The major findings of this study include: 1) a simple equation relating altered genetic output from variant alleles and RAFs to the total number of risk genes required to reach threshold for schizophrenia, 2) support for the occurrence of protective variants undergoing positive selection and a means to deal with them computationally and 3) the discovery of extensive gene-gene interactions among schizophrenia risk genes and a strategy for including them in genetic liability calculations. Finally, we provide a quantitative basis for harmonizing views about the relative contributions of SNPs and CNVs to the genetic architecture.
Wray and Visscher (2010) made impressive headway in characterizing the genetic architecture of schizophrenia and our analysis generally complements theirs. Specifically, we have developed a straightforward mathematical model for estimating risk gene burden with explicit terms for RAF, the effect size due to risk-gene variation (VAO), gene-gene interactions and inclusion of protective variants. The model provides a quick snapshot of the different factors that influence the genetic liability for schizophrenia. Parameters can be varied to take into account as much complexity as desired.
Analysis of the PGC data revealed that there was an equal number of schizophrenia risk variants with increased vs. decreased occurrence (negative vs. protective effects?) in the case population. Previous studies have revealed evidence for both positive and negative selection of variants that show differences in frequency comparing control subjects with those affected with schizophrenia (Liu et al. 2019; Polimanti and Gelernter 2017; Pardiñas et al. 2018; Yao et al. 2020). For an optimized phenotype, most mutations are expected to produce a negative effect because there is a statistically greater chance of an adverse outcome the closer the phenotype is to optimum. Since this trend was not observed in the PGC dataset, it suggests that most of the traits underlying schizophrenia are not optimized at this point or perhaps are not optimizable due to pleiotropy.
The work of Nishino et al. (2018) and Hess et al. (2021) suggest that many protective variants will be expressed in the genomes of the case population. These protective (resilience) variants counterbalance the adverse effects of risk variants (Hess et al. 2021). Therefore, the balance between adverse and protective alleles must tilt substantially toward the former to produce symptoms of schizophrenia. Taken together with the observation that risk variants for schizophrenia are enriched in essential genes (Kasap et al. 2018), these findings suggest that negative selection of adverse risk variants is minimal, perhaps because the associated genes have been optimized for critical pleiotropic purposes. Furthermore, SNPs associated with schizophrenia risk may affect the expression of multiple genes, sometimes with opposing functions (Peng et al. 2021). Therefore, the net result on relevant behavior may reflect the sum of effects on many genes (Peng et al. 2021), which can be accounted for with our concept of functional locus output (FLO).
A special subset of variants associated with syntenic blocks of genes exhibited intermediate RAFs and average OR scores under little apparent selective pressure. We previously speculated that the process of gene amalgamation to form these syntenic blocks may have been accompanied by creation of recombination coldspots, which would impede selection but also preserve weak risk variants in the DNA blocks (Kasap et al., 2018). The data presented here support this earlier suggestion.
Consistent with studies of gene interactions in bipolar disorder and depression (Franklin and Dwyer 2020; Sall et al. 2021), candidate risk genes for schizophrenia showed greater network interaction than random gene sets of the same size. These gene-gene interactions may cause network ripple effects among the connected genes that amplify the impact of small individual VAOs. Others have observed a similar phenomenon among risk genes for schizophrenia (Cheah et al. 2016; Su et al. 2017). Furthermore, this notion is similar to that of Greenspan (2001) who described the functional connectivity of gene networks in terms of flexibility and pleiotropy. Despite some overlap with the omnigenic model of Boyle at al. (2017), there are also important distinctions. At some level, all genes are interconnected based on how the genome evolved (Dwyer, 2020). This common origin can obscure real differences between genes and disorders. For example, the risk genes for bipolar disorder and depression are much more interconnected (2-3-fold higher number of gene interactions) than the schizophrenia risk genes. If the various risk genes participated in omnigenic interactions, we would not expect to observe these large differences, especially in psychiatric disorders with overlapping symptoms and involving similar cell types.
Our studies focused on gene-gene interaction networks, whereas others have explored pathway networks to gain insights into schizophrenia (Willsey et al. 2018). The networks have been derived from gene function analysis, protein-protein interactions, co-expression data and other sources (Gilman et al. 2012; Schwarz et al. 2016; Walker et al. 2019; Willsey et al. 2018). These studies have been very informative about possible mechanisms contributing to pathogenesis; however, they do not directly address the genetic architecture. Because the pathways, proteins and genes involved in schizophrenia are organized into tangible networks, genetic risk calculations need to reflect this inherent connectivity among networked genes. Here, we have attempted to quantitatively assess the impact of gene interactions on overall risk burden.
Multiple factors contribute to the total risk burden for schizophrenia: risk variants (SNPs, CNVs), protective variants, gene interactions, allele frequency and the overall effect size of the genetic mutation. We have represented these various factors in a simple equation that can reconcile relative contributions from common non-coding SNPs and CNVs. CNVs, null mutants and major functional mutations (loss or gain of function) will produce significant liability for schizophrenia; however, their effects manifest on a background of significant risk alleles (Bassett et al. 2017; Bergen et al. 2019; Tansey et al. 2016). Using estimated parameters, we calculated that a deletion CNV is roughly equivalent to around 15% of the total single-nucleotide variants required to reach threshold for disease. By itself, a single CNV in a risk gene would not be sufficient to cause schizophrenia without the contributions of many additional background risk variants.
Interpretation of the genetic architecture in schizophrenia must be considered cautiously due to various inherent limitations. GWAS and CNV studies have likely identified instances of false positive candidate genes, whereas some actual risk genes may have been missed. Rare MAFs have thus far been largely neglected; however, these rare alleles are likely to have large effects based on previous work (Bergen et al. 2019; Suárez-Rama et al. 2015; The International Schizophrenia Consortium 2008). Environmental factors and epigenetic changes will also complicate interpretation of genetic influences on disease susceptibility. We have limited our simulations to average values for parameters such as RAFs and VAO, so the real situation is much more complicated; however, the equation can manage greater complexity than presented here. Nevertheless, overall trends such as a requirement for a substantial number of risk variants to reach threshold and the potential significance of genetic interactions are based on solid observations and logic. Finally, there may be alternative ways to handle gene-gene interactions mathematically; however, this work provides a useful conceptual framework to the problem.
According to the model, non-affected individuals must harbor a non-trivial complement of risk alleles that experience little selection. Furthermore, threshold combinations of risk alleles will be inherited at a set frequency in the population – a phenomenon previously described as inevitable bad luck (Kasap et al. 2018). Therefore, schizophrenia differs significantly from pathological genetic conditions such as inherited metabolic disorders or rare Mendelian diseases. Instead, schizophrenia risk variants via their different VAOs and RAFs determine whether certain quantitative traits fall in the normal range. In our distant ancestors, concurrent expression of various suboptimum traits may have carried little penalty for the individual. Without the complexity, artificiality and stress of modern society, someone showing a collection of traits that would be diagnosed today as schizophrenia may still have been largely functional when living a simpler existence in nature. This emerging view of schizophrenia has important implications for diagnosis and treatment.