Multivariate pattern analysis of schizophrenia-associated genetic variants in relation to cognitive endophenotypes

Abstract

BACKGROUND: Polygenic risk scores for schizophrenia – aggregate measures of genetic risk –predict a variety of endophenotypes, suggesting that risk genes confer vulnerability via intermediate levels of impairment, like cognition. However, this method offers no insight into which genes or combinations of genes drive these effects. Furthermore, it cannot capture complex, non-linear effects (i.e., multi-gene interactions), which are known to play a role in the etiology of schizophrenia.

METHODS: We used random forest, an ensemble machine-learning approach, to predict six cognitive phenotypes in a sample of patients and controls (N = 739) using genetic variants previously identified in a large genome-wide association study of schizophrenia. Cross-validation was applied to the discovery sample and results were validated in an independent sample of similar ancestry (N = 364). For comparison, we ran generalized linear models (GLM) including all variants, as well as calculated polygenic scores, to predict each of the same outcomes.

RESULTS: Random forest models for visual memory, verbal memory, and processing speed replicated in an independent sample. Conversely, no GLM models replicated and polygenic scores for only one cognitive outcome significantly predicted performance in an independent sample.

CONCLUSIONS: Identifying genotypic patterns can increase the predictive power of genotypes on cognitive phenotypes, suggesting that genetic effects on cognition may depend on particular combinations. As such, pattern analysis can provide an advance in genotype-phenotype mapping that may aid in delineating the neurobiological pathways by which known risk loci for schizophrenia confer vulnerability.