Know the past to see the future: Memory and prediction in in-vitro cortical neurons

Memory plays a crucial role in processing and retaining information, and it has been studied extensively at both the behavioral and cellular levels. It can be subdivided into short-term memory, where information is retained for a few seconds, and long-term memory, where memories are stored for minutes to years. Episodic memories are first encoded, then consolidated, and finally retrieved. This study focuses on the process of memory consolidation, where new information in the hippocampus are replayed and transferred to neocortical areas for long-term storage. This process, known as systems consolidation, occurs during slow-wave sleep (SWS) and is thought to rely on spike-timing-dependent plasticity (STDP), a mechanism by which the timing of neuronal spikes adjusts synaptic strength. Knowledge about specific processes involved in memory consolidation are still scarce.

Another critical brain function is prediction, which is hypothesized to be essential for the efficient execution of actions and may be linked to memory. Unlike memory, prediction has been primarily studied theoretically. Open questions remain regarding the mechanisms of prediction, whether it is a general capability of neural networks, and the extent to which it relies on memory.

Here, we used in-vitro networks of dissociated cortical neurons from rats cultured on microelectrode arrays (MEAs) to investigate both memory and prediction mechanisms. We employed experimental protocols that used electrical and optogenetic stimulation to study memory consolidation and prediction. Mutual information was used to quantify prediction, while changes in functional connectivity, measured using Conditional Firing Probability (CFP), were used to study memory. Maximum Entropy (MaxEnt) models were also tested as a possible CFP alternative.

Our findings show that background input, applied during slow-wave sleep conditions, hamper memory consolidation in vitro, while network excitability is critical for this function. Furthermore, we demonstrated that prediction depends on short-term memory but becomes less dependent as long-term memory traces form. Notably, blocking NMDA receptors impaired long-term memory without affecting prediction, confirming STDP’s role in long-term memory and prediction's reliance on short-term memory. This study establishes an in-vitro model for investigating memory and prediction and introduces mathematical tools for probing their underlying mechanisms.