Abstract
Neural activity in the primate lateral prefrontal cortex (LPFC) has been causally linked to working memory (WM). Some researchers have described this pattern of activity as persistent firing. One issue that remains unclear is whether persistent firing has a spatiotemporal structure within populations of LPFC neurons. Here, we hypothesize that precise temporal sequences of individually active neurons enable persistent firing and thus coding of WM in LPFC. We implanted two 96-channel Utah microelectrode arrays in each animal (areas 8A/ 9/46). During our virtual working memory task, a target was presented in 1 of 9 locations in a virtual arena. The target then disappeared and following a 2 second delay period, subjects navigated to the cued target location using a joystick. Spiking activity showed clear temporal structure. When ordered by peak firing time, these short bursts formed clear “bands” spanning the duration of the trial. We next sought to understand whether these temporal sequences of bursts were related to memory content, so we developed a novel computational method to analyze spike sequences across trials. This method works by representing individual sequences as a complex-valued vector and performing dimensionality reduction on the resulting covariance matrix. We find that sequence structure, independent from firing rate information, is related to target location in the WM task. Consistent with this, many individual neurons also fired at consistent times during the WM trial. The relation between temporal structure and WM content was less consistent on incorrect trials and following subanesthetic injections of ketamine. These results reveal a striking temporal organization of spiking activity during WM in LPFC neuronal ensembles. They also suggest that persistent firing in LPFC has a precise temporal structure that allows neuronal ensembles to robustly maintain and manipulate task relevant information.