Computational dissection of human episodic memory reveals mental process-specific genetic profiles

Significance

Human memory is a highly heritable and complex trait that is the outcome of several neurobiologically distinct mental processes, such as learning, memory maintenance, and emotional modulation of memory strength. Such processes, which are not always amenable to direct observation, can be dissociated using computational models. Here we provide converging evidence linking computational model parameters and related behavioral and neuroimaging phenotypes to sets of biologically related genes in several populations of healthy young and elderly adults. Our findings suggest that distinct genetic profiles underlie specific mental processes of human episodic memory. Understanding the genetic and neural bases of such processes is essential for the development of novel therapeutic approaches that could lead to a better treatment of neuropsychiatric disorders.

Abstract

Episodic memory performance is the result of distinct mental processes, such as learning, memory maintenance, and emotional modulation of memory strength. Such processes can be effectively dissociated using computational models. Here we performed gene set enrichment analyses of model parameters estimated from the episodic memory performance of 1,765 healthy young adults. We report robust and replicated associations of the amine compound SLC (solute-carrier) transporters gene set with the learning rate, of the collagen formation and transmembrane receptor protein tyrosine kinase activity gene sets with the modulation of memory strength by negative emotional arousal, and of the L1 cell adhesion molecule (L1CAM) interactions gene set with the repetition-based memory improvement. Furthermore, in a large functional MRI sample of 795 subjects we found that the association between L1CAM interactions and memory maintenance revealed large clusters of differences in brain activity in frontal cortical areas. Our findings provide converging evidence that distinct genetic profiles underlie specific mental processes of human episodic memory. They also provide empirical support to previous theoretical and neurobiological studies linking specific neuromodulators to the learning rate and linking neural cell adhesion molecules to memory maintenance. Furthermore, our study suggests additional memory-related genetic pathways, which may contribute to a better understanding of the neurobiology of human memory.

Human memory is a highly heritable and complex trait that is the outcome of numerous neurocognitive processes (1, 2). Although many genetic variations associated with human memory performance have been identified (3–9), most were based on general memory scores, such as the number of recalled items. Such scores may serve as good correlates of overall memory ability, but they offer little insight into the genetic underpinnings of specific memory-related mental processes, such as encoding, memory maintenance, modulation by emotional arousal, or guessing strategies. Importantly, neurobiological and molecular profiles of such processes are known to be partly distinct (10–12). Because some of the mental processes involved in memory are not readily amenable to direct observation, computational modeling can be used to make inferences about them (13, 14). Model-based analyses provided insights into neurocomputational mechanisms of reward-based learning and decision making (15–17), related model parameters such as the learning rate to genetic polymorphisms (18, 19), and provided a computational explanation for the inverted-U-shaped relation between stress intensity and behavioral performance (20). As a relatively recent development, the model-based analysis approach has largely been missing from studies of human episodic memory and genome-wide association studies (GWAS).

GWAS already had identified a number of associations between single genetic markers and cognitive traits, such as episodic memory (7–9) and modulation of memory strength by negative emotional arousal (21). However, because many such traits are complex and multigenic, it is unlikely that a single genetic variant will explain a sufficient amount of behavioral variability. As a statistically more powerful and biologically meaningful alternative, gene set-based methods of linking sets of biologically related genes with traits of interest have been developed (22, 23). They test whether a group of genes (gene set) is more enriched in associations with a studied phenotype than would be expected by chance. Such studies identified gene sets related to neuropsychiatric phenotypes (24, 25) and to cognition in healthy individuals, including working memory (26) and emotionally aversive memory (27).

In the present study, we performed a gene set enrichment analyses (GSEA) of specific parameters of episodic memory performance estimated from a computational model of verbal memory in young adults (21). We found four distinct gene sets associated with different computational parameters in independent samples, three of which also were linked to related measures of picture memory in young adults and verbal memory in elderly individuals. The L1 cell adhesion molecule (L1CAM) interactions gene set also was associated with differences in picture recognition-related brain activity in a number of frontal cortical areas.

Results

The verbal memory task contained neutral, positive, and negative words, which the participants had to recall freely first after their presentation and later after an ∼5 min delay. Participants in the discovery sample (n = 1,239; Table 1) and the replication sample (n = 526) performed the identical verbal memory task. Eight performance measures (PM_1–8; Table 2) indicated the number of correctly recalled words in each valence category (i.e., neutral, positive, and negative) and confabulative errors (words not on the learning list) in the immediate and the delayed recall. Because performance measures were partly correlated, the effective dimensionality of data was smaller: principal component analysis (PCA) revealed five major components (Fig. S1).

Table 1.

Samples and core phenotypes used

Sample name (location)	Phenotypes of interest	Sample size
Discovery (Zurich and Basel)	Computational model parameters	1,239
Replication (Basel)	Computational model parameters	526
Words/pictures* (Basel)	Free recall of pictures	493
Pictures/fMRI (Basel)	Free recall of pictures	835
	Picture recognition	822
	Working memory	825
	fMRI data on picture recognition	795
	fMRI data on working memory	797
AgeCoDe elderly (Bonn)	Free recall of words	743

^*This is a subsample of the Replication sample. All other samples are independent.

Table 2.

Description of performance measures (PM_1–8) in the verbal memory task and their population statistics

Performance measure	Discovery sample, mean ± SEM	Replication sample, mean ± SEM
PM1: Positive words correctly recalled immediately	7.666 ± 0.042	7.645 ± 0.058
PM2: Negative words correctly recalled immediately	8.318 ± 0.036	8.364 ± 0.050
PM3: Neutral words correctly recalled immediately	7.392 ± 0.047	7.345 ± 0.065
PM4: Mistakes made immediately after encoding	1.302 ± 0.038	1.164 ± 0.050
PM5: Positive words correctly recalled after 5 min	2.948 ± 0.043	3.055 ± 0.061
PM6: Negative words correctly recalled after 5 min	2.645 ± 0.041	2.882 ± 0.059
PM7: Neutral words correctly recalled after 5 min	2.448 ± 0.041	2.618 ± 0.056
PM8: Mistakes made 5 min after encoding	1.681 ± 0.048	1.727 ± 0.071

Fig. S1.

PCA of eight performance measures (PM1–PM8, described in Table 2) in the discovery sample (A) (modified from ref. 21) and the replication sample (B). Percentages denote variance explained by each component. The first five principal components (PC_1–5) explain more than 80% of the variance in the data; their loadings suggest that the first component (PC₁) can be approximately related to general learning ability and the second one (PC₂) to delayed memory recall (compared with immediate recall). The latter three components could be interpreted only in the discovery sample: PC₃ related approximately to errors, PC₄ to the recall of negative words, and PC₅ to the recall of positive minus neutral words.

Computational Modeling.

To study specific cognitive processes relevant for memory performance, we used a simple computational model of learning in the verbal task (21). Our model had the following parameters, which we aimed to infer from the behavioral data: the learning rate α and the Gaussian noise σ, the repetition-based memory improvement c (related to memory maintenance) and the forgetting rate γ, the modulation of memory strength by positive or negative emotional arousal ε_pos and ε_neg, the decision threshold β, and the sigmoidal steepness s (for a detailed description see Materials and Methods). Because the effective dimensionality of the behavioral data was lower than the number of model parameters, we empirically selected five of these parameters (α, β, c, ε_pos, ε_neg) to be estimated individually, keeping the three remaining ones fixed. After the estimation (Fig. S2), more than 99% of the resulting individual parameter sets passed the χ² test of goodness-of-fit [mean χ²(discovery) = 1.5057 and χ²(replication) = 1.4955], indicating that the model was sufficiently flexible. High correlations [mean ρ_Spearman(discovery) = 0.981 and ρ_Spearman(replication) = 0.986] among the 10 best parameter sets (hill climbing end points in stage 2 of the estimation) suggested high estimation reliability, and the presence of 99.8% of individual parameter sets within the middle 90% of the value ranges (except the bound of c = 1, because repetition should not weaken memories) indicated that parameter estimation bounds did not constrain the results.

Fig. S2.

Parameter estimation results for the discovery sample (A–D) (modified from ref. 21) and the replication sample (E–H). (A and E) Stage 1: Based on a moderately accurate estimation of model parameters, two selections of fixed parameters (γ, σ, s and σ, c, s) are chosen for hill climbing. The best 10 of all 56 possible three-of-eight combinations are shown. (B and F) Stage 2: The hill-climbing results of estimating three fixed parameters for the best model (with Gaussian noise σ, sigmoidal steepness s, and forgetting rate γ fixed) are shown, with bigger circles and lighter colors indicating better goodness-of-fit; the 10 best hill-climbing points (biggest light yellow circles) were selected for evaluating averages of all their possible combinations. (C and G) Stage 3: The 10 combinations with the best goodness-of-fit (χ²) are displayed. Overall, 267 of 1,023 combinations had better χ² than the best hill-climbing point (χ² = 1.522) for the discovery sample, as did 77 of 1,023 combinations for the replication sample (χ² = 1.507), suggesting that averaging parameters helps overcome gaps in step size and leads to refined parameter values. (D and H) Histograms of the best-fitting individual parameters show distributions with the following means: c = 1.95, α = 1.93, β = 1.27, ε_neg = 1.12, ε_pos = 1.09 for the discovery sample and c = 2.23, α = 2.04, β = 1.22, ε_neg = 1.13, ε_pos = 1.11 for the replication sample.

GSEA.

We first performed a GWAS with the estimated model parameters (α, β, c, ε_pos, ε_neg) with 1,239 individuals in the discovery sample (for quality control and statistics, see Materials and Methods). Based on the resulting P values of association with these parameters, we performed a GSEA using MAGENTA (23). Of the 1,411 gene sets used for input, we found false-discovery rate (FDR)-corrected enrichment for seven gene sets (Table 3). A second GSEA was performed for these gene sets in an independent sample of 526 individuals (Table 1), again using GWAS-derived P values of association with the estimated model parameters. Of the seven gene sets significantly enriched in the discovery sample, four largely nonoverlapping gene sets were also enriched in the replication sample (Table 3): amine compound solute-carrier (SLC) transporters associated with the learning rate α, L1CAM interactions associated with the repetition-based memory improvement c, and collagen formation and transmembrane receptor protein tyrosine kinase activity associated with ε_neg. Although behavioral measures of verbal memory were correlated with model parameters (Table S1), none of the gene sets reported above would have been detected if these behavioral measures had served as starting points for GSEA (using the same multiple testing correction threshold as for the model parameters); however, five of the seven gene sets identified in the discovery sample were nominally associated with the relevant behavioral measures.

Table 3.

Associations between computational model parameters for the verbal task and different gene sets

Model parameter	Database	Gene set (number of genes in gene set)	Discovery sample (n = 1,239)		Replication sample (n = 526)
			Pnominal	PFDR	Pnominal
Learning rate α	Reactome	Amine compound SLC transporters (27)	2.0 × 10−4	0.0405	0.0031
Decision threshold β	KEGG	Axon guidance (129)	4.4 × 10−5	0.0067	0.405
Repetition-based memory improvement c	Reactome	L1CAM interactions (86)	3.0 × 10−6	0.0011	0.0231
		Interaction between L1 and ankyrins (23)	1.0 × 10−3	0.0365	0.176
Modulation by negative emotional arousal εneg	Reactome	Collagen formation (58)	1.0 × 10−4	0.0129	0.0025
	Gene Ontology	Transmembrane receptor protein tyrosine kinase activity (43)	4.1 × 10−5	0.0124	0.0212
		Transmembrane receptor protein kinase activity (51)	1.0 × 10−4	0.0398	0.059

Only the gene sets surviving MAGENTA FDR correction in the discovery sample are shown. The replicated gene sets are shown in bold. Interaction between L1 and ankyrins is a subset of L1CAM interactions, and transmembrane receptor protein tyrosine kinase activity is a subset of transmembrane receptor protein kinase activity. L1CAM interactions contains four genes [EGFR, FGFR1, and neuropilin 1 and 2 (NRP1, and NRP2)] shared with transmembrane receptor protein tyrosine kinase activity.

Table S1.

Correlations of estimated model parameters with related behavioral measures of verbal task performance in the discovery sample and picture task performance in the replication (words/pictures) sample

Model parameter	Relevant phenotype in the verbal task	Correlation, Spearman’s ρ	Relevant free recall phenotype in the picture task	Correlation, Spearman’s ρ
Learning rate α	Percentage of correctly remembered words in the immediate recall	0.722	Percentage of correctly remembered pictures	0.152
Decision threshold β	Number of mistakes in the delayed recall	−0.825	Number of mistakes in the picture recall	−0.204
Repetition-based memory improvement c	Percentage of correctly remembered words in the delayed recall	0.607	Percentage of correctly remembered pictures	0.098
Modulation by positive emotional arousal εpos	Percentage of correctly remembered positive minus neutral words in the delayed recall	0.737	Percentage of correctly remembered positive minus neutral pictures	0.133
Modulation by negative emotional arousal εneg	Percentage of correctly remembered negative minus neutral words in the delayed recall	0.708	Percentage of correctly remembered negative minus neutral pictures	0.074

Correlations with verbal task performance are quite high because computational modeling was performed for the verbal task. Correlations with picture task performance suggest that the percentage of correctly remembered pictures in the free recall is related more to the learning rate α than to the repetition-based memory improvement c and hence is used for the replication of gene sets associated with α. All correlations are statistically significant.

We next examined if the four replicated gene sets showed enrichment in associations with related, albeit less specific, behavioral measures in different tasks and in additional populations (Table 1). Computational modeling was not practical for these tasks because of the low dimensionality of the behavioral data. In the sample of nondemented elderly individuals from the German Study on Aging, Cognition and Dementia in Primary Care Patients (AgeCoDe) (n = 743), participants performed a word-list verbal recall task from the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) neuropsychological test battery (28). In a picture task, performed by two populations of healthy young adults, participants were presented neutral, positive, and negative pictures, which they had to recall freely 10 min after presentation. After a longer delay, participants of the Basel pictures/functional MRI (fMRI) sample also were asked to perform picture recognition: provided with both previously shown and new pictures, they had to answer whether they had seen each picture. For all of these tasks, we selected behavioral measures most closely related to each computational parameter. For the learning rate α, immediate recall is more appropriate than delayed recall, because the latter also accounts for forgetting. However, because immediate recall was not available in the picture task, we used free delayed recall (see also Table S1). For the repetition-based memory improvement c, only delayed recall was appropriate, because immediate recall did not account for memory maintenance. In the picture task both free recall and recognition measured memory over delays; however, because we also aimed to study the neural correlates of memory maintenance using functional neuroimaging, which were available only for picture recognition data, we chose the latter phenotype. For negative emotional modulation factor ε_neg, only picture data were available; therefore we chose the difference between freely recalled negative and neutral pictures as a relevant phenotype. Based on the GWAS results of these phenotypes (population statistics are shown in Table 4), we used MAGENTA for testing the enrichment of the four gene sets associated with different computational parameters of the (original) verbal task. We found that amine compound SLC transporters were associated with immediate recall of words in the elderly sample and with free recall of pictures in both young samples; L1CAM interactions were associated with delayed recall of words in the elderly sample and with picture recognition in the pictures/fMRI sample; collagen formation was associated with free recall of negative minus neutral pictures in the words/pictures sample; and there were no significant associations with transmembrane receptor protein tyrosine kinase activity (Table 5). Because the arousal of negative pictures [2.36 (mean) ± 0.06 (SEM)] was higher than that of positive pictures (1.94 ± 0.04; Student t test; P = 0.000002), the association of collagen formation with picture recall may be mediated by arousal rather than by negative valence. We also checked if sex was correlated with any of the significant SNPs (one per gene) in any of the significantly enriched gene sets (Tables 3 and 5). No significant correlations, corrected for the number of SNPs per gene set, were found (Ps > 0.05).

Table 4.

Population statistics of relevant phenotypes for the picture task and for the word task used in the German elderly sample

Phenotype	Data sample	Mean ± SEM
Percentage of correctly remembered pictures in the free recall	Words/pictures (n = 493)	39.58 ± 0.49
Percentage of correctly remembered negative pictures minus neutral pictures in the free recall	Words/pictures (n = 493)	17.17 ± 0.59
Percentage of correctly remembered pictures in the free recall	Pictures/fMRI (n = 835)	42.57 ± 0.39
Percentage of correctly remembered negative pictures minus neutral pictures in the free recall	Pictures/fMRI (n = 835)	17.67 ± 0.45
Picture recognition	Pictures/fMRI (n = 822)
Percentage of correctly recollected pictures*		80.74 ± 0.55
Percentage of previously seen pictures rated as familiar		15.68 ± 0.49
Percentage of previously seen pictures rated as new		3.57 ± 0.15
Percentage of new pictures rated as known		1.02 ± 0.07
Percentage of new pictures rated as familiar		9.41 ± 0.28
Percentage of new pictures rated as new		89.57 ± 0.31
Number of correctly remembered words in the immediate recall	AgeCoDe elderly (n = 743)	19.55 ± 0.13
Number of correctly remembered words in the delayed recall	AgeCoDe elderly (n = 743)	5.88 ± 0.07

^*Previously seen pictures rated as known.

Table 5.

Association between the four replicated gene sets and other related phenotypes

Database	Gene set (number of genes in gene set, associated model parameter)	Relevant phenotype	Data sample	Association P value
Reactome	Amine compound SLC transporters (27, α)	Percentage of correctly remembered pictures in the free recall	Words/pictures	0.0302
			Pictures/fMRI	0.0001
		Number of correctly remembered words in the immediate recall	AgeCoDe elderly	0.032
	L1CAM interactions (86, c)	Percentage of correctly recollected pictures in picture recognition	Pictures/fMRI	0.0419
		Number of correctly remembered words in the delayed recall	AgeCoDe elderly	0.0068
	Collagen formation (58, εneg)	Percentage of correctly remembered negative pictures minus neutral pictures in the free recall	Words/pictures	0.0206
			Pictures/fMRI	0.145
Gene Ontology	Transmembrane receptor protein tyrosine kinase activity (43, εneg)	Percentage of correctly remembered negative pictures minus neutral pictures in the free recall	Words/pictures	0.147
			Pictures/fMRI	0.153

Neural Correlates of the Discovered Gene Sets.

Because the amine compound SLC transporters and L1CAM interactions gene sets were associated with free recall and recognition of pictures, respectively, in the pictures/fMRI sample, we examined neural correlates of these associations using functional neuroimaging. Participants in this sample performed the encoding and recognition parts of the picture task in the fMRI scanner; free recall was performed outside the scanner. We studied association with amine compound SLC transporters (related to the learning rate α) by using two encoding contrasts: all pictures vs. scrambled controls and the difference due to memory (Dm, i.e., subsequently remembered vs. not remembered pictures; 29). Because neural correlates of memory maintenance may not be apparent during encoding, we examined the association with L1CAM interactions (related to the repetition-based memory improvement c) using the recognition contrast of previously seen vs. new pictures, a task that engages the brain areas involved in memory maintenance and retrieval, such as prefrontal cortex and medial temporal lobe (30, 31). To capture multiallelic effects of each gene set on the differences in brain activity, we generated individual multilocus genetic scores (Materials and Methods). For L1CAM interactions, the score was built using 28 significant SNPs from the same number of genes (one most significant SNP per gene; Table S2); the genetic score for amine compound SLC transporters comprised 14 SNPs.

Table S2.

Composition of the L1CAM interactions genetic score: SNP IDs, tagged genes, statistics of association with picture recognition in the pictures/fMRI sample, reference allele with direction of association with picture recognition memory, major allele

SNP rsID	Gene symbol	Association P value	Reference allele	Direction of association	Major allele
rs9459771	RPS6KA2	0.0011	T	+	C
rs11715984	DLG1	0.0017	G	+	A
rs13258436	KCNQ3	0.0020	T	+	C
rs197721	ITGA9	0.0049	T	+	G
rs7153298	RPS6KA5	0.0071	T	−	T
rs17210194	CNTN6	0.0110	G	−	G
rs3506	AP2B1	0.0113	T	+	T
rs17547685	SH3GL2	0.0130	T	−	T
rs17333297	NRCAM	0.0149	G	−	A
rs6983315	FGFR1	0.0185	G	+	G
rs11009157	ITGB1	0.0195	G	−	G
rs4948381	ANK3	0.0235	G	+	A
rs442103	DPYSL2	0.0247	G	−	G
rs10504047	ANK1	0.0254	G	−	C
rs1478438	ITGA1	0.0267	T	−	T
rs7781264	EGFR	0.0281	T	+	C
rs645078	RPS6KA4	0.0311	C	−	A
rs11647503	CSNK2A2	0.0318	T	+	C
rs949215	LAMA1	0.0318	C	−	C
rs6580842	SCN8A	0.0335	G	+	A
rs883501	NRP2	0.0335	T	+	A
rs1421933	ITGA2	0.0352	G	−	G
rs17115275	NCAM1	0.0352	G	−	A
rs7643664	ALCAM	0.0360	C	+	C
rs16943748	CLTC	0.0369	G	+	G
rs6793943	SCN5A	0.0407	T	+	C
rs11240351	CNTN2	0.0407	T	+	T
rs1331325	NRP1	0.0407	T	−	C

Genetic score-dependent analysis revealed no whole-brain familywise error (FWE)-corrected correlations with differences in brain activity for amine compound SLC transporters. However, the L1CAM interactions genetic score was negatively correlated at the whole-brain FWE-corrected level with differences in the activity of numerous frontal regions: large clusters in the left superior frontal cortex and the left inferior frontal gyrus (Fig. 1, A and B and Table 6) and small clusters in the bilateral orbitofrontal cortices and insulae (Table 6). The results were very similar if the two most significant SNPs per gene were used to build a genetic score (Table S3) or if only correctly rated pictures were included in the contrast (recollected previously seen pictures vs. new pictures rated as new). We also performed a genetic score-independent analysis of brain activity, which revealed substantial differences in activity at all cluster peaks dependent on the L1CAM interactions genetic score (Fig. 1, C and D), suggesting that higher values of the genetic score reflect smaller increases in frontal activity during the recognition of previously seen pictures as compared with the viewing of new pictures. To ensure that the observed effect specifically reflected differences in the episodic memory of previously seen pictures, we also used the contrast between recollected previously seen pictures vs. previously seen pictures rated as familiar or new (n = 786). Of the 238 voxels significant in the contrast between previously seen vs. new pictures, the great majority also passed small-volume correction (SVC) for the latter contrast [162 voxels in superior frontal cortex, peak coordinates (0, 17, 48), T = −4.93; four voxels in left lateral orbitofrontal cortex, peak coordinates (−30, 22, −8), T = −5.07], with the peak voxels also surviving whole-brain correction.

Fig. 1.

Functional neuroimaging of picture recognition (the contrast between previously seen vs. new pictures). (A and B) Differences in brain activity dependent on the L1CAM interactions genetic score. (C and D) Genotype-independent differences in brain activity. Color-coded t values are shown. The maps are centered at (−3, 28, 40) in the left superior frontal cortex (A and C) and at (−47, 44, −8) in the left inferior frontal gyrus (B and D). Only voxels surviving whole-brain FWE correction are shown (P whole-brain FWE-corrected < 0.05; |T| > 4.77; n = 795). Activations are overlaid on sections of the study-specific group template.

Table 6.

Association between the L1CAM interactions genetic score and brain activity differences during picture recognition

Regional correspondence of the maximum	Left/right	No. of voxels	Peak MNI coordinates			Peak statistics		Empirical P value
			X	Y	Z	T	Pnominal
Superior frontal cortex (79%), caudal anterior cingulate cortex (2%)	Left	187	−3	28	40	−6.19	9.4 × 10−10	0.0001
Pars orbitalis cortex (34%), pars triangularis cortex (28%), rostral middle frontal cortex (16%)	Left	39	−47	44	−8	−5.70	1.6 × 10−8	0.0021
Lateral orbitofrontal cortex (20%), insula (23%)	Right	6	30	22	0	−5.16	3.2 × 10−7	0.018
Lateral orbitofrontal cortex (73%), insula (8%)	Left	4	−28	22	−4	−4.87	1.4 × 10−6	0.053

Previously seen vs. new pictures fMRI contrast was used. Only clusters with at least three voxels surviving whole-brain FWE correction are shown (P whole-brain FWE-corrected < 0.05; |T| > 4.77; n = 795). Region names are in accordance with the FreeSurfer nomenclature; probabilities are in accordance with the in-house atlas. For each cluster we also provide empirical P values of the peak voxel based on 10,000 randomly generated genetic scores (null distribution).

Table S3.

Association between differences in brain activity during picture recognition and the L1CAM interactions genetic score that uses the two most significant SNPs per gene

Regional correspondence of the maximum	Left/right	No. of voxels	Peak MNI coordinates			Peak statistics
			X	Y	Z	T	Pnominal
Superior frontal cortex (80%)	Left	76	−3	30	40	−5.77	1.1 × 10−8
Pars orbitalis cortex (56%), pars triangularis cortex (22%), rostral middle frontal cortex (4%)	Left	23	−47	41	−12	−5.42	8.0 × 10−8
Pars orbitalis cortex (35%), rostral middle frontal cortex (1%)	Right	23	41	44	−12	−5.37	1.0 × 10−7
Lateral orbitofrontal cortex (20%), insula (23%)	Right	21	30	22	0	−5.32	1.3 × 10−7
Superior frontal cortex (84%)	Right	3	11	28	60	−5.02	6.2 × 10−7

The fMRI contrast between previously seen vs. new pictures was used. Only clusters with at least three voxels surviving whole-brain FWE correction are shown (P whole-brain FWE-corrected < 0.05; |T| > 4.77; n = 795). Region names are in accordance with the FreeSurfer nomenclature (71); probabilities are in accordance with the in-house atlas.

Because the differences in brain activity at all cluster peaks dependent on the genetic score of L1CAM interactions also were negatively associated with recognition performance (Table S4, n = 795), we performed a mediator analysis to exclude the possibility that the association between the genetic score and the differences in brain activity could be explained fully by the correlated behavioral phenotype. We examined whether the differences in activity in these brain regions were mediating the correlation between the genetic score and picture recognition performance. We found that all four clusters acted as significant mediators of the genetic score–behavior relationship (for more details about mediation, see Materials and Methods and Fig. S3).

Table S4.

Association between differences in brain activity during picture recognition (fMRI contrast of previously seen vs. new pictures) and the percentage of correctly recollected pictures

Regional correspondence of the maximum	Left/right	MNI coordinates at the maximum			T value
		X	Y	Z
Superior frontal cortex (79%), caudal anterior cingulate cortex (2%)	Left	−3	28	40	−12.13
Pars orbitalis cortex (34%), pars triangularis cortex (28%), rostral middle frontal cortex (16%)	Left	−47	44	−8	−9.11
Lateral orbitofrontal cortex (20%), insula (23%)	Right	30	22	0	−9.08
Lateral orbitofrontal cortex (73%), insula (8%)	Left	−28	22	−4	−9.65

Test statistics are shown at the maxima of clusters associated with the L1CAM interactions genetic score (Table 6). They indicate that differences in brain activity at these clusters are also associated with picture recognition performance in the same direction as the L1CAM interactions genetic score. Region names are in accordance with the FreeSurfer nomenclature (71); probabilities are in accordance with the in-house atlas.

Fig. S3.

Scheme and results of the mediation analysis. (A) The overall predictor–outcome relationship is denoted as “c,” and the direct effect controlling for the mediator is denoted as “c′.” The a × b effect tests the significance of mediation (c–c′). For more details see Materials and Methods. (B) Mediation results at (−3, 28, 40) and (−47, 44, −8), depicting voxels for which the contrast in brain activation related to the previously seen vs. new pictures was considered a significant mediator of the relationship between the genetic score and recognition performance (blue scale), at a significance threshold of P(FDR) = 0.0075. The value of this threshold for whole-brain correction might be unusually high because of correlations between the genetic score and the behavioral phenotype. As a control, random shuffling of fMRI images between subjects was performed 10 times, always leading to no voxels surviving the FDR significance threshold. We also found that behavioral phenotype was a significant mediator of the relationship between the genetic score and differences in brain activity. Results are shown based on the red-to-yellow scale, superimposed on the original ones, at a significance threshold of P(FDR) = 0.0117. Effect sizes with neuroimaging and behavioral phenotypes as mediators were comparable in the four clusters associated with the genetic score. For the coronal and horizontal sections left in the picture refer to the right side of the brain and vice versa.

As two additional controls, we first extracted average beta values for the two largest clusters, whose activity was associated with the picture recognition phenotype at the whole-brain FWE-corrected level [cluster 1: peak coordinates (0, 28, 40), T = −12.36, number of voxels = 2,917; cluster 2: peak coordinates (−47, 41, 16), T = −10.57, number of voxels = 1,541] and subjected them to GSEA using MAGENTA. The L1CAM interactions gene set was significantly enriched in both cases, with nominal P = 0.013 for each cluster. Second, we generated 10,000 genetic scores based on randomly selected SNPs (28 SNPs per genetic score, as for L1CAM interactions) and the picture recognition phenotype, using the same procedure as for the real genetic scores. None of the random scores showed statistically significant, whole-brain-corrected correlation with the neuroimaging phenotype (the contrast between previously seen vs. new pictures) after correction for the number of genetic scores (threshold P_nominal = 2.2 × 10⁻¹⁰). Based on the null distribution computed with the 10,000 random scores, we also provide empirical P values for the four clusters, which describe the probabilities of their peak voxel significance being reached for the most significant voxel across the whole brain using the random scores (Table 6).

To evaluate the specificity of our imaging genetics finding, we examined if the correlation between the L1CAM interactions genetic score and differences in frontal activity also could be observed during encoding (the subsequent memory contrast, Dm). No such correlation was found at either the whole-brain FWE-corrected level or when applying SVC for 238 voxels significant in the recognition contrast, suggesting either that the relation between L1CAM interactions and differences in frontal activity was not evident during encoding and developed later or that it was specific to the cognitive demands of the recognition task. Because the participants in the pictures/fMRI sample also performed a working memory task (N-back), we examined if the reported association with L1CAM interactions was related to working memory. The MAGENTA analysis performed using P values of GWAS of N-back performance revealed significant enrichment for L1CAM interactions (P = 0.023, n = 825). The multilocus genetic score based on N-back performance was negatively correlated with the one based on picture recognition performance (ρ_Spearman = −0.140, P = 5.9 × 10⁻⁵, n = 815), suggesting that, with regard to this gene set, better working memory was related to worse picture recognition. Because common frontal mechanisms may mediate both effects, we examined the neural correlates of working memory using fMRI (2-back vs. 0-back contrast, n = 797) and found no correlation between differences in brain activity and the working memory-based L1CAM interactions genetic score at either the whole-brain or SVC levels of significance (number of voxels for the latter, 238). We also examined if the multilocus genetic score based on N-back performance was related to differences in brain activity in the recognition of previously seen vs. new pictures. Again no significant voxels were found at either whole-brain FWE-corrected or SVC thresholds, suggesting that there was no neural evidence implicating L1CAM interactions in working memory or common mechanisms with picture recognition.

Discussion

In this study we show robust associations between multiple gene sets and different computational parameters of verbal memory as well as other relevant neurobehavioral traits in four samples of young and elderly individuals (Table 1). Specifically, amine compound SLC transporters were associated with the learning rate α in two samples of young individuals performing the verbal task, with free recall of pictures in two samples of young individuals, and with immediate recall of words in an elderly sample. L1CAM interactions were linked to the repetition-based memory improvement c (reflecting memory maintenance) in two young samples, to picture recognition and large clusters of differences in recognition-related frontal activity in a third young sample, and to delayed recall of words in the elderly sample. Finally, collagen formation and transmembrane receptor protein tyrosine kinase activity were associated with modulation of verbal memory strength by negative emotional arousal (ε_neg) in two young samples, with the former gene set also linked to free recall of negative minus neutral pictures in young individuals.

At this point, it is important to specify the term replication in the context of gene set-based association studies. In contrast to SNP-based studies, in which the object of replication is the significant SNP (or variants in high linkage disequilibrium), the object of replication in gene set-based studies is the significant gene set. Across different samples, replicated gene sets typically are composed of different tagging SNPs (32, 33), a phenomenon related to the allelic and locus heterogeneity of complex traits (22, 34). In our study we did not observe substantial tagging SNP overlap between replicated gene sets across samples (Tables S5 and S6).

Table S5.

Tagging SNPs of the amine compound SLC transporters gene set

Gene symbol	Tagging SNP in the discovery sample	Tagging SNP in the replication sample and its LD with the discovery SNP		Tagging SNP in the pictures/fMRI sample and its LD with the discovery SNP		Tagging SNP in the words/pictures sample and its LD with the discovery SNP		Tagging SNP in the AgeCoDe sample and its LD with the discovery SNP
SLC6A11	rs6442209	rs746278	0.0054	rs9868537	0.1251	rs1601370	0.0192	rs966030	0.0055
SLC14A2	rs7232775	rs7226686	0.0068	rs7232775	1	rs17670646	0.0845	rs4890292	0.0256
SLC6A20	rs17279437	rs7653640	0.0077	rs2251109	0.097	rs17213127	0.0058	rs11718841	0.017
SLC6A2	rs734980	rs1861646	0.004	rs36016	0.0106	rs879522	0.0118	rs10521329	0.0006
SLC44A3	rs7550014	rs11165265	0.0191	rs3849308	0.0297	rs859064	0.3058	rs7550014	1
SLC6A6	rs2164576	rs17308574	0.016	rs3773167	0.0162	rs4685160	0.1064	rs6774615	0.0685
SLC6A7	rs3776087	rs9885318	0.2705	rs3776087	1	rs4601032	0.0252	rs3776087	1
SLC6A18	rs13361701	rs6554678	0.0822	rs13361701	1	rs7704728	0.0489	rs6554677	0.1335
SLC18A2	rs363334	rs363224	0.0007	rs363223	0.0006	rs363278	0.0023	rs363271	0

For each gene listed in Table 7, the corresponding tagging (i.e., the most significant) SNP in each sample is shown together with its linkage disequilibrium (LD) r² value with the tagging SNP from the discovery sample. Gene descriptions, chromosome numbers, and association P values are provided in Table 7.

Table S6.

Tagging SNPs of the L1CAM interactions gene set

Gene symbol	Tagging SNP in the discovery sample	Tagging SNP in the replication sample and its LD with the discovery SNP		Tagging SNP in the pictures/fMRI sample and its LD with the discovery SNP		Tagging SNP in the AgeCoDe sample and its LD with the discovery SNP
RPS6KA2	rs2281143	rs1894660	0.0009	rs9459771	0.0005	rs9459777	0.0008
KCNQ3	rs2597348	rs2721905	0.1884	rs13258436	0.0091	rs1457784	0.0006
DPYSL2	rs7829347	rs17322275	0.0244	rs442103	0.0083	rs9314325	0.0935
ITGA9	rs399703	rs197763	0.8689	rs197721	0.057	rs6782856	0
ANK3	rs11593565	rs4948254	0.0053	rs4948381	0.001	rs12355908	0.0107
CNTN6	rs1353825	rs9880762	0.0007	rs17210194	0.0102	rs3772295	0.1717
SH3GL2	rs10491540	rs1998247	0.0414	rs17547685	0.0092	rs3808776	0.0101
SCN2A	rs353128	rs13025009	0	rs13023748	0.1728	rs10497258	0.0004
FGFR1	rs11777067	rs2288696	0.0958	rs6983315	0.3326	rs2978073	0.003
NRCAM	rs6466228	rs425013	0.0114	rs17333297	0.0019	rs404287	0.0313
ITGA1	rs870992	rs7718758	0.0224	rs1478438	0.0518	rs2938793	0
LAMA1	rs11661841	rs10163962	0.1539	rs949215	0.0009	rs8084092	0.0085
EGFR	rs6593210	rs11238349	0.0017	rs7781264	0.0123	rs4947986	0.0606
NCAM1	rs17115275	rs884110	0.0109	rs17115275	1	rs1245135	0.0038
SPTBN1	rs6545424	rs10166797	0.0988	rs6715538	0.0069	rs17343939	0.1586
CSNK2A2	rs1393203	rs37358	0.0706	rs11647503	0.0063	rs2550362	0.0203
ANK1	rs7823955	rs11784321	0.0266	rs10504047	0.5597	rs2304872	0.0052
ITGA2	rs153141	rs1421940	0.6648	rs1421933	0.0069	rs1421933	0.0282
CNTN1	rs1119060	rs10784951	0.0219	rs12829787	0.0237	rs7970805	0.0116
NRP2	rs849581	rs849570	0.7338	rs883501	0.0027	rs3771000	0.0004
NRP1	rs2506150	rs1331312	0.0144	rs1331325	0.005	rs2243668	0.0008

For each gene listed in Table 8, the corresponding tagging (i.e., the most significant) SNP in each sample is shown together with its LD r² value with the tagging SNP from the discovery sample. Gene descriptions, chromosome numbers, and association P values are provided in Table 8.

The correspondence of significant genes between different samples and phenotypes shows that for amine compound SLC transporters genes encoding GABA, norepinephrine, choline, and vesicular amine transporters are significant in multiple samples, including the discovery sample (Table 7 and Table S5). These genes have been implicated in epilepsy, substance abuse, attention deficit hyperactivity disorder, schizophrenia, and neurodegenerative disorders (35–37). On a neurocomputational level, acetylcholine plays an important role in the encoding of new memories, may act as a switch between encoding and retrieval (38), and has been suggested as a neural correlate of the learning rate (39). Variations in several dopaminergic genes have been linked to the learning rate in the reinforcement learning framework (18, 19) and in economic games (40). Therefore, our results support the notion that the relations between certain neuromodulators and the learning rate generalize for various types of learning models.

Table 7.

Correspondence of significant amine compound SLC transporters genes among samples

Gene name	Chromosome	Description	Discovery sample	Replication sample	Pictures/fMRI sample	Words/pictures sample	AgeCoDe sample
SLC6A11	3	GABA transporter 3	0.0184	0.0167	0.0084	0.0004	0.0707
SLC14A2	18	Urea transporter 2	0.0179	0.0159	0.0141	0.0049	0.0619
SLC6A20	3	X transporter protein 3	0.0059	0.1808	0.0092	0.2848	0.0024
SLC6A2	16	Norepinephrine transporter	0.0491	0.1976	0.0141	0.0097	0.0162
SLC44A3	1	Choline transporter-like protein 3	0.0184	0.0529	0.0235	0.0047	0.2009
SLC6A6	3	Taurine transporter	0.0287	0.0371	0.0437	0.0932	0.0235
SLC6A7	5	Brain-specific l-proline transporter	0.0184	0.0275	0.0295	0.1263	0.1680
SLC6A18	5	Sodium channel-like protein	0.0377	0.0482	0.0417	0.0234	0.1872
SLC18A2	10	Vesicular amine transporter 2	0.0491	0.1651	0.0388	0.0205	0.1940

Only genes with significant SNPs in the discovery sample and at least two other samples are shown (the SNP rsIDs and linkage disequilibrium between them are provided in Table S5). The P values of association with relevant phenotypes (Table 5) are provided. Genes are sorted by the product of these values.

For L1CAM interactions, genes encoding fibroblast growth factor receptor (FGFR), ankyrin at node of Ranvier (ANK3), contactins, integrins, and important neural cell adhesion molecules (NRCAM and NCAM1) were significant in multiple samples including the discovery sample (Table 8 and Table S6). These molecules have been implicated in neuronal development and signaling and are essential for synaptic plasticity and its stability in many different brain regions, spatial learning, and memory consolidation (41–44), supporting our discovered association with a memory maintenance-related computational parameter. Some of them, e.g., ANK3 and NCAM1, also have been related to neurodevelopmental and psychiatric disorders (45) and to chronic stress-related cognitive disturbances (41).

Table 8.

Correspondence of significant L1CAM interactions genes between samples

Gene name	Chromosome	Description	Discovery sample	Replication sample	Pictures/fMRI sample	AgeCoDe sample
RPS6KA2	6	Ribosomal S6 kinase 3	0.0005	0.0177	0.0011	0.0020
KCNQ3	8	Potassium channel subunit Kv7.3	0.0017	0.0044	0.0020	0.0048
DPYSL2	8	Collapsin response mediator protein 2	0.0059	0.0062	0.0247	0.0009
ITGA9	3	Integrin alpha-9	0.0027	0.0008	0.0049	0.0858
ANK3	10	Ankyrin 3, node of Ranvier	0.0023	0.0197	0.0235	0.0013
CNTN6	3	Contactin 6	0.0009	0.0131	0.0110	0.0235
SH3GL2	9	Endophilin A1 BAR domain	0.0032	0.0031	0.0130	0.0578
SCN2A	2	Sodium channel, type II, α SUBUNIT	0.0024	0.0072	0.0676	0.0087
FGFR1	8	Fibroblast growth factor receptor 1	0.0125	0.0018	0.0185	0.0260
NRCAM	7	Neuronal cell adhesion molecule	0.0004	0.0409	0.0149	0.0578
ITGA1	5	Integrin, alpha 1	0.0079	0.0345	0.0267	0.0024
LAMA1	18	Laminin, alpha 1	0.0310	0.0879	0.0318	0.0003
EGFR	7	Epidermal growth factor receptor	0.0011	0.0328	0.0281	0.0459
NCAM1	11	Neural cell adhesion molecule 1	0.0144	0.0037	0.0352	0.0840
SPTBN1	2	Beta-II spectrin	0.0273	0.0045	0.1314	0.0162
CSNK2A2	16	Casein kinase II subunit alpha	0.0059	0.0807	0.0318	0.0190
ANK1	8	Ankyrin 1, erythrocytic	0.0152	0.0114	0.0254	0.0788
ITGA2	5	Integrin, alpha 2	0.0161	0.0312	0.0352	0.0247
CNTN1	12	Contactin 1/F3	0.0140	0.0290	0.0632	0.0235
NRP2	2	Neuropilin 2	0.0266	0.0181	0.0335	0.0504
NRP1	10	Neuropilin 1	0.0240	0.0275	0.0407	0.0823

Only genes with significant SNPs in the discovery sample and at least two other samples are shown (the SNP rsIDs and linkage disequilibrium between them are provided in Table S6). P values of association with relevant phenotypes (Table 5) are provided. Genes are sorted by the product of these values.

Among collagen formation genes significant in both discovery and replication samples, TLL2 was linked to mouse avoidance behavior and human bipolar disorder (46), and ADAMTS14 was linked to multiple sclerosis (47). Collagens are essential for connective tissue and mobility, but they also have been implicated in anxiety- and fear-related behaviors in humans and mice (48–50). Interestingly, we found that the collagen formation gene set was significantly enriched (P = 0.005) in the AgeCoDe elderly sample with Geriatric Depression Scale score (51) as phenotype. Finally, for the transmembrane receptor protein tyrosine kinase activity genes significant in both discovery and replication samples, we did not find references implicating them in emotional memory, anxiety, or affective disorders.

Despite the convergence of our results, particularly for amine compound SLC transporters and L1CAM interactions, the exact processes and gene systems influencing the learning rate and memory maintenance cannot be inferred directly, because GSEA are fundamentally based on gene sets defined in different databases [Reactome, Kyoto Encyclopedia of Genes and Genomes (KEGG), Gene Ontology, BioCarta] that are in the process of being refined and extended. Our use of related behavioral phenotypes rather than computational model parameters in the picture task and the AgeCoDe elderly sample also is a notable limitation. Because a variety of relevant behavioral measures are necessary for fitting nontrivial computational models, it is essential for modelers and experimentalists to collaborate during the process of task design. Furthermore, we note that none of our previously reported GWAS SNPs associated with memory performance, such as those of KIBRA (7) and CTNNBL1 (9), was significantly associated with the model parameters. Memory recall performance and model parameters are not identical phenotypes; thus SNPs significantly associated with a model parameter can be missed when a behavioral measure serves as a starting phenotype, and vice versa.

Functional neuroimaging results revealed a robust negative correlation between the L1CAM interactions genetic score and the differences in activity in left superior and inferior frontal gyri, which have been implicated in the recognition of words and pictures and in working memory in healthy individuals (31, 52–54) and Alzheimer’s patients (55, 56). The superior frontal gyrus has been implicated in cognitive control (57), and the inferior frontal gyrus has been implicated in retrieval attempt and effort (58, 59) and depth of memory (60). Reverse inference analysis using Neurosynth (a database of fMRI coordinates and associated keywords automatically extracted from the literature; 61) showed that peak coordinates of the largest cluster dependent on the L1CAM interactions genetic score (−3, 28, 40) were related mostly to the term “response conflict” (Z-score = 5.03, posterior probability = 0.82). Thus, one explanation for the negative correlation between the differences in the activity in these regions and the genetic score is that these frontal activity patterns reflect increased effort or conflict between different options in the case of poorer recognition performance and less effort/conflict in the case of good memory. The negative correlation between the L1CAM interactions genetic score and the differences in activity in the left superior and inferior frontal gyri also may indicate that the reported regions code specifically for familiarity rather than recollection of pictures, as suggested by dual-process models (62, 63). However, because of the ceiling effect in our picture recognition data (with less than 4% of previously seen pictures rated as new and more than 80% as recollected), the percentages of correctly recollected pictures were strongly inversely correlated with percentages of previously seen pictures rated as familiar (ρ_Spearman = −0.96). Therefore, based on our data, it was impossible to dissociate recollection and familiarity effectively at either the behavioral or neural level (where contrasts based on familiar pictures were significantly underpowered), which remains an important question for future studies.

In conclusion, by computationally dissecting episodic memory performance into specific components, we were able to identify distinct genetic profiles that underlie specific mental processes of human episodic memory.

Materials and Methods

Participants, Data Preprocessing, and GWAS Quality Control.

We had three samples of healthy, young Swiss subjects: the discovery sample (used for associations with model parameters in the verbal task) consisted of the Zurich subsample (192 males, 514 females, age mean ± SD = 21.92 ± 2.95 y), and the Basel subsample (261 males, 504 females, age 22.47 ± 3.62 y). The Basel words/pictures sample (200 males, 410 females, age 22.48 ± 3.51 y) was used for the direct replication of associations with model parameters in the verbal task and for testing associations with performance in the picture task; the Basel pictures/fMRI sample (404 males, 588 females, age 22.47 ± 3.35 y) was used for testing associations with performance in the picture task and studying their neural correlates with neuroimaging. Participants’ physical and mental health was assessed by standard questionnaires and telephone interviews. Indications for exclusion were the recent use of medications influencing the central nervous system, pregnancy, visual system disorders, hormonal irregularities, a body mass index outside the 18.0–35.0 range (20.0–35.0 for men), recent or regular use of cannabis, excessive use of alcohol, use of other psychoactive drugs, history of mental illness, cancer, brain surgery, and for the fMRI sample, the presence of metal implants and claustrophobia. Written informed consent was obtained from the subjects after a complete description of the study. The ethics committees of the Cantons of Zurich and Basel, Switzerland approved the study protocols.

A total of 930,856 SNPs were genotyped. For association testing, markers with call rate less than 0.95, with minor allele frequency less than 0.05, and with Hardy–Weinberg equilibrium P < 0.05 were excluded. Population stratification was assessed by analyzing all genomewide, array-based autosomal SNPs passing the quality-control (QC) criteria with EIGENSTRAT (64). PCA was applied to each population (Zurich words subsample, Basel words subsample, Basel words/pictures sample, and Basel pictures/fMRI sample) to reduce genetic variation to a few dimensions. For PCA, default parameters were used (i.e., definition of 10 principal components in five iterations; the outlier criterion was 6 SDs). EIGENSTRAT identified a number of individuals deviating from a large, genetically homogenous population cluster: 112 in the Zurich words subsample, 68 in the Basel words subsample, 55 in the Basel words/pictures sample, and 112 in the Basel pictures/fMRI sample. These outliers were removed from further analyses. Because learning and memory abilities deteriorate with age (65), we also excluded the outliers (individuals with age more than 2 SDs from the mean in each population) from further analyses: 18 in the Zurich words subsample, 44 in the Basel words subsample, 29 in the Basel words/pictures sample, and 47 in the Basel pictures/fMRI sample. After outliers based on population stratification and age were excluded, the following numbers of participants remained for the final analysis: 1,239 in the discovery sample (582 in the Zurich words subsample, 657 in the Basel words subsample), 526 in the Basel words/pictures sample (picture data available for 493 subjects), and 835 in the Basel pictures/fMRI sample [picture recognition data available for 822 subjects and fMRI data for 795 subjects (for the remaining subjects the fMRI data were corrupted); working memory data were available for 825 subjects, and fMRI data were available for 797 subjects; picture recognition and working memory data were available for 815 subjects].

GWAS Statistics and GSEA.

All GWAS were run under the assumption of an additive model and using the Spearman rank correlation test. Sex chromosomes were excluded from the analysis. The resulting P values of autosomal SNPs in the GWAS (one list per sample and phenotype) served as input for the GSEA, which was performed using MAGENTA (23). Briefly, the method first maps SNPs to genes (each SNP can be counted toward only one gene) and then assigns each gene an SNP association score (i.e., the maximum SNP P value within ± 0 kb of the annotated gene; 26). By applying a stepwise multiple linear regression, the analysis is corrected for the following confounders: gene size, number of SNPs, number of independent SNPs, number of recombination hotspots, linkage disequilibrium, and genetic distance (measured in centimorgans). Last, a gene set enrichment-like statistical test is applied to determine whether a gene set is enriched for highly ranked P values compared with a gene set of identical size randomly drawn from the genome. Based on recommendations in the literature (23, 26), we used a 75th percentile cutoff to determine the significance of a gene, because that cutoff has the optimal power for weak genetic effects that are expected for complex, polygenic traits. The gene sets used were extracted and curated from the MSigDB v3.1 database (www.broadinstitute.org/gsea/msigdb), including gene sets from different online databases (KEGG, Gene Ontology, BioCarta, and Reactome). We used a gene set size of 20 to 200 genes to avoid both overly narrow and overly broad functional gene set categories (26), resulting in 1,411 gene sets to be analyzed. Furthermore, to reduce bias, genes mapping in the high linkage disequilibrium (LD) extended major histocompatibility complex were excluded from the respective gene sets (66). FDR multiple testing correction was used to correct for the number of gene sets.

Memory Testing—the Verbal Task.

Subjects viewed six series of five semantically unrelated nouns presented at a rate of one word/s with the instruction to learn the words for immediate free recall after each series. The words were taken from the collections of Hager and Hasselhorn (67) and consisted of 10 neutral words such as “angle,” 10 positive words such as “happiness,” and 10 negative words such as “poverty.” The order of words was pseudorandom, with each group of five words containing no more than three words per valence category. In addition, subjects underwent an unexpected delayed free-recall test of the learned words after ∼5 min. The free recall of a word was considered successful only if it was spelled correctly or a with single-letter typo that did not make it become a different valid word. The relevant performance measures (PMs) are described in Table 2. In the Basel words/pictures sample, the picture task was performed more than 3 h after the end of the verbal task. Word arousal ratings were available in the Zurich sample (on a scale of 1–5), and the mean arousal ratings did not differ significantly between positive and negative words: arousal(positive) = 3.41 (mean) ± 0.44 (SEM); arousal(negative) = 4.10 ± 0.16; Student t test; P = 0.16.

Computational Model for the Verbal Task.

To dissociate specific mental processes involved in learning and memory, we used a computational model to describe individual performance in the verbal memory task (21). Depending on how well individuals remember a word, they may or may not try to write it down in the free recall, and if they try, their recall may or may not be correct. The probability that the attempted recall is correct depends on memory strength m of each word as follows: $p_{correct}\left(m,s\right)=1/\left(1+\exp \left(-s\left(m-m_{50\%}\right)\right)\right)$, with sigmoidal steepness s and center of the sigmoid m_50% = 1. The decision to attempt the recall of weak memories depends on the participant’s willingness to risk making errors. We modeled this decision-making aspect using the decision threshold β, where recall was attempted for words with memory strength m > β but not for those with m < β. During the encoding, initial memory strength for each word was assigned as m = α × ε + N(0, σ), with α being the learning rate, ε the emotional modulation of memory strength (ε = ε_neg for negative words, ε = ε_pos for positive words, ε = 1 for neutral words), and N(0, σ) the Gaussian noise with mean 0 and SD σ, reflecting randomness in learning different words. Because the memory strength of words that have been recalled correctly in the immediate recall is likely to increase through repetition (i.e., the participants writing them down), we multiplied the memory strength m of immediately recalled words by the repetition-based memory improvement c (c≥1). To model forgetting during the 5-min delay, we multiplied all memory strengths by the forgetting rate γ (γ<1). It is important to note that, because no external measure could distinguish the repetition-based memory improvement from the forgetting rate in our experiments, these two parameters were closely related and together reflected memory maintenance.

The described model had eight parameters: α, β, γ, ε_pos, ε_neg, s, c, and σ. Because our behavioral phenotype consisted of only eight partly correlated measures per individual, it was impossible to estimate all eight parameters reliably. Because the PCA indicated five substantial components (Fig. S1), we chose to keep five of the eight parameters in the model flexible (different between individuals) and the three remaining ones fixed (same for all individuals). The selection of the most appropriate model (i.e., which parameters would be flexible and which fixed) was performed empirically, based on the corresponding mean goodness-of-fit values. Such selection ensured that closely related parameters, such as c and γ, would not both be flexible, whereas relatively independent parameters, which can account for more variance, would.

Estimation and Evaluation of Best-Fitting Model Parameters.

To estimate the best-fitting model parameters, we computed the expected values of all performance measures (PM_1–8; Table 2) as a function of eight model parameters (α, β, γ, ε_pos, ε_neg, s, c, σ). We numerically computed integrals over probability distributions of memory strength m, because this method was more efficient and robust than simulating the model with random numbers and computing averages over multiple simulation runs. As a control, we also simulated the model stochastically: The averages of PMs over 100,000 simulations were almost exactly the same as those using the expected value-based method.

To evaluate how well the model with a particular set of parameters fits individual behavioral performance, we used the following goodness-of-fit function (20, 21, 68):

$${\mathit{\chi }}^2={\displaystyle \sum _{i=1}^8\frac{{\left(P{M}_i^{\exp }-P{M}_i^{\mathrm{mod}}\left[parameters\right]\right)}^2}{{\left({\mathit{\sigma }}_i^{\exp }\right)}^2}}$$

where PM_i^exp and PM_i^mod are experimental and modeled performance measures of that individual, respectively, and (σ_i^exp)² is the variance in the experimental data of PM_i.

With χ² as the objective function to minimize, we performed the estimation of best-fitting parameters in several stages (21):

• i)Model selection: To determine which five parameters should be flexible, we evaluated all 56 possible five-of-eight combinations. Because of the high computational cost of running 56 full estimation procedures, at this stage we performed only a moderately accurate estimation of the three fixed parameters.
• ii)Using the two best models, we performed a more refined estimation of fixed parameters, thereby improving the χ² values. We note that although the improvements in χ² values were substantial, they were small compared with the differences between the initial χ² values of the two best models and those of the other, worse models; therefore, it is very unlikely that any of those other models would become comparatively better with refinement.
• iii)For the final refinement, we evaluated the averages of all 2¹⁰ − 1 = 1,023 combinations of the 10 best parameter sets for each model, thereby further improving the χ² values. Finally, parameter sets from the model with the best goodness-of-fit were used for GWAS.

In all parameter estimation steps the search was performed in the following ranges: (α, β, ε_pos, ε_neg, σ) ∈ (0.3, 3.5), c ∈ (1, 4.2), γ ∈ (0, 0.8), and s ∈ (0, 16). In choosing the ranges we balanced two partially opposing aims: to keep them as similar as possible to avoid bias to the estimation results and to keep them as close as possible to the likely distribution of each parameter to maximize estimation accuracy. The default range of 0.3–3.5 ensured that fewer than 1% of estimated parameter values were near the boundaries. Other ranges were used either because of fundamental constraints (c ≥1 and γ<1) or because the likely spread of parameter values was very different from the default. To evaluate how well the model fits individual data, we used the χ² test with ν = 8 − 5 = three degrees of freedom (five flexible parameters and eight PMs). Values of P > 0.05 indicate no statistical difference between modeled and observed PMs, meaning that the model fits the data well.

The Picture Task and Controls.

After receiving general information about the study and giving their informed consent, participants were instructed and trained on the picture task as well as the N-back working memory task. They performed the encoding of pictures for 20 min and then performed the N-back task for 10 min.

In the picture task, stimuli consisted of 72 pictures that were selected from the International Affective Picture System (IAPS; 69) and from in-house standardized picture sets that allowed us to equate the pictures for visual complexity and content (e.g., human presence). Participants were instructed to view the pictures passively and subsequently rate them according to emotional valence/arousal. Pictures received from IAPS were classified according to the IAPS valence rating; the remaining eight neutral pictures were rated based on an in-house valence rating. Positive stimuli were selected initially to match the arousal ratings of negative stimuli based on data from a pilot study in 20 subjects. Examples of pictures are as follows: erotica, sports, and appealing animals for the positive valence; bodily injury, snake, attack scenes for the negative valence; and neutral faces, household objects, and buildings for the neutral condition. On the basis of normative valence scores (from 1 to 9), pictures were assigned to the emotionally negative (2.3 ± 0.6), emotionally neutral (5.0 ± 0.3), and emotionally positive (7.6 ± 0.4) conditions, resulting in 24 pictures for each emotional valence. Four additional pictures showing neutral objects were used to control for primacy and recency effects in memory. Two of these pictures were presented in the beginning and two at the end of the picture task. They were not included in the analysis. In addition, 24 scrambled pictures were used. The background of the scrambled pictures contained the color information of all pictures used in the experiment (except primacy and recency pictures), overlaid with a crystal and distortion filter (Adobe Photoshop CS3). In the foreground, a mostly transparent geometrical object (rectangle or ellipse of different sizes and orientations) was shown. The pictures were presented for 2.5 s in a quasi-randomized order so that no more than four pictures in the same category occurred consecutively. A fixation-cross appeared on the screen for 500 ms before each picture presentation. The stimulus onset time was jittered within 3 s [1 repetition time (TR)] per valence category with regard to the scan onset. After a picture was presented, participants subjectively rated the picture according to valence (negative, neutral, positive) and arousal (large, medium, small) on a three-point scale (self assessment manikin) by pressing a button with a finger of their dominant hand. For scrambled pictures, participants rated the form (vertical, symmetric or horizontal) and size (large, medium, small) of the geometrical object in the foreground. A maximum 6 s was allowed for rating each picture. Participants were not told that they should remember the pictures for later recall.

The participants also performed the 0- and 2-back versions of the N-back task (70). The task consisted of 12 blocks (six 0-back, six 2-back) in which 14 test stimuli (letters) were presented. The 0-back condition required participants to respond to the occurrence of the letter “x” in a sequence of letters (e.g., N-l-X-g ...), and the 2-back condition required subjects to compare the currently presented letter with the penultimate letter to decide whether they are identical or not (e.g., S-f-s-g ...) Performance was recorded as the number of correct responses, and the difference between the 2-back and the 0-back conditions served as a control phenotype of working memory.

After finishing the tasks, participants performed free recall of the pictures, which required them to write down a short description (a few words) of the previously seen pictures. Remembered primacy and recency pictures and training pictures were excluded from the analysis. No time limit was set for this task. Two trained investigators independently rated the descriptions for recall success (interrater reliability >99%). Depending on the computational model parameter with which each gene set was associated, different behavioral phenotypes were used. For the learning rate α, the relevant phenotype was the percentage of pictures remembered in the free recall (Table S1). For the negative emotional modulation parameter ε_neg, the relevant phenotype was the percentage of negative pictures minus the percentage of neutral pictures remembered in the free recall.

Multilocus Genetic Scores.

To capture the multiallelic effect of gene sets associated with different model parameters, we generated individual multilocus genetic scores for each of these gene sets. The score comprised SNPs from an equal number of genes (i.e., one most significant SNP per gene) that were associated with the relevant phenotype of the picture task (or the control N-back task) in the pictures/fMRI sample. The score was computed by summing up the individual number of reference alleles over all of these SNPs, weighted by the direction of effect on the phenotype with “1” (the reference allele positively correlated with the phenotype) or “−1” (the reference allele negatively correlated with the phenotype). Each missing SNP was assigned the average value of that SNP over all participants of the sample for whom it was not missing. There were 0.41% of SNPs missing for L1CAM interactions, and 0.69% were missing for amine compound SLC transporters. To test the robustness of the fMRI results, we also built a genetic score using two most significant SNPs per gene.

Construction of a Population-Average Anatomical Probabilistic Atlas.

Automatic segmentation of 1,000 subjects’ T1-weighted images, including the participants from the current study, was used to build a population-average probabilistic anatomical atlas. More precisely, each participant's T1-weighted image first was automatically segmented into cortical and subcortical structures using FreeSurfer (version 4.5, surfer.nmr.mgh.harvard.edu; 71). Labeling of the cortical gyri was based on the Desikan–Killiany Atlas (72), yielding 35 regions per hemisphere.

The segmented T1 image then was normalized to the study-specific anatomical template space using the subject’s previously computed warp field and was affine-registered to the MNI space. Nearest-neighbor interpolation was applied to preserve labeling of the different structures. The normalized segmentations finally were averaged across subjects to create a population-average probabilistic atlas. Each voxel of the template consequently could be assigned a probability of belonging to a given anatomical structure, based on the individual information from 1,000 subjects.

Functional Neuroimaging and Picture Recognition.

Participants of the Basel pictures/fMRI sample performed the picture and N-back tasks in the scanner. They were positioned in the scanner after training. The participants received earplugs and headphones to reduce scanner noise. Their heads were fixated in the coil using small cushions, and they were told not to move their heads. Functional MRIs were acquired during the performance of the picture task for ∼20 min, followed by the 10-min N-back task. After that task, participants were taken from the scanner to a different room for free recall of the presented pictures. Pictures were presented in the scanner using MR-compatible LCD goggles (Visuastim XGA; Resonance Technology). Eye correction was used when necessary.

Approximately 80 min after the presentation of the last picture during encoding, participants were repositioned in the scanner and performed a recognition task for 20 min. The recognition task consisted of two sets of stimuli that were either new (i.e., not presented before) or old (i.e., presented during the picture encoding task). Each of the two sets contained 72 pictures (24 pictures of each valence). During recognition, the pictures were presented for 1 s in a quasi-randomized order so that no more than four pictures of the same category appeared consecutively. A fixation cross appeared on the screen for 500 ms before each picture presentation. The stimulus onset time was jittered within 3 s (1 TR) per valence and previously seen/new category with regard to the scan onset. After picture presentation, participants subjectively rated the picture as recollected, familiar, or new by pressing a button with a finger of their dominant hand. Picture rating was possible in a time window of maximum 3 s.

Depending on the computational model parameter with which each gene set was found to be associated, different fMRI contrasts were used. For the learning rate α, we used two encoding contrasts: pictures of all valence categories vs. scrambled controls and subsequently remembered pictures (of all valence categories) vs. subsequently not-remembered pictures; the latter contrast is known as difference due to memory (Dm; 29). For the repetition-based memory improvement c, which reflects memory maintenance, the relevant behavioral phenotype was the percentage of recollected pictures that had been seen previously, and the appropriate fMRI contrast was previously seen pictures vs. new pictures during recognition. Recognition data were used for this parameter because free recall was not performed in the scanner, and memory maintenance processes may not be reflected during the encoding. To ensure that the effects observed using contrast between the previously seen vs. new pictures reflected the episodic memory component of previously seen, not new, pictures, we also used the contrast of previously seen pictures that were recollected vs. those rated as familiar or new. To investigate neural correlates of working memory, the 2-back vs. 0-back contrast was used.

The contrasts described above were calculated individually using a fixed-effects model (first-level analysis). Because we were using such contrasts, we controlled for all factors that were constant between the two conditions, whether of neuronal or nonneuronal origin. The resulting contrast parameters then were used for genotype-dependent analyses in a random effects model (second-level analysis). Specifically, we used a regression model to analyze gene-dose–dependent differences in brain activity, with the multilocus genetic score, based on the related behavioral phenotype, as covariate. We controlled for the effect of sex by including it as a covariate. The significance threshold for score-dependent analysis was set at P < 0.05, FWE-corrected for multiple comparisons in the whole brain. We also performed second-level analysis with the related behavioral phenotype itself as a covariate.

Mediation Analysis.

To test whether the observed effect of the L1CAM interactions gene set on behavioral performance was mediated by the activation of specific brain regions, we ran a whole-brain mediation analysis using the Multilevel Mediation and Moderation toolbox (wagerlab.colorado.edu/wiki/doku.php/help/mediation/m3_mediation_fmri_toolbox; 73). We looked for brain regions whose activation during recognition (the contrast between previously seen vs. new pictures) mediated the association between the multilocus genetic score (called the “predictor”) and recognition performance. To be considered as mediator, a voxel had to satisfy the following criteria: the predictor variable must be related to the mediator (path a); the mediator must be directly related to behavior, controlling for the predictor (path b); the mediation effect must be significant (effect a*b), meaning that the predictor–behavior relationship is reduced significantly by including the mediator in the model. Bootstrapping was used to obtain corresponding P values, and an FDR threshold of q < 0.05 across all effects of interest (a, b, and a*b coefficients) was applied to correct for multiple comparisons. Furthermore, sex was included as covariate.

Because causality between correlated brain activity and behavior cannot be inferred, we also tested using the same procedure to determine whether the association between the multilocus genetic score and differences in brain activity during recognition (the contrast between previously seen vs. new pictures) was mediated by picture recognition.

Sample from the German AgeCoDe Study.

The sample consisted of elderly participants in the German AgeCoDe study, an ongoing primary care-based prospective longitudinal study on the early detection of mild cognitive impairment and dementia established by the German Competence Network Dementia. The sampling frame and sample selection process of the AgeCoDe study have been described in detail previously (74). After the application of the selection criteria (SI Materials and Methods), a total of 1,244 subjects remained in the sample. Sufficient DNA samples for genomewide genotyping were available for 782 subjects. Additionally, 39 subjects were excluded because of sex-check inconsistencies, low call rate, or extreme values of heterozygosity rate, which may indicate a genotyping bias. The final sample comprised 743 subjects (241 males and 502 females; mean age: 79.53 ± 3.21 y). As part of the structured clinical interviews, subjects were presented a list of 10 words from the German version of the CERAD Word List Learning Task (28) three times (presentation per word: 2 s); each trial was presented in a different order. Participants were asked to read each word aloud as it was presented. After each trial, subjects were asked to recall freely as many words as possible. The number of correctly remembered items over the three trials (immediate recall) served as a relevant phenotype for the gene set associated with the learning rate α, whereas the number of correctly recalled items after a 10-min delay (filled with other questions of the structured interview) was relevant for the gene set associated with the repetition-based memory improvement c.

SI Materials and Methods

Array-Based SNP Genotyping.

Blood was drawn by using 2 × 9 mL EDTA tubes (Sarstedt). Saliva was collected with the Oragene DNA sample collection kit (DNA Genotek Inc.). DNA isolation was done according to standard protocols. Samples were processed as described in the Genome-Wide Human SNP Nsp/Sty 6.0 User Guide (Affymetrix). Briefly, the genomic DNA concentration was determined by using a Nano-Drop ND-1000 and was adjusted to 50 ng/µL in water. DNA (250 ng) was digested in parallel with 10 units of StyI and NspI restriction enzymes (New England Biolabs) for 2 h at 37 °C. Enzyme-specific adaptor oligonucleotides then were ligated onto the digested ends with T4 DNA Ligase for 3 h at 16 °C. After adjustment to 100 µL with water, 10 µL of the diluted ligation reactions were subjected to PCR. Three PCR reactions of 100 µL were performed for StyI-digested products, and four PCR reactions were performed for NspI-digested products. PCR was performed with Titanium Taq DNA Polymerase (Clontech) in the presence of 4.5 μM PCR primer 002 (Affymetrix), 350 μM each dNTP (Clontech), 1 M G-C Melt (Clontech), and 1× Titanium Taq PCR Buffer (Clontech). Cycling parameters were as follows: initial denaturation at 94 °C for 3 min, amplification at 94 °C for 30 s, 60 °C for 45 s, and extension at 68 °C for 15 s, repeated a total of 30 times, with a final extension at 68 °C for 7 min. Reactions then were verified to migrate at an average size of 200–1,100 bp using 2% TBE gel electrophoresis. PCR products were combined and purified with the Filter Bottom Plate (Seahorse Bioscience) using Agencourt Magnetic Beads (Beckman Coulter). Purified PCR products were quantified on a Zenyth 200rt microplate reader (Anthos Labtec). On average, 4–5 µg/µL were obtained for each sample. From this stage on, the SNP Nsp/Sty 5.0/6.0 Assay Kit (Affymetrix) was used. Around 250 µg of purified PCR products were fragmented using 0.5 units of DNaseI at 37 °C for 35 min. Fragmentation of the products to an average size of less than 180 bp was verified using 4% Tris-borate-EDTA gel electrophoresis. Following fragmentation, the DNA was end-labeled with 105 units of terminal deoxynucleotidyl transferase at 37 °C for 4 h. The labeled DNA then was hybridized onto a Genome-Wide Human SNP 6.0 Array at 50 °C for 18 h at 60 rpm (Affymetrix GeneChip Hybridization Oven 645, Rotisserie 00-0331). The hybridized array was washed, stained, and scanned according to the manufacturer’s (Affymetrix) instructions using Affymetrix GeneChip Command Console (AGCC version 3.0.1.1214). Generation of SNP calls and Array QC were performed using the command line programs of the Affymetrix Power Tools package (version: apt-1.14.2). According to the manufacturer’s recommendation, Contrast QC was chosen as the QC metric, using the default value of greater or equal than 0.4. The mean call rate for all samples averaged >98.5%. All samples passing QC criteria were subsequently genotyped using the Birdseed (v2) algorithm.

fMRI Data Acquisition and Preprocessing.

Measurements were performed on a Siemens MAGNETOM Verio 3-T whole-body MR unit equipped with a 12-channel head coil. Functional time series were acquired with a single-shot echo-planar sequence using parallel imaging (GRAPPA). We used the following acquisition parameters: echo time (TE) = 35 ms, field of view (FOV) = 22 cm; acquisition matrix = 80 × 80; interpolated to 128 × 128; voxel size: 2.75 × 2.75 × 4 mm³; GRAPPA acceleration factor R = 2.0. Using a midsagittal scout image, 32 contiguous axial slices placed along the anterior–posterior commissure (AC–PC) plane covering the entire brain with a TR = 3,000 ms (α = 82°) were acquired using an ascending interleaved sequence. A high-resolution T1-weighted anatomical image was acquired using a magnetization prepared gradient echo sequence (MPRAGE, TR = 2,000 ms; TE = 3.37 ms; TI = 1,000 ms; flip angle = 8°; 176 slices; FOV= 256 mm).

Preprocessing and data analysis were performed using SPM8 (Statistical Parametric Mapping, Wellcome Trust Centre for Neuroimaging; www.fil.ion.ucl.ac.uk/spm/) implemented in MATLAB R2011b (MathWorks Inc.). Volumes were slice-time corrected to the first slice and realigned using the “register to mean” option. A mean image was generated from the realigned series and coregistered to the structural image, ensuring that functional and structural images were spatially aligned.

The functional images and the structural images were spatially normalized by applying DARTEL, which leads to an improved registration between subjects. Normalization incorporated the following steps: (i) Structural images of each subject were segmented using the “New Segment” procedure in SPM8. (ii). The resulting gray- and white-matter images were used to derive a study-specific group template. The template was computed from a large population of 1,000 subjects that included the 795 subjects from this study. (iii) An affine transformation was applied to map the group template to MNI space. (iv) Subject-to-template and template-to-MNI transformations were combined to map the functional images to MNI space. The functional images were smoothed with an isotropic 8-mm FWHM Gaussian filter.

For each subject, analyses were conducted in the framework of the general linear model (GLM). Regressors modeling the onsets and duration of stimulus events were convolved with a canonical hemodynamic response function. More precisely, the model comprised regressors for button presses modeled as stick/delta functions, picture presentations modeled with an epoch/boxcar function (duration: encoding: 2.5 s; retrieval: 1 s), and rating scales modeled with an epoch/boxcar function of variable duration (depending on when the subsequent button press occurred). Serial correlations were removed using a first-order autoregressive model, and a high-pass filter (128 s) was applied to remove low-frequency noise. Six movement parameters were also entered as nuisance covariates. Pictures accounting for possible primacy and recency effects were excluded from statistical analysis.

fMRI Group Statistics.

EPI sequences suffer from signal loss in the presence of magnetic field inhomogeneities that can occur close to air–tissue boundaries. The normalization procedure applied in DARTEL accurately transforms both voxels with signal and voxels with signal loss to MNI space. In SPM8, signal loss at an MNI coordinate in a functional image of only one subject leads to the exclusion of the voxel at this coordinate from the group-level analysis. Therefore, the probability of a voxel being excluded increases with sample size. GLM Flex circumvents this problem by allowing a variable number of subjects at each voxel (Martinos Center, Massachusetts General Hospital; mrtools.mgh.harvard.edu). The minimum number of subjects per voxel was set to 400.

German AgeCoDe Sample: Selection of Participants.

Participants were recruited between January 2003 and November 2004 in six German study centers (Bonn, Duesseldorf, Hamburg, Leipzig, Mannheim, and Munich) via general practitioners (GPs) connected with the respective study sites. Inclusion criteria were age of 75 y or older, absence of dementia (according to the GP’s judgment), and at least one contact with the GP within the last 12 mo. Exclusion criteria were GP consultations by home visits only, residence in a nursing home, presence of a severe illness with an anticipated fatal outcome within 3 mo, insufficient German language abilities, deafness or blindness, lack of ability to provide an informed consent, and status as being only an occasional patient of the participating GP. A total of 3,327 subjects were contacted successfully and assessed with structured clinical interviews at their homes. A total of 110 individuals were excluded after the first interview because of the presence of dementia or an actual age below 75 y (falsely classified as age 75 y or older in the sample-selection process). For the present analyses, data from baseline and three follow-up measurements at 18-mo intervals were available. In a primary care-based sample of older individuals, conditions can be present that affect cognition and the reliability of neuropsychological tests. To generate a sample of healthy elderly individuals, we further used the following selection criteria at baseline: age between 75 and 90 y, German as native language, at least a school-leaving certificate, absence of severe hearing or vision impairments, absence of insufficient test motivation as judged by the interviewer, absence of disturbing factors during neuropsychological testing, and absence of all the following comorbid conditions: Parkinson’s disease, epilepsy, alcohol abuse, stroke, multiple sclerosis, evidence of depression (a score of 6 or higher on the Geriatric Depression Scale), traumatic brain injury with unconsciousness of more than 30 min, visible neurological malfunctions, and dementia according to Diagnostic and Statistical Manual of Mental Disorders-IV criteria. In addition, we excluded subjects who converted to dementia before the third follow-up (i.e., within 4.5 y after baseline) or for whom neuropsychological test data were not available at baseline and all follow-up visits.