Abstract
Aging is associated with cognitive impairment, but there are large individual differences in these declines. One neural measure that is lower in older adults and predicts these individual differences is moment-to-moment brain signal variability. Testing the assumption that GABA should heighten neural variability, we examined whether reduced brain signal variability in older, poorer performing adults could be boosted by increasing GABA pharmacologically. Brain signal variability was estimated using fMRI in 20 young and 24 older healthy human adults during placebo and GABA agonist sessions. As expected, older adults exhibited lower signal variability at placebo, and, crucially, GABA agonism boosted older adults' variability to the levels of young adults. Furthermore, poorer performing older adults experienced a greater increase in variability on drug, suggesting that those with more to gain benefit the most from GABA system potentiation. GABA may thus serve as a core neurochemical target in future work on aging- and cognition-related human brain dynamics.
SIGNIFICANCE STATEMENT Prior research indicates that moment-to-moment brain signal variability is lower in older, poorer performing adults. We found that this reduced brain signal variability could be boosted through GABA agonism in older adults to the levels of young adults and that this boost was largest in the poorer performing older adults. These results provide the first evidence that brain signal variability can be restored by increasing GABAergic activity and suggest the promise of developing interventions targeting inhibitory systems to help slow cognitive declines in healthy aging.
Introduction
Functional magnetic resonance imaging (fMRI) has become a predominant method for noninvasively estimating brain activity in human beings. Most fMRI studies treat moment-to-moment variability in the blood oxygenation level-dependent (BOLD) fMRI signal as noise, but research demonstrates that such variability is associated with better behavioral performance and is a more powerful predictor of cognitive abilities than mean BOLD signal (McIntosh et al., 2008; Garrett et al., 2011, 2013b; Grady and Garrett, 2014; Burzynska et al., 2015). Consistent with these findings, theoretical, experimental, and computational modeling work suggests that greater brain signal variability typifies younger, higher performing adults and well-functioning cortical networks capable of greater complexity and flexibility, increased dynamic range and information transfer, and stronger long-range functional connectivity (Li et al., 2006; Faisal et al., 2008; McIntosh et al., 2008, 2010; Shew et al., 2009; Deco et al., 2011; Garrett et al., 2011, 2013b; Misic et al., 2011; Vakorin et al., 2011; Beharelle et al., 2012; Grady and Garrett, 2014; Nomi et al., 2017). However, little is known about the underlying basis of performance-related deficits in brain signal variability, and even less is known about how to reverse these deficits.
GABA, the brain's major inhibitory neurotransmitter, plays a role in many of the same functions brain signal variability has been associated with, including cortical plasticity (Jones, 1993; Hensch et al., 1998; Fagiolini et al., 2004), the synchronization of neural oscillations (functional connectivity; Fingelkurts et al., 2004; Bonifazi et al., 2009; Kapogiannis et al., 2013), information capacity (Shew et al., 2011; Puzerey and Galán, 2014), efficiency (Sengupta et al., 2013; Zhou and Yu, 2018), pattern complexity (Monteforte and Wolf, 2010; Lajoie et al., 2014; Agrawal et al., 2018), and dynamic range (Shew et al., 2009; Agrawal et al., 2018) of neural networks. Manipulating the strength of inhibitory connections in artificial neural networks also dramatically influences the number of different states the network can sample (Agrawal et al., 2018). Crucially, decreasing GABA activity pharmacologically in healthy young rats and monkeys has been found to decrease network signal variability and to reduce the number of states that can be visited by the cortical network (Shew et al., 2011). Inspired by these results and by previous work showing that older adults express lower GABA levels (Gao et al., 2013; Porges et al., 2017; Cuypers et al., 2018; Cassady et al., 2019; Chamberlain et al., 2019; Lalwani et al., 2019) and lower brain signal variability in a host of cortical regions (Garrett et al., 2011, 2013a; Grady and Garrett, 2014, 2018; Waschke et al., 2021), we hypothesized that pharmacologically increasing GABA activity might causally reverse deficient brain signal variability levels in older adults. Furthermore, given that lower brain signal variability is typically associated with poorer cognitive performance even in older adults (Garrett et al., 2011, 2013a; Grady and Garrett, 2014; Burzynska et al., 2015), we also hypothesized that GABA agonism-related upregulation of brain signal variability should be largest in poorer performing older adults. However, some studies suggest that region- and measure-specific brain signal variability can also be higher in older, poorer performing adults (Samanez-Larkin et al., 2010; Boylan et al., 2021). We thus examined whether there were any regions exhibiting higher signal variability with older adult age and the role of GABA agonism in the whole brain regardless of sign using multivariate partial least squares (PLS; McIntosh et al., 1996).
In the current study, we analyzed data from 21 young (age 18–25 years) and 25 older (age 65–85 years) adults who had previously participated in the Michigan Neural Distinctiveness (MiND) study (Gagnon et al., 2019). Specifically, we investigated (1) the effect of age on resting-state brain signal variability, (2) the effect of a pharmacological manipulation of GABA activity on brain signal variability, and (3) the association among individual differences in composite cognitive scores and changes in brain signal variability on drug compared with placebo.
Materials and Methods
This dataset was collected as part of the MiND study. Here, we only describe the portions of the study that are relevant to this analysis. Gagnon et al. (2019) has details about the entire study protocol. The ethical approval for the study was granted by the Institutional Review Board of the University of Michigan (HUM00103117).
Participants
We analyzed data from 25 young (age 18–29 years) and 21 older (age 65 years and above) male and female human adults who completed the entire MiND study and received the drug manipulation. All participants were recruited from Ann Arbor and the surrounding area and were right-handed, native English speakers, and had normal or corrected-to-normal vision. Participants completed an initial telephone screening interview and were determined to be eligible. We screened out participants who scored 23 or lower on the Montreal Cognitive Assessment (Carson et al., 2018). All sessions described took place at the University of Michigan Functional MRI Laboratory. We present data collected during two sessions, each on a separate day.
Behavior testing
Participants completed an extensive cognitive and behavioral task battery including tasks from the National Institutes of Health (NIH) Toolbox for Assessment of Neurologic and Behavioral Function (Weintraub et al., 2014). The toolbox tasks were administered using an iPad, and the associated software automatically generates a standardized composite cognitive score for each participant. Here, we provide a brief description of the tasks that contributed to this composite measure (Weintraub et al., 2014).
Pattern Comparison Processing Speed Test
Two simple side-by-side pictures were presented on an iPad, and participants were instructed to discern, as fast as they could, whether the two pictures were the same or different. Participants pressed buttons on the iPad screen to indicate their response. The score was calculated based on the number of items they correctly answered in 85 s.
List Sorting Working Memory Test
Participants were presented with a few pictures from a specific category (e.g., animals) one at a time on the iPad. Participants were then asked to list the items in increasing order of size. Participants' response was marked correct if they listed all the items in the correct order.
Flanker Inhibitory Control and Attention Test
Participants were presented with a row of arrows on the iPad and instructed to indicate the direction of the middle arrow as quickly as they could. They pressed a left or right arrow button located on the iPad screen to indicate their response. The middle arrow could point in the same direction as the arrows surrounding it (congruent trials) or in the opposite direction (incongruent trials). There was a total of 20 trials, 40% of which were incongruent. The score was based on a combination of reaction time and accuracy.
Dimensional Change Card Sort Test
In this task, participants were presented with one target image and two response images that matched the target image in either shape or color. During each trial, participants were first presented with the word “shape” or “color” and were asked to choose the response image that matched the target image based on that dimension. There were 30 trials, and 23% of these were color trials. The score computed for this task was also based on a combination of reaction time and accuracy.
Picture Sequence Memory Test
Participants first had to recall the order of 15 images displayed in sequence on the iPad screen. They moved images on the screen to match the order they remembered them being presented. Participants were then presented with 18 images, including the first 15 images and three new images presented in the middle of the sequence. Again, they were asked to recall the order of the 18 images. The score was based on the total number of correct adjacent pairs.
Picture Vocabulary Test
Participants listened to an audio recording of a word and four pictures displayed on the iPad screen. They were instructed to select the picture that best matched the meaning of the word they heard. This test used the Computerized Adaptive System, whereby the difficulty of the next question was determined by the previous answer. A raw score was computed using Item Response Theory.
Oral Reading Recognition Test
Participants were presented a word on the iPad screen and asked to read the word out loud. Using a pronunciation guide, the examiner scored the response as correct or incorrect. This test also used Computerized Adaptive Testing, and the score was computed using Item Response Theory.
fMRI scans
Participants were given a low-dose benzodiazepine (lorazepam) or a placebo pill ∼1 h before the scan on 2 separate days. The order of the sessions (on and off drug) was counterbalanced across participants. Participants were not told which pill they received on which day; they were blind to the drug administration order. During the drug session, participants were administered a 0.5 or 1 mg oral dose of lorazepam. The drug dosage (0.5 or 1 mg) was randomly assigned across participants for the following reasons: (1) to maximize the chances of including a dose strong enough to produce observable effects without producing significant sedation and (2) to make it possible to analyze the effect of dosage. The participants were screened for use of medications that might interact with lorazepam or affect GABA levels. They also had no history of claustrophobia or mood disorders.
The functional scanning parameters (detailed below) were identical during both sessions. Functional MRI data were collected using a 3T General Electric Discovery MRI system with a volumetric quadrature birdcage head coil and two 32-channel receive arrays. The functional scans were T2*-weighted images collected with a 2D gradient echo pulse sequence with the following parameters: repetition time = 2000 ms, echo time = 30 ms, flip angle = 90°, field of view (FOV) = 220 × 220 mm, 43 axial slices with thickness of 3 mm and no spacing, collected in an ascending sequential sequence (voxels were 3 × 3 × 3 mm). The total acquisition time for the resting-state functional scan was 8 min 10 s, with 245 volumes. Participants were instructed to relax, keep their eyes open, and focus on a fixation cross presented for the duration of the scan. Using an eye-tracking system (in view mode only), we ensured that participants indeed had their eyes open and fixated. Heart rate was collected via a pulse oximeter placed on the left middle finger. We also obtained a T1-weighted image using the spoiled gradient-recalled (3D BRAVO) echo sequence during this session, with the following parameters: inversion time = 500 ms, flip angle = 15°, FOV = 256 × 256 mm.
fMRI data preprocessing
The fMRI data were preprocessed and analyzed using a combination of Functional MRI of the Brain (FMRIB) Software Library (FSL), SPM12, and MATLAB-based scripts. The first five volumes of each scan were discarded. Heart rate was collected via the pulse oximeter placed on the left middle finger, and the data were physio corrected during preprocessing. We performed first-level preprocessing using the FSL FEAT (FMRIB fMRI Expert Analysis Tool; Woolrich et al., 2001) with default parameters for motion correction, normalization, and smoothing (7 mm). We used the SPM12 function spm_detrend to remove linear, quadratic, and cubic trends in the time series and also applied a Butterworth filter (0.01–0.1 Hz). We then ran FSL MELODIC (Multivariate Exploratory Linear Optimized Decomposition into Independent Components) to perform independent component analysis. Three separate raters identified noise components through manual visual inspection based on Kelly et al. (2010). These components reflected noise related to sinus activity, vascular and ventricle activations, and motion. We then removed the components identified as noise by at least two of the three raters using the FSL regfilt function. Subsequently, we performed linear registration of the functional and anatomic images of each participant and the Montreal Neurological Institute (MNI)152 template using the FSL FLIRT (FMRIB Linear Image Registration Tool).
Quantification of brain signal variability (SDBOLD)
After preprocessing the fMRI data, the SD in the fMRI signal during the entire resting-state scan was computed at each voxel for each participant and scanning session separately. We also computed SDBOLD-Change at each voxel as the difference between SDBOLD-Drug and SDBOLD-Placebo. All analyses were restricted to voxels in a MNI152 gray matter volume mask.
Statistical analysis
To examine the regional distribution of the effects on brain signal variability, we used multivariate PLS analyses (McIntosh et al., 1996).
We employed behavior-PLS for investigating effects of age group on SDBOLD-Placebo. In this simple one-condition, one-behavior PLS, a correlation matrix between age group and the signal of each voxel (SDBOLD-Placebo) was first computed across subjects. This correlation matrix was then decomposed using singular value decomposition (SVD). This resulted in a singular (S; reflecting the correlation strength) value, and brain voxel (V; i.e., a weighting pattern across brain voxels that optimally expresses the correlation) weights—in this case, the original voxelwise correlations with age but scaled to be unit length. We then calculated individual brainscores by taking the dot product of the brain voxel weights and a given subject's brain measures. Thus, brainscores indicate the degree to which a subject expresses the multivariate spatial pattern captured by an age-driven latent variable.
Next, we used task-PLS with two groups for investigating the effects of drug on SDBOLD. The task-PLS is similar to the behavior-PLS but involves a singular value decomposition of a between-subject covariance (COV) matrix instead of the correlation matrix. We first computed a COV matrix between drug and placebo conditions and the SDBOLD of each voxel within each age group. Then, using SVD, we estimated a left singular vector of experimental condition (U) weights for each age group, a right singular vector of brain V weights, and a diagonal matrix of S values. This produced four latent variables (LVs). Only the first LV was significant and represented greater variability during the drug condition than during placebo in both age groups. Brainscore was computed separately for each condition as the dot product of brain voxel weights and each subjects' SDBOLD-Placebo and SDBOLD-Drug.
We used simple task-PLS for investigating effects of drug on SDBOLD within each age group separately. This is similar to the previously described task-PLS except there are no age groups. Thus, SVD results in only two latent variables. The first LV, representing greater variability during the drug condition compared with placebo, was significant in older adults. The LV in younger adults alone was not significant.
We used rank-based behavior-PLS with two groups for investigating effects of cognitive processing on SDBOLD-Change. Similar to the previously described one-condition, one-behavior PLS, first a rank-based correlation matrix was computed between the composite cognitive scores and the signal of each voxel (SDBOLD-Placebo) across subjects within each age group and then stacked into a single matrix. An SVD of this matrix results in two latent variables defined by S value and brain V weights. This PLS identifies a single latent space in the brain between the two groups that best captures the relationship between voxel signals and the behavioral measure for each of the groups. Thus, the relationship between SDBOLD-Change and cognitive processing could be different in older adults and younger adults. Brainscores were computed as a dot product between V and SDBOLD-Change at each voxel for each subject.
We used a rank-based behavior PLS for investigating the effect of baseline composite cognitive performance on variability during all three conditions (placebo, drug, and change). Like the previously described one-condition, one-behavior PLS, a rank-based correlation matrix is computed based on baseline composite cognitive task scores and the signal of each voxel (SDBOLD) across subjects for each of the three conditions. An SVD on this matrix resulted in three (equal to the number of conditions) latent variables defined by S values, brain V weights, and U weights. Only one latent variable was significant. Brainscore for each condition was computed as the dot product between V weights and variability in each condition (SDBOLD-Placebo, SDBOLD-Drug, SDBOLD-Change) for each subject.
Finally, we used a rank-based behavior-PLS for investigating the effects of several cognitive processing tasks on SDBOLD-Change in older adults. It is very similar to the previously described one-condition, one-behavior PLS. First a rank-based correlation matrix is computed based on each of the cognitive task scores and the signal of each voxel (SDBOLD-Placebo) across subjects. An SVD is performed on this matrix. In our model it resulted in seven (equal to the number of cognitive tasks) latent variables defined by S values, brain V weights, and U weights. Only one latent variable was significant. All the U weights of this LV were negative, suggesting a negative correlation between change in variability and all cognitive tasks. Similar to brainscore, a cognitive score was computed as the dot product between U weights and the cognitive score in each task for each subject.
For all the PLS models, significance of the detected relations was assessed using 1000 permutation tests of the singular value, and the robustness of voxel saliences was computed using 1000 bootstrapped resamples. By dividing the mean salience of each voxel by the bootstrapped SE, we obtained bootstrap ratios (BSRs) as normalized estimates of robustness. We thresholded the BSRs at a value of ≥3.00, which approximates a 99.9% confidence interval. We then used the Harvard-Oxford cortical atlas to identify the regional identity of the significant clusters (Table 1–Table 6) in the cortical regions and the automated anatomical atlas (AAL) to identify the regional identity of significant clusters in the subcortical regions. All the other statistical analyses were conducted using R (https://www.r-project.org/). The lme4 package was used to perform the linear mixed-effects analyses, and figures were plotted using the ggplot2 package (Wickham, 2016) (https://cran.r-project.org/web/packages/lme4/index.html).
Brain regions that exhibited a reliable association between age and SDBOLD on placebo
Data availability
Data were collected as a part of a bigger study in Gagnon et al. (2019) and will be publicly available within 1 year of completion of the study. The code used for analysis is available from GitHub (https://github.com/umich-tpolklab/DynamicRecoveryPaper).
Results
Aging and brain signal variability
Consistent with previous results (Grady and Garrett, 2014), resting-state brain signal variability (operationalized as the SD of BOLD fMRI signal, SDBOLD) was significantly lower in older versus younger adults (Welch two-sample t test, t(35.1) = 3.98, p = 0.0003). There were two outliers (one young and one older adult) with a Cook's distance >0.087 (four/sample size). Even after excluding these two subjects, brain signal variability was significantly lower in older adults (permuted p = 0.02) compared with younger adults (Welch two-sample t test of brainscore, t(32.4) = 3.96, p = 0.0004; Fig. 1, Table 1). All the results presented below are based on excluding the two outlier subjects.
Effect of age on SDBOLD. A, Resting-state SDBOLD at placebo is significantly lower in older adults (purple) compared with younger adults (green). B, Spatial pattern expressing the effect of age. Yellow/red regions showed a reliable decrease in variability with age, whereas blue regions showed a reliable increase. Bootstrap ratios increase from red to yellow and from dark to light blue and are thresholded at a value of ≥3.00.
Increasing GABA activity boosts brain signal variability in older adults
We also investigated the effect of a small dose of a positive allosteric modulator (lorazepam) of the GABAA receptor on brain signal variability. We used PLS with younger and older adults as two separate groups and found one significant latent variable (permuted p < 0.001) showing higher brain signal variability on drug compared with placebo. The spatial pattern indicated a reliable increase in variability on drug in several regions including the cingulate gyrus, cerebellum, frontal, temporal, sensorimotor, and occipital regions. No regions showed a reliable reduction in variability on drug compared with placebo (Figure 2B, Table 2).
Brain regions that exhibited a reliable increase in variability on drug compared with placebo, dominantly in older compared with younger adults
Effect of GABA agonism on SDBOLD. A, SDBOLD increases on drug significantly in older adults but not in younger adults. Error bars indicate SD. B, Spatial pattern expressing the effect of drug on variability. Yellow/red regions exhibited a reliable increase in variability on drug (bootstrap ratio increases from red to yellow). No regions showed a reliable decline in variability on drug (absence of blue/green regions). Bootstrap ratios are thresholded at a value of ≥3.00.
We used a linear mixed-effects model to investigate the effect of age group, drug, and the age group × drug interaction on brain signal variability estimates produced from the above model after accounting for the subjects as random effects. Brain signal variability was significantly larger on drug than on placebo (F(1,42) = 14.8, p = 0.0004), and the age group × drug interaction was also significant (F(1,42) = 5.6, p = 0.02). On investigating the interaction further, we found that there was a significant increase in variability on drug within older adults (t(23) = −4.04, p = 0.0005), but not within younger adults (t(19) = −0.97, p = 0.34; Fig. 2A). Furthermore, variability was not significantly different between older adults on drug and younger adults on placebo (t(42) = −1.6, p = 0.11), consistent with the hypothesis that the drug restored the older adults' brain signal variability to young adult levels.
Given that the drug effect was only robust within the older group, we then reran an older-adult-only PLS model to isolate relevant brain regions (Fig. 3A). We found that variability reliably increased on drug (permuted p < 0.001) in several regions including the parahippocampal gyrus, fusiform cortex, sensorimotor regions, cerebellum, cingulate gyrus, and frontal and occipital regions (Fig. 3B, Table 3). No regions showed a reliable reduction on drug compared with placebo. As a further set of controls for this older-adult-only PLS result, we used a linear mixed-effects model to account for main effects of session order (counterbalanced across participants), gray matter volume, days between sessions (mean, 16 d; range, 2–79 d), dosage (0.5 vs 1 mg), self-reported drowsiness before and after the drug session, age (mean, 69.9; range, 65–81), and drug on brain signal variability estimates from this model, as well as the age × drug interaction after accounting for subjects as random effects. Only the main effect of drug (F(1,23) = 16.62, p = 0.0005) was significant, and this remained the only significant effect after excluding one participant with a Cook's distance >0.17 (four/number of older adults; F(1,22) = 18.14, p = 0.0003).
Brain regions that exhibited a reliable increase in variability on drug compared with placebo in older adults alone
Effect of GABA agonism on SDBOLD in older adults. A, Resting-state SDBOLD in older adults was significantly higher on drug (red) compared with placebo (blue). B, Spatial pattern expressing the effect of the drug on brain signal variability. Yellow/red regions exhibited a reliable increase in variability on drug versus placebo, whereas blue regions exhibited a reliable decrease. Bootstrap ratios were thresholded at a value of more than or equal to ±3.00.
Increasing GABA activity leads to greater boost in brain signal variability in poorer performing older adults
To examine the relationship between cognitive performance and change in brain signal variability on drug, we obtained the composite cognition score from the NIH Toolbox for the Assessment of Neurologic and Behavioral Function (Weintraub et al., 2014) and computed change in variability at each voxel (SDBOLD-change = SDBOLD-Drug – SDBOLD-Placebo). Consistent with the previous literature (Garrett et al., 2011, 2013a; Grady and Garrett, 2014; Burzynska et al., 2015), we found that the composite cognition scores were significantly lower in older adults compared with those of young adults (t(40.7) = 4.7, p <0.001). Using PLS with two age groups modeled separately, we found a single significant latent variable (permuted p = 0.03) that captured the relationship between change in brain signal variability on drug versus placebo (SDBOLD-change) and the composite cognitive score from the NIH battery. The relationship was negative and significant in older adults (r(22) = −0.51, bootstrap CI: 0.46, 0.85; Fig. 4A) but not in younger adults (r(18) = 0.17, bootstrap CI: −0.1, 0.6). The age group × cognition interaction (computed using ANOVA on the brainscore estimates) was also significant (F(1,40) = 7.2, p = 0.01). There were no outliers based on Cook's distance. The negative association between cognitive performance and GABA drug-related change in variability in older adults suggests that poorer performers experience a greater drug-related boost in SDBOLD than higher performers, an effect that was reliable in several brain regions, including the cingulate and middle frontal gyrus, thalamus, superior parietal lobule, sensorimotor, lateral occipital, and temporal regions (Fig. 4B, Table 4).
Brain regions that exhibited a reliable association between overall cognitive processing score and drug-related shifts in variability, mainly expressed within the older adult group
Cognition and boost in variability on drug. A, Change in SDBOLD on drug is negatively correlated with composite cognitive score in older adults but not in younger adults. Error bars indicate bootstrapped 95% confidence intervals (see above, Materials and Methods). B, Spatial pattern expressing the relationship between change in variability on drug and baseline composite cognitive score. Yellow/red regions exhibited a reliable negative relationship (bootstrap ratio increases from red to yellow). No regions exhibited a reliable positive relationship (absence of blue/green regions). Bootstrap ratios were thresholded at a value of more than or equal to ±3.00.
These results support the hypothesis that older adults with poorer cognitive processing benefit more from the drug-induced increase in GABA activity. Within older adults only, we further investigated the relationship between composite cognitive score and variability during all three conditions, namely, (1) baseline variability (SDBOLD-Placebo), (2) variability on drug (SDBOLD-Drug), and (3) change in variability on drug from baseline (SDBOLD-Change; Fig. 5A,B). We found that change in variability (SDBOLD-Change) explained significant variance in cognition even after accounting for baseline variability (SDBOLD-Placebo; F(1,21) = 17.4, p = 0.0004) and variability on drug (SDBOLD-Drug; F(1,21) = 9.1, p = 0.007).
Cognition and SDBOLD in older adults. A, Baseline variability is positively correlated with baseline cognitive performance (but not significantly), whereas variability on drug and change in variability from placebo to drug are negatively correlated with baseline cognitive performance. Error bars indicate bootstrapped 95% confidence intervals (see above, Materials and Methods). B, Spatial pattern expressing the relationship between variability and baseline cognitive performance in all three conditions. Yellow/red regions exhibited a reliable effect (bootstrap ratio increases from red to yellow). No regions exhibited a reliable effect in the opposite direction (absence of blue/green regions). Bootstrap ratios were thresholded at a value of more than or equal to ±3.00.
Finally, we examined the role of change in variability on drug and performance on the individual cognitive tasks (during baseline) that make up the composite cognitive score from the NIH toolbox. We found that all the tasks were negatively associated (significantly in five of seven using 1000 bootstraps) with change in variability on drug (Fig. 6A). This relationship was reliable in several regions of the brain including the middle frontal gyrus, superior parietal lobule, thalamus, parahippocampal gyrus, cingulum, sensorimotor, lateral occipital, and temporal regions (Fig. 6C, Table 5). Figure 6B shows the overall latent relationship between cognitive performance and SDBOLD-change derived from this PLS model (r(22) = −0.58, p = 0.003). This relationship was also significant, even after accounting for self-reported drowsiness before and after the on-drug scan, psychomotor vigilance score before and after the on-drug scan, age, and dosage (0.5 vs 1 mg; F(1,16) = 10.03, p = 0.006). This correlation was negative, suggesting that older adults with high cognitive processing (during baseline) showed the smallest changes on drug, whereas those with the lowest scores experienced a greater boost.
Brain regions that exhibited a reliable association between cognitive processing score from the NIH toolbox and brain signal variability measure during placebo and drug condition as well as drug-related change in brain signal variability in older adults alone
Brain regions that exhibited a reliable association between various cognitive processing tasks from the NIH toolbox and drug-related shifts in brain signal variability in older adults alone
Cognition and boost in variability on drug in older adults. A, Drug-related change in SDBOLD in older adults is negatively correlated with baseline performance on several cognitive tasks. Error bars indicate bootstrapped 95% confidence intervals (see above, Materials and Methods). B, Drug-related change in SDBOLD in older adults is negatively correlated with baseline cognitive score. C, Spatial pattern expressing the relationship between change in variability on drug and baseline cognitive performance. Yellow/red regions exhibit reliable effects (bootstrap ratio increases from red to yellow). No regions exhibited reliable effects in the opposite direction (absence of blue/green regions). Bootstrap ratios were thresholded at a value of more than or equal to ±3.00.
GABA-driven upregulation of brain signal variability in poorer performing older adults is present in regions showing robust age reductions in variability
To examine whether the drug-related boost in brain signal variability was directly present in brain regions showing age-related reductions in SDBOLD, we created a mask using clusters showing robust age-related declines in brain signal variability (Fig. 1, Table 1). We then computed average brain signal variability within this mask during both the placebo and drug conditions. We found that brain signal variability in these regions was significantly higher on drug compared with placebo in older adults (t(23)=3.7, p = 0.001), effects similar to those found using the whole-brain PLS model (t(23) = 4.1, p = 0.0004). Likewise, there was a negative correlation between composite cognitive score and average change in variability within the mask in older adults (r(22) = −0.48, p = 0.02), similar to that found using the whole-brain PLS model (r(22) = −0.58, p = 0.002).
Discussion
In the present study, we found that brain signal variability was significantly lower in older adults than in younger adults, that increasing GABA activity pharmacologically increased brain signal variability in older adults to the level of healthy young adults, and that GABA-related boosts in variability were largest in the poorest performing older adults. These results provide evidence that GABA may provide a crucial basis for understanding associations among brain signal variability, aging, and cognition in humans.
Our finding that brain signal variability was significantly reduced in older adults (e.g., in the superior frontal regions, bilateral superior parietal lobule, and occipital, sensorimotor, and auditory regions) is consistent with a number of previous studies. For example, fMRI signal variability in cortex is often found to be lower in older adults compared with young adults in various studies using fixation baseline periods as a resting-state proxy (Grady and Garrett, 2014) and in those using entire resting-state data (Kielar et al., 2016; Nomi et al., 2017; Grady and Garrett, 2018). Age-related differences in resting-state variability are also robust to multiple vascular controls at the voxel level (Garrett et al., 2017). Further in line with previous work (Garrett et al., 2011; Nomi et al., 2017), relatively few clusters (∼6% of the overall spatial pattern in cerebellum and inferior temporal cortex; Table 1) exhibited greater variability in the older adults. These results add to a growing body of literature suggesting that brain signal variability levels are lower overall in older versus younger adults. However, other studies using different preprocessing pipelines, experimental designs, and variability measures such as the mean squared successive differences approach (Samanez-Larkin et al., 2010; Boylan et al., 2021) have found mainly positive effects between BOLD variability and adult age. Future work could investigate such differences by aggregating various datasets, performing direct comparisons of effects, and unifying the preprocessing pipelines.
GABA levels have been found to decline with age in human visual (Chamberlain et al., 2019), sensorimotor (Cassady et al., 2019; Cuypers et al., 2020), auditory (Lalwani et al., 2019), parietal (Gao et al., 2013), and frontal cortices (Porges et al., 2017), and are associated with individual differences in cognitive and sensorimotor abilities (Porges et al., 2017; Simmonite et al., 2019; Cassady et al., 2019; Levin et al., 2019). Moreover, the amplitude of low-frequency fluctuations, which is mathematically equivalent to SDBOLD when computed on the exact same band-limited time series (median correlation >0.92 in the present dataset), is also positively associated with GABA-binding potential in healthy young adults (Nugent et al., 2015). In the present work, we found that agonizing GABA activity led to an increase in cortical brain signal variability, especially in older, poorer performers.
So how does boosting GABA lead to higher moment-to-moment variability in brain activity? Computational modeling suggests a possible explanation. Having sufficient inhibitory activity to offset excitatory activity has been found to be crucial to allow artificial neural networks to operate near so-called criticality, an operating point near the edge of instability where it is easier to switch from one network state to another (Shew et al., 2011; Deco and Jirsa, 2012; Poil et al., 2012; Agrawal et al., 2018). If inhibitory connections are too weak, then excitatory activity dominates and many neurons fire synchronously, resulting in redundant coding and deep attractor states that are very stable and harder to transition from (Agrawal et al., 2018). In this hyperexcited regime, the network only samples a few synchronous configurations and therefore does not exhibit as much variability. Conversely, if inhibitory connections are too strong, then fewer neurons fire, which can also lead to a reduction in the number of states the network visits. Indeed, increasing GABA activity using propofol (a general anesthetic that also modulates activity at GABAA receptors) leads to a decrease in power spectral density (another measure of signal variance) in monkey electrocorticography recordings (Gao et al., 2017). In short, inhibitory connections should be strong enough to balance the excitatory connections but not so strong that they dampen the entire network. Networks function at criticality and optimally when network dynamics are stabilized by sufficient inhibition (for review, see Sadeh and Clopath, 2021); that is, there is an inverted-U relationship between brain dynamics and GABA that is similar to that observed with other neurotransmitters.
It is plausible that older adults with poorer cognitive performance may reside on the lower left half of a GABA-variability inverted–U function, whereas younger and better performing older adults may be closer to the peak. Increasing GABA activity in older adults with worse cognitive performance should then allow their neural networks to operate nearer to criticality and visit different states more frequently, leading to increased brain signal variability. Conversely, younger adults and older adults with better cognitive performance would not be expected to show as much change in network dynamics/variability on drug. Consistent with these expectations, we found that older adults showed robust increases in variability through GABA agonism and reached levels comparable to younger adults' variability levels at placebo. Additionally, within older adults, change in variability explained significant variance in cognition even after accounting for variance explained by variability on placebo or drug alone. Finally, we found that the poorest performers were most likely to have their signal variability levels boosted on drug, and that association could not be attributed to individual differences in drowsiness, age, or drug dosage. These results provide the first evidence that brain signal variability can be restored by increasing activity of the GABAergic system, particularly for older, poorer cognitive performers.
Limitations and future work
It is important to note that this study was cross-sectional rather than longitudinal, and so the observed age differences could be influenced by cohort or period effects (Hofer et al., 2002). Another limitation was the absence of a middle-age group, making it difficult to disentangle whether the age-related changes in variability were because of aging versus maturation (Editorial Board of the Journal of Neuroscience, 2019). However, previous research has shown that brain signal variability increases during development (from ages 8 to 15 years; McIntosh et al., 2010), and the role of individual differences in variability on cognition among older adults indicates that it is more likely to be an aging process. Additionally, we do not have behavioral data on drug and therefore cannot assess how manipulating GABA activity affected behavior in our sample. However, promisingly, several drugs targeting the GABAergic system have been shown previously to attenuate or even reverse features and symptoms of Alzheimer's disease (for review, see Guzmán et al., 2018). Further research in healthy older adults is thus needed to assess whether cognitive function can also be jointly boosted with brain signal variability via GABA agonism. Finally, we need to better understand how the GABA system is associated with other candidate neurotransmitter systems that have also been proposed previously as plausible bases of moment-to-moment brain signal dynamics (e.g., noradrenaline, dopamine; Garrett et al., 2015; Alavash et al., 2018; Kosciessa et al., 2021) and, in particular, whether the inhibitory system is a more or less effective pharmacological target than the excitatory (glutamatergic) system in future work linking aging, cognition, and neural variability in humans.
Conclusion
Overall, we found that GABA agonism can increase brain signal variability in older poorer performing adults. These results suggest the critical role of the GABAergic system in neural variability and the importance of both in aging and cognition. Potentiating GABAergic signaling represents a potentially promising direction to pursue in efforts to mitigate age-related deficits in brain function and behavior.
Footnotes
This work was supported by a grant from the National Institutes of Health to T.A.P (R01AG050523). We thank Holly Gagnon for participant recruitment and data collection.
The authors declare no competing financial interests.
- Correspondence should be addressed to Thad A. Polk at tpolk{at}umich.edu
