Seventy-five healthy young adults (49 females and 26 males; age range = 18–38) who participated in six separate studies at UC Berkeley were included in these analyses, three of which have been published (Jacobs and D’Esposito, 2011; Lee and D’Esposito, 2012; Gallen et al., 2016b). Participants were screened to exclude those with any history of neurologic or psychiatric disorders and, for Study 3, for abnormal or infrequent menstrual cycles or use of hormonal birth control. Here, we included participants with a minimum of four blocks of each cognitive task condition (see below). This resulted in 26 participants from Study 1 (Gallen et al., 2016b), 10 participants from Study 2 (Lee and D’Esposito, 2012), 11 participants from Study 3 (Jacobs and D’Esposito, 2011), and 28 participants from unpublished data. Informed consent was obtained from participants in accordance with the Committee for the Protection of Human Subjects at the University of California, Berkeley.

The cognitive task performed during fMRI scanning was a working memory (WM) N-back task with face and scene stimuli (Figure 1), in which participants were instructed to identify repetitions 1-back from the current image (1-back condition) or respond to the current image (0-back condition). The task consisted of either 16 or 20 (N = 17 or 58 participants, respectively) total blocks lasting approximately 2 min each (Chen et al., 2011). Separate analyses using this task have been previously published for data collected in Studies 1–3 (Jacobs and D’Esposito, 2011; Lee and D’Esposito, 2012; Gallen et al., 2016b), including a re-analysis of data collected in Studies 1–3 (Cohen et al., 2014). Each task block contained 20 consecutively presented face and scene stimuli (10 of each, pseudo-randomly ordered to limit predictability of the upcoming stimuli), presented for 600 ms, with a 2.4, 4.4, or 6.4 s jittered delay (randomly ordered) between each stimulus presentation. Participants responded to each stimulus with one of two buttons using their right index or middle fingers.

Each block had one of four task conditions that varied both task goals and WM demands. We refer to these conditions as: “Categorize,” “Remember Scenes,” “Remember Faces,” and “Remember Both.” In this study, we analyzed data from the Categorize, Remember Scenes, and Remember Faces conditions. In the Categorize condition, participants indicated whether the current image was a face (index finger) or a scene (middle finger) and did not need to remember the image (i.e., a 0-back condition). As such, “Categorize” is a control condition that matches sensory information without WM demands. In the Remember Scenes (Ignore Faces) 1-back WM condition, participants were instructed to selectively attend to and remember stimuli from the relevant category (scenes) and respond to stimuli from the irrelevant category (faces) as distractors. Conversely, in the Remember Faces (Ignore Scenes) condition, participants were instructed to selectively attend to and remember face stimuli and respond to irrelevant scene stimuli as distractors. In these conditions with varying stimulus relevance (Remember Scenes and Remember Faces), participants indicated if the currently attended image matched the previous image (i.e., 1-back) of the same category (matches: index finger, non-matches: middle finger); participants also responded to all irrelevant items with the non-match button. Finally, there was a “Remember Both” condition that was not analyzed in this study, in which participants were instructed to attend to and remember both the face and scene stimuli. Depending on the study, participants completed four or five blocks of each condition during the scanning session (pseudo-randomly ordered), yielding a maximum of 40 or 50 face and scene stimuli for each task condition.

Imaging data were collected on 3T Siemens Trio scanners at the University of California, Berkeley Henry H. Wheeler, Jr. Brain Imaging Center, or at the University of California, San Francisco Neurosciences Imaging Center (Study 3 only). T1-weighted structural and T2*-weighted echo-planar images (EPIs) were collected with a 12-channel head coil for all studies. Functional data were collected using parallel imaging (GRAPPA) with acceleration factor 2 in Study 3 and unpublished Studies 1 and 3, and two participants in unpublished Study 2 (TE = 27 ms). GRAPPA was not used in Study 1 (TE = 24 ms), Study 2 (TE = 32 ms), and three participants in unpublished Study 2 (TE = 24 ms). All studies used 18 5-mm axial slices with a 0.5-mm gap (descending slices for Study 1 and three participants in unpublished Study 2; interleaved slices for Studies 2, 3, and unpublished Studies 1, 3, and two participants in unpublished Study 2). All studies collected functional volumes with TR = 1,000 ms and a 3.5-mm² in-plane resolution. Four or five 114-volume runs of each of the four task conditions were collected, yielding a total of 30.4 or 38 min of task data. A high-resolution axial MP-RAGE T1-weighted sequence was used to acquire structural images for all studies (TR = 2,300 ms, TE = 2.98 ms, FA = 9°, 1 mm³ voxels).

Standard preprocessing of EPI data was carried out with AFNI (Analysis of Functional NeuroImages) and FSL (FMRIB Software Library). EPI volumes were slice-time, and motion corrected, co-registered to the T1-weighted structural image, and warped to MNI (Montreal Neurological Institute) template space using FSL’s non-linear registration. Intensity spikes were removed and interpolated with AFNI after slice-time and motion correction. Functional data were resampled to 2-mm isotropic voxels, combining motion correction and atlas transformation in a single interpolation. MNI-warped functional data were spatially smoothed with a 6-mm full width at half maximum Gaussian kernel and scaled so that each voxel’s run mean was equal to 100.

To quantify task-related functional connectivity, we implemented a beta-series correlation approach (Gazzaley et al., 2004; Rissman et al., 2004). As with prior work using this task (Lee and D’Esposito, 2012; Cohen et al., 2014; Cohen and D’Esposito, 2021), we focused on examining stimulus-evoked responses (i.e., activity during the processing of task stimuli). To estimate trial-wise evoked responses in the BOLD signal, a voxel-wise regression was performed. Here, task-dependent changes were modeled with independent regressors for correct trials for each task condition (Categorize, Remember Scenes, Remember Faces, Remember Both) and stimulus type (faces or scenes). These regressors were created by convolving a canonical double-gamma hemodynamic response function (HRF) with the onset times for each trial. We also included separate regressors that modeled incorrect and missed trials, motion parameters, run means, and linear trends (as “nuisance regressors” of no interest). This resulted in a parameter estimate, or beta value, for each voxel and trial of the task. Only correct trials for the Categorize, Remember Scenes, Remember Faces conditions were included in subsequent analyses. To interrogate large-scale brain network properties, we next parcellated the brain into a set of 264 atlas regions of interest [ROIs; (Power et al., 2011)] each comprised of 81 voxels, representing the network nodes, and averaged the beta values across all voxels within each ROI.

As we were interested in examining the effects of (1) task conditions (e.g., 0-back vs. 1-back) as well as (2) trial-level task goals for each stimulus, we sorted the beta values into their separate task conditions (i.e., Categorize, Remember Scenes, Remember Faces) and stimulus types (i.e., faces or scenes). First, we examined effects of task condition collapsed across stimulus types (i.e., 0-back Categorize condition and 1-back Remember Faces and Remember Scenes conditions). To do so, we concatenated all beta values from both face and scene trials for a given task condition to create 3 separate condition-level “beta-series” for each atlas ROI: Categorize, Remember Scenes, and Remember Faces. Second, we examined effects of stimulus goals (i.e., categorize, irrelevant, or relevant) for trials within each task condition. To do so, we concatenated beta values separately for face and scene stimuli, depending on the task condition and stimulus goals. This resulted in 6 beta-series for each ROI: categorized scenes (Categorize condition), categorized faces (Categorize condition), relevant scenes (Remember Scenes condition), irrelevant faces (Remember Scenes condition), relevant faces (Remember Faces condition), and irrelevant scenes (Remember Faces condition).

Finally, task-related functional connectivity matrices, representing the network edges, were created for each participant by correlating the beta-series between each pair of ROIs using Pearson’s correlation coefficient and applying a Fisher z-transform. Connectivity matrices were created separately for each of the 9 beta-series types described above (e.g., a separate matrix was created for remember scenes, relevant scenes, and irrelevant faces). The range of beta values for each of the stimulus goals and stimulus types across participants were as follows: categorized scenes (35–50), categorized faces (37–50), relevant scenes (32–50), irrelevant faces (39–50), relevant faces (33–50), and irrelevant scenes (36–50). ROIs were excluded from all analyses if they were missing EPI coverage due to incomplete sampling of the whole brain during fMRI scanning in 90% or more voxels from the original 81-voxel atlas ROI across scanning runs in any participant. A total of 193 ROIs were included in the final network analyses.

The 193 × 193 connectivity matrices were binarized to create adjacency matrices that indicate the presence or absence of a connection, or network edge, between a pair of regions. Matrices were binarized over a range of connection density thresholds, where thresholding of the matrices was achieved by matching the number of network connections across participants (here, the top 2–10% of all possible connections in the network in 2% increments) (Gallen et al., 2016a; Baniqued et al., 2017, 2019). Each of these thresholded matrices was used to create unweighted, undirected whole-brain graphs (defined as a set of nodes, or ROIs, and the edges, or task-related connections, between them) with which network metrics were examined. Network metrics were created separately for each connection threshold and are presented here as the average across connection density thresholds.

First, we examined changes in whole-brain network modularity, a global network measure that compares the number of connections within to the number of connections between sub-networks, or “modules,” across the entire brain network (Newman, 2004; Newman and Girvan, 2004). Modularity will be 1 if all connections fall within sub-networks, and it will be 0 if there are no more connections within sub-networks than would be expected by chance. To do so, we assigned each atlas ROI to a sub-network according to the sub-network partition identified with these nodes, as described in Power et al. (2011). We then quantified whole-brain network modularity for each participant and beta-series type, resulting in 9 modularity values for each participant.

After identifying global changes in network modularity due to goals (e.g., relevant or irrelevant), we examined which sub-networks contributed most to the observed whole-brain modularity effects, which is the sum of contributions from each individual sub-network. To do so, we quantified the modularity of each of the sub-networks described in Power et al. (2011) and identified which showed similar patterns to the whole-brain modularity effects. Note that the coverage of our functional data did not include the cerebellum, thus only permitting analysis of 13 of the 14 brain sub-networks identified in Power et al. (2011). We also examined the reliability of our whole-brain modularity findings by using a spectral algorithm (Newman, 2006) to identify the most optimal modular partition (i.e., maximal modularity) for each participant and condition separately, rather than imposing the Power et al. (2011) modular partition across all networks.

To examine effects of stimulus goals on task performance, we conducted repeated-measures ANOVAs with within-subject factors of stimulus goals (Categorize, Irrelevant, Relevant) and stimulus type (Faces, Scenes), separately for accuracy (percent correct trials) and response time (RT; average response time for correct trials).

To examine changes in whole-brain modular network organization, we conducted two separate analyses, examining (1) effects of task condition and (2) effects of stimulus goals on whole-brain, global network modularity. First, the effect of task condition on global modularity was assessed with a repeated measures ANOVA with a within-subject factor of task condition (Categorize, Remember Scenes, Remember Faces). Second, the effect of stimulus goals on global modularity was assessed with a repeated measures ANOVA with within-subject factors of stimulus goals (Categorize, Irrelevant, Relevant) and stimulus type (Faces, Scenes). We then conducted follow-up ANOVAs to identify which sub-networks contributed most to the observed global modularity stimulus goal effects. Finally, we quantified relationships between task performance and goal-related changes in modularity using Pearson’s correlations, focusing on RT of all relevant trials that placed demands on WM in line with previous work (Kitzbichler et al., 2011; Vatansever et al., 2015). For all analyses, we set a significance threshold of p < 0.05.

Accuracy and response times (RT) were significantly influenced by both stimulus goals and stimulus type (Figure 2). An ANOVA with within-subject factors of stimulus goals (Categorize, Irrelevant, Relevant) and stimulus type (Scenes, Faces) on accuracy and RT showed main effects of goals [accuracy: F(2,148) = 71.59, p < 0.001; RT: F(2,148) = 136.07, p < 0.001] and type [accuracy: F(1,74) = 15.35, p < 0.001; RT: F(1,74) = 40.08, p < 0.001]. For accuracy, participants were less accurate for relevant stimuli used for WM performance, compared to categorized and irrelevant stimuli that only required a judgment on the sensory information (e.g., a face or a scene). Participants were also less accurate for categorized stimuli compared to irrelevant stimuli (Figure 2A). For RT, participants were slower for relevant stimuli compared to categorized and irrelevant stimuli (Figure 2B). Participants were also overall more accurate and faster for face stimuli compared to scene stimuli.

We also found a significant interaction between stimulus goals and type for accuracy [F(2,148) = 7.12, p = 0.001] and RT [F(2,148) = 49.90, p < 0.001], suggesting that the effects of stimulus type depended on the stimulus goals. Follow-up ANOVAs comparing face and scene stimuli for each condition showed that participants were more accurate and faster for faces than scenes when they were categorized or irrelevant [accuracy: F(1,74) = 23.98, p < 0.001; F(1,74) = 10.98, p = 0.001; RT: F(1,74) = 115.96, p < 0.001; F(1,74) = 39.47, p < 0.001]; however, we found no evidence for a difference in accuracy between faces and scenes [F(1,74) = 0.17, p = 0.68] when they were relevant, and participants were slower for faces than scenes when they were relevant [F(1,74) = 16.10, p < 0.001].

We next investigated how whole-brain modularity changed depending on the task condition and stimulus goals. First, we examined changes in modularity for overall task conditions: Categorize, Remember Scenes, and Remember Faces. We found that whole-brain modularity differed between the task conditions [F(2,148) = 5.73, p = 0.004], such that modularity was lower for the 1-back conditions (Remember Scenes and Remember Faces) compared to the 0-back (Categorize) condition (Figure 3A; p = 0.007; p = 0.003), but was not significantly different between the Remember Scenes and Remember Faces conditions (p = 0.72). These findings suggest that modularity was lower during the more demanding 1-back conditions compared to a baseline 0-back condition.

Next, we examined how modular brain network organization changed on a trial-wise basis during the task conditions, depending on the stimulus goals (i.e., categorize, irrelevant, relevant). We found a main effect of stimulus goals [Figure 3B; F(2,148) = 19.04, p < 0.001], such that modularity was selectively lower during goal-directed processing of relevant stimuli compared to irrelevant and categorized stimuli (both p < 0.001), but was not significantly different between processing irrelevant and categorized stimuli (p = 0.30). We did not detect either a significant main effect of stimulus type [F(1,74) = 0.01, p = 0.92] or, importantly, a significant interaction between stimulus goals and type [F(2,148) = 0.10, p = 0.90], suggesting that the observed reductions in modularity for relevant stimuli unlikely depended on whether they were faces or scenes. These findings suggest that modularity was lower during goal-directed processing of task-relevant stimuli, compared to irrelevant or categorized stimuli.

We examined the robustness of our findings by confirming that our results were similar using a spectral partitioning algorithm to identify “optimal” sub-networks (sub-network partitions with the maximum modularity) for each participant and condition rather than using the Power et al. (2011) sub-network definitions. For effects of task condition, modularity was lower for the 1-back Remember Faces condition compared to the 0-back (Categorize) condition at a “trend” level (p = 0.08), but was not statistically different for the 1-back Remember Scenes condition compared to the 0-back (Categorize) condition (p = 0.22). Modularity was not significantly different between the 1-back conditions (p = 0.64). For effects of stimulus goals, modularity was lower during processing of relevant stimuli compared to irrelevant and categorized stimuli (p = 0.001 and p = 0.006, respectively), but was not significantly different between processing irrelevant and categorized stimuli (p = 0.70).

Further, we confirmed that reductions in modularity during the processing of task-relevant stimuli were not related to potential confounds. First, reductions in modularity were not related to participant age [relevant–irrelevant: r(53) = −0.19, p = 0.17; relevant–categorize: r(53) = −0.19, p = 0.16; note that participant-specific age data was only available for 55 participants]. Second, reductions in modularity were not related to participant in-scanner motion, quantified as the Euclidean norm of motion parameters using AFNI [relevant–irrelevant: r(73) = 0.06, p = 0.60; relevant–categorize: r(73) = 0.12, p = 0.32].

As whole-brain modularity is quantified as the sum of each sub-network’s modularity values, we next sought to examine individual sub-network contributions to the observed whole-brain modularity findings. We conducted an ANOVA with within-subject factors of stimulus goals (Categorize, Irrelevant, Relevant) and sub-network (N = 13). For this analysis, the modularity values were averaged for face and scene stimuli, as there was no main effect of stimulus type or interaction between stimulus goals and type for whole-brain modularity.

We found a significant interaction between stimulus goals and sub-network [F(24,1776) = 6.75, p < 0.001], suggesting that the effect of stimulus goals on modularity differed across sub-networks. To further interrogate this interaction, we conducted separate ANOVAs with a factor of stimulus goals (Categorize, Irrelevant, Relevant) for each sub-network to identify those with effects that mirrored the whole-brain modularity effects (i.e., reduced modularity for relevant stimuli compared to irrelevant and categorized stimuli; Supplementary Table 1). We found a significant effect of stimulus goals for the default mode network [DMN; F(2,148) = 12.38, p < 0.001] and visual [F(2,148) = 4.94, p = 0.008] sub-network (Figures 4A, B, respectively). In both sub-networks, modularity was reduced during processing of relevant stimuli compared to irrelevant (DMN: p < 0.001; visual: p = 0.001) and categorized stimuli (DMN: p < 0.001; visual p = 0.02), but was not significantly different between irrelevant and categorized stimuli (DMN: p = 0.11; visual: p = 0.58).

Finally, we examined relationships between goal-related changes in network modularity and task performance to assess the cognitive relevance of such alterations in network organization. As prior work investigating behavioral relationships with modularity changes due to N-back demands has focused on RT (Kitzbichler et al., 2011; Vatansever et al., 2015), we similarly used RT for all relevant stimuli (trials with high WM demands) as our behavioral metric to index WM performance (quantified as the mean RT across relevant face and scene stimuli).

We first examined how the reductions in whole-brain modularity for relevant stimuli (compared to categorized and irrelevant stimuli) were related to task performance. We found that individuals with lower modularity for relevant stimuli compared to irrelevant stimuli had faster WM task performance [Figure 5A; r(73) = 0.34, p = 0.003]. However, we did not detect a significant relationship between task performance and the change in modularity between relevant and categorized stimuli [r(73) = 0.04, p = 0.72]. Moreover, we did not detect a significant relationship between changes in task performance (relevant–irrelevant RT) and changes in network modularity [r(73) = 0.11, p = 0.36].

We examined similar relationships at the sub-network level. We focused on modularity changes between relevant and irrelevant stimuli within the DMN and visual sub-networks, as they showed the most prominent effects of goal-related stimulus goals on modularity. To do so, we quantified the change in modularity between relevant and irrelevant stimuli for each sub-network and averaged these values across (1) DMN and the visual sub-networks and (2) all other sub-networks (N = 11). We found that individuals with lower DMN and visual modularity for relevant stimuli (compared to irrelevant stimuli) had faster task performance [Figure 5B; r(73) = 0.33, p = 0.004]. Correlations between changes in modularity and performance separately for DMN and the visual sub-network showed a similar relationship [r(73) = 0.23, p = 0.049; r(73) = 0.21, p = 0.068, respectively]. Importantly, we did not detect a significant relationship between changes in modularity and task performance for the remaining 11 sub-networks that did not show effects of stimulus goals similar to the whole-brain modularity findings (average modularity across the 11 sub-networks); [r(73) = 0.16, p = 0.17].