Intermittent rate coding and cue-specific ensembles support working memory

The study addresses whether working memory is maintained by persistent, cue-specific spiking in prefrontal cortex or by activity-silent synaptic mechanisms. Decades of work in primates and other species established persistent delay-period activity as a dominant model of working memory. More recent proposals suggest that short-term synaptic plasticity can transiently store information without sustained spiking, potentially explaining variability and intermittency observed in single-trial activity. The authors aim to resolve this controversy by directly measuring single-trial population dynamics from large, simultaneously recorded ensembles in lateral prefrontal cortex during spatial working memory tasks, testing whether mnemonic information persists continuously in firing rates or alternates with activity-silent states and whether cue-specific synaptic ensembles carry information when spiking codes lapse.

Classic primate prefrontal studies reported sustained, cue-specific delay activity during working memory, across spatial and feature domains and multiple brain areas and species, supporting persistent spiking as the dominant model. However, recent theories and evidence support ‘activity-silent’ models whereby short-term synaptic plasticity maintains mnemonic traces without continuous spiking. Shortcomings in the persistent model include modest or absent delay activity in some contexts, strong single-trial variability, and bursty high-gamma suggesting irregular population spiking. Prior evidence for activity-silent traces has come from indirect measures, serial biases linked to functional connectivity, task-set maintenance, and TMS/reactivation in humans, but it remains unclear whether cue-specific synaptic mechanisms operate during canonical memory delays. The present work leverages large-scale simultaneous recordings to directly assess both spiking and functional connectivity-based signatures at single-trial resolution.

Subjects: Three monkeys (A, H, J). Recorded in lateral prefrontal cortex (areas 8 and 9/46). Tasks: Two interleaved spatial working memory tasks with eight possible cue locations. In both tasks, a brief (50 ms) spatial cue was followed by a memory delay (1,400–1,600 ms) during fixation. In match-to-sample (MTS), two targets appeared and monkeys saccaded to the target at the cued location. In memory-guided saccade (MGS), no targets appeared and monkeys saccaded to the previously cued blank location. Performance was high across sessions. Recording: High-density Neuropixels NHP probes were used across 25 sessions, yielding hundreds of units per session (mean 329 ± 46; total 8,225 single and multi-units). This enabled measurement of local population dynamics and a large number of simultaneously recorded neuronal pairs per session (mean 52,314 ± 11,252 pairs). Decoding and confidence: For each session, a leave-one-trial-out binary classifier (logistic regression) was trained at each time point to discriminate the test cue location versus its opposite location. Analyses of the memory delay were restricted to 500–1,400 ms after cue onset to avoid visually evoked activity. Classification accuracy across time and cue locations was assessed; classifier confidence (posterior probability of the correct class) provided a time-resolved single-trial measure of mnemonic information. State identification: To determine whether confidence reflected discrete states, confidence distributions per session were fit with a single beta distribution (single-state) versus a mixture of two beta distributions (two-state) using cross-validation. For state labeling on trials, the classification procedure was repeated 50 times with shuffled condition labels to obtain a null distribution of confidence values. Contiguous time points with confidence significantly above null were labeled On states; non-significant periods (P > 0.20) were labeled Off states; periods significantly below null were rare. Cross-validated single-neuron analyses: To test how states reflected neuronal properties, half the recorded neurons were used to label On/Off states and held-out neurons were analyzed for spatial tuning and firing rate changes during the delay, with train/test halves swapped for full coverage. Functional connectivity analyses: During the delay, cross-correlograms (CCGs) were computed for all neuronal pairs and cue conditions, thresholded and normalized using established methods. Analyses focused on low-latency (<10 ms), non-zero-lag peaks consistent with synaptic interactions. Cue-specific ensembles were defined by comparing the sets of significant CCGs across cue conditions. Dissimilarity was quantified as the Manhattan distance (count of condition-unique connections), compared against null distributions from condition-shuffled data to obtain z-scores. On vs Off ensemble specificity: The Manhattan distance analysis was repeated separately restricted to On or Off epochs to test whether cue-specific connectivity persists when rate codes lapse. Temporal generalization and code stability: Classifiers trained at specific times (e.g., cue-evoked responses) were tested across time to assess code stability (cluster-mass test with multiple-comparisons correction). Demixed PCA and single-neuron analyses further tested stability. Joint selectivity and connectivity: For each cue, the proportion of neuronal pairs jointly selective during the evoked period (0–400 ms post-cue) and the proportion exhibiting significant delay CCGs were measured. The observed conjunction rate was compared to the product of marginals to test overrepresentation. Statistical tests included sign-rank tests, chi-square tests, correlations, and ANOVAs as reported.

• Decoding accuracy: Mean binary classification accuracy for cue location was significantly above chance across the memory delay in all sessions (all P < 0.001, sign-rank), with session-wise mean accuracies ranging from 59% to 89%. Accuracy was similar across cue locations (range 67–74%, all P < 0.001).
• Single-trial dynamics: Classifier confidence frequently lapsed to chance during the delay on single trials, showing sharp transitions between high and low confidence that were not aligned across trials. Confidence correlated with reaction times on correct trials.
• Two-state model: A cross-validated two-beta mixture model fit confidence distributions better than a single-state model in 19/25 sessions (χ²(1) = 7, P = 0.009). Mixture-of-Gaussians analyses yielded similar results.
• On/Off states: From cue onset to end of delay, trials exhibited on average 2.35 ± 0.04 On states (median 2) and 3.74 ± 0.04 Off states (median 4). Mean durations: On 192.4 ± 2.1 ms (median 150 ms); Off 146.1 ± 1.7 ms (median 100 ms). Periods of confidently incorrect decoding were rare (mean 0.11 ± 0.01 per trial). State at go cue predicted accuracy and reaction time. Classifiers trained and tested only on Off states did not perform above chance, arguing against an alternative rate code during Off periods. On/Off dynamics were not explained by microsaccades, LFP phase, or recording noise.
• Single-neuron properties across states: In held-out neurons, spatial tuning during the delay was strong in On states and nearly abolished in Off states (two-way repeated-measures ANOVA interaction P < 0.001). Cue location explained ~8.2% of firing-rate variance during On states vs 0.7% during Off states (~12-fold decrease). Mean firing rates during On states were elevated above spontaneous baseline (mean baseline 5.8 ± 0.4 Hz; P < 0.001), whereas Off-state rates were indistinguishable from baseline (P = 0.192) and significantly lower than On (P = 0.001).
• Functional connectivity and cue-specific ensembles: Significant low-latency, non-zero-lag CCG peaks were observed in 1.36 ± 0.22% of pairs (12,740 total significant CCGs). The probability of significant CCGs decreased with inter-neuronal distance (r(39) = −0.42, P = 0.007). Connectivity patterns differed by cue: Manhattan distance between cue-specific connectivity maps exceeded chance across sessions (P < 0.001), including when matching firing rates across conditions (P < 0.001).
• Ensembles during Off vs On: Cue-specific Manhattan distance was not greater than chance during On states (P = 0.129) but was significantly greater during Off states (P = 2.5 × 10⁻5; also P = 0.001 with firing-rate-matched conditions), indicating mnemonic information in connectivity patterns when rate codes lapse.
• Stable rate code across time: Classifiers trained on cue-evoked responses generalized across the delay, indicating a stable memory code (P < 0.001, cluster-mass test). Demixed PCA and single-neuron analyses supported code stability.
• Joint selectivity and connectivity: For any given cue, 1.47 ± 0.4% of pairs were jointly selective in the evoked period, and 1.4 ± 0.2% showed significant delay CCGs; their conjunction occurred 2.5× more often than expected by independence (P < 0.001). Overall, 55% of neurons showed a cue-specific evoked response, 81% participated in a cue-specific functional connection, and 10.4% were in jointly selective-and-connected pairs.

The results show that mnemonic information in lateral prefrontal cortex does not persist continuously in population spiking during memory delays. Instead, information alternates between coordinated On states—characterized by elevated firing rates and strong spatial tuning—and Off states—marked by firing rates at baseline and loss of tuning. Despite these lapses in rate coding, cue-specific functional connectivity patterns among neuronal pairs persist and carry information during Off states, consistent with synaptic models in which short-term plasticity maintains latent mnemonic traces. The stable generalization of classifiers from cue-evoked responses across the delay supports a stable spiking code when present, while the overrepresentation of pairs that are both jointly cue-selective and functionally connected links evoked spiking responses to subsequent cue-specific synaptic interactions. Together, these findings reconcile persistent and activity-silent accounts by demonstrating an interplay: intermittent rate-based coding coexists with synaptic ensembles that preserve information during silent epochs, supporting robust working memory at single-trial resolution.

This study demonstrates that working memory in primate lateral prefrontal cortex is supported by intermittent, coordinated rate coding interleaved with activity-silent periods during which cue-specific synaptic ensembles maintain information. Large-scale Neuropixels recordings and single-trial decoding revealed discrete On/Off states; functional connectivity analyses showed cue-specific ensembles particularly evident during Off states; and temporal generalization confirmed a stable spiking code when active. These findings integrate persistent-spiking and synaptic mechanisms into a unified account of working memory. Future work should test these dynamics across additional brain regions (e.g., parietal cortex), broader spatial scales, and various task demands (spatial vs object, manipulation load), and investigate neuromodulatory influences (e.g., dopaminergic tone) on state transitions and mnemonic stability.

Observations are from local populations in lateral prefrontal cortex during spatial working memory tasks; generalization to other cortical and subcortical areas, object-based tasks, and different manipulation demands remains to be established. The study infers synaptic mechanisms from functional connectivity (CCGs) rather than direct synaptic measurements. On/Off state heterogeneity across trials suggests potential modulation by global arousal or neuromodulatory state, which was not directly manipulated. Analyses were constrained to the memory delay window to avoid visual response contamination, and significant CCGs represent a small fraction of all pairs.