Frontoparietal Correlation Dynamics Reveal Interplay between Integration and Segregation during Visual Working Memory

Frontoparietal correlations during working memory

To determine the patterns of functional connectivity across the frontoparietal network, we used cross-correlation analysis to estimate the incidence, magnitude, and relative phase angle between all combinations of simultaneously recorded LFPs during the delay period of the task. We found significant correlations in 60% and 48% of the signal pairs in Monkey A and Monkey B, respectively, resulting in 1644 significantly correlated signal pairs from 63 different intra-areal and interareal combinations (Table 2; for selection criteria, see Materials and Methods). Figure 4 shows four examples of the signal correlations we observed in the data from Figure 1c. During the task, the correlation magnitude peaks before the sample, transiently declines, and then increases again during the delay (Fig. 4a–d). These magnitude changes are accompanied by subtle shifts in the relative phase angle, especially during the response to the sample when the correlation magnitude is at its minimum. Notably, the correlation magnitudes and phase relationships are related to the cortical locations of the recording sites. In these examples, the PEC-PEC signal pair, recorded medial to the intraparietal sulcus (IPS), exhibits a strong, in-phase correlation (Fig. 4a,e), whereas the PG-PEC signal pair, recorded from opposite sides of the IPS, displays a weaker, but nearly anti-phase, correlation (Fig. 4b,f). Similar phase relationships are present in the long-distance frontoparietal signal pairs. The dPFC-PEC (Fig. 4c,g) and dPFC-PG (Fig. 4d,h) signal pairs display weaker correlations while maintaining a striking difference in relative phase.

Analysis of the incidence and magnitude of correlation with respect to spatial separation revealed distinct patterns of functional connectivity. We evaluated these effects by grouping signal pairs into six categories: (1) within PFC, (2) PPC signals medial to the IPS, (3) PPC signals lateral to the IPS, (4) PPC signals on opposite sides of the IPS, (5) PFC and medial PPC pairs, and (6) PFC and lateral PPC pairs. As shown in Table 3, the incidence and magnitude of correlation were highest among signal pairs recorded on the same side of the IPS, somewhat lower within PFC, and markedly lower for signal pairs separated by larger distances (i.e., mediolateral parietal, frontoparietal). To assess the effect of interelectrode separation on correlation magnitude within PFC and PPC, we plotted the correlation magnitude versus distance for all pairs within PFC, and together for all pairs on the same side of the IPS. This revealed a fall off of correlation magnitude as a function of distance for electrode separations <6 mm (Fig. 5). Thus, the magnitude of correlation decreases with distance within and between adjacent areas of the PFC and PPC (Murthy and Fetz, 1996; Destexhe et al., 1999), whereas both the incidence and magnitude of correlation plateau to lower levels over distances that span widely separated areas of cortex (Roelfsema et al., 1997).

Consistent with the results in Figure 4, we found a clear bimodal distribution of relative phase values among significantly correlated pairs in both monkeys (Fig. 6). With all data combined, the distributions of relative phases were split into approximately in-phase (near 0°) and anti-phase (near 180°) modes (Fig. 6a,e). To identify the source of these distinct modes, we separated the data into PFC, PPC, and frontoparietal categories. Within PFC, the vast majority of phase values are centered near 0° in both monkeys (Fig. 6b,f). Within PPC, the relative phase distributions are distinctly bimodal in both monkeys, with the dominant mode centered near 0° and the secondary mode centered near 180° (Fig. 6c,g). These bimodal phase distributions also exist in the frontoparietal correlations in both animals (Fig. 6d,h): The dominant mode, centered near −16° in both monkeys, displays a prefrontal phase lead, as previously reported (Salazar et al., 2012), whereas the secondary mode is more broadly distributed and centered at −180° and −148° in Monkey A and Monkey B, respectively (Table 4).

To gain further insight into these processes, we subdivided the correlation analysis with respect to each intra-areal and interareal combination of signals. The magnitude, relative phase, and frequency values of the correlations are shown in Figure 7. Because the sign of the relative phase angle can be ambiguous, we report the absolute value of the relative phase. The results reveal that the patterns of correlation are anatomically specific. Within PFC, the correlations vary widely in magnitude with respect to areal combination (and thus distance) and are dominated by in-phase relationships, with a small percentage of values scattered ∼100° in Monkey B (Fig. 7a). Within PPC, signal pairs recorded on either the medial (Fig. 7b) or the lateral (Fig. 7c) side of the IPS have a broad distribution of correlation magnitudes and are highly clustered in-phase, with only one value exceeding 90°. In contrast, the long-distance correlations between signals recorded on the medial and lateral sides of the IPS, and between these areas and the PFC, are smaller in magnitude and exhibit a clear bimodal distribution of phase values (Fig. 7d–f). Signal pairs recorded on opposite sides of the IPS are dominated, with some exceptions, by near anti-phase relationships (Fig. 7d). Signal pairs recorded in PFC and the medial side of the IPS (Fig. 7e) are dominated by in-phase relationships, with the exception of signals sampled from areas 8B and MIP in Monkey B that are almost entirely out-of-phase. In contrast, signal pairs recorded in PFC and the lateral side of the IPS (Fig. 7f) exhibit both unimodal and bimodal phase distributions that are anatomically specific. In Monkey A, all signal pairs, except one, have phase values >90°. In Monkey B, correlations between prefrontal and lateral parietal signals are either exclusively out-of-phase (8B-LIP), exclusively in-phase (8AD-PG), or bimodally distributed (8B-PG, dPFC-PG, dPFC-LIP). Figure 7 (bottom row of plots) shows the corresponding frequencies of correlation derived from the fitted functions. The frequencies vary within the band of interest, display some subtle difference with respect to areal combination, and differ between the two animals. Moreover, there is no apparent difference in correlation magnitudes between the two animals (also see Fig. 5), even though the dominant frequencies are different. Clear examples of this can be seen by examining the distributions for the dPFC-PEC and PEC-PEC pairs. These subtle differences in frequency may provide another means for balancing integration and segregation.

In addition to the abundance of correlated interactions, we also found a number of interareal pairings with significant correlation occurring ≤10% of the time (Fig. 8; Table 2). Although this effect might be related to small sample sizes (e.g., 6DR-PG) and thus reflect an experimental bias, other interareal combinations have ≥50 signal pairs and show a very low probability of correlation. For example, none of the cross-correlograms computed from the 60 MIP-LIP signal pairs in Monkey B reached significance. This indicates that these and other areas are functioning largely independently under the conditions of this experiment.

To further illustrate the specificity of these interactions, Figure 9 shows the correlation patterns obtained from a single recording session in Monkey B. Within PFC, the signal correlations have values <0.5 and, with one exception, display relative phase values <45° (Fig. 9c). The intraparietal signals exhibit stronger correlations having relative phase values <45° when the signals are recorded on the same sides of the IPS, and weaker correlations with relative phases >135° for recordings on opposite sides of the IPS (Fig. 9d). The frontoparietal correlations are weaker and display relative phase values that are clustered into two groups (Fig. 9e): those <45° and those >135°. This segregation of phases, as in Figure 4, displays some anatomical boundaries. Signals in area 8B exhibit near anti-phase relationships with signals recorded in areas LIP and MIP, even though the signals in LIP and MIP are statistically independent (Fig. 8). The remaining prefrontal signals in this example in areas 8AD, dPFC, and vPFC have in-phase relationships with parietal signals recorded in areas that are medial to the IPS but are statistically independent from signals recorded lateral to the IPS, even though significant anti-phase correlations occur between signals that span the IPS. These patterns of correlation suggest that in-phase and anti-phase relationships are established at different times among different sites in the network.

Task-dependent correlation patterns

Next, we examined the task dependence of frontoparietal correlations by comparing the presample period (0.1–0.5 s) with a window of the same size in the delay period that is locked to the first possible match (1.4–1.8 s). Signal pairs were selected using the same criteria, but for both task periods, resulting in a new set of signal pairs (Table 2). We tested whether the correlation increased or decreased between the presample and delay periods for each areal combination (sign test, paired, two-sided) and plotted the difference distributions (minimum of 5 signal pairs) pooled across both animals (Fig. 10). Several trends emerged from this analysis. Within PFC, a small, but significant, increase in correlation occurred during the delay for areal combinations 8B-8B, 8B-8AD, and 8B-dPFC. Within PPC, correlations displayed a distinct pattern of changes: areas medial to the IPS displayed a robust increase in correlation during the delay (Fig. 10b), those lateral to the IPS showed no change (Fig. 10c), and areas separated by the IPS tended to decrease during the delay, but only significantly for PE-LIP (Fig. 10d). A related pattern of changes was apparent in the frontoparietal correlations. Significant increases occurred between dPFC and medial parietal areas PEC, PE, and MIP (Fig. 10e) and between prefrontal areas dPFC and 8B and lateral parietal areas PG and LIP (Fig. 10f). In addition, we found substantial task-dependent changes in relative phase (Fig. 10, bottom plots). Shifts in relative phase between the two task periods ranged from 0° to 45°, with the largest shifts occurring most often between frontoparietal areal combinations (Fig. 10e,f).

To further explore the potential for these activity patterns to reorganize, we compared the delay period of the task (1.4–1.8 s) to the ITI shortly after the match period on error trials in Monkey B (For technical reasons, the ITI period was only recorded after incorrect trials in Monkey B; for further explanation, see Materials and Methods.) During the ITI of error trials, we observed a strong oscillatory response, visible in the LFP, which showed a marked change in relative phase among the signals (Fig. 11a). Analysis of the peak amplitude of these oscillatory bursts revealed that they occurred earlier in PPC (671 ± 96 ms, after match onset) than PFC (769 ± 137 ms, after match onset) (Wilcoxon rank sum test, p ≪ 0.001). Using this timing information, we centered the ITI analysis epoch over the peak amplitudes of the bursts (0.5–0.9 s after match onset; see Materials and Methods). Signal pairs were once again subjected to the fitting criteria for both task periods, resulting in a slightly different set of pairs (Table 2). Figure 11b, c shows the time course of the trial-averaged cross-correlation for two of the signal pairs from Figure 11a. In these examples, the relative phase is near 180° for the duration of the task and then abruptly flips to an in-phase relationship during the ITI. This effect, accompanied by an increase in correlation strength, occurs between PFC and PPC signals (Fig. 11b) and between the parietal signals on opposite sides of the IPS (Fig. 11c), demonstrating its widespread nature. Figure 11d shows the third frontoparietal pair, which shows a weak in-phase correlation before the sample, is largely uncorrelated during the delay, and then is highly correlated during the ITI, with no change in relative phase.

We quantified these task-dependent changes in the full dataset using the same method shown in Figure 10. Figure 12a–f (top plots) shows the difference distributions in correlation magnitude for each areal combination (outliers not shown). In general, correlation strength increased during the ITI, but the range of change was much larger than that seen during the task, indicating widespread activation and, with few exceptions, increased synchronization across the frontoparietal network. Correlation strength increased within and between several combinations of prefrontal areas 8AD, 8B, dPFC, and vPFC (Fig. 12a), among parietal areas medial to the IPS (Fig. 12b), and somewhat heterogeneously among the other prefrontal and parietal areas (Fig. 12c–f). We also found a striking change in relative phase relationships that occurred predominantly among signals separated by large distances (Fig. 12a–f, bottom plots). Whereas the changes in phase during the task were all <50° and often much closer to 0° (Fig. 10), during the ITI following error trials we found numerous changes in phase exceeding 90°. Also, changes in correlation strength and relative phase were not obligatorily coupled. We found multiple areal combinations where correlation strength changed while relative phase remained stable, and vice versa.

A potential concern is that the patterns of correlation during the delay period may differ substantially between correct and incorrect trials, leading to the effects reported in Figures 11 and 12. We examined this possibility and found no significant difference in correlation magnitude across the population of areal combinations (paired sign test, p = 0.18), and the mean difference in relative phase angle was <4.5°.

Methodological considerations

Because we used monopolar recording methods, volume-conducted signals may have contributed to our estimates of correlation. However, several factors suggest that volume conduction effects are minimal. First, intracortical microelectrodes are known to record highly local LFPs (Gray and Singer, 1989; Katzner et al., 2009; Wang and Roe, 2012; Xing et al., 2012) that exhibit minimal volume conduction (Engel et al., 1990). Thus, given that most of our recordings were separated by >2 mm, we expect volume conduction effects to be small. Second, if volume conduction were contributing significantly to our findings, we would not expect to find task- and time-dependent phase differences between sites (Salazar et al., 2012). Third, the presence of anti-phase correlations and the absence of correlations among some signals spanning the IPS indicate that even closely spaced electrodes, lying in separate cortical areas, record distinct neural signals. Together, this evidence argues against a significant contribution of volume conduction at all spatial scales measured.

It could be argued that the phase differences we observe depend on the cortical depth of our recordings. Cortical LFP signals exhibit phase differences, and polarity reversals, as a function of depth (Mitzdorf and Singer, 1978; Murthy and Fetz, 1996; Bollimunta et al., 2008) that could lead to dispersion of our measured phase values. Several of our findings, however, indicate that sampling bias across a cortical dipole field cannot account for the anti-phase relationships we observe. For example, in order for laminar sampling bias to account for the anti-phase correlations between areas PE and LIP (Fig. 7d), and the exclusively in-phase correlations within areas PE (Fig. 7b) and LIP (Fig. 7c), it would require that signal measurements were made on one side of the dipole (i.e., superficial layers) in one area and the opposite side of the dipole (i.e., deep layers) in the other area. Similar arguments apply to all the paired recordings made on the same side of the IPS (Fig. 7b,c). Those phase distributions are somewhat dispersed, but there is no evidence of anti-phase relationships as might be expected from a dipole field. Furthermore, any doubt that the anti-phase relationships are functional, and not due to a sampling bias across a cortical dipole, is dispelled by the rapid and pronounced changes in relative phase following error trials (Figs. 11 and 12).

Functional role of frontoparietal phase relationships

How then should we interpret the bimodal distribution of intraparietal and frontoparietal phase relationships in Figures 6 and 7? The simplest interpretation is that during the task two large-scale frontoparietal networks are established: one involving in-phase coupling (|θ| < 90°) between medial parietal and prefrontal areas, and the other involving anti-phase coupling (|θ| > 90°) between lateral parietal and prefrontal areas. However, this interpretation may be too simplistic. For instance, signals in areas 8B and MIP display largely anti-phase correlations, whereas signals in areas PG and LIP display differing combinations of in-phase and anti-phase correlations with prefrontal areas 8AD, 8B, and dPFC. There are also a small percentage of intraprefrontal correlations with relative phase values near or >90°. These data suggest that the phase of correlated activity reflects specific patterns of coupling among many of the prefrontal and parietal areas with a strong bias for anti-phase coupling across the intraparietal sulcus.

The question also arises as to how to interpret the specific relative phase values. Presumably in-phase and anti-phase coupling have differing functional consequences. However, given the large distances separating the recorded sites, with transmission delays ranging from ∼5 to 10 ms, there is no simple functional interpretation. One possibility is that the anti-phase correlations occurring across the IPS, and between certain frontoparietal areal combinations, as well as the low incidence of correlation between certain areal combinations, such as MIP and LIP, reflect functional segregation that is imposed by the oculomotor and visual mnemonic nature of the task. This seems likely, given the differing functional properties of neurons in these various prefrontal (Petrides, 2005; Lanzilotto et al., 2013) and posterior parietal (Breveglieri et al., 2008; Andersen and Cui 2009) cortical areas.

Other aspects of the frontoparietal phase relationships require explanation. The data in Figure 4 suggest that in-phase and anti-phase coupling between PFC and medial and lateral parietal areas occur simultaneously, whereas the results in Figure 9 suggest some degree of independence in the frontoparietal networks. However, the interpretation of these findings must be gauged by the fact that the median value of the long-distance correlation coefficients is <0.2. This indicates that most of the time the signals in widely separated areas of the frontoparietal network are independent and that coupling between distant areas is likely to be highly transient (Gray et al., 1992). Thus, analysis of the temporal properties of the correlations on single trials will be necessary to reveal the underlying spatiotemporal organization.

Our previous findings that frontoparietal coherence in the 12–22 Hz frequency range is content specific (Salazar et al., 2012) imply that the corresponding phase relationships may also contribute to short-term representations. Determining if and how this occurs will require a more extensive analysis than we have conducted here. The relative phase and frequency of coherent activity may vary systematically with different aspects of working memory, such as content, load, and task dependence, thereby adding further dimensions to the representations. Another possibility, discussed above, may be that the bimodal distribution of relative phases provides a mechanism for task-dependent segregation of activity into distinct networks, thereby reducing potential interference among competing representations. This could be tested by switching from a saccade-based task to one in which choices are based on varying combinations of reach and saccade. In such an experiment, the phase relationships between different components of the frontoparietal network may change as the task switches from choices based on saccades to reach to a combination of the two (Dean et al., 2012).

Our finding of abrupt and marked changes in frontoparietal correlations following error trials (Figs. 11 and 12) suggests a link to the cognitive processing of errors (Larson et al., 2014). The widely distributed and highly synchronized bursts of activity may correspond to the well-documented error-related negativity (Falkenstein et al., 1991; Gehring et al., 1993) and reflect processes related to reinforcement learning (Holroyd and Coles, 2002) mediated by midbrain dopamine systems (Schultz et al., 1997). However, the present data are limited, and further experiments are required to test these hypotheses.

Coordination in large-scale networks

The phase of rhythmic correlation has been postulated to be a critical variable controlling the selective transmission, or communication, of signals between separate neuronal populations (Fries, 2005; Womelsdorf et al., 2007). Communication between neuronal groups is thought to be most effective when they oscillate at the same frequency and are synchronized at relative phases near 0°. This model is consistent with our understanding of synaptic integration (Abeles, 1982; König et al., 1995; Azouz and Gray, 2000, 2003) and is supported by experimental evidence (Engel et al., 1991a, b; Roelfsema et al., 1997; Salazar et al., 2004). Other studies have found substantial phase differences between distant neuronal populations (Gregoriou et al., 2009; Liebe et al., 2012) and suggested that these effects reflect directional influences mediated through relatively fixed conduction delays. However, neither of these perspectives is fully consistent with our findings. We find robust correlations in a common frequency range over a wide span of spatial separations that exhibit a bimodal distribution of relative phases. Moreover, the phase relations are not fixed. They vary during the task and exhibit large transitions during the intertrial interval following error trials. A potentially more favorable interpretation of these findings is that these dynamics are a product of global cooperation across the brain and that changes in correlation magnitude and phase reflect shifts in global coordination states that are not predictable on the basis of expected transmission delays (Bressler and Kelso, 2001; Tognoli and Kelso, 2014).