Persistently active neurons in human medial frontal and medial temporal lobe support working memory

Abstract

Persistent neural activity is a putative mechanism for the maintenance of working memories. Persistent activity relies on the activity of a distributed network of areas, but the differential contribution of each area remains unclear. We recorded single neurons in the human medial frontal cortex and medial temporal lobe while subjects held up to three items in memory. We found persistently active neurons in both areas. Persistent activity of hippocampal and amygdala neurons was stimulus-specific, formed stable attractors and was predictive of memory content. Medial frontal cortex persistent activity, on the other hand, was modulated by memory load and task set but was not stimulus-specific. Trial-by-trial variability in persistent activity in both areas was related to memory strength, because it predicted the speed and accuracy by which stimuli were remembered. This work reveals, in humans, direct evidence for a distributed network of persistently active neurons supporting working memory maintenance.

Figure 1. Task, recording locations, and behavioral results

(a) The task. Upper row represents examples of the screens presented to the subjects during an example trial (with load 2). The lower row represents the lengths of time for which each screen was shown. Each trial consisted of 1–3 sequentially presented pictures (encoding), followed by a variable delay (holding or maintenance period). After the delay, a probe image was shown and patients indicated whether the probe was or was not shown during the immediately preceding encoding period. (b) Location of recording sites in MNI152 space (see methods). Recording locations are indicated by different colors (red is pre-SMA, blue is dACC, yellow is hippocampus, and cyan is amygdala). (c–d) Behavioral results. (c) Accuracy of all sessions, rank-ordered. (d) Median reaction time (relative to onset of the probe image) as a function of load. Each dashed line connects an individual session. (b–c) Thick and light blue lines represent the mean and s.e.m across all sessions, respectively.

Figure 2. Stimulus-selective concept cells

(a) Example concept cell recorded from the amygdala. Upper panel shows the Post-stimulus time histogram (PSTH; binsize 200ms, stepsize 2ms). Colors denote the different images (shown on the right). Shaded areas represent ±s.e.m across trials. Middle panel marks periods of significance (1×5 ANOVA; corrected for multiple comparisons using a cluster-size correction, see methods). Bottom panel shows raster with trials re-ordered according to image identity for plotting purposes only. Image onset is at t=0 (gray bar). The right inset shows the mean extracellular waveform ±s.e.m of all spikes associated with this cell. (b) Percent of all recorded cells who qualified as concept cells in each area during the screening and WM (“Sternberg”) tasks. The numbers below each area label denotes the number of neurons associated with each bar. We test if the observed percentage is higher than that expected by chance by comparing with a null distribution estimated after scrambling the condition labels randomly (repeated 500 times; * denotes p<0.05, ** p<0.01, and *** denotes p<=0.002). (c) Percentage of images shown for which we observed at least one concept cell in both tasks (p=0.0014). Each dashed line connects an individual session. This shows that the screening task successfully identified responsive neurons. Thick and thin blue lines represent the mean and ±s.e.m, respectively.

Figure 3. Concept cells are persistently active during WM maintenance

(a,b) Two example concept cells recorded from the amygdala (a) and hippocampus (b). For each, the upper panel shows the PSTH (binsize 200ms, stepsize 2ms). Shaded areas represent ±s.e.m across trials. Middle panel marks periods of significance between preferred vs. not preferred stimuli (corrected for multiple comparisons using a cluster-size correction, see methods). Bottom panel shows raster with trials re-ordered according to condition for plotting purposes only. Both neurons show both visually evoked selective activity (red) and sustained activity (blue) during maintenance. Note how during maintenance, concept cells have elevated activity only when their preferred stimulus was held in memory (blue vs. gray). Also, note how the sustained activity (blue) was suppressed during encoding of the non-preferred image (i.e. encoding 3) when the preferred stimulus was already held in memory. (c,d) Maintenance activity of the same neurons shown in (a,b), but only for the subset of trials with load 2 (two items held in memory). There was no significant difference in activity between trials where the preferred image was shown first vs. second (encoding 1 or 2; middle panel marks periods of significance). See Fig. 4d for a similar analysis at the population level. See Fig. S3 for further single-cell examples.

Figure 4. Population analysis of MTL concept cells

(a) Average firing rate of all concept cells identified in the amygdala (n=57) in the different phases of the task. Shaded areas represent ±s.e.m across neurons. Gray vertical bars mark periods of time during which an image was on the screen. Bottom panel marks points of time during which the activity of the cells was significantly different between trials when a preferred image was in memory vs. when it was not (corrected for multiple comparisons based on cluster size, see methods). Colors mark different trials as indicated. For subplots a-d, only correct trials were used. (b) Picture selectivity index (PSI) during encoding, maintenance, and retrieval for all identified concept cells (each data point is one neuron; data points are sorted according to the encoding phase of the task). Neurons in both amygdala and hippocampus, but not dACC, maintained their selectivity throughout the task and showed persistent activity. Significance was computed against chance (PSI=0). (c) PSI for different load conditions indicates that neurons maintained persistent activity for loads 1–3 in amygdala and 1–2 in hippocampus, but not for dACC. (d) PSI for loads 2 and 3 as a function of whether the preferred image was shown first, second, or third during encoding. This shows that images which were shown directly before the maintenance period did not have greater selectivity (load 2 P=0.702; load 3 P=0.873). (e) Relationship between firing rate of concept cells in the MTL and behavior. The firing rate was significantly higher for correct compared to incorrect trials only when the preferred stimulus was held in memory. For (c–e), PSI and firing rate was calculated for the entire maintenance period. Throughout, * denotes p<0.05, ** p<0.01 and *** p<0.001 as estimated with permutation tests. Pre-SMA is not shown in this figure because we did not identify any concept cells in this area.

Figure 5. Persistent activity of maintenance cells in MFC

(a) Average firing rate of all maintenance neurons (n=54) in pre-SMA for different load conditions and (b) for trials that ended with a fast or slow response (median split of RT). Shaded areas represent ±s.e.m across neurons. Gray vertical bars mark image presentation. Black bars indicate significance at P<0.05 of a 1×3 (top, permutation ANOVA) and 1×2 (bottom, permutation t-test) with load as dependent variable for (a) and response time for (b). Multiple comparisons were corrected for using a cluster-size approach (see methods). (c) Percentage of all recorded cells identified as maintenance neurons in each area. Numbers below the area label denotes number of cells. The medial frontal areas (dACC, pre-SMA) contained significantly higher proportions of maintenance neurons (chi²[1]=21.1; P=4.353e-6) compared to areas in the MTL. (d) Percentage of maintenance neurons whose firing rate during maintenance differed as a function of load. Notably, in pre-SMA, 37% of cells decreased their firing rate as a function of load. (e) Percentage of maintenance neurons whose activity during maintenance differed as a function of response time. (f) The firing rate of maintenance neurons in dACC and amygdala differed as a function of whether stimuli were later remembered or forgotten. (c–e) Significance was assessed by comparing with a null distribution estimated using a bootstrap (see methods). For (f) we used permutation test. (f) Boxplot represents quartiles (25%, 75%), line is median, whiskers show range up to 1.5 times the interquartile range, and dots above whiskers show outliers. Throughout, * denotes p< 0.05, ** p< 0.01 and *** p<= 0.002.

Figure 6. Probe neurons reflect WM-retrieval related evoked activity in MFC

(a) Firing rate of an example probe neuron recorded from the pre-SMA, shown separately for trials where the probe was held in (IN,cyan) or not (OUT,magenta) in memory. Upper panel shows the PSTH (binsize 200 bins, stepsize 2 ms, shaded areas represent ±s.e.m). Middle panel marks points of time with a significant difference between IN and OUT trials, corrected for multiple comparisons using a cluster-size approach. Bottom panel shows raster with re-ordered trials. Gray vertical bars mark image presentation. (b) Same neuron as in (a), but aligned the response (button press was at 1.2 sec after image presentation) to button press. Note the much reduced peak response (1.44 Hz vs 0.45 Hz, permuted t-test: P=0.005). (c) Percentage of probe neurons in each area. Probe neurons were most prominent in pre-SMA, followed by dACC. (d) Probe neurons elevated their firing rate only during retrieval, but not encoding (P=0.0002, permuted t-test). (e–f) Percentage of probe neurons in each area whose firing rate during probe (−800-0 ms relative to button press) differed as a function of IN vs. OUT (e) or as a function of button press (f). Most cells showed no difference, i.e. they responded equally strongly (but selectively) to the probe stimulus. (c,e,f) Significance was assessed based on a null distribution estimated based on permuted labels. * denotes p< 0.05, ** p< 0.01 and *** p< 0.001.

Figure 7. Population decoding from all recorded neurons during the maintenance period

(a) During maintenance, picture identity could be decoded from the activity of amygdala and hippocampus, but not dACC and pre-SMA, neurons. Decoding performance was maintained when only cells identified as concept cells were considered (magenta). One vs. all denotes average accuracy of decoders trained to distinguish between a given image and all the others (50% chance level). (b) Picture identity decoding using different time windows for training and testing. Shown is the test – retest decoding performance for load 1 trials (chance level is 20%). (c) The activity of neurons in the pre-SMA during maintenance was predictive of how many items were held in memory (load, 1–3). Decoding only from maintenance neurons (magenta) was sufficient. (d) The activity of neurons in the pre-SMA and amygdala during maintenance was predictive of later response speed. * denotes significance at p< 0.05, ** p< 0.01 and *** p<= 0.002. Markers below bars indicate significance vs. chance performance, estimated by randomly scrambled labels. Significance of pairwise tests was estimated by comparisons with a null distribution of the same differences estimated from decoders trained on data with randomly scrambled labels. (a,c,d) Boxplots represent quartiles (25%, 75%), line indicates the median, whiskers show range up to 1.5 times the interquartile range, and dots above whiskers show outliers.

Figure 8. Persistent activity during maintenance forms attractors

(a) Illustration of the mean trajectories in neuronal state space formed by the three demixed principal components (dPCs) associated with picture identity during encoding (thin line) and maintenance (thick line). The dot indicates the point of time 200 ms after image onset and the arrow indicates the direction of change. Colors mark different images (only 4 of the total 5 are shown for clarity). (b) Different view of (a). (c) Multidimensional velocity of the population in the different phases of the task (shown for all load 1 trials). The velocity during maintenance was significantly slower compared to encoding and was not significantly different from that during baseline. (d) Multidimensional pairwise distance between all possible pairs of attractors during maintenance (load 1). The distance during maintenance was significantly larger compared to that during baseline (P=0.0005). Together (c,d) are indicative of attractors. The significance of the population-metrics shown in (c–d) was computed by randomly subsampling a subset of trials and neurons (see methods). (e) Schematic representation of the distances between the attractors (attractors are defined based on correct load 1 trials) for each image (small filled dots) and the average position in state space for image D for two behaviors: remembered (correct) and forgotten (incorrect) computed for all loads separately and averaged. This representation was determined based on multidimensional scaling of the state space (see methods). Isolines depict areas of equal distance from the attractor for image D. (f) The distance to the attractor (DA) was significantly smaller for correct compared to incorrect trials only when concept cells were part of the population. Note that DA <1 indicates that the trajectory is closer to the correct attractor than all the other attractors. (g) The distance to the attractor (DA) corresponding to the remembered image was indicative of the speed of the response. This relationship was observed only for concept cells. (h) Distance from attractor predicts performance on individual trials. Each dot is one trial. (c,d) Boxplots represent quartiles (25%, 75%), line indicates the median, whiskers show range up to 1.5 times the interquartile range, and dots above whiskers show outliers. * denotes p< 0.05, ** p< 0.01 and *** denotes p<=0.002.