Uncovering the complex genetics of human temperament

Abstract

Experimental studies of learning suggest that human temperament may depend on the molecular mechanisms for associative conditioning, which are highly conserved in animals. The main genetic pathways for associative conditioning are known in experimental animals, but have not been identified in prior genome-wide association studies (GWAS) of human temperament. We used a data-driven machine learning method for GWAS to uncover the complex genotypic-phenotypic networks and environmental interactions related to human temperament. In a discovery sample of 2149 healthy Finns, we identified sets of single-nucleotide polymorphisms (SNPs) that cluster within particular individuals (i.e., SNP sets) regardless of phenotype. Second, we identified 3 clusters of people with distinct temperament profiles measured by the Temperament and Character Inventory regardless of genotype. Third, we found 51 SNP sets that identified 736 gene loci and were significantly associated with temperament. The identified genes were enriched in pathways activated by associative conditioning in animals, including the ERK, PI3K, and PKC pathways. 74% of the identified genes were unique to a specific temperament profile. Environmental influences measured in childhood and adulthood had small but significant effects. We confirmed the replicability of the 51 Finnish SNP sets in healthy Korean (90%) and German samples (89%), as well as their associations with temperament. The identified SNPs explained nearly all the heritability expected in each sample (37-53%) despite variable cultures and environments. We conclude that human temperament is strongly influenced by more than 700 genes that modulate associative conditioning by molecular processes for synaptic plasticity and long-term memory.

Simulation of model overfit in variance explained with genetic data

“…The result of these simulations are presented in Figure 1. Within sample scoring, where the identification of SNPs and their effects was estimated and tested within a single sample as described [1,4,5] yielded an average estimate of 60% variance explained (SD = 0.008). This is not less than the variance explained reported by the authors in their non-random datasets (57% and 48% for the Finnish samples applied to character and temperament, respectively [4,5]).…”

“…However, the extent of inflation in the described estimate of "heritability" (the within-sample variance explained; as reported in the abstracts of both original papers [4,5]) is so substantial as to be uninterpretable.…”

“…To demonstrate the extent of effect size inflation resulting from the above-described method, I present a simulation in which random genotypes and phenotypes were generated and Steps 1 and 3 were applied as described by the authors [1,4,5]. Here, however, Step 2 was instead executed as random selection of SNPs for inclusion in the score.…”

“…An author response [1] to my previous commentary [2] claimed that the application of PGMRA [3] in parallel papers on the genetics of character [4] and temperament [5] was not subject to concerns about overfitting. Unlike polygenic scoring, the goal was not to identify a generalizable combination of genetic variants to apply to outside samples.…”

“…The authors interpreted the linear regression model R 2 obtained from Step 3 as a measure of successfully identified heritability. Regardless of the method applied to select SNPs in Step 2whether it is global or local, data-driven [4,5] or based on theory [6,7] -the combination of Steps 1 and 3, as described, causes inflation of the estimated effect size.…”