Overlapping brain correlates of superior cognition among children at genetic risk for Alzheimer’s disease and/or major depressive disorder

Early life adversity (ELA) tends to accelerate neurobiological ageing, which, in turn, is thought to heighten vulnerability to both major depressive disorder (MDD) and Alzheimer’s disease (AD). The two conditions are putatively related, with MDD representing either a risk factor or early symptom of AD. Given the substantial environmental susceptibility of both disorders, timely identification of their neurocognitive markers could facilitate interventions to prevent clinical onset. To this end, we analysed multimodal data from the Adolescent Brain and Cognitive Development study (ages 9–10 years). To disentangle genetic from correlated genetic-environmental influences, while also probing gene-adversity interactions, we compared adoptees, a group generally exposed to substantial ELA, with children raised by their biological families via genetic risk scores (GRS) from genome-wide association studies. AD and MDD GRSs predicted overlapping and widespread neurodevelopmental alterations associated with superior fluid cognition. Specifically, among adoptees only, greater AD GRS were related to accelerated structural maturation (i.e., cortical thinning) and higher MDD GRS were linked to delayed functional neurodevelopment, as reflected in compensatory brain activation on an inhibitory control task. Our study identifies compensatory mechanisms linked to MDD risk and highlights the potential cognitive benefits of accelerated maturation linked to AD vulnerability in late childhood.


Perinatal Adversity
An index of perinatal adversity was extracted through principal components analysis from caregiver responses on the Developmental History Questionnaire [4], which was completed The ABCD version of the SST comprises two runs of 180 trials each: 150 "Go" trials, 15 "Stop" trials expected to be unsuccessful and 15 "Stop" trials expected to be successful.
To maintain the breakdown of the successful/unsuccessful "Stop" trials, a tracking algorithm was implemented to alter the interval between the presentation of the 'Go stimulus' and the onset of the 'Stop' signal based on the participant's performance. Each run was restricted to begin with a 'Go' trial and stop trials were separated by a minimum of one 'Go' trial.
Each trial lasted 1000 ms and began with the presentation of a black rightward-(50% of the time) or leftward-facing arrow ('Go stimulus') displayed on a mid-grey background.
Participants were asked to indicate the arrow direction "as quickly and accurately as possible" using a response panel consisting of two buttons. Participants used their dominant hand to respond to the 'Go' stimuli and this was mapped congruently with handedness. On 'Stop' trials the 'Go' stimulus was unpredictably followed by a 'Stop' signal in the form of an upward facing arrow presented for 300 ms, which indicated to the participants that they should inhibit their response to the previously presented 'Go' signal. The presentation of the "Go" (on the "Go" trials) and "Stop" cue was followed by a fixation cross varying in duration based on participant's reaction time for a total trial duration of 1000 ms.

MRI Data Acquisition
Scanning was performed across 21 US sites, with a protocol harmonised for Siemens thick). The fMRI data were acquired with a multiband EPI sequence (TR=800 ms, TE=30 ms, flip angle=52°, FOV = 216 x 216 mm, 60 slices of 2.4 × 2.4 mm in-plane resolution, 2.4 mm thick, multiband acceleration factor of 6).
Four resting state fMRI scans (eyes open with passive crosshair viewing), lasting 20 minutes in total, were collected in order to ensure at least 8 minutes of low-motion data. Two SST runs were also acquired for a total duration of 11:40 minutes (6).

MRI Data Preprocessing
Our analyses used tabulated sMRI and fMRI data available as part of the ABCD Study Curated Annual Release 4.0. The main processing steps applied to these data by the ABCD study team are outlined below (for further details, see [6]).

fMRI
Preprocessing of all functional images involved correction for head motion, spatial and gradient distortions, bias field removal, elimination of initial volumes ( reach steady state equilibrium, normalisation of the voxel time series and co-registration of the functional images to the participant's T1 -weighted structural image. Linear regression was used to remove from each voxel's time course quadratic trends, as well as the six motion parameters, their first derivatives, and squares (24 motion terms in total). Estimated motion time courses were filtered to attenuate signals related to respiration.

Resting-state (RS)
The following preprocessing steps were specific to the RS data (1) regression of the mean time courses of cerebral WM, ventricles, whole brain, and their first derivatives, (2) bandpass filtering of the residual time series between .009-.08 Hz; exclusion of (3) time points with framewise displacement (FD) greater than .20 mm, (4) those that were outliers in standard deviation (SD) across ROIs (i.e., SD > three times the median absolute deviation below or above the median SD for a given participant), and of (5) time periods with fewer than 5 contiguous volumes with FD smaller than .20 mm.
Temporal variance. Based on the preprocessed data, averaged time courses were computed for cortical ROIs from an anatomically defined parcellation. Temporal variance was estimated for each ROI as an amplitude index of low frequency fluctuations, which is assumed to reflect spontaneous neural activity and is predictive of task-related responsiveness.

Task
The following preprocessing steps were specific to the task data: (1) regression of the baseline, and (2) removal of time points with FD > .90 mm. Task-specific activation strength was estimated for each individual participant using a general linear model in AFNI's 3dDeconvolve [7]. The baseline model ("null model") included regressors for average signal, quadratic trend and motion (i.e., 24 motion regressors in total, specifically, the linear and quadratic motion parameters and their derivatives). The GLM included the stimulus time series convolved with the hemodynamic response function (HRF). The latter was modelled with a gamma variate function and its temporal derivative in AFNI's SPMG option within 3dDeconvolve. Events were modelled as instantaneous. Both sets of analyses detailed below were based on the linear contrast reflecting successful inhibition (correct Stop > correct Go [baseline]).

Task-related activation. Our tests examined (1) the difference in BOLD signal
between the second and the first run of the SST task, with lower values indicating decreased activation and, thus, increasing neural efficiency, potentially reflective of training effects (i.e., less activation is needed to support correct Stop performance), and (2) average BOLD signal across the two runs of the SST task, with lower values indicating overall greater neural efficiency, likely indicative of functional maturation.

Task-related variability. Our analyses focused on responses in each of the ROIs in
the Destrieux anatomical atlas, specifically, (1) the difference in the standard error (SEM) of the GLM beta coefficient between the second and the first SST run, with lower values indicating greater stabilisation of the task-related response, and (2) the average SEM of the GLM beta coefficients estimated across the two SST runs, with lower SEM values indicating a more consistent response to the task-relevant information.

sMRI
The sMRI preprocessing pipepline included removal of non-brain tissue, corrections for gradient non-linearity distortions and intensity inhomogeneity, intensity normalisation, as well as rigid resampling and alignment to an averaged brain image in standard space. Cortical reconstruction and subcortical segmentation were performed using FreeSurfer version 5.3, where estimates of cortical thickness and volume were computed for each of the 148 ROIs from the Destrieux atlas.

Genetic Risk Scores (GRS)
MDD and AD GRSs were each computed as the weighted sum of risk alleles, as derived from the summary statistics of two large genome-wide association studies (GWASs) focused on each disorder ( [8,9] for MDD and AD, respectively), which had been made available by the original authors via the Public Results tab on the FUMA website (https://fuma.ctglab.nl/browse, 10). For AD, we computed both a composite (i.e., full) GRS and a separate Apolipoprotein E (APOE) region (chromosome 19:44.4-46.5 Mb)-vs no-APOE region GRS (cf. 11). To compute MDD and AD GRSs based on the *.genotype ABCD data, we used the PLINK genetic analysis toolset [12] with single nucleotide polymorphisms (SNPs) significant at GWAS level p ≤ 5x10 -8 because these are the SNPs likely to make the most robust contribution to genetic vulnerability. However, in supplemental analyses (see Figures S3-4), we verified that all our results are replicated when using more lenient significance thresholds for the GRS-contributing SNPs.

Residualisation for Confounding Variables
To minimise bias in our multivariate brain-behaviour analyses [13], only the nonimaging variables were residualised for the following confounders: sex, race (separate dummy-coded variables for "Black", "Asian", 'Mixed Race" regressed simultaneously from the non-imaging variables to account for potential differences between these racial groups and White participants), handedness, serious medical problems, scanner site, material deprivation, family conflict, neighbourhood crime, age at adoption, average modality-specific motion per participant, and chronological age (in order to estimate accelerated/decelerated neurodevelopment relative to the other participants). Due to data (un)availability, only the non-adoptee data were residualised for perinatal adversity. The adoptee and non-adoptee data were residualised separately.

Partial least squares analysis (PLS)
To provide a comprehensive description of the relationship between genetic risk for AD/MDD and neurodevelopmental timing among adoptees versus non-adoptees, we used partial least squares correlation often referred to as PLS [14], a multivariate technique that can identify in an unconstrained, data-driven manner, neural patterns (i.e., latent variables or LVs) related to different conditions (i.e., task PLS) and/or individual differences variables (behavioral PLS). PLS was implemented using a series of Matlab scripts, which are available for download at https://www.rotman-baycrest.on.ca/index.php?section=345.
We thus conducted two behavioural PLS analyses featuring MDD GRS (both analyses) and either the composite AD GRS (analysis 1) or the APOE-vs no-APOE-based GRSs (analysis 2) in the "behavioural" set. In each analysis, each type of data (cortical thickness, resting state BOLDSV, as well as BOLDM and BOLDSV, averaged across the two SST runs and its difference between run 2 and run 1 of the SST) was modelled as a separate condition, whereas the adoptees and non-adoptees were modelled as separate groups. Within each group, the brain matrix contained the participants' concatenated scores for all the data types across the 148 Destrieux ROIs. The design matrix contained a number of dummy coded variables corresponding to each condition within each group (e.g., adoptees' cortical thickness scores). By entering both groups within the same PLS analysis we were able to identify group-specific associations between genetic vulnerability and neurodevelopmental timing.

Significance and reliability testing
In all the reported PLS analyses, the significance of each LV was determined using a permutation test (5000 permutations). In the permutation test, the rows of the brain data are randomly reordered [14]. In the case of our present analyses, PLS assigned to each ROI a weight, which reflected the contribution of the respective ROI to a specific LV. The reliability of each ROI's contribution to a particular LV was tested by submitting all weights to a bootstrap estimation (1000 bootstraps) of the standard errors (SEs) (the bootstrap samples were obtained by sampling with replacement from the participants, [14]). In order to increase the stability of the reported results, we used a number of permutations/bootstraps greater than the standard ones (i.e., 500 permutations/100 bootstrap samples), as recommended by [15] for use in PLS analyses of neuroimaging data. A bootstrap ratio (BSR) (weight/SE) of at least 3 in absolute value (approximate associated p-value < .0001) was used as a threshold for identifying those ROIs that made a significant contribution to the identified LVs. The BSR is analogous to a z-score, so an absolute value greater than 2 is thought to make a reliable contribution to the LV [14], although for neuroimaging data BSR absolute values of at least 3 are recommended for use [15].

Mediation analyses
To test whether genetic risk for AD/MDD is linked to distinct patterns of brain and cognitive development among adoptees versus non-adoptees, we conducted three moderated mediation analyses for MDD, composite AD and no-APOE-based AD GRS using Hayes' PROCESS 3.5 macro for the Statistical Package for the Social Sciences (SPSS, [16]).
PROCESS is an ordinary least squares (OLS) and logistic regression path analysis modelling tool, based on observable variables. Mediation models were tested employing 95% confidence intervals (Cis) with 50000 bootstrapping samples. In line with extant guidelines on balancing Type I and Type II errors in mediation analyses [17], the CIs for indirect effects was estimated using percentile bootstrap, which is the default option in PROCESS 3.5. As recommended [18], a heterodasticity consistent standard error and covariance matrix estimator was used. Bootstrapping-based 95% CIs for the indirect effects and for the moderation mediation index (cf. [19,20]), as outputted by PROCESS, were used as effect size estimates.

Supplementary PLS Analyses (involving only White participants)
To confirm that racial differences in genetic architecture and risk loci did not accounted for 54 % of the covariance in the GRS-brain data and distinguished brain markers of genetic risk for AD from those linked to vulnerability to MDD ( Figure S1-b). The brain LV was most strongly expressed in frontal, insular and inferior temporal areas (see Figure   S1-a). The associated neurogenetic pattern replicated the one observed in the full sample (cf. Figure S1-c, d).

PLS 2: APOE-vs. non-APOE-based genetic vulnerability to AD is linked to
distinct neurodevelopmental markers. The second PLS analysis identified a sole significant LV (p = .002), which accounted for 51% of the covariance in the GRS-brain data and differentiated among brain markers of genetic risk for MDD, as well as APOE-vs non-APOE-linked vulnerability to AD ( Figure S2-b). The associated brain LV was most strongly expressed in the frontal, insular and parietal areas ( Figure S2-a). As in the analyses involving the full sample, the neural markers of MDD GRS tended to comprise the same functional data types as those observed in the first PLS analysis. The neurogenetic pattern associated with APOE-and non-APOE-based GRS replicated the one observed with the full sample (cf. Figure S2-c,d).