Dynamic frontotemporal systems process space and time in working memory

Abstract

How do we rapidly process incoming streams of information in working memory, a cognitive mechanism central to human behavior? Dominant views of working memory focus on the prefrontal cortex (PFC), but human hippocampal recordings provide a neurophysiological signature distinct from the PFC. Are these regions independent, or do they interact in the service of working memory? We addressed this core issue in behavior by recording directly from frontotemporal sites in humans performing a visuospatial working memory task that operationalizes the types of identity and spatiotemporal information we encounter every day. Theta band oscillations drove bidirectional interactions between the PFC and medial temporal lobe (MTL; including the hippocampus). MTL theta oscillations directed the PFC preferentially during the processing of spatiotemporal information, while PFC theta oscillations directed the MTL for all types of information being processed in working memory. These findings reveal an MTL theta mechanism for processing space and time and a domain-general PFC theta mechanism, providing evidence that rapid, dynamic MTL-PFC interactions underlie working memory for everyday experiences.

Fig 1. Working memory task design and electrode coverage.

(A) Single-trial working memory task design. Following a 1-s pretrial fixation interval (−250 to −50 ms pretrial epoch used as baseline), subjects were directed to focus on either “identity” or “relation” information. Then, 2 common shapes were presented for 200 ms each in a specific spatiotemporal configuration (i.e., top/bottom spatial and first/second temporal positions). After a 900- or 1,150-ms jittered precue fixation delay, the test cue appeared (i.e., one word presented on screen for 800 ms), followed by a postcue fixation delay of the same length. Working memory was tested in a two-alternative forced choice test (0.5 chance rate). In the identity test (top), subjects indicated whether the pair was the “same” pair they just studied (correct response in this example: “no”). In the spatiotemporal relation test (bottom), subjects indicated which shape fit the top/bottom spatial or first/second temporal relation cue (correct response for cue “top” or “second”: “circle”). (B) Reconstruction of electrode coverage for all subjects. Electrodes are overlaid on the left hemisphere, displayed in 4 views. Red = MTL; green = PFC; blue = OFC. Underlying data can be found in University of California, Berkeley, Collaborative Research in Computational Neuroscience database (http://dx.doi.org/10.6080/K0VX0DQD). MTL, medial temporal lobe; OFC, orbitofrontal cortex; PFC, prefrontal cortex.

Fig 2. Task-induced power at encoding and delay.

(A) Task-induced power over encoding, precue, and postcue intervals in representative MTL electrodes from 2 Ss. The MTL showed sustained theta band (3–7 Hz) activity, narrowband alpha desynchronization, and variable, transient activities above 16 Hz (|z| > 1.96, p < 0.05). Anatomy, S2: CA1 (hippocampus); S3: perirhinal cortex. (B) Equivalent to panel A: the PFC showed sustained theta band activity and variable, transient activities above 9.5 Hz (|z| > 1.96, p < 0.05). Anatomy, S2 and S3: middle/superior frontal gyrus (dorsolateral PFC). (C) Equivalent to panel A: the OFC showed sustained theta band activity, periodic narrowband alpha, and variable, transient activities above 16 Hz (|z| > 1.96, p < 0.05). Anatomy, S2 and S3: medial OFC. Underlying data can be found in University of California, Berkeley, Collaborative Research in Computational Neuroscience database (http://dx.doi.org/10.6080/K0VX0DQD). MTL, medial temporal lobe; OFC, orbitofrontal cortex; PFC, prefrontal cortex; S, subject.

Fig 3. Theta oscillations link frontotemporal regions during working memory.

(A) PLVs over pre- and postcue delays within all MTL electrodes (top) and between all MTL and PFC (middle) and OFC (bottom) electrodes in a representative S. Theta and alpha peaks were observed within the MTL and between MTL and frontal regions over delay. Data are represented as the temporal mean of each delay interval, normalized by the maximum PLV across all electrodes. The shaded area indicates the theta range used for PSI analysis, as depicted in panel B. (B) Theta band PSI shifted from a unidirectional, frontal-driven network during the precue delay to a bidirectional, MTL-frontal network during the postcue delay (p < 4 × 10⁻²⁷). Data are represented as mean ± SEM per S; positive values indicate that the MTL leads the PFC (top) or OFC (bottom), and negative values indicate that the PFC/OFC leads the MTL. * = significant effect. Underlying data can be found in University of California, Berkeley, Collaborative Research in Computational Neuroscience database (http://dx.doi.org/10.6080/K0VX0DQD). MTL, medial temporal lobe; OFC, orbitofrontal cortex; PFC, prefrontal cortex; PLV, phase-locking value; POST, postcue delay; PRE, precue delay; PSI, Phase Slope Index; S, subject.

Fig 4. Theta oscillations direct cross-spectral activity during information processing.

(A) Theta phase–amplitude directionality shifted from pre- to postcue delay both locally and across regions (p < 7 × 10⁻²⁶). Data are represented as mean ± SEM per S; positive values indicate that phase leads amplitude, and negative values indicate that amplitude leads phase. * = significant effect. (B) Percentage of individual electrodes and electrode pairs in which theta phase directed higher-frequency amplitudes during the postcue delay. Data are represented as mean ± SEM. (C) Schematic of theta phase–led bidirectional frontotemporal interactions during the postcue delay. Underlying data can be found in University of California, Berkeley, Collaborative Research in Computational Neuroscience database (http://dx.doi.org/10.6080/K0VX0DQD). AMP, amplitude; MTL, medial temporal lobe; OFC, orbitofrontal cortex; PFC, prefrontal cortex; POST, postcue delay; PRE, precue delay; S, subject.

Fig 5. MTL theta PAC dynamics track space.

(A) PAC by condition during the postcue delay in a representative MTL electrode. The MTL showed variable, transient PAC across the spectrum of amplitudes (z > 1.96, p < 0.05). PAC was greater for spatial than identity information, with peak differences at 22.5- and 64-Hz amplitudes (p < 3 × 10⁻⁵). The black block indicates the amplitude data range depicted in panel B. ** = significant condition and condition × frequency effects. (B) The distribution of raw higher-frequency amplitudes across 18 theta phase bins, by condition, normalized by the maximum amplitude across all of the phase bins. AMP frequency range: broadband gamma (centered at 64 Hz). Underlying data can be found in University of California, Berkeley, Collaborative Research in Computational Neuroscience database (http://dx.doi.org/10.6080/K0VX0DQD). AMP, amplitude; DIFF, difference (i.e., spatial–identity, temporal–identity); FREQ, frequency; MTL, medial temporal lobe; NORM, normalized; PAC, phase-amplitude coupling; RAD, radians.

Fig 6. MTL→PFC theta PAC dynamics track both space and time.

(A) PAC by condition during the postcue delay in a representative MTL–PFC electrode pair. MTL→PFC (top) and PFC→MTL (bottom) pairs showed variable, transient PAC across the spectrum of amplitudes (z > 1.96, p < 0.05). MTL→PFC PAC was greater for spatial than identity information, with peak differences at 13.5- to 16-Hz amplitudes during the 100- to 250- and 800- to 900-ms epochs (p < 2 × 10⁻⁴). MTL→PFC PAC was also greater for temporal than identity information (p < 9 × 10⁻⁵). No condition differences were observed in the PFC→MTL direction. The black block indicates the amplitude data range depicted in panel B. * = significant condition effect; *** = significant condition, condition × frequency, condition × time, and 3-way interaction effects. (B) The distribution of raw higher-frequency amplitudes across 18 theta phase bins, by condition, normalized by the maximum amplitude across all of the phase bins. Underlying data can be found in University of California, Berkeley, Collaborative Research in Computational Neuroscience database (http://dx.doi.org/10.6080/K0VX0DQD). AMP frequency range: beta (centered at 13.5 Hz). AMP, amplitude; MTL, medial temporal lobe; PAC, phase-amplitude coupling; PFC, prefrontal cortex; RAD, radians.

Fig 7. MTL→OFC theta PAC dynamics track space but not time.

(A) PAC by condition during the postcue delay in a representative MTL–OFC electrode pair. MTL→OFC (top) and OFC→MTL (bottom) pairs showed variable, transient PAC across the spectrum of amplitudes (z > 1.96, p < 0.05). MTL→OFC PAC was greater for spatial than identity information, with peak differences at 9.5-Hz amplitudes (p < 5 × 10⁻³). No condition differences were observed in the OFC→MTL direction. The black block indicates the amplitude data range depicted in panel B. ** = significant condition and condition × frequency effects. (B) The distribution of raw higher-frequency amplitudes across 18 theta phase bins, by condition, normalized by the maximum amplitude across all of the phase bins. AMP frequency range: alpha (centered at 9.5 Hz). Underlying data can be found in University of California, Berkeley, Collaborative Research in Computational Neuroscience database (http://dx.doi.org/10.6080/K0VX0DQD). AMP, amplitude; MTL, medial temporal lobe; OFC, orbitofrontal cortex; PAC, phase-amplitude coupling; RAD, radians.

Fig 8. Bidirectional MTL–PFC theta networks for working memory.

(A) PAC grand means by condition and direction for MTL–PFC (top) and MTL–OFC (bottom) networks. Condition moderated PAC in the MTL→PFC direction so that MTL→PFC PAC was greater for spatiotemporal than identity information (p < 2 × 10⁻⁵), while PFC→MTL PAC was greater than MTL→PFC PAC overall (p < 2 × 10⁻⁹), revealing bidirectional PAC for processing space and time. In contrast, MTL→OFC PAC was greater than OFC→MTL PAC, which was moderated by condition so that MTL→OFC PAC was greatest for spatial information. Data are represented as mean ± SEM. ** = significant direction and condition × direction effects. (B) Schematics of spatial (left) and temporal (right) information processing. The bidirectional MTL–PFC network subserves spatial and temporal information processing, while a relatively unidirectional MTL→OFC network is also involved in spatial information processing. Underlying data can be found in University of California, Berkeley, Collaborative Research in Computational Neuroscience database (http://dx.doi.org/10.6080/K0VX0DQD). MTL, medial temporal lobe; OFC, orbitofrontal cortex; PAC, phase-amplitude coupling; PFC, prefrontal cortex.