A cholinergic hub in the nucleus accumbens gates opioid-reward learning.

The classic definition of mood by Jean Delay refers to “a basic affective disposition […] that gives each of our states of mind a pleasant or unpleasant tone, oscillating between the extreme poles of pleasure and pain” []. While emotion is thought to be closely tied to a specific event, mood is thought to reflect the cumulative effect of multiple stimuli [–]. How mood fluctuates in response to (or independently of) life events is a crucial question for understanding both everyday affective fluctuations and mood disorders. Empirical observations of the effect of life events on mood fluctuations are common [–], in real life as well as in lab experiments [,], but only recently have formal theories been proposed [–]. An important step in this direction has been the use of computational models to analyse mood fluctuations on a short time scale, approached through a series of self-report mood ratings collected during behavioural tests [–]. In short, such computational model conceive mood as an affective state that integrates over time the positive and negative outcomes. Several variations have been suggested around this idea. A first group of models conceives mood as a “representation of momentum.” Mood is updated via a delta-rule driven by reward prediction error rather than direct reward values. As a result, mood converges toward the temporal derivative of reward availability in the environment. Some of these models include an effect of mood on the perception of obtained outcomes (thereby facilitating reinforcement learning []) and/or an effect on the perception of prospects [–] (thereby facilitating the decision to engage in reward-effort trade-offs). A variation on this idea is that mood is not just a representation of the external world, but rather captures an agent’s ability to achieve positive outcomes in that world []. In a recent study, we proposed a simple mechanism to implement this idea at the algorithmic level, dubbed the MAGNETO model []. In this model, mood is updated through a delta-rule using the net value of actions (reward minus cost) as the driving outcome signal. Consequently, mood converges toward a global net expected trade-off (how much rewards are expected to outweigh the time and effort invested in upcoming potential actions). Beyond integrating costs, mood becomes a representation of average reward availability rather than its temporal momentum. The critical idea underlying these models is that such a process could be adaptive in an autocorrelated environment (e.g., in early spring, the appearance of low-hanging fruits on a given tree predicts the appearance of fruits on other trees within the next few weeks, and vice versa when “winter is coming” [,]). Experimentally, several tasks have been proposed to induce short-term (i.e., < 2 hours) mood changes. The most frequently used are gambling/lottery tasks [], in which participants receive positive and negative feedback as a result of a random draw. The outcome could be completely independent of the agent’s will (as in a wheel of fortune), or it could depend on the agent’s choice to participate in the lottery (e.g., a choice between 1 euro for sure or a 50% chance of winning 2 euros). More recently, we and others have used a quiz task designed to induce episodes of high correct response rate (leading to higher mood) vs. low correct response rate (leading to lower mood) [–]. The idea behind this task was to maximize the sense of agency, which might arguably be minimal in a lottery task where participants could rely on simple decision heuristics. The underlying hypothesis is that mood might be more responsive to outcomes when they depend on the agent’s behaviour, than when they depend solely on the environment [–]. However, this increased sense of agency comes at the cost of reduced experimental control: the course of the task, including the sequence of outcomes, inherently depends on participants’ choices and responses rather than being fully determined by the experimenter. This reduced control may affect the validity and reliability of inferred computational phenotypes.