Explaining the missing heritability of psychiatric disorders

Evidence from family, twin and adoption studies indicates that psychiatric disorders are substantially heritable. Heritability is usually expressed as the proportion of trait variance attributable to additive genetic factors (narrow sense heritability: h²). The h² estimates for schizophrenia, attention‐deficit/hyperactivity disorder, autism spectrum disorder and bipolar disorder are all >0.66, and are substantial for a range of other psychiatric conditions ¹ .

This evidence has motivated the application of increasingly sophisticated genomic approaches, including genome‐wide association studies (GWAS) and next generation sequencing, that have identified a large number of genetic risk factors across a range of psychiatric conditions ² . These studies revealed that psychiatric disorders are highly polygenic, with the major component of the heritability captured so far coming from common alleles (population frequency >0.01) detected in GWAS.

While this is extremely encouraging, and has set up an empirical platform upon which future progress towards precision psychiatry can be built ² , estimates of h² accounted for by the genetic variants identified in GWAS have always been substantially lower than the estimates of h² from family, twin and adoption studies. This shortfall is not a peculiarity of psychiatric disorders; it is also seen in many polygenic diseases and traits, and has been termed the “missing heritability”.

Three main explanations for this missing heritability have been proposed3, 4. First, it is possible that the estimates of h² from family, twin and adoption studies were inflated due to confounding factors such as shared environment. Second, estimates of h² from genomic studies may be deflated as they do not account for non‐additive genetic effects such as dominance and gene‐gene interactions. Finally, it may be the case that many risk alleles have simply not been identified by GWAS, either because their effects are too small or because they are too uncommon.

While all of these hypotheses remain plausible, the last one has received support from recent studies of polygenic traits and diseases, suggesting that many causal variants remain unidentified. In order to understand this, a brief explanation of GWAS is required. These studies involve genotyping single nucleotide polymorphisms (SNPs) that are common in the population (typically 500,000 ‐ 1 million SNPs with a population frequency >5%). Because common SNPs tend to be correlated with their neighbours – a phenomenon known as linkage disequilibrium (LD) – the genotypes of additional SNPs can be inferred through a statistical process known as “imputation”. This greatly increases the number of SNPs available to GWAS (typically >10 million SNPs with a population frequency >1%). When researchers seek associations in GWAS, they need to correct for the large number of statistical tests by taking a stringent threshold for statistical significance (known as genome‐wide significance). This greatly reduces the occurrence of false positives, but at the expense of causing many real associations to be missed.

Early studies that revealed the missing heritability focused only on SNPs that met genome‐wide significance. Subsequent studies have shown that more accurate and larger estimates of h² can be obtained by considering all available SNPs together, including imputed as well as directly genotyped SNPs, and by using data from reference samples that have undergone whole‐genome sequencing (WGS) to allow better imputation of rare variants.

When these approaches are implemented, the proportion of h² that is captured increases to around one‐ to two‐thirds of that expected in polygenic traits and diseases ⁴ , with h² estimates for schizophrenia, bipolar disorder and autism being 0.23, 0.25 and 0.17, respectively ⁵ . This indicates that a proportion of the missing heritability was carried by SNPs that currently lie below the genome‐wide significance threshold and also those that were insufficiently correlated with common SNPs to allow accurate imputation. It is, therefore, anticipated that the increased power of GWAS obtained from a substantial increase in both the number of common SNPs and the sample size will result in many more risk variants of small effect meeting genome‐wide significance, as well as improving estimates of heritability ⁴ .

However, the ability of common SNPs used in GWAS to capture the effects of variants with which they are in low LD is limited. The application of exome sequencing and WGS to complex disease cohorts has confirmed the presence in the human genome of a large number of rare genetic variants (defined as having a population frequency <1%). Importantly, these are not well correlated through LD with common SNPs and are therefore not accurately imputed in GWAS.

Recent work applying WGS to a large population cohort ⁶ has shown that estimates of heritability made using rare as well as common variants are much closer to those predicted from family studies for both height and body mass index, with much of the increase coming from SNPs that could not be accurately imputed from GWAS.

It is well recognized that, when compared to height and body mass index, many psychiatric disorders are under greater negative selection, and this is expected to result in a greater contribution from rare risk alleles. It is, therefore, plausible that rare genetic variants could be particularly relevant to psychiatric disorders, meaning that future WGS studies in large samples could prove to be particularly fruitful.

The prospect of large scale WGS studies in psychiatry is certainly exciting and will likely reveal much about genetic architecture and biology, as well as delivering better predictive tools. Short‐read sequencing (SRS), based on compiling reads from <150bp segments, is currently the most widely used approach to WGS, because of its low cost and high throughput. It is particularly powerful in identifying rare single nucleotide variants and small insertion/deletions ⁷ . Robust approaches have been recently introduced to detect structural variants such as duplications, deletions, inversions, and other changes involving larger DNA segments (generally greater than 50‐100 bases long) that are likely to be relevant to psychiatric disorders ⁸ .

While SRS will undoubtedly be increasingly and fruitfully applied in psychiatric genomics in the coming years, it has limitations imposed by the fact that it works by stitching together short reads in silico. This means that there are regions of the genome which are difficult or impossible to read, such as those containing large structural variants, repetitive sequences, extreme guanine‐cytosine content, or sequences with multiple homologous elements within the genome. This is sometimes known as the “dark genome”.

There are now a number of long‐read sequencing (LRS) platforms that allow the analysis of segments of the human genome up to 200kb, and these are capable of shining a light into the dark genome. Emerging studies using LRS are identifying larger, more harmful structural variants and long repetitive elements7, 9, both of which are candidates for involvement in psychiatric disorders.

Psychiatric genomics is a work in progress. GWAS have been hugely successful in identifying the role of multiple common variants, but recent work on missing heritability suggests a need to focus now on rare variants, and in the next few years we can expect studies based upon both SRS and LRS technologies to do this.

Fully characterizing the genetic architecture of psychiatric disorders is likely to improve polygenic risk prediction for both screening and stratification, allow a better understanding of the underlying biological mechanisms of disease, and broaden the landscape of pharmaceutical targets ² .