Unmasking Latent Inhibitory Connections in Human Cortex to Reveal Dormant Cortical Memories

Abstract

Balance of cortical excitation and inhibition (EI) is thought to be disrupted in several neuropsychiatric conditions, yet it is not clear how it is maintained in the healthy human brain. When EI balance is disturbed during learning and memory in animal models, it can be restabilized via formation of inhibitory replicas of newly formed excitatory connections. Here we assess evidence for such selective inhibitory rebalancing in humans. Using fMRI repetition suppression we measure newly formed cortical associations in the human brain. We show that expression of these associations reduces over time despite persistence in behavior, consistent with inhibitory rebalancing. To test this, we modulated excitation/inhibition balance with transcranial direct current stimulation (tDCS). Using ultra-high-field (7T) MRI and spectroscopy, we show that reducing GABA allows cortical associations to be re-expressed. This suggests that in humans associative memories are stored in balanced excitatory-inhibitory ensembles that lie dormant unless latent inhibitory connections are unmasked.

Figure 1

Indexing Cortical Associations in the Human Brain using Cross-stimulus Adaptation Immediately after Learning (A) Left: stimuli are associatively paired: A-B and C-D. Middle and right: due to repetition suppression, the predicted BOLD response to activation of associated but different stimuli, A followed by B, was reduced relative to consecutive unrelated stimuli, A followed by C. (B) Before entering the scanner, participants learned to associate pairs of stimuli using a three-alternative forced-choice task. On each trial, in response to a test shape, the participant had to select the associated stimulus from the full set. (C) During scanning, two stimuli were presented in short succession on each trial. (D) Using the task shown in (B), one set of participants learned to pair colored shapes (experiment 1), A with B and C with D. (E) Using the stimuli shown in (D), the BOLD response to consecutive presentation of two unrelated stimuli (AC, A followed by C) was contrasted against the BOLD response to consecutive presentation of two associated stimuli (AB, A followed by B): “unrelated” minus “associated,” and the contrast thresholded at p < 0.05 uncorrected for display purposes. (F) Parameter estimates (mean ± SEM) were extracted from an orthogonal ROI (see Table S1) in occipital and temporal cortices, for trials where stimuli were associated (AB, A followed by B) and trials where stimuli were unrelated (AC, A followed by C). The difference in parameter estimates for these two trial types (AC-AB, shown on the right) gave a significant cross-stimulus adaptation effect within this ROI (p = 0.043). (G) A second set of participants learned to associate rotationally invariant gray shapes (experiment 2), pairing A with B and C with D. (H) Using the stimuli shown in (G), the BOLD response to consecutive presentation of two unrelated stimuli (AC, A followed by C) was contrasted against the BOLD response to consecutive presentation of two associated stimuli (AB, A followed by B): “unrelated” minus “associated,” and the contrast thresholded at p < 0.05 uncorrected for display purposes. (I) Parameter estimates (mean ± SEM) were extracted from an orthogonal ROI (see Table S1) in right temporal cortex, for trials where stimuli were associated (AB, A followed by B) and trials where stimuli were unrelated (AC, A followed by C). The difference in parameter estimates for these two trial types (AC-AB, shown on the right) gave a significant cross-stimulus adaptation effect within this ROI (p = 0.024). (J) Cross-stimulus adaptation can be observed across cortex, in the anatomical regions that encode features specific to the associated stimuli. Blue region: colored shape associations as shown in (E). Green region: rotationally invariant stimulus associations as shown in (H).Purple region: stimuli associated with food reward (p = 0.032 within ROI). Pink region: associated imaginary food reward (p = 0.014 within ROI, see also Figure 4C of Barron et al., 2013).

Figure 2

Cortical Associative Memories Are Silenced with Time (A) Left: example stimuli that were associatively paired: A-B and C-D. Middle and right: after inhibitory rebalancing had occurred, cross-stimulus adaptation between associated stimuli, A followed by B, was no longer predicted in the BOLD response as new inhibitory connections quench excitatory coactivation. Therefore activation of associated but different stimuli, A followed by B, was expected to be equivalent to activation of consecutive unrelated stimuli, A followed by C. (B) One set of participants (experiment 1) were scanned on a second occasion 24 hr after the initial scan and a significant reduction in cross-stimulus adaptation (measured with “associated” minus “not”) was observed across days (p = 0.045) (shown: mean ± SEM for each day).

Figure 3

The Latent Cortical Associations Are Uncovered in the Human Brain via Local Modulation of GABA (A) Following downregulation of cortical GABA, cross-stimulus adaptation between associated stimuli, A followed by B, was once again predicted in the BOLD response relative to the control condition A followed by C. (B) Rotationally invariant shapes were used as the stimuli for the associative learning task (as in Figure 1G). (C) The protocol used to test for evidence of inhibitory rebalancing of cortical associations in the human brain. Participants completed the associative learning task shown in Figure 1B, before completing two fMRI task blocks. Returning 24 hr later, the fMRI task was repeated in conjunction with MRS and tDCS. The first fMRI task block was followed by a baseline MRS measurement. Twenty minutes of tDCS commenced, and a “during tDCS” MRS measurement simultaneously acquired. The second fMRI task block started half way through the tDCS session, followed by a final “post-task” MRS measurement. After exiting the scanner, participants were given a surprise memory test to check they still knew the paired associations. (D) The mean tDCS electrode location, with x-coordinate defined using the peak x-coordinates from Figure 1H. (E) By comparing MRS measurements acquired before and during tDCS (shown: mean ± SEM), a significant reduction in GABA concentration was observed (“baseline” stimulation minus “during” stimulation, p = 0.006). (F) B1 corresponds to block 1, and B2 to block 2. Parameter estimates were extracted to obtain a measure of cross-stimulus adaptation for each scanning block (mean ± SEM). As in Figure 1I, significant cross-stimulus adaptation was observed immediately after learning (Day1 B1, p = 0.044), and, as in Figure 2B, there was a significant reduction in cross-stimulus adaptation across days (Day1 B1 minus Day2 B1, p = 0.034). On day2, following tDCS, there was a significant increase in cross-stimulus adaptation (Day2 B2 minus Day2 B1, p = 0.006) and the interaction between this effect and day 1 was also significant (day ^∗ block: [(Day2 B2 minus Day2 B1) minus (Day1 B2 minus Day1 B1)], p = 0.010). (G) The change in GABA concentration before versus during tDCS correlated with the change in cross-stimulus adaptation from Day2 B1 to Day2 B2 (with effects due to glutamate removed, r₁₇ = 0.486, p = 0.041).

Figure 4

Memory Accuracy Predicts Cross-stimulus Adaptation (A) There was no significant difference between participants’ accuracy on the associative learning task performed on day1 and the surprise memory test performed after scanning on day2 (p = 0.821) (shown: mean ± SEM for each day). (B) During periods of EI imbalance (Day1-block1 and Day2-block2), the average cross-stimulus adaptation significantly correlated with memory performance on the surprise memory test (r₂₀ = 0.57, p = 0.007). (C) During periods of EI balance (Day1-block2 and Day2-block1), the average cross-stimulus adaptation did not correlate with memory performance on the surprise memory test (r₂₀ = 0.016, p = 0.946).

Figure 5

Cortical Excitability and Changes in GABA and Glutamate Concentration (A) The concentration of GABA for each MRS acquisition, averaged across the group (mean ± SEM). As shown in Figure 3E, a significant reduction in GABA concentration was observed when comparing MRS measurements acquired before and during tDCS (p = 0.006). There was no significant difference between these GABA concentration measurements and the GABA concentration measured after the fMRI task block (p = 0.114). (B) The concentration of glutamate for each MRS acquisition, averaged across the group (mean ± SEM). There was no significant difference between glutamate concentration measured before versus during tDCS (p = 0.872). However, there was a significant increase in glutamate after the final fMRI task block (p = 0.020). (C) The region of interest used to assess changes in raw BOLD following application of tDCS. To avoid confounding our analysis with adaptation effects this ROI was defined from the average BOLD response to pairs of unrelated stimuli across all task blocks (see Supplemental Experimental Procedures). (D) Parameter estimates (mean ± SEM), extracted from the ROI shown in (C), revealed a significant increase in the raw BOLD response to nonadapting stimuli following application of tDCS (block2 – block1: p = 0.043). (E) The increase in BOLD response, shown in (D), was predicted by the post-task increase in cortical excitability, measured using MRS (change in glutamate concentration contrasted with change in GABA concentration using multiple regression: p = 0.024). This result is illustrated here by the positive correlation between the change in BOLD and post-task change glutamate concentration (r₁₇ = 0.488, p = 0.0398, with effects due to GABA removed) (left), and the negative trend between the change in BOLD and the post-task change in GABA concentration (r₁₇ = −0.424, p = 0.080, with effects due to glutamate removed) (right).

Figure 6

Neural Network Model Showing How Latent Cortical Associations Can Be Uncovered by Downregulating the Efficacy of Inhibitory Neurons (A–D) Four snapshots of recurrent network activity in response to stimulating one of four embedded cell assemblies. In the first row, each panel features a schematic of the parameter conditions of the network. The assemblies are pictured as colored squares. Excitatory and inhibitory connections are drawn in orange and gray, respectively. The second row shows the average firing rate over 1 s of every excitatory neuron in the network, assembled on a square grid. The third row visualizes the average firing rate of all excitatory neurons in each (red, green, yellow, or blue colored) assembly, averaged over 5 trials. (A) In the initial, balanced state, activation of the upper left (red) cell assembly leads to high firing rates in the activated neuron group, but not in other neurons (cf. Figure S4A). (B) After excitatory connections between associated cell-assemblies were selectively enhanced, the activation of the same assembly coactivates the associated green cell-assembly. (C) After disynaptic inhibition has been strengthened to balance the surplus excitation, the stimulation no longer resulted in coactivation of the associated green cell assembly. (D) Reducing the efficacy of all inhibitory synapses in the balanced network restored coactivation of the associated cell assembly (green) in response to driving the red cell assembly. (E) Complete simulation of all stages of the protocol (A) through (D) in 80 min and accordingly adjusted learning rate $\eta$. Solid lines show the average activity of the red and green cell assemblies over 2 s, and the activity of all background neurons is plotted in black. Circles show the average firing rate of red and green assembly neurons when they are stimulated (solid circles) or when the other assembly is stimulated (open circles), at 40 s intervals. Open black circles show the firing rates of un-stimulated background neurons during stimulations. The simulation begins with a naive network without assembly structure, firing at 5 Hz. After four cell assemblies are introduced (t = 7 min) the firing rate of assembly and background neurons increases, but inhibitory synaptic plasticity re-stabilizes network activity at 5 Hz. Red and the green cell assemblies can be individually activated, as shown in (A). When “associative” connections between the red and the green, and the blue and yellow (data not shown) cell assemblies are introduced (t = 23.5 min), high firing rates (maximum 136 Hz) of the unstimulated network are adjusted over the course of several minutes, but the associated cell assemblies coactivate in response to stimulation of either assembly, as shown in (B). Over time, inhibitory plasticity refines the disynaptic inhibitory inputs to each assembly so that coactivation between associated assemblies is reduced, as shown in (C). By reducing the efficacy of all inhibitory synapses, as thought to occur during tDCS (t = 74 min), the coactivation between associated cell assemblies is recovered, as shown in (D).