Identification of a genetic cluster influencing memory performance and hippocampal activity in humans

First stage.

To investigate the complex genotype–phenotype relationship, we conducted a linkage-disequilibrium-based multilocus genetic search (17). First, we selected the human homologues of 47 genes with well established molecular and biological functions in animal memory (1–5, 18–35) (Table 1, which is published as supporting information on the PNAS web site). Depending on genomic length, we mapped these genes with 160 SNPs commonly observed in Caucasians to enable haplotype reconstruction and, thereby, to maximize the coverage of a gene’s naturally occurring variability. Guided by recent recommendations for the large-scale analysis of complex traits (17), we used a two-stage approach. In the first stage, we used logistic regression models to extract a set of loci influencing human episodic memory under rather liberal statistical criteria (α = 5%). Sixteen variations fulfilled this first criterion. Because nonrandom genetic heterogeneity (i.e., population structure) in outbred populations may lead to spurious associations, we calculated the level of genetic heterogeneity in our study population by genotyping each subject for 318 SNPs located in both genic and nongenic regions and distributed over all autosomes. Structured association analysis (36) revealed that the allele-frequency divergence in our population was low, and that the individual genetic background values were normally distributed (Fig. 1). Eight subjects were identified as outliers (i.e., beyond the 25% or 75% limits of the normal distribution curve). Importantly, the findings of stage 1 remained unchanged after exclusion of these subjects from the statistical analyses. In addition, we included each subject’s genetic background value as a covariate in the logistic regression model. Again, the results remained unchanged.

Second stage.

In the second stage, the 16 significant variations from stage 1 were stringently evaluated by a permutation-based analysis, which reduced the multiple testing burden and allowed for between-loci interactions (37). This method, termed set association, evaluates sets of polymorphic markers and provides a cluster of significant markers with a single test statistic (37, 38). Seven variations finally contributed to a cluster with significant impact on episodic memory (P = 0.00008, Fig. 2). Association due to linkage disequilibrium with SNPs in adjacent genes was excluded (data not shown). We also performed post hoc analyses with 53 SNPs deliberately selected because of their a priori unlikely involvement in memory (i.e., SNPs in regions not harboring any genes or SNPs within genes hitherto unrelated to known signaling pathways of memory; Table 2, which is published as supporting information on the PNAS web site). These SNPs were treated statistically in exactly the same manner as our hypothesis-driven SNPs. In the first liberal stage, one of 53 SNPs exceeded the 0.05 significance level. After set-association analysis (stage 2), this SNP was no longer significant (Table 2).

The seven-SNP cluster was further used for the calculation of an individual’s memory-related genetic score, termed individual memory-associated genetic score (IMAGS). Briefly, IMAGS corresponds to the number of memory-associated genetic variations present in a subject weighted by the variations’ effect size. The genetic score of individuals carrying none of the seven variations associated with better episodic memory was set equal to 0, which is the minimal possible IMAGS value. Accordingly, IMAGS of individuals possessing one variation was set equal to the corresponding set-association value for the particular variation (e.g., 0.20 for ADCY8, 0.12 for PRKCA). For two or more variations associated with better episodic memory, set-association statistics of the corresponding variations were added (38). In our study population, IMAGS correlated positively and significantly with memory performance (β = 0.232, T = 4.23, P = 0.000009, linear regression analysis).

IMAGS represents physiological variability in genes encoding such well characterized memory-related molecules as adenylyl cyclase (18, 19), PKA (20–22), CAMKII (23–26), NMDA receptor (27–31), metabotrobic glutamate receptor (32, 33), and PKC (34, 35). Importantly, IMAGS did not correlate significantly with immediate recall (r = −0.058, P = 0.3). Performance in this task requires high levels of attention and motivation along with well functioning working memory. Therefore, the genotype-dependent differences in delayed episodic memory were not caused by genotype effects on confounding factors such as motivation, attention, or working memory. To further demonstrate the specificity of the IMAGS, we randomly selected seven SNPs (rs3744215, rs892200, hCV2522892, hCV347736, hCV1521704, rs3763679, and hCV1341801) from the group of genetic variations that did not emerge as memory-related in the first stage and calculated a negative-control IMAGS (ncIMAGS) in exactly the same way as we did for the IMAGS. We did not observe any significant correlations between ncIMAGS and delayed memory performance (β = 0.026, T = 0.473, P = 0.6, linear regression analysis).

Stage 1.

Genetic variability in 47 memory-related genes was assessed by genotyping 160 SNPs and reconstructing haplotypes. Analysis of linkage disequilibrium and haplotype reconstruction was done with powermarker Version 3.22 (www.powermarker.net). We used forward and backward logistic regression models, controlled for age, gender, education, and immediate recall performance, to examine the influence of SNPs and haplotypes on episodic memory performance. SNPs were entered by using the additive model, and haplotypes were included as binomial categorical variables (42). At this stage, 16 SNPs and haplotypes fulfilled the criterion of P ≤ 0.05.

Stage 2.

Variations fulfilling the criterion of stage 1 were further evaluated by the sumstat program (Rockefeller University, New York), which has been developed for statistical analysis with the set-association method (37). Unlike some conventional methods of data reduction (e.g., factor analysis), the set-association method uses relevant sources of genetic information such as allelic association and Hardy–Weinberg disequilibrium (HWD). Information is combined over multiple markers and genes in the genome, quality control is improved by trimming SNPs with high HWD values, and permutation testing limits the overall false-positive rate (37). Specifically, for each marker, a contingency χ² statistic testing the association of genotype with phenotype status is calculated. These χ² statistics are then ranked from largest to smallest. Progressively larger sums (S_j) are then calculated over the j largest χ² statistics. For example, S₁ is the largest χ² statistic of association. S₂ is the sum of the largest and second largest. S₃ is the sum of the largest, second largest, and third largest, and so on. The empirical significance level (P_j) for each S_j is evaluated by permutation methods carried out under the null hypothesis of no genetic association. The smallest of the empirical significance levels (i.e., P_jmin) identifies the best and most parsimonious model predicting phenotype status. Importantly, the set-association method has been shown to be of superior power compared with conventional locus-by-locus analyses and to successfully capture statistical interactions among genes (37, 38, 43–45). In this study, we used the maximum possible number of permutation tests (n = 50,000) to calculate the genetic cluster’s significance.

Calculation of population structure was done by using the structure program (46), following the developers’ instructions (36).

Power of the sample.

The power of our sample to detect correlations of |R| ≥ 0.2 at α ≤ 0.05 is >90% (47).

Subjects.

Twenty-one females and 11 males, mean age 22.5 ± 2.5 (standard error) years, participated in this study.

Episodic memory task.

Stimuli consisted of 16 face-profession pairs for associative learning and 24 head contours for the baseline condition. The learning condition consisted of four blocks with four trials of 6 s each. Consequently, each block lasted 24 s and was announced by an instruction slide. The instruction for associative learning of the face-profession pairs was to imagine the presented person acting in a scene of the written profession. Subjects indicated by button press whether they found it easy or difficult to imagine a scene. For recall testing of the learned face-profession associations, the faces were presented alone with the instruction to recall each person’s occupation. The baseline condition consisted of four blocks with six trials of 4 s each. The task was to decide whether the area of the left or right ear of a head contour was larger. The sequence of conditions was counterbalanced across subjects.

Working memory.

The experiment included one fMRI time series with a two-back task for the assessment of working memory and a baseline task (x-target) for the assessment of concentration. The two-back task required subjects to respond to a letter repeat with one intervening letter (e.g., S –f –s –g). The x-target task required subjects to respond to the occurrence of the letter x. Each task was given in five blocks of 26 s each. Blocks were announced by an instruction slide. Stimuli were 50 upper- or lowercase letters typed in black on a white background. Thirteen upper- or lowercase letters were presented per block for the duration of 2 s each.

Data acquisition.

Magnetic resonance (MR) measurements were performed on a 3-T Philips (Eindhoven, The Netherlands) Intera whole-body MR scanner equipped with a transmit–receive body coil and a commercial eight-element head coil array. Functional data were obtained from 32 transverse slices parallel to the AC–PC plane covering the whole brain with a measured spatial resolution of 2.8 × 2.8 × 4 mm³ (acquisition matrix 80 × 80) and a reconstructed resolution of 1.7 × 1.7 × 4 mm³. Data were acquired by using a SENSE-sshEPI (48) sequence with an acceleration factor of R = 2.0. Other scan parameters were Echo time = 35 ms, Repetition time = 3,000 ms, and θ = 82°.

Analysis of fMRI data.

Image pre- and postprocessing and the statistical analyses were performed with spm2 (Wellcome Department of Imaging Neuroscience, University College London, London). Standard preprocessing procedures were applied, i.e., realignment, normalization, and spatial smoothing (8 mm) (49). On the single-subject level, data were analyzed according to the fixed-effects model (spm2). The six head movement parameters were included in the model as confounding factors. Data were high-pass-filtered with a specific filter value for each fMRI time series. This value was determined according to 2× (stimulus onset asynchrony)× repetition time. On the second level, within-subject contrasts were entered into a random effects analysis (spm2), which accounts for variance among subjects (50). We also analyzed the correlation between the within-subject encoding contrasts (learning–baseline) and IMAGS (simple regression, spm2). Threshold was set at a P < 0.001 level, uncorrected for multiple comparisons. spm2 coordinates refer to standard brains from the Montreal Neurological Institute.

Neuropsychology.

In addition to the fMRI experiment, subjects underwent detailed neuropsychological testing. Memory functions were assessed with the Wechsler Memory Scale Revised in German; intelligence with the Hamburg Wechsler Intelligence Test; spatial thinking with the Luria Mental Rotation Test; and executive functions with a verbal (S-words) and a nonverbal (five-point) fluency task, the Kramer Card Sorting Test, and the Stroop Test.

Supporting Tables