Intelligence differences are substantially heritable, but studies to date have failed to find replicable associations between SNPs and cognitive traits that account for variation in intelligence in the normal population 
. One potential source of the missing heritability is copy-number variation. Following a similar approach to Williams et
, in which the combined effect of rare CNVs on variation in ADHD was investigated in a sample of 366 cases and 1047 controls, we examined whether variation in rare CNVs had any effect within older cohorts with intelligence distributions in the normal range had any effect on the variation in cognitive ability. No significant combined effect of rare CNVs was found on intelligence in our combined sample of over 3,000 elderly individuals.
We found that the total load of CNVs longer than 500 kb per individual was not significantly associated with fluid- or crystallized-type intelligence phenotypes. Neither was total length of copy number variants, nor the total number of genes disrupted by rare CNVs. Testing for differences in intelligence between individuals carrying CNVs known to effect neurocognitive phenotypes, and non-carriers found a suggestive effect of SHANK3
on fluid intelligence scores. SHANK3
is a post-synaptic density protein involved in the regulation of synaptic transmission, and has been implicated in both autism and schizophrenia. SHANK3
is within the region of the chromosome 22q13.3 deletion syndrome, which is characterized by neonatal hypotonia, global developmental delay, severe cognitive deficits, normal to accelerated growth, absent to severely delayed speech, autistic behaviour, and minor dysmorphic features 
. Haploinsufficiency of this gene as a major causative factor in the neurologic symptoms of 22q13 deletion syndrome 
, and Gauthier et al. 
identified two de novo mutations (R1117X and R536W) in two families with schizophrenia, in patients also displaying borderline or mild mental retardation. A recent GWAS study on cognitive phenotypes which used SNPs to tag common CNVs 
, found no significant association between these tagging markers and any of the measured cognitive phenotypes after correcting for multiple testing. This study by Need et al.
focussed on associations between cognitive phenotypes and specific copy-number variants associated with psychiatric illnesses in a sample of 1,000. Similarly, Saus et al.
, using a candidate gene approach, found no significant differences between rare variants in cases and controls in major depressive disorder, bipolar disorder, schizophrenia or anxiety disorders, testing a control group of 341 individuals against case samples of ~200 per disorder.
Comparing our results to Yeo et al.
, who found significant associations between rare deletions and variation in intelligence in a sample of 77 individuals, we failed to replicate their observed significant association between the extent of rare deletions and general intelligence, even when applying the same set of CNV QC criteria to our larger data set There are several possible reasons why this may be so. Whereas the phenotypic measures of intelligence that were used are broadly comparable, there are slight differences in our CNV calling and QC methods, the major one being our use of a minimum 500 kb cut-off in CNV length, but no such cut-off was used by Yeo et al.
This is apparent from the overall rate of CNVs per individual in the two studies, with Yeo et al.
finding an average of 17.42 CNVs per individual, with an average deletion rate of 10.95 (SD
5.48) and insertion rate of 6.47 (SD
9.82); compared to our observations of ~0.05 in total, a substantial disparity between the two rates, and between Yeo et al.
and other studies of ‘rare’ CNVs which include a length cut-off 
. The rationale behind such a cut-off is that the calls for longer CNVs are more accurate than shorter variants, and including short variants in an association analysis may increase the probability of type-I errors. When applying the QC criteria of Yeo et al.
to our data (removing the length restriction, and excluding common CNVs at the 5% threshold), the observed average rate of CNVs called per individual was 15.48 (SD
7.59). None of the regression analyses performed on CNVs called using these criteria showed a significant effect in our sample (Tables S3 and S4).
The samples studied in the present report comprise individuals within the normal cognitive range and we would expect the rate of CNVs detected to be similar to that obtained in other groups of healthy individuals that have undergone the same CNV quality control and selection procedures. Several studies have investigated the effect of long, rare CNVs on disease susceptibility, using the same length and frequency criteria employed here. Comparing our rates of CNV detection to the control groups of these studies, our observed values of 0.052 and 0.053 are comparable to the values of 0.05 observed by Blauw et al.
, slightly lower than the value of 0.075 reported by Williams et al.
, and lower than the values of 0.12 reported by the ISC 
, 0.17 in Pinto et
, and 0.1924 in Bochukova et al.
. Some of these discrepancies can partially be accounted for by differences in genotype platforms and CNV calling algorithms. Blauw et al.
genotyped their control samples using a number of different platforms, but analysed only CNVs called by markers common to all: these were effectively the markers present on the HumanHap 300 array, comprising ~300 K SNP markers and lacking CNV specific probes. The ISC study 
used genotypes derived from Affymetrix 5.0 and 6.0 arrays, and noted differences between these two arrays within their study. Bochukova et al.
also used the Affymetrix 6.0, whereas Pinto et al.
used the Illumina 1M chip for genotyping, and both observed a higher rate of CNVs per sample. Whenever a minimum number of SNPs is used as a criterion to define a CNV, there will be de facto
more CNVs called on chips with higher marker densities.
Although the absolute rates of CNVs called differ between studies, the proportions of deletions detected compared to duplications are more consistent between studies. Of our total observed variants, 24.6% (41/167),are deletions in, compared to 28.0% in Bochukova et al.
, 28.3% in ISC, and 29.2 % in Pinto et al
. The anomalously large proportion of 72.0% deletions in Blauw et al.
may be due to the set of markers used detecting more deletions. The 16.7% duplications called in Williams et al.
is smaller than other studies, perhaps due to the small sample size.
There are several other reasons why we may have failed to detect any effect of rare CNVs on cognitive ability in the cohorts analysed here. Primary amongst them, we are unlikely to have captured all of the genetic variation present within our samples. The Illumina 610 Quad chip used for genotyping in this study contains several non-polymorphic markers in known copy number variable regions, but will still not capture all of the variation present. With subsequent developments in microarrays, such as the Illumina 1M array used by Yeo et al.
, and the advent of reasonably priced whole-genome sequencing, we could capture a higher proportion of the actual variation. Other factors beyond genetic and structural variation may also contribute towards variation in general cognitive ability, including environmental variation and gene methylation or other epigenetic effects. Significant regions for rare CNVs reported in other neurocognitive disorders are not found significant by genome-wide associations using SNPs. Need et al.
suggest that looking for the effect of rare variants enriched in schizophrenia patients within a healthy population may reveal an association. Of the regions we examined, SHANK3
remained significant following permutation analysis, suggesting that this gene may be involved in normal variation in fluid intelligence. However, of the three CNV carriers identified in our samples that overlap the SHANK3
region, two carried duplications (copy number 3) and one a deletion (copy number 1), suggesting that imbalance rather than copy number per se
may be important, but this counter-intuitive observation requires further investigation.
Many illnesses are associated with lower cognitive ability, and there is evidence that these states often involve structural genetic variation. Cognitive impairment is often a symptom in genomic syndromes associated with specific large-scale structural variation, including Williams-Beuren, Smith-Magenis, and Velo-Cardio-Facial Syndrome amongst others 
. These syndromes are characterised by recurrent deletions or duplications at specific loci, which are large enough to be detected using Fluorescence In-Situ Hybridisation (FISH) or other microscopic techniques. Recent advances in microarray technology have allowed detection and characterisation of sub-microscopic structural variants, and found them to be ubiquitous throughout the genome. These copy-number variants are a major source of normal human genetic variation 
, but have also been found to be associated with complex disorders, including many neuro-psychiatric conditions, (for example, ADHD, depression, schizophrenia and autism-spectrum disorders).
To conclude, we find that, within the analytical limitations of the detection system available to us, there is no evidence for the effect of total CNV load on intelligence within the normal older population. Looking at specific CNV regions, we find evidence to suggest that copy number variation in the SHANK3,region where copy number variation has been previously associated with susceptibility to autism and schizophrenia, is associated with normal variation in fluid intelligence. However, our study does not preclude further contributions of CNVs at either extremes of the normal range of intelligence, or indeed on an individual by individual basis. New tools, including whole genome resequencing of individuals and their relatives with life-course measures of intelligence would be valuable in further resolving the important issue of identifying genetic contributions to individual differences in intelligence.