Functional specialization of parallel distributed networks revealed by analysis of trial-to-trial variation in processing demands

Abstract

Multiple large-scale networks populate human association cortex. Here, we explored the functional properties of these networks by exploiting trial-to-trial variation in component-processing demands. In two behavioral studies (n = 136 and n = 238), participants quantified strategies used to solve individual task trials that spanned remembering, imagining future scenarios, and various control trials. These trials were also all scanned in an independent sample of functional MRI participants (n = 10), each with sufficient data to precisely define within-individual networks. Stable latent factors varied across trials and correlated with trial-level functional responses selectively across networks. One network linked to parahippocampal cortex, labeled Default Network A (DN-A), tracked scene construction, including for control trials that possessed minimal episodic memory demands. To the degree, a trial encouraged participants to construct a mental scene with imagery and awareness about spatial locations of objects or places, the response in DN-A increased. The juxtaposed Default Network B (DN-B) showed no such response but varied in relation to social processing demands. Another adjacent network, labeled Frontoparietal Network B (FPN-B), robustly correlated with trial difficulty. These results support that DN-A and DN-B are specialized networks differentially supporting information processing within spatial and social domains. Both networks are dissociable from a closely juxtaposed domain-general control network that tracks cognitive effort.NEW & NOTEWORTHY Tasks shown to differentially recruit parallel association networks are multifaceted, leaving open questions about network processes. Here, examining trial-to-trial network response properties in relation to trial traits reveals new insights into network functions. In particular, processes linked to scene construction selectively recruit a distributed network with links to parahippocampal and retrosplenial cortices, including during trials designed not to rely on the personal past. Adjacent networks show distinct patterns, providing novel evidence of functional specialization.

Keywords: association cortex; cognitive control; frontoparietal control network; hippocampus.

Graphical abstract

Figure 1.

Behavioral ratings illustrate unique and reliable strategy use patterns. Mean strategy ratings from independent groups of behavioral participants show striking similarity (red: exp 1, blue: exp 2). Four example trials are displayed, chosen from the original “target” conditions designed to demand episodic projection (e.g., remembering and imagining the future). Above each plot is the actual question the participants viewed; below is the measured strategy pattern. The strategies plotted on the x-axis are listed in Table 1. The trials share high ratings for strategies relevant to episodic memory, as intended, such as consideration of the personal past (Personal_Past_Exper), events (Sequence_Events), and mental scenes (Visual_Imagery, Loc_Obj_Places). High intertrial variability on other strategy dimensions highlights the exploratory opportunity (see also Fig. 2). For example, Difficulty was low for some trials and higher for others. Each point shows a mean strategy rating across participants with standard error bars. Sample sizes for estimating each trial varied, always with N > 16. In the present four example trials, N = 136 (exp 1) and N = 238 (exp 2) for trial 29; N = 17 and N = 39 for trial 51; N = 22 and N = 42 for trial 110, and N = 19 and N = 40 for trial 157. Exper, Experiences; Pers, Personal. *A repeated trial with a larger sample size.

Figure 2.

Strategy use patterns differ markedly among the control trials. Strategy patterns are shown for four trials taken from the control conditions. For each trial, mean strategy ratings from independent groups again reveal notable overlap (red: exp 1, blue: exp 2). The four trials were selected from the original “control” conditions, designed to minimize demands on episodic projection. Most of these trials show lower reliance on the personal past (Pers_Past_Exper) and greater use of facts than the target trials in Fig. 1. The control trials also reveal marked variability. Multiple trials involve strategies related to mental scenes (Visual_Imagery, Loc_Obj_Places), for example, or to considering unfolding events (Sequence_Events). Each point shows a mean strategy rating across participants with standard error bars. As in Fig. 1, sample sizes for estimating each trial varied. In the present four example trials, N = 117 (exp 1) and N = 238 (exp 2) for trial 4; N = 18 and N = 37 for trial 16; N = 19 and N = 40 for trial 71, and N = 21 and N = 40 for trial 124. Exper, Experiences; Pers, Personal. *A repeated trial with a larger sample size.

Figure 3.

Reliable strategy clusters emerge that capture trial-to-trial variation. Left: a correlation matrix illustrates the relations among the 16 scaled strategy probes, using data from all 180 trial questions in behavioral exp 1. Strong correlations emerge between subsets of strategy probes indicating that individual trials have distinct rating combinations. For example, trials high in use of visual imagery (Visual_Imagery) also tend to be high in reports of imagining the locations of objects and places (Loc_Obj_Places) and, to a lesser degree, the locations of people (Loc_People). Right: an independent correlation matrix from exp 2 reveals that the strategy relations are reliable (ordered here as in exp 1 for visualization). The added boxes around correlation clusters in exp 2 reveal the groupings that were selected based on hierarchical clustering as shown in Fig. 4. Exper, Experiences; Pers, Personal.

Figure 4.

Hierarchical clustering yields five distinct strategy composite scores. Hierarchical clustering identified strategies that could be combined into composite scores for functional network analysis. Top: the dendrogram from exp 1 displays composites using a cut point above “Facts” and “Difficulty,” preserving all clusters at least as strong as this pair. The cut point is noted by a red triangle on the y-axis. Dashed boxes show the cluster groupings. Bottom: the independently estimated dendrogram from exp 2 reveals a similar structure. Again using a cut point above “Facts” and “Difficulty” leads to similar clusters that include a core set of strategy probes converged upon across both experiments. These 5 strategy composites were carried forward for analysis of the functional MRI data. They are heuristically labeled as (I) Difficulty, (II) Autobiographical, (III) Scene Construction, (IV) Others-Relevant, and (V) Self-Relevant. Strategy probes that were not consistent between the two experiments (or weakly associated) were not included in the final composite scores, allowing only the most robust and stable strategy probes to be incorporated into the final 5 composite scores.

Figure 5.

Difficulty composite scores track trial-to-trial variation in response times supporting validity. Response time (RT) estimates provided an opportunity to validate subjective ratings of Difficulty. Mean RTs were calculated for each trial (y-axis) and plotted against the Difficulty composite scores (x-axis) from exp 2. The observed strong positive relation provides evidence for the validity of the Difficulty composite, even though it is based on participant self-report. The Pearson’s correlation value is shown in the bottom left. The line represents a linear model predicting Difficulty scores by RT across trials.

Figure 6.

Distributed networks estimated from functional connectivity within individuals: left hemisphere. Whole brain estimates of the 6 networks are displayed for each of the 10 extensively sampled individuals (identical to Ref. ; see also Ref. 26). For each individual, the k-means solution featuring the fewest clusters that differentiated the 6 target networks was chosen. Networks are shown in the left hemisphere and include default network A (DN-A, red), default network B (DN-B, pink), a language network (LANG, yellow), two candidate frontoparietal control networks (FPN-A, light blue; FPN-B, dark blue), and a candidate for a cingulo-opercular network (CING-OPER, green). These networks were defined independently of assessments of functional response properties. Brains in Figs. 6, 7, and 13 are shown at a slightly rotated angle (orientation: x = −0.93, y = 61.00, z = 89.64 in Connectome Workbench v1.3.2; Ref. 68).

Figure 7.

Distributed networks estimated from functional connectivity within individuals: right hemisphere. The networks from Fig. 6 are displayed for the right hemisphere, including default network A (DN-A, red), default network B (DN-B, pink), a language network (LANG, yellow), two candidate frontoparietal control networks (FPN-A, light blue; FPN-B, dark blue), and a candidate for the cingulo-opercular network (CING-OPER, green).

Figure 8.

Strategy composites are associated with differential and selective network activity. For each strategy composite, mean scores from all 180 trials were correlated with mean response estimates for each of the 6 networks. The strategy is labeled at the top of each plot; the six colored bars reflect the Pearson’s correlations with 95% confidence intervals. Scene Construction: a particularly striking and selective relation is observed between Scene Construction composite scores and DN-A activity. Difficulty: the Difficulty composite scores show a strong positive correlation to FPN-B and a strong but weaker relation to FPN-A. Others-Relevant: the Others-Relevant composite scores reveal a modest association with DN-B. Autobiographical and Self-Relevant: results between the Autobiographical and Self-Relevant composite scores are more ambiguous (partially related to confounding effects of Difficulty; see Fig. 12). Networks, from left to right: default network A (DN-A, red), default network B (DN-B, pink), a language network (LANG, yellow), two candidate frontoparietal control networks (FPN-A, light blue; FPN-B, dark blue), and a candidate for the cingulo-opercular network (CING-OPER, green).

Figure 9.

Scene construction is selectively related to DN-A but not DN-B activity. Scatter plots of individual trial activity levels within DN-A illustrate a strong relation to the Scene Construction composite. Top: the mean activity level for each of the 180 trials is plotted for DN-A (y-axis) against the mean Scene Construction composite scores (x-axis). Note that each separate point represents the mean behavioral score for that unique trial from 37–42 participants and the mean functional MRI response for that unique trial averaged across 10 participants. There is a striking linear relationship between the Scene Construction composite and DN-A activity. Bottom: the mean activity level for each trial is similarly plotted for the adjacent network DN-B against the same Scene Construction composite scores (x-axis). There is minimal relation. Pearson’s correlation values are shown in the bottom left corners. DN-A, default network A; DN-B, default network B.

Figure 10.

Trial-to-trial variation in Scene Construction tracks DN-A activity levels even for trials that do not involve episodic remembering or prospection. Scatter plots are displayed for DN-A split by whether the trials originated from the target conditions constructed to demand episodic projection (top) or were originally included within control conditions constructed to minimize such demands (bottom). The correlation with Scene Construction is present in both sets of trials with a strong and clear linear association within the original control trial conditions. Points are colored based on their condition origin: Past Self (pink), Future Self (red), Present Self (blue), Past Non-Self (dark green), and Future Non-Self (light green). Each point represents a single trial; Pearson’s correlation values are shown in the bottom left corners. Regression lines in color include only the trials from target (top) or control (bottom) conditions; black dashed regression lines include all 180 trials. Triangles indicate the mean composite score across trials in each plot, highlighting condition-level differences. Scene Construction tracks DN-A response both for trials originally constructed to target DN-A and for trials constructed to minimize demands on episodic projection. DN-A, default network A.

Figure 11.

Contrasting Scene Construction and Difficulty reveals a double-dissociation between DN-A and FPN-B. Scatter plots contrast the differential relations of DN-A and FPN-B with the Scene Construction and Difficulty composite scores. Top, left: the mean activity level for each of the 180 trials is plotted for DN-A (y-axis) against the mean Scene Construction composite scores (x-axis). Bottom, left: the mean activity level for each trial is plotted for FPN-B (y-axis) against the mean Scene Construction composite scores (x-axis). Note the absence of a relation. Top, right: the mean activity level for each trial is plotted for DN-A (y-axis) against the mean Difficulty composite scores (x-axis). Bottom, right: the mean activity level for each trial is plotted for FPN-B (y-axis) against the mean Difficulty composite scores (x-axis) revealing a strong, positive relation. Pearson’s correlation values are shown in the bottom left corners. Scene Construction scores track DN-A activity and Difficulty scores track FPN-B activity, but not vice versa, illustrating a functional double-dissociation between these two closely juxtaposed networks. DN-A, default network A; FPN-B, frontoparietal control network B.

Figure 12.

Strategy composites are associated with differential and selective network activity after regression of Difficulty. Given the possibility of a confounding effect of Difficulty on estimates of network selectivity, the network correlation bar plots were recomputed after regressing the Difficulty composite scores (see text). Scene Construction: Scene Construction composite scores maintained a strong and selective correlation to DN-A activity. Difficulty: as mandated by the analysis, variation related to the Difficulty Composite scores was removed. Since correlations could not be estimated to a vector of zeros, the original Difficulty plot (from Fig. 8, faded here) is shown for reference. Others-Relevant: the Others-Relevant composite scores maintained a weaker but selective relation to DN-B activity. Autobiographical: of interest, the Autobiographical composite scores, once Difficulty was regressed, revealed a pattern similar but weaker to that of Scene Construction. Self-Relevant: the Self-Relevant composite scores were nonspecific. Networks, from left to right: default network A (DN-A, red), default network B (DN-B, pink), a language network (LANG, yellow), two candidate frontoparietal control networks (FPN-A, light blue; FPN-B, dark blue), and the cingulo-opercular network (CING-OPER, green).

Figure 13.

Post hoc contrasts of extreme trials can recapitulate networks within the individual. As a confirmation of the discovery that Scene Construction and Difficulty are associated with functional MRI response levels in distinct networks, within-individual contrast maps are displayed. Left: for 5 selected individuals, maps illustrate the contrast of the 10 trials with the lowest Scene Construction composite scores versus the 10 trials with the highest scores, selected only from the original control conditions. Note that even with this relatively small amount of data per individual (from trial conditions originally selected to minimize demands on episodic projection), functional MRI differences emerge across the distributed regions that comprise DN-A. The black outlines show the independently estimated boundaries of DN-A for each participant. Right: maps illustrate the contrast of the 10 trials with the highest Difficulty composite scores versus the 10 trials with the lowest scores. Differences emerge that fall within FPN-B. These contrasts were not possible in initial condition-level analyses (26) and provide evidence of novel processing insights. Maps are shown for the right hemisphere. DN-A, default network A; FPN-B, frontoparietal control network B.

Figure 14.

Further exploration of Autobiographical and Scene Construction scores in relation to DN-A response. Following regression of Difficulty, the trial-to-trial correlation between Autobiographical and Scene Construction scores strengthened (r = 0.58, top). Even for trials originally treated as controls (from Present Self and Past and Future Non-Self conditions), this correlation remained strong (r = 0.49, middle). But among control trials, subsets of questions were identified that showed higher scores either on the Autobiographical (blue) or the Scene Construction composite (red). Plotting the DN-A response values for each trial in these subsets (bottom) revealed higher DN-A activity during trials in the Scene Construction subset. Mean DN-A response in each trial subset is shown by a diamond. DN-A, default network A.

Figure 15.

Individual strategy probes for scene construction processes show the strongest relation to DN-A response. For each individual strategy probe from the RSS, trial-level correlations to each network’s response pattern are plotted. Gray circles show correlations prior to Difficulty regression and black circles postregression. Dotted lines demarcate composites (see Fig. 4). The strategy probe for Locations of People (Loc_People, marked with a *) is grouped with the Scene Construction probes for visualization. DN-A (top) shows high correlations to Scene Construction probes, as well as to the Sequence_Events probe from the Autobiographical composite. DN-A is more weakly correlated to the Pers_Past_Exper probe. This pattern supports DN-A’s role in mental construction of scenes and events. Patterns across other networks show strong correlations between the FPNs and Difficulty probes (in gray, prior to regression) and a weaker but unique relation between DN-B and Others-Relevant strategies, which survives Difficulty regression. The Difficulty composite comprised Facts and Difficulty strategies, so only preregression correlations to these strategies are shown (in gray). CING-OPER, cingulo-opercular network; DN-A, default network A; DN-B, default network B; FPNs, frontoparietal control networks; LANG, language network; RSS, response strategies scale.