6
Third Research Session: Behavioral Genetics
The third research panel was moderated by Steven Hyman (Harvard University) and showcased cutting-edge work in the area of behavioral genetics. Panelists included Benjamin Neale, assistant professor in medicine at Harvard Medical School, and David Cesarini, associate professor in economics at New York University. Each panelist presented an overview of his research program and highlighted key findings, methodologies, data considerations, and relevance to the work of analysts in the intelligence community (IC).
Hyman said a discussion on behavioral genetics is worthwhile to acknowledge the rapidly accelerating research advances in the field and closely look at claims and possibilities. “Genetics gives clues to biology,” he said. If a trait is at least partly heritable, whether a physical characteristic, an illness, or a behavioral variation, then markers for traits of interest can be found in DNA variations. These provide tools for biological investigation and stratification of human subjects for study. DNA is already widely used in forensic science and for genealogy to identify people’s relations to other individuals and among diverse human populations.
Hyman pointed out that the nucleus of human cells has 3 billion base pairs of DNA that make up a human genome. DNA makes messenger RNAs, which are turned into proteins that are the building blocks of cells. According to Hyman, it is understood that differences in proteins encoded by variations in DNA sequence influence behavioral tendencies through the ways that brains develop, process information, and change as one grows and learns. Hyman emphasized up front that DNA is not a deterministic blueprint of behaviors. Its influences are exerted in the context of diverse
environmental signals, and there is a certain amount of stochasticity or noise in the reading out of DNA during development and in response to the environment throughout life.
Physical characteristics (e.g., adult height and body mass index), normal behavioral tendencies, cognitive ability, and many diseases and behavioral disorders are highly influenced by genes.1 These traits are the manifestation of many slight differences in the DNA sequences of people’s genomes. Hyman noted that a very small number of genetic illnesses are caused by a single highly penetrant gene. Huntington’s disease is one of those; people with the Huntington’s risk gene will get the disease if they live long enough. Most other diseases are much more complex, as is all normal behavioral variation, resulting from an aggregate of variations in many genes that appear to contribute small individual effects. There is no single gene responsible for them (schizophrenia, for example). Rather, according to Hyman, hundreds of loci in the genome can influence such traits.
Hyman recognized that identifying the many signals of very small effect has been difficult because they must be discovered against a huge background of neutral human variation. However, advances in technologies have created a genetic revolution; gene chips now allow scientists to determine a person’s genotype (DNA sequence) at a million or more places in the genome very cheaply, and the cost of sequencing DNA has come down significantly,2 making it feasible to study the entire genomes of many individuals. According to Hyman, as more and more DNA samples are analyzed and convergent information about genetic variance accumulates, scientists are expanding their understanding of disease mechanisms, including disorders of cognition and behavior, but also biological contributors to many normal physical and behavioral traits. Predicting the occurrence of a disorder or some behavioral trait is probabilistic. However, Hyman noted that when combined with other sources of information, prediction, although still probabilistic, gains in power.
BEHAVIORAL GENETICS AND POLYGENIC INHERITANCE
Benjamin Neale summarized the state of the field of genetics in understanding certain kinds of behavior. He reviewed the idea of polygenic inheritance in order to bound enthusiasm on what can be predicted. Science and medicine have known of a biological basis for physical traits and
___________________
1 Hyman pointed out that autism, schizophrenia, and bipolar disorder are highly heritable.
2 The cost of sequencing DNA has decreased faster than Moore’s law for microprocessors. Moore’s law, named after Intel cofounder Gordon Moore, predicts that the number of transistors per square inch on integrated circuits will continue to double each year.
disease within families (heritability) for a long time but have not figured out the mechanisms driving these traits.
According to Neale, “in the last 15 years, [science has advanced to] mapping of the human genome outright, characterization of genetic variation in populations, in particular the single nucleotide polymorphisms, the single base pair changes, and the correlation across those different genetic variants . . . to genome-wide association arrays.” Neale pointed out that these arrays allow scientists the opportunity to study genetic variation on hundreds of thousands of individuals simultaneously.
Neale further noted that many of the 18,000 to 20,000 genes in the human genome are involved in a variety of different disorders. Specific genes have been found to be involved in multiple different diseases that are related. For example, the same types of genes that are markers for Crohn’s disease are also markers for rheumatoid arthritis. Neale acknowledged that growing awareness that genes influence multiple things has changed how scientists think about prediction of traits.
He provided an example of investigating schizophrenia. Schizophrenia is a behavioral phenotype because it is diagnosed based on psychiatric interview and includes symptoms like hallucinations, excess paranoia, and disorganized thought. Neale reported that genetic studies have looked at whether certain gene variants are more common in cases with schizophrenia than in control samples without schizophrenia. These studies have moved toward large-scale investigations of genome-wide associations on schizophrenia. One of the first large studies had sample sizes of 2,600 cases and 3,300 controls; however, noted Neale, the evidence was not strong enough to be sure that certain variants were associated with schizophrenia.
To make headway on suspected associations, according to Neale, the research community working on schizophrenia genetics formed the Psychiatric Genomics Consortium (PGC) to share data. Through the consortium, sample sizes for studies were tripled, and scientists became confident they had sufficient evidence to identify five genome-wide regions where genetic variant was influencing the occurrence of schizophrenia. As cases and data have been added, samples have reached over 25,000 cases (with 28,000+ controls), and scientists have identified about 100 to 150 genome-wide significant loci for schizophrenia. While these discoveries are exciting, Neale reported that the science is far from interpreting whether these signals are the biological mechanisms for schizophrenia. However, the signals do help scientists decide where in the genome to invest their time.
In terms of figuring out how much risk for schizophrenia these genetic variants confer, Neale reported that “typical risk estimates for these kinds of effects range in odds ratios of 1.05 to maybe 1.2, which is a very small perturbation in risk when . . . a baseline rate for schizophrenia [is only] 1 percent [of the population].” However, it may be possible, Neale suggested,
to create a tailored score (a polygenic risk score) for an individual, summing up the effect sizes of the genetic variation across his or her genome. The degree of prediction, according to Neale, is improving but remains low at this time. Studies developing polygenic risk scores have found that individuals who would score in the top 10 percent in polygenic risk score have an absolute risk of getting schizophrenia of around 3 percent, compared with the 1 percent observed in the overall population.
Neale said prediction could improve if tools can be combined. For example, risk-factor inventories have been useful for some time. One for predicting the risk of coronary heart disease is the Framingham Risk Score, recently updated by the American College of Cardiology, which creates a score based on a combination of behaviors and characteristics like smoking, LDL cholesterol, HDL cholesterol, diabetes, hypertension, whether treated, and age. Current research is creating an index to increase prediction potential by reviewing genetic risk in conjunction with information from the Framingham Risk Score or the American College of Cardiology Risk Score.
Neale stated that all behavioral traits have some genetic basis, and combining genetic and phenotypic information will help build better predictors. Genetic prediction continues to improve as sample sizes increase and polygenic risk scores are created. He cautioned that taking a purely genetic deterministic point of view is not an accurate reflection of the current science; genetic information needs to be considered with phenotypic information. Prediction is still challenging, especially with rare events.
PREDICTING BEHAVIORAL TRAITS FROM GENOMIC DATA
David Cesarini expanded on Neale’s presentation about predicting outcomes from genotypic data, but focused on behavioral traits. He began with background on the kind of evidence available before the explosion of genomic data. Earlier genetic studies compared outcomes among twins and adoptees. Cesarini pointed out that, with the steadily increasing technological advances and dramatic falls in the cost of measuring DNA, progress in the field of behavioral genetics has been quite rapid.
He referenced a study of Swedish brothers, born between 1950 and 1970, which looked at sibling correlation for five different outcomes: (1) height, (2) body mass index (BMI), (3) years of schooling, (4) cognitive skills as measured by a test like the Arms Force Qualifying Test, and (5) socio-emotional skills from a military psychologist’s assessment of their ability to deal with wartime stress. Correlations on these outcomes among siblings were considered on seven different sibling types (in declining “genetic relatedness” order): (1) monozygotic twins, (2) dizygotic twins, (3) full siblings living together, (4) full siblings living apart, (5) half siblings
living together, (6) half siblings living apart, and (7) adoptees. Cesarini presented a graph illustrating that as the degree of genetic relatedness declines among different types of siblings, the observed amount of similarity declines.
Cesarini pointed out that the focus of behavioral genetics is also about looking for correlations to make inferences about heritability.3 The term “gene discovery,” according to Cesarini, refers to a process of considering an outcome of interest and identifying genetic variants that are statistically distinguishable among people with different genotypes.
Cesarini reviewed how gene discovery studies have changed. Historically, the most common research approach has been what is known as a candidate-gene study. The candidate-gene studies would test hypotheses of biological function for particular genetic variants (i.e., a small set of SNPs4) on specific outcomes. According to Cesarini, these studies have found some associations between gene variants and outcomes; however, they have not been easily replicated for a number of reasons. Cesarini identified some concerns with this methodology: (1) the hypotheses may have some face value but not the biological underpinning; (2) small sample sizes do not afford enough statistical power to draw strong conclusions; (3) the methodology does not deal effectively with confounding variables; and (4) there was a lot of undisclosed hypothesis testing in the field that made studies difficult to interpret and replicate.
A different and emerging research approach to gene discovery, noted Cesarini, is what is known as a genome-wide association study (GWAS). Used in the schizophrenia studies presented by Neale, this approach consists of atheoretically testing a large number of genetic variants for association with some outcome. Cesarini pointed out three advantages of GWAS: (1) an up-front understanding that the methodology is testing about 1 million independent hypotheses; (2) genome-wide data help deal with confounding variables and ensure that analyses are conducted in a genetically homogenous sample; and (3) it follows from the logic of Bayes’ rule. Results from GWAS research have helped scientists, according to Cesarini, understand why the link to heritability is indiscernible; that is, the effect sizes have been “hiding” across a number of genetic variants. He noted that the GWAS literature is still in its infancy, but the method shows promise for studying many personal traits.
Cesarini summarized the current state of GWAS findings for three
___________________
3 Cesarini referred to heritability as an R squared from a regression of outcome on all kinds of genetic variables, which indicates its predictive power on the outcomes.
4 SNPs (pronounced as “snips”) stand for Single Nucleotide Polymorphisms. A SNP is the replacement of a nucleotide (a DNA building block) in a DNA segment and the most common form of genetic variation among people. For more information, see https://ghr.nlm.nih.gov/primer/genomicresearch/snp [December 2016].
traits: height, BMI, and years of education. He pointed out that like the work with schizophrenia, “as the discovery samples get bigger and bigger, the number of independent variants that reach genome-wide significance increases.” Cesarini reported that the science has currently identified over 700 independent loci at genome-wide significance associated with height; about 100 associated with BMI; and about 160 associated with years of education. In addition, the replication record is quite encouraging. However, noted Cesarini, the predictive power is quite small. Currently, about 14 percent of variation in height and about 7 percent of variation in BMI and education can be explained by genetic data. If sample sizes were quadrupled (about 2 million people), Cesarini suggested that a quarter of the variation in height and maybe 10 to 12 percent of the variation in BMI and education could be explained by genetic data. He pointed out that these figures are not huge but are also not negligible. He estimated that this future predictive power on height from genetic information would be roughly the same as the predictive power from knowing the average height of two parents.
According to Cesarini, in the shift in how gene studies are conducted, scientists have learned that candidate-gene studies with small samples have a weak replication track record; the replication track record of GWAS research is a lot stronger; and the number of associations identified and the predictive power of scores increase as the sample sizes have increased. Cesarini remarked that the GWAS work has illuminated the principle of polygenicity: that is, human traits are associated with very many genetic variants, each one accounting for a small percentage of the observed variability.
In closing, Cesarini noted that the field of behavioral genetics will expand the number of phenotypes for which predictions become feasible beyond the three traits illustrated in his presentation. He suggested that gene discovery is useful in several ways. It helps elucidate biological mechanisms, aids empirical research, and may lead to polygenic scores that provide information on individuals who are at risk for various outcomes.
Hyman asked Cesarini if behavioral prediction from genetic data could be imagined in 5 years if many hundreds of thousands of people have been studied for behavioral phenotypes of interest. He replied that behavioral genetics will become increasingly valuable in some settings. The power of prediction will vary across the complexity of the phenotype. He explained that eye color is comparatively easier to determine from DNA samples because there are just a handful of specific genetic variants that govern the vast majority of variability in eye color in the population, in contrast to height, for which there are thousands of genetic effects, each contributing a very little bit to the overall expected physical trait.
Neale interjected a point to consider about the nature of predictions: average is a good bet if additional information is not available.
DISCUSSION
George Gerliczy (Central Intelligence Agency) pointed out that the intelligence community studies individuals a great deal, but it usually looks at foreign leaders, leaders of countries or militaries, or terrorist organizations and makes any predictions about tendencies from a distance. Gerliczy pointed out that analysts in the IC are generally not psychologists, although there are some medical professionals conducting medical and psychiatric assessments of select foreign leaders.
Gerliczy offered that, like many other uses of advancing technologies, the IC is concerned with understanding how other nations or adversaries apply the insights from behavioral genetics in their actions. Such considerations require that the IC has enough analysts who are sufficiently familiar with the science and the capabilities to be able to speak intelligently about the policy implications.
Sallie Keller (Virginia Polytechnic Institute and State University) asked the panelists how privacy and privacy consent are managed in the international studies. Neale reported that their strategy is to divorce every individual identifier from the genetic data and phenotypic information, with the exception of individuals who are under an appropriate institutional review board (IRB) approval to maintain their information. He said he recognized the many ways of identifying individuals if their genomes or set of phenotypes could be picked out from records, and as such policies for safeguarding records become important considerations.
Hyman pointed out that he had just hired a global compliance officer for his institution to address concerns about the uses of personal data, including genomic data. He noted that people and governments are concerned about research data being accessed by the IC or by private companies like Google and Amazon. He emphasized the importance of issues about data handling and privacy for the research community.
A summit attendee asked if collecting personal information (genetic data and phenotypes) became ubiquitous from a young age, could it lead to predictions of a future leader’s behaviors. Neale responded that it would be technologically feasible, but it was beyond his expertise to discuss the privacy implications. He said in terms of health care, it is less controversial to keep genotypic information as part of medical records and handled with safeguards to use as part of improving public health.
One of the summit attendees expressed concern that the presentations missed the idea that behavior is a fuzzy concept. For example, the framework or diagnostic category for schizophrenia changed between 2009 and 2014. In addition, there are differences internationally around diagnostic categories—there are some cultures where hearing voices is appreciated. The attendee encouraged the SBS Decadal Survey to pay serious attention
to studies on the role of the environment and not overemphasize phenotype genetic models.
Neale clarified that the core phenotypic definition of schizophrenia is the same. There is a group of people who lose their cognitive abilities at a very early age, have an erosion of their ability to structure their thoughts and emotions, and have hallucinations and delusions. Neale reiterated that there is clear evidence of genetic influences from a number of research approaches. He agreed that behavioral boundaries are fuzzy. For example, the genetic evidence from studies of schizophrenia and the genetic evidence from studies of bipolar disorder show a very high degree of overlap from a common variant genetic risk point of view.
Charles Gaukel (National Intelligence Council) followed up on Neale’s comment that most people are average by definition. For many questions the IC considers, it is useful to know average conditions, and in many cases this information is not clear or available. Gaukel remarked that it would be useful to have models or mechanisms to characterize average conditions and also identify factors that would trigger someone or something to deviate from the average. For example, for many years, according to Gaukel, the Central Intelligence Agency has had a useful program called the Political Instability Task Force, which periodically assembles correlates to a country’s vulnerability to instability and identifies countries at risk of domestic instability.
He said it would be helpful if science could address the time horizon challenge. As an analogy, he noted that correlates that put people more at risk for coronary disease are known (diet, lack of exercise); however, many of those in the populations at relatively high risk will not have a heart attack in any given year. Gaukel suggested the IC would be interested in learning more about indicators that suggest one is moving toward a tipping point or a position of higher risk.
Amy Kruse (Cubic Global Defense) pointed to research on how changes in brain chemistry are correlated with aspects of decision-making and risk-taking as well as susceptibility to persuasion. She asked whether this area of research could help advance predictions of behaviors and whether wearables and other physiological sensors might play a role in data collection. Hyman acknowledged this area of research could be useful, but current studies using brain imaging and electroencephalograms are very expensive, and so progress is advancing at a slower rate than the genetics work. The genetics work has progressed rapidly in the past decade because of decreasing costs of microarrays, computation, and sequencing, thus permitting the study of large numbers of humans to achieve statistical power. Hyman said there would have to be a similar cost revolution in the neurobiological tools to get to the levels of certainty needed.
Neale added that another reason for the success in the area of genetics is that the field has been concerned about and taken steps to control biases and confounding variables. He recognized that additional data from wearables or other devices could potentially be useful, but the biases and confounding variables that troubled earlier phenotypic studies will continue to be applicable and should be taken into consideration. Lessons learned from the genetics work include focusing on scale, consistent effects, statistical standards, and reproducibility.
Paul Glimcher stressed that even if every genome from every person on Earth was included in GWAS research, traits could not be predicted with 100 percent accuracy. Statistically, there is no linear end point. With such a complete database, the science could tell how much of the variability in the world was due to genetics. Glimcher emphasized it is important to recognize that the phenotype, or trait, is a combination of genetics and environment, and no amount of genetic data will ever eliminate that fact. He added that there may even be some variability that a complete knowledge of environment and a complete knowledge of genotype would not eliminate.
Hyman added that another reason research in genetics has been successful is that it is easier than working on the environmental contribution to phenotypes. People live for a long time and go through all kinds of experiences, making it difficult to understand the effect of environment and any gene-environment interactions.
As the discussion concluded, Keller noted the importance of the day’s discussions in setting up issues for the SBS Decadal Survey to consider. She noted that analysts in the intelligence community are operating at a rapid pace with real problems. They want findings from research to support their work and to communicate these findings to policy makers effectively. She said the SBS Decadal Survey will need to examine the pace at which science is progressing and find ways to separate out knowledge relevant to the intelligence community and its operations.
This page intentionally left blank.
