxel in the image, and (2) data reduction methods, that attempt to reduce the search space by reducing the number of features in the image, or the genome, or both. Data reduction methods may include classical methods, such as canonical covariates analysis, or independent components analysis, or modern variants such as sparse coding, compressive sensing, or “deep learning” for feature discovery (see for a review of multivariate imaging genomics methods). Among the “brute force” methods,, detail a practical method whereby several sites run a voxel-based morphometric analysis independently, using a GWAS or other covariate-based analysis at each voxel, and later communicate their findings to a central site for meta-analysis (see Fig. 5). This approach was able to map out in the brain and meta-analyze the effects of the top SNP from the ENIGMA2 study, which screened the genome for variants associated with the size of subcortical structures (Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,, ). To avoid re-computing everything when a new site joins, this “meta-morphometry” approach allows cohorts to align their data to their own brain templates, which are later aligned to an overall mean template for meta-analysis. Such a distributed effort offers many advantages for imaging genomics, due to the vast number of predictors: as new cohorts join, each site's computational hardware can be leveraged by all the others. Such an approach allows cooperative computation on data without requiring all the data to be shared or ever transferred. This is an interesting area of cooperative machine learning that can also increase “buy-in” — opening up participation to countries with stricter data transfer laws.

50As part of ENIGMA3, a genome-wide screen of the cortex, one subproject will adopt “genetic clustering” methods to identify coherent patterns of gene effects in the brain. Based on the notion of genetic correlation, brain regions or sets of voxels can be grouped into clusters with similar genetic determination. The standard decomposition of the brain into regions may be adapted to include genetic clusters, or new regions where genome-wide association may be more efficient. This approach has already been applied to create genetic partitions of the cortex; initial work in ENIGMA will overlay pre-made partitions on the cortical data from each site. Genetic correlations can now be computed rapidly from GWAS summary statistics making it feasible to compute and perform clustering on matrices of “genetic connectivity” whose entries are genetic correlations. The ENIGMA-GCTA Working Group is currently studying these methods, in multisite data.

51Many disorders affect the brain's white matter and connectivity. Using diffusion tensor imaging (DTI), ENIGMA's disease working groups have begun to compile evidence across cohorts for differences in a range of DTI measures, which reflect white matter integrity and microstructure. Several years of work went into harmonizing ENIGMA's DTI analysis protocols, to study which metrics are consistently heritable and reproducible across multiple twin and family cohorts worldwide (,; Kochunov 2014; ). These DTI protocols have been carried forward into ongoing GWAS and disease studies, and initial genome-wide screens of the structural connectome. On the genetic side, ENIGMA working groups have also formed to assess other kinds of genetic variation, including copy number variants (CNVs), where abnormalities have been reported in autism, schizophrenia, and learning disabilities. The ENIGMA CNV helpdesk is now supervising supervising an initial analysis of CNV data in 13,057 people from 24 cohorts worldwide, after developing harmonized protocols for CNV “calling” and quality control. Participating cohorts include groups from Japan, Mexican-Americans, and people of Western European, Nordic or Swedish ancestry. Initial efforts are evaluating known “psychiatric” CNVs as predictors of MRI and DTI phenotypes computed in other ENIGMA projects. Challenges include the pooling of data from genotyping chips with different coverage; some have sparse coverage of SNPs in regions with segmental duplications or complex CNVs.

52In a complementary initiative, the ENIGMA-Epigenetics working group is studying epigenetic processes such as methylation, which is an index of biological aging and lifecourse ‘stress’ that may explain an important proportion of the gene-environment contribution to expression of many common diseases such as stroke and dementia. The group is now performing epigenome-wide association studies (EWAS), across 14 cohorts from Asia, Australia, North America, and Western Europe, to test associations between DNA methylation and brain measures, initially focusing on total brain volume, subcortical volumes and cortical thickness and surface areas. The working group is analyzing methylation data from 9,000 people, of whom 5,000 have both methylation data and MRI. In addition, the ENIGMA-Epigenetics group is prioritizing the analysis of DNA methylation sites based on their effects on gene expression or association with stress- and anxiety-related phenotypes. There is some evidence of early life changes in stress response genes through methylation, just as early life events influence later life disease expression — notably stroke, white matter hyperintensities, and cognitive impairment. Of great interest are epigenetic changes throughout the life span, and with aging, which may predict mortality from all causes, as well as physical and cognitive performance. Associations are being tested first for brain phenotypes that are known to change the most across the lifespan, based on incoming information from ENIGMA's Lifespan study in over 10,000 individuals.

53ENIGMA was not designed to make predictions about individuals based on their scans and genomic data. As in most epidemiological studies, the power lies in aggregating so much individual data that subtle effects on the brain can be detected, including findings that each cohort's data were insufficient to detect. In other words, its primary goal has been to relate brain measures to disease and treatment effects, and to variants in the genome. With the aggregated data, it has been possible to determine how reproducible these patterns are worldwide. Also, for the study of treatment effects, ENIGMA does not have the ideal design. Ideally, one would prefer to have pre–post treatment longitudinal designs instead of the cross-sectional comparisons in ENIGMA, where medication status is often confounded by age, disease duration, comorbidity and disease severity.

54Even if a large data sample is needed to discover a factor that influences the brain, it does not mean that it is irrelevant to individuals; APOE is one such example, discovered in 1993 by linkage analysis in pedigrees. More recently, a rare variant in the TREM2 gene was found to affect Alzheimer's disease risk and accelerate brain tissue loss as we age — perhaps doubling loss rates in old age and increasing AD risk by a factor of 2–4. This gene variant is undoubtedly important for those who carry it: it is found in a little under 1% of controls and a little over 1% of AD patients.

55In current clinical practice, it is not recommended to notify a research participant of their APOE status, and most ethics boards clearly define the circumstances in which incidental findings or health-relevant information is communicated back to a research participant. In the case of APOE, participants are not typically informed of their genetic status, as there are no effective treatments for late–onset Alzheimer's disease. Still, discovering predictors of more rapid decline is useful for the pharmaceutical industry for understanding the behavior of participants in clinical trials, and can greatly improve drug trial design, reducing costs. Enrichment approaches use some characteristic of a patient to select them for a clinical trial — this may be prior response to a certain drug, or it also may be a prediction that they are more likely to decline (FDA, 2013). In the AD field, some clinical trials now select patients based on having a PiB-positive PET scan – as evidence of incipient AD pathology – and the APOE4 risk genotype, as carriers are more likely to develop AD. This selective enrolment allows faster, less costly, and more well powered clinical trials, with demonstrable reductions in the number of patients needed to show treatment effects (Hua et al., submitted for publication).

56ENIGMA's disease working groups are likely to broaden the set of known factors that help predict recovery or decline. In ENIGMA-HIV, for example, a key goal is to understand predictors of resilience — factors that might forecast healthy brain development after the use of antiretroviral treatment. Crucially, it is important to know if a predictor of decline is specific to one cohort or likely to generalize to others, or if it is applicable in a limited set of situations. Understanding how APOE4 and other major risk genes shift the lifetime trajectory of brain measures will also help determine how much they will help when used for clinical trial stratification. This is a goal of the ENIGMA-Lifespan group. Clearly, any predictors of suicidal behavior would be very important in the management and follow-up of patients with psychiatric disorders, and a secondary project on suicidality was started within the ENIGMA-Depression working group (Rentería et al., submitted for publication). Similarly, factors that predict whether ADHD in a child will persist into adulthood, will have clinical utility. Ultimately, the stratification or clustering of ENIGMA cohort data into subtypes, based on imaging, clinical or behavioral data, may point to distinctions that help us understand the heterogeneity of these disorders. This heterogeneity, without models to disentangle it, makes individual patient predictions harder to make.

57One effort where ENIGMA may contribute to individual prediction and evaluation – albeit with some caveats – is the ENIGMA-Lifespan project. In this work, ENIGMA cohorts are invited to contribute volumetric measures from normal individuals in their samples, which span the age range from 2 to 92 years of age. Although some cohort studies focus on children or the elderly, many scan people across the lifespan, allowing the computation of age-trajectories for several key brain measures; the results show a remarkable difference in the maturational trajectory of different structures, supporting many earlier neurodevelopmental reports on the sequence of brain development. To cope with the non-uniform sampling density of the cohorts, these overall trajectories must be interpreted cautiously; clearly some parts of the lifespan are better sampled than others, and unmodeled effects of scan site, demographics, and even cultural or environmental differences may drive some of the effects. Clearly, disentangling the driving factors is statistically complex, but the potential is there, to derive normative measures and models of our path through life, in cohort studies as diverse as ENIGMA. The life span analyses (and normative curves) are also highly relevant for neurodevelopmental disorders such as OCD, ADHD, autism, etc. — for early detection, and secondary prevention in at-risk populations. Eventually, there may even be efforts to train individuals in specific domains, to stimulate the maturation of specific brain areas that appear to be deviant from the norm curves.

58Such normative data have possible applications for individual assessment, if used judiciously. In pediatrics, growth charts for height and weight offer metrics of where a child stands relative to others of the same age, as a Z-score for example. Similar metrics for brain structure, among others, may help in studies of neurodevelopment where interventions and treatments are used to promote healthy maturation, or recovery, as in the case of brain trauma, for example. Similarly, better trajectories to chart loss of brain volume with advancing age help in routine diagnosis of the individual with possible cognitive problems, by indicating first if their brain is within normal limits for age, and secondly the precise centile on which it lies – much more data is needed to populate these graphs, but (much like child growth charts) they have the potential to be highly valuable in routine clinical practice as well as research. Original scan data are being collected to expand these templates (e.g., www.brainsimagebank.ac.uk).

59Norming of brain measures also has commercial applications. ENIGMA relies heavily on developments in software for imaging and genotype acquisition, quality control, and analysis, that make standardized assessment possible. In some regions of the world, such as Thailand and Cambodia, ENIGMA has contributors who are interested in whether it makes sense to use brain development norms from Western cohorts, or build their own. By comparing developmental trajectories across very diverse multi-cohort data, better answers to these and other practical questions are within reach.

60With the advent of very large neuroimaging datasets, we can fit predictive models to the data and test them for their robustness. Our models of how diseases and genes affect the brain are constantly being tested and improved, especially in situations where statistical effects have previously been too small to discover, or have been confounded by factors that cannot be adjusted for. In GWAS for example, there are known genetic differences in allele frequencies across populations, and if these are not accurately modeled based on much larger datasets, and adjusted for using multidimensional scaling, they will confound the analysis and lead to spurious results - many more SNPs will show “effects” on the brain, ultimately turning out to be false positives. Years of “false alarms” led the genomics community to adopt strict standards for reporting effects, including a standard genome-wide significance threshold (described above). In addition, independent replication of effects is required. In imaging, a somewhat more flexible approach has been used, with approaches from FDR to random field theory and permutation all co-existing in the literature; the use of candidate brain regions or prior hypotheses in functional imaging studies is encouraged, but the use of candidate regions in genomics is sometimes hotly debated as leading to many false positive effects. argued that the growing flexibility in analyses used in neuroimaging is increasing the reporting of false positive results, and meta-analyses may offer better estimates of the validity of claims regarding brain differences in major depression and bipolar illness, fields for which they meta-analyzed the neuroimaging literature.

61Given the sample sizes attained, ENIGMA offers a framework not only for unrestricted searches, but also to test more focused hypotheses and provide internal replication using, for example, cross-validation methods. So far, the Working Groups have over 30 “secondary proposals”: many study clinical measures, disease subtypes, and patterns of behavior such as suicidality or negative symptoms, or other differences that might contribute to the heterogeneity of brain disease and outcomes. One such project, in the ENIGMA-Major Depression group, assesses the effects of childhood trauma on depression-related brain measures, a factor that may be modeled effectively by comparisons with data from the ENIGMA-PTSD group, where childhood trauma is also a major predictive factor. Partnerships between ENIGMA groups may resolve some sources of brain differences that are difficult to disentangle. In HIV+ people who abuse stimulant drugs, for example, white matter inflammation is commonly reported, while patterns of accelerated atrophy are often seen in HIV+ people who do not use intravenous drugs, especially in those carrying the APOE4 genotype. These and other predictors can be assessed in partnerships between the ENIGMA-Addictions and ENIGMA-HIV groups, by determining a common core of predictor variables that can be harmonized.

62More refined models are also needed: we now know that the profile and extent of brain differences in disease may depend critically on a patient's age, duration of illness and course of treatment, as well as adherence to the treatment, polypharmacy and other unmeasured factors. Differences in ancestral background, as determined based on genotype, are strongly related to systematic differences in brain shape. Any realistic understanding of the brain imaging measures must take all these into account, as well as acknowledge the existence of causal factors perhaps not yet known or even imagined. The quest to identify individual predictors is therefore more likely to succeed in finding factors that affect aggregate risk and outcome in groups of individuals, rather than offer firm predictions regarding an individual.

63A more immediately achievable goal, for ENIGMA, is to rank brain measures in terms of how well they do predict individual decline, or diagnosis. Predictors of imminent brain decline are already used to boost the power for clinical trials in Alzheimer's disease, by over-enrolling, or separately analyzing patients whose brain measures, or clinical and genomic measures, suggest that they will decline faster. In ENIGMA, the ENIGMA-Plasticity group is evaluating the genetic influences on measures of brain change, in a meta-analytic setting. If reproducible drivers of brain decline could be found by screening brain data worldwide, they would help in planning enrichment approaches for drug trials. Several major initiatives have this goal (e.g., ADNI; ). Currently, the only genetic marker used for enrichment is APOE, but this may change as more information accumulates (see Lupton et al., submitted for publication). The complex pattern of association between brain measures and SNPs across the APOE gene (Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,, ) suggests that future polygenic predictors based on machine learning may better predict clinical decline, and decline in brain measures, than the standard APOE genetic test, which is based on just 2 SNPs.

64Innovations in machine learning make it possible to build robust predictive models from millions of predictors, often using dimension reduction techniques to home in on more efficient sets of variables that explain the most variance in the data; this vast field, including sparse learning and compressive sensing, is especially valuable in imaging genomics, with millions of predictors in both the images and the genome. Several machine learning developments have been applied to connect genomic and imaging measures, using methods such as parallel ICA, elastic net, sparse reduced rank regression (sRRR; ), among others. ENIGMA is beginning to test some of these models, specifically in the disease working groups, for case-control differentiation and differential diagnosis. Past efforts to combine imaging and genomic data for outcome prediction suggest that imaging measures may be much more predictive of future clinical decline than genomic measures, but both are complementary (Peters and the Alzheimer's Disease DREAM Challenge, submitted for publication). Predictive models should improve as they draw on more data, and the larger ENIGMA GWAS studies are now discovering more genetic markers that can be used in predictive models for brain measures (Hibar and the CHARGE and ENIGMA2 Consortia, submitted for publication;,,; Adams and the CHARGE and ENIGMA2 Consortia, submitted for publication). However, compelling as these approaches are and not wishing to dampen the enthusiasm for these very promising techniques, the image measurements being predicted generally require a human check and correction if necessary, particularly in datasets with complex imaging features such as occur in older patients with stroke – machine learning analysis algorithms still cannot reliably separate the hyperintensity due to a small cortical infarct from that due to a white matter hyperintensity or artifact, reliably. Also, the variants driving the heritability of disease risk are only just beginning to be discovered for many of the major brain diseases studied within and outside of ENIGMA. Unsupervised learning is also relevant for understanding the heterogeneity of diseases, which has made it harder to discover their causes and mechanisms. argued that one could use unsupervised learning on imaging, clinical and genetic data to see whether subtypes (or clusters) can be identified within a disease, and whether these data cluster together in agreement (or disagreement) with current diagnostic classifications.

65In conclusion, we have reviewed current work by the ENIGMA Consortium. ENIGMA began in 2009, and is now a distributed effort, with over 30 working groups (see Table 1), coordinated from many centers worldwide. As we noted, ENIGMA's main goals have been to detect effects of disease and genetic variants on the brain, to see how consistent these effects are worldwide, and to study what modulates these effects. On the genetic side, it may soon be possible for polygenic scoring to produce predictors that are routinely used in brain imaging studies, explaining some of the observed variance. This may make other effects easier to detect. On the disease side, we are beginning to identify and confirm distinctive patterns of brain differences in each of a range of brain diseases, along with a better understanding of which patterns are specific to given disorders, which patterns tend to generalize, and what factors account for the heterogeneity across cohorts. This will help us understand the situations where predictive models can be used, for diagnostic classification, outcome prediction, and norming of individual data against appropriate reference populations.

66We end with a note in praise of small studies. Like any consortium, ENIGMA would be impossible without the cohort studies and all the individuals who contribute; most of the data analyzed in ENIGMA came from cohorts with relatively modest sample sizes. Inevitably, many hypotheses are not addressable on a large scale, and some questions - especially causal questions - involve targeted interventions or phenotypic assessments with a depth or sophistication not likely to be attained at every site. As Aristotle said, “Nobody has the ability to work everything out, but everyone has something useful to say; working together, the whole vast world of science is within our reach.” (ἐκ πάντων δὲ συναθροιζομένων γίγνεσθαί τι μέγεθος; Aristotle, Metaphysics α, c. 350 BCE). This is the ENIGMA motto: http://enigma.ini.usc.edu/about-2/.