Overnight fasting facilitates safety learning by changing the neurophysiological response to relief from threat omission

Reviewer #1 (Public Review):

Questions and concerns:

The abstract is hard to follow. The authors there refer to a previous experiment showing that "overnight fasting diminishes excessive avoidance and speeds up fear extinction by decreasing subjective relief during threat omissions" (L26). They go on to say that "relief tracks the reward prediction error signal that governs safety learning" (L28). This is puzzling. While getting less relief/safety from avoidance actions will surely diminish avoidance (because avoidance actions are less reinforced), getting less relief/safety from omissions of an unconditioned stimulus (US) in fear extinction should slow down (not speed up) fear extinction. In the same vein, why are "lower activations [in fMRI] in the ventromedial prefrontal cortex and nucleus accumbens in response to threat omissions signaled by a safe cue" (L34) associated with "increased effective avoidance and sped up fear extinction" (L33)? This clearly goes against the existing literature on reward prediction errors (PEs) in fear learning paradigms, where these PEs in the mesolimbic dopamine system drive extinction, that is, they are associated with better extinction (and should therefore also be associated with more avoidance). For instance, in the rodent, Luo et al., 2018 (DOI: 10.1038/s41467-018-04784-7) and Salinas-Hernandez et al., 2018 (DOI: https://doi.org/10.7554/eLife.388181 of 25RESEARCH ARTICLE) and 2023 (https://doi.org/10.1016/j.neuron.2023.08.025ll) have in various constellations optogenetically enhanced and diminished, respectively, the PE signal at the time of US omission in extinction in either VTA or nucleus accumbens and thereby sped up and slowed down, respectively, extinction learning. If the results of the current experiment contradict established knowledge, the reader must be clearly informed about this. By contrast, the abstract gives the impressions as if the current results were to be expected and in line with the literature ("since relief tracks the reward prediction error signal ..., we hypothesized ...").

It would also help the reader if it was clarified that the finding of "increased effective avoidance" (L33) went counter to the hypothesis, e.g., by saying "Contrary to our hypothesis, we observed ...".

Introduction:

L51: The presentation of exposure therapy is a bit misleading and may create confusion. While it is probably correct that exposure works by "promoting safety learning", this is generally thought to be the case only for Pavlovian associations (CS-US), that is, for extinction (where safety learning creates the new association of CS and "no US"). It is, however, not generally considered to be the case for the instrumental action-outcome associations that underlie avoidance learning ("I do this or that, then I do not have to experience the feared object or situation"). Therapists try to prevent this type of learning from happening, exactly by promoting the confrontation with fear objects or situations in the absence of any avoidance action.

Generally, I think the introduction suffers from the absence of a short explanation of what avoidance and extinction learning are, behaviorally, and what types of mechanisms are believed to drive them, and that the one (avoidance) is thought to contribute to the maintenance of fears whereas the other (extinction) reduces fear. The non-specialist reader is somehow left in the dark.

In the same vein, on L63, presenting the results of their previous fasting study that serves as a discovery study for the present experiment, the authors make a distinction between "unnecessary avoidance during a signal of safety" and "effective avoidance during a signal of upcoming threat". It is really expecting too much from the reader that they will understand at this stage that a CS can become a signal of safety through extinction or that a CS not paired with a US during conditioning (a "CS-") is a safety signal and that it is not necessary to avoid such a signal, whereas a non-extinguished CS (signaling threat) may well be avoided. (At least, this is how I understood the distinction.)

I was then really confused by the following statement (L65) that "the decrease in unnecessary avoidance was mediated by lower levels of relief ... during omissions of threat". If a CS is already extinguished (has no remaining or only little threat value, that is, is a safety stimulus), there is no longer threat omission when the US does not occur, and no relief. There should also be no relief to US omission after a CS-. More importantly even, if fasted participants reported lower levels of relief from threat omission, why did they not also show less effective avoidance (which is driven by the reinforcement provided by the relief that occurs when a successful avoidance action has prevented a US from occurring after or during the CS)?

L69: Also the statement "a faster decline in relief ... ratings during ... extinction, suggesting faster decrease of threat expectancies" can only be understood by the reader if they already know what a PE is and by what rules PE-driven learning is governed (that is, essentially, if they know Rescorla-Wagner). I think the authors must explain, in order to allow a non-specialist reader to follow their text, that the PE (supposed to be indexed by the relief rating) reflects the discrepancy between the magnitude of an outcome expectation (e.g., here, expectation of the US) and the obtained outcome (here, US or not); that, therefore, a PE is generated when a subject expects a US (as a result of prior conditioning) but does not get it; that this leads to a proportional update (reduction) of the US expectation in the next trial; and that this in turn leads to a diminished PE when the US again does not occur. Notably, the reader must be made aware that the higher the PE, the higher the reduction and the faster the extinction (proportionality).

The reader must also be made aware that the update is additionally determined by some multiplicatory "transmission" function or constant (e.g., learning rate in Rescorla-Wagner) that defines the size of the relationship between the magnitude of the PE and the magnitude of the update (reduction). Hence, in two individuals, even if the magnitude of the PE is identical, the magnitude of the update may differ because of individual differences in the learning rate (to take the Rescorla-Wagner implementation). The authors, however, seem to ignore the possibility that fasting changes the learning rate.

Both the dynamics of the PE and the learning rate, of course, add complexity to the interpretation of the past and present data. But I think the authors cannot avoid this when they want to make sense of a treatment (fasting) that they believe affects safety learning. Speaking of "lower levels of relief" (L66) must be qualified by whether these lower ratings were observed initially (when the first PEs were registered at initial threat omissions, meaning that safety learning should be relatively slowed down by fasting) or on average or later during a safety learning experiment (which could indicate that learning under fasting was relatively quicker/more successful).

Following upon this, in L74, the conclusion from observations of lower levels of relief during avoidance and faster decline in relief during extinction in the previous study that "overnight fasting decreased the reward value of safety (less relief pleasantness)" may be wrong if the faster decline and the resulting lower average levels of relief were the consequence of a higher initial PE in the fasting group, as would be expected from the Rescorla-Wagner rule. If the latter were the case, this would suggest that subjects actually registered more safety (a higher discrepancy to their threat expectation) in early trials. This could also explain why fasting sped up extinction in that study (see Abstract). It might also explain why "effective avoidance" (L64) was at least maintained (although it should actually also be sped up). It might make less parsimonious explanations ("fasting biases .. to focus on food at the expense of safety", L79), requiring the presence of a food source and a utility function of accepting a threat in the obtainment of food, unnecessary.

All this, however, rests on whether I think I have understood what the authors want to say about their relief measurements and the way the operationalized avoidance in the previous study.

More unclarities due to not giving full information: L91: "... extinction and avoidance learning. Accordingly, human fMRI studies have found ... activations in the ventral striatum and the VTA during threat omissions that might contribute to establishing a new safety CS-->noUS memory that reduces the initial fear response." However, in avoidance, it is an action that is reinforced by the US omission and hence an action-->noUS memory that is being formed. The CS keeps its threat value acquired during the preceding conditioning phase, and the reduction of fear during CS presentations is contingent upon the exertion of the avoidance action.

L99: "Because overnight fasting decreased relief rating particularly during omissions after safety signals". Again, if a US is omitted after a safety signal (an extinguished CS or a CS-), there should be no PE and no relief. If there were still relief ratings at US omission after a safety signal, this would suggest extinction did not fully work or differential conditioning was not successful. In any case, it is not clear at all why relief was specifically decreased during omissions after safety signals and not (and much more so) during omissions after threat signals, where there is clearly a PE. If this was not the case, one has to wonder if something went wrong in the discovery study.

The paragraph starting L103 and the associated figure 1 could be a bit more precise and give a bit more information in order to provide the reader a proper understanding of key experimental manipulations, in particular the ART task. Please define abbreviations "CS+unav", "CS+av". L108 ff.: One gets the impression there is only one CS+, whereas there are two. Say explicitly that one CS+ remains unavoidable during the Avoidance phase (CS+unav). What is the purpose of this stimulus? Do participants learn during the Avoidance phase that the CS+unav is unavoidable and the CS+av is avoidable or is this instructed? Do participants have to press the button within a certain time after CS+unav onset in order to avoid the US, or with a certain force? Is avoidance in case of successful button pressing deterministic or probabilistic? Say that the frame with the non-lit lamp is the ITI.

Relief ratings (Figure 1b): The rating says "How pleasant was the relief that you felt?". That is, the experimenter insinuates that the participant will have felt relief and only wants to know how pleasant that relief was. The subjects has no chance to indicate there was no relief. This may be the reason why, in the discovery study, subjects indicated relief to safe stimuli, see above. Why did the authors not simply ask about the degree of relief felt, which would give a subject the chance to say there was no relief? I think this is a major flaw.

L119: "We previously found that overnight fasting reduces avoidance and relief mostly to a safe CS-." If this is really the only thing that the authors found, then the fasting manipulation in their previous study failed to modulate avoidance of CS+s and the PE signaling at the time of US omissions after CS+s, that is, after actual threat stimuli. The procedure then clearly is not suited to study influences of fasting on avoidance learning. Whatever it does manipulate, it is not relief-based avoidance learning.

L130: It makes absolutely no sense to hypothesize that a manipulation reducing relief in extinction learning will decrease activation in the neural PE circuitry at the time of US omission more after the CS- than after the CS+. Of course, the PE is highest when the US is not given after the CS+, and this is where any relief manipulation should have an effect. As said above, the authors must also specify their hypothesis with respect of timing (early or late extinction? See the animal papers cited above.)

Reviewer #2 (Public Review):

The authors repeated a previous behavioural study on the effects of overnight fasting on avoidance and extinction learning in healthy female participants in the 3T MRI scanner. Previous behavioural findings were replicated only in part. Fasting related changes of fMRI signals were less than expected.

This paper is not without interest. Anxiety disorders are very frequent, and there is still a need to better understand ways to improve extinction and reduced avoidance. The authors follow up on previous observations of their group using overnight fasting. The findings, however, were largely negative, and it is difficult to tell how robust the observed positive findings are. The paradigm did not work as well as expected in the MR scanner.

Introduction/main hypothesis: The reviewer does not understand why a smaller reward prediction error should result in faster extinction learning? The opposite should be the case. Plus, how much of a reward prediction error is expected in the CS- condition in extinction training? Here the US omission is expected. The reviewer may miss a key concept of the study.

Results: A major part of the behavioural data of a previous pure behavioural study was not reproduced (avoidance learning), plus many of the MRI findings did not show a difference between the fasting and re-feed groups. Given the large amount of comparisons it makes one wonder how robust the presented findings are. The advances to the field are therefore limited.

Author response:

We thank the editor and reviewers for the time they spent reviewing our manuscript entitled ‘Overnight fasting facilitates safety learning by changing the neurophysiological response to relief from threat omission’ which was sent as an original paper for a potential publication in eLife.

Since we take the reviewer comments at heart and recognize the very complex scenario of our previous and current results we will take more time to re-think the paper. This time will serve us to look back to the interpretation of the results of our previous behavioral study, to the preregistration plan as well as findings of our current fMRI (replication) study.

We aim to address the fundamental issues indicated by the reviewers as soon and as clearly as possible.