The challenges of psychiatric genetics

Back in April I posted on the elusive genetics of bipolar disorder, a crippling psychiatric condition affecting over 2% of the population in any given year.

The major message from that article is that although bipolar disorder is massively influenced by genetic factors (around 85% of the variation in risk is thought to be due to genetics) we still don't really have the faintest idea exactly which genes are involved. This is despite three reasonably large genome-wide association studies involving over 4,000 bipolar patients in total, which generated weak and contradictory results and failed to provide a single compelling candidate for genetic variation underlying this disease.

This disappointing result has also held largely true for other psychiatric conditions with strong genetic components, such as schizophrenia, major depression and autism. Genetic studies of these conditions have had some success identifying rare mutations that underlie severe cases, but the vast majority of the genetic variants contributing to risk remain undiscovered.

There are several reasons why genome-wide association studies can fail to yield significant harvests of disease-associated genes. I summed these up with respect to bipolar disease as follows:

The researchers are surely hoping that small effect sizes are the major problem, since this is the easiest problem to remedy (simply increase sample sizes). Disease heterogeneity - in other words, multiple diseases with distinct causes that all converge on a bipolar end-point - also seems like a particularly plausible explanation given the complexities of mental illness. It's also likely that various types of genetic variants that are largely invisible to existing SNP chips, like rare variants and copy-number variation, are important.

The same story probably holds largely true for other psychiatric conditions. In this week's issue of Nature, a news article and an editorial both tackle the challenges of psychiatric genetics, and lay out the ambitious strategies currently being pursued by researchers around the world to overcome them.

Small effect sizes
The first hurdle that I describe above is the fact that most of the variants underlying these conditions probably have very small effect sizes (only increasing risk by less than 20%). Such variants will only be identified by cranking up sample sizes immensely, an approach that has yielded some limited success for other genetically complex traits such as height and obesity. The Nature news feature has a table listing some of the major collaborative efforts currently collecting genetic information from the very large cohorts required to dissect out the basis of these conditions:



In most cases, these samples are being built up by pooling results from multiple different studies, often gathered by groups from around the world. As sample sizes increase the power of studies to detect small-effect variants grows. The effect of sample size on the power of genome-wide association studies is illustrated in the graph below from a recent review by Peter Visscher*:

Take a single genetic variant that explains just 0.5% of the variance in the risk of a psychiatric disorder. With a sample size of 5,000 individuals with that disorder you still have a mere 50% chance of detecting that variant. Double your sample size, and that probability jumps to a near-certainty of detection - and your power of detecting even smaller-effect variants (explaining, say, 0.2% of the risk) starts to climb to respectable levels.

By staring at those curves for a while, and bearing in mind that many of the variants found by recent genome-wide association studies explain well under 0.2% of the risk variance, you will quickly start to appreciate why researchers are pushing for ever-larger disease populations to work with. With truly enormous samples on the order of 50 to 100 thousand patients - not out of the question for international consortiums studying reasonably common diseases such as bipolar - the power to detect even very weak risk variants becomes reasonable.

If there are common genetic variants contributing to the risk of these diseases, such large collaborative studies will eventually find them; so long, of course, as they can tackle the next (and potentially far more serious) problem of disease heterogeneity.

Complex, heterogeneous diseases
The second major problem I mentioned with analysing the genetic basis of these diseases is that they are complex, multifactorial, and extremely difficult to diagnose and classify. Psychiatric conditions are probably the most difficult area of medicine to draw hard boundaries: many symptoms are shared by multiple conditions, and many patients display a diffuse constellation of clinical signs that makes a clean diagnosis impossible.

This complexity and heterogeneity is the basis of considerable tension between geneticists and neuroscientists, which is explored in the Nature editorial. Basically, to build up those massive sample sizes shown above geneticists are forced to lump together patients with a variety of clinical symptoms, thus essentially ignoring the complexity inherent in these conditions - a failure that neuroscientists find inexcusable. In turn, geneticists (like myself) get seriously annoyed by the tendency of neuroscientists to make big, bold claims about disease mechanisms based on studies with tiny sample sizes.

Both sides make reasonable criticisms. As I said in the quote from my previous article above, it seems likely that disease heterogeneity - that is, multiple diseases states with the same broad end point being simplistically lumped together - plays a major role in the failure of genome-wide association studies of psychiatric conditions; at the same time, the scientific value of much of the "sexy" neurobiology currently being published (e.g. functional MRI finds that conservatives have lower activity in "compassion" centres of the brain, or whatever) is sometimes highly questionable. Both sides of this scuffle have something to learn from their opponents.

The editorial argues, sensibly, that geneticists and neuroscientists just need to start getting along. The ideal situation is one in which rigorous clinical assessments are used to generate patient cohorts that are as homogeneous as possible that can then be subjected to large-scale genetic analysis. One especially promising avenue is the use of "endophenotypes" - that is, simple and easily quantifiable traits that are sometimes but not always associated with a particular disease. Cleanly defined endophenotypes, such as very specific dysfunctions of brain activity, may prove much more amenable to genetic dissection than the larger, more complex diseases they are associated with.

Comprehensively tackling the genetic of psychiatric conditions will require a forceful and combined approach drawing on the clinical expertise of neuropsychiatrists and the experience of geneticists in unravelling the genetic mechanisms of complex traits. To some extent this is happening already (no large genetics consortium would be naive enough to embark on a multi-million dollar project without consulting clinical experts) - but obviously there is considerable room for improvement.

Moving beyond common SNPs
Current genome-wide association studies currently rely largely on the use of single-letter variations in DNA called single nucleotide polymorphisms (SNPs), mainly because these are easy to analyse and can be simultaneously analysed in their hundreds of thousands using chip-based assays. For various reasons almost all of the SNPs on current genome-wide association chips are common sites of variation, present at a frequency of 5% or more in the population. However, recent studies have made it look increasingly likely that a large proportion of the genetic risk of common diseases lies in types of genetic variation that cannot be detected using common SNPs: rare variants, and large-scale rearrangements of DNA known as structural variation.

The approaches required to capture these variants are already pretty well-known, although they remain expensive and technically challenging. In an ideal world, genome-wide association studies would be truly genome-wide - in other words, they would utilise the entire DNA sequence of all of the patients and controls in the sample to find every possible genetic variant that might contribute to disease. Unfortunately, such an approach is currently out of reach, for several reasons:

1. The cost of DNA sequencing is still too high;
   
2. The computing power required to analyse the unbelievable volumes of data generated by such a project would be astronomical;
   
3. Statistical issues associated with examining so many data-points from each patient and control would greatly increase the required sample sizes, driving costs and computational requirements up even higher; and
4. Our ability to predict the effects of most genetic variants on human biology - which would be important for understanding which of the millions of rare variants found in such a study are actually harmful - is still far too weak.
   

Each of those challenges (particularly the costs of sequencing) will ultimately be overcome, but this will take time. In the meantime, there are some less powerful but technically feasible approaches that are likely to yield some useful results over the next few years. One of these is carefully targeted sequencing of a small fraction of the genome, consisting of candidate genes considered a priori likely to have some role in the disease in question (an approach strongly advocated by Walter Bodmer in a recent perspective article in Nature Genetics), which will identify at least some of the rare variants that underlie disease risk. To nail the structural variation, new chips that can pick up even small regions of variable copy number (that is, duplications and deletions) have already had some success in identifying potential genetic changes underlying sporadic cases of schizophrenia.

Both approaches have their limitations. The success of the candidate gene approach will be constrained by researchers' ability to identify the genes most likely to be involved in a particular disease - but in fact our currently severaly limited understanding of disease genetics is precisely why we need to study this issue in the first place! (In the Nature news piece, Harvard's Steven Hyman memorably describes this approach as "like packing your own lunch box and then looking in the box to see what's in it.") And while chip-based detection of structural variation is rapidly increasing in resolution, it's extremely difficult to determine which of the variants identified in a study are disease-causing and which are harmless polymorphisms - this is currently done probabilistically, by showing that there is an enrichment of new variants in disease cases compared to controls, but this approach cannot tell you which of the identified variants are actually causative.

From psychiatric genetics to genetic psychiatry?
There are several important reasons researchers are interested in the genetics of mental illness: identifying causal genes helps to dissect out the molecular pathways involved in disease, and may help to pull out otherwise invisible sub-types of a disease; studying "extreme" mental phenotypes may illuminate the genetic basis of variation in cognition and personality traits in "normal" people; and, perhaps most importantly, by identifying the genes underlying psychiatric diseases we may be able to target at-risk individuals for monitoring and intervention, potentially heading off severe disease before it takes hold.

In the headlong pursuit of these goals the field of psychiatric genetics has developed an unfortunate reputation built on bold claims made with limited evidence, and literally hundreds of reported associations that have completely failed to stand up to replication. Just a couple of years ago the shiny new tools of large-scale genomics promised an end to this ignoble period in the history of the field; unfortunately, the introduction of larger samples, higher genomic coverage and increased statistical rigour has not brought the desired clarity to the field, but rather seems to have increased the levels of confusion and uncertainty.

If anything, that crucial third goal - using genetic to predict the risk of mental illness - now appears further away than it did just a couple of years ago. Back in early 2007 we didn't have many convincing genetic predictors of mental illness, but at least it was possible to imagine that emerging genomic technologies might identify a small core set of large-effect variants that would help clinicians to predict disease risk. Right now we still don't have many useful genetic predictors, and that illusion of hope is gone.

In summary: while there's no doubt that these conditions do have a strong genetic basis, it's now abundantly clear that this basis is frighteningly complex, with common variants of moderate-to-large effect - the types of variants that would be most useful for risk prediction - being essentially absent. It's going to take many years, massive cohorts, the clever application of new genomic technologies, and a willingness from both neuroscientists and geneticists to listen to one another to move this field forward.

(Brain scan image from Science Photo Library.)

* Thanks to reader Chris for providing me with the citation, which I had carelessly misplaced!


 Subscribe to Genetic Future.