Memory-specific encoding activities of the ventral tegmental area dopamine and GABA neurons

Reviewer #1 (Public Review):

Midbrain dopamine neurons have attracted attention as a part of the brain's reward system. A different line of research, on the other hand, has shown that these neurons are also involved in higher cognitive functions such as short-term memory. However, these neurons are thought not to encode short-term memory itself because they just exhibit a phasic response in short-term memory tasks, which cannot seem to maintain information during the memory period. To understand the role of dopamine neurons in short-term memory, the present study investigated the electrophysiological property of these neurons in rodents performing a T-maze version of short-term memory task, in which a visual cue indicated which arm (left or right) of the T-maze was associated with a reward. The animal needed to maintain this information while they were located between the cue presentation position and the selection position of the T-maze. The authors found that the activity of some dopamine neurons changed depending on the information while the animals were located in the memory position. This dopamine neuron modulation was unable to explain the motivation or motor component of the task. The authors concluded that this modulation reflected the information stored as short-term memory.

I was simply surprised by their finding because these dopamine neurons are similar to neurons in the prefrontal cortex that store memory information with a sustained activity. Dopamine neurons are an evolutionally conserved structure, which is seen even in insects, whereas the prefrontal cortex is developed mainly in the primate. I feel that their findings are novel and would attract much attention from readers in the field. But the authors need to conduct additional analyses to consolidate their conclusion.

Reviewer #2 (Public Review):

The authors phototag DA and GABA neurons in the VTA in mice performing a t-maze task, and report choice-specific responses in the delay period of a memory-guided task, more so than in a variant task w/o a memory component. Overall, I found the results convincing. While showing responses that are choice selective in DA neurons is not entirely novel (e.g. Morris et al NN 2006, Parker et al NN 2016), the fact that this feature is stronger when there is a memory requirement is an interesting and novel observation.

I found the plots in 3B misleading because it looks like the main result is the sequential firing of DA neurons during the Tmaze. However, many of the neurons aren't significant by their permutation test. Often people either only plot the neurons that are significant, or plot with cross-validation (ie sort by half of the trials, and plot the other half).

Relatedly, the cross-task comparisons of sequences (Fig, 4,5) are hampered by the fact that they sort in one task, then plot in the other, which will make the sequences look less robust even if they were equally strong. What happens if they swap which task's sequences they use to order the neurons? I do realize they also show statistical comparisons of modulated units across tasks, which is helpful.

Overall, the introduction was scholarly and did a good job covering a vast literature. But the explanation of t-maze data towards the end of the introduction was confusing. In Line 87, I would not say "in the same task" but "in a similar task" because there are many differences between the tasks in question. And not clear what is meant by "by averaging neuronal population activities, none of these computational schemes would have been revealed. " There was trial averaging, at least in Harvey et al. I thought the main result of that paper related to coding schemes was that neural activity was sequential, not persistent. I think it would help the paper to say that clearly. Also, I'm not aware it was shown that choice selectivity diminishes when the memory demand of the task is removed - please clarify if that is true in both referenced papers. If so, an interpretation of this present data could be found in Lee et al biorxiv 2022, which presents a computational model that implies that the heterogeneity in the VTA DA system is a reflection of the heterogeneity found in upstream regions (the state representation), based on the idea that different subsets of DA neurons calculate prediction errors with respect to different subsets of the state representation.

I am surprised only 28% of DA neurons responded to reward - the reward is not completely certain in this task. This seems lower than other papers in mice (even Pavlovian conditioning, when the reward is entirely certain). It would be helpful if the authors comment on how this number compares to other papers.