Neuromodulatory systems partially account for the topography of cortical networks of learning under uncertainty

Reviewer #1 (Public review):

Summary:

This study investigates whether the distribution of receptors and transporters of neurotransmitters accounts for the topography of cortical activity of confidence and surprise in probability learning. The authors first examined the invariance of functional correlates of confidence and surprises with multiple fMRI studies and then investigated whether 20 PET-derived receptor and transporter density maps account for this cortical invariant activity of confidence and surprise in probabilistic learning. Beyond these specific findings, the main novelty of this study lies in its attempt to bridge neuromodulatory systems and cognitive processes using neuroimaging data. This integrative approach is particularly valuable, as it showcases a framework to combine neurochemical architecture and cognitive computations.

Strengths:

This study attempts to link neuromodulatory systems with cognitive processes involved in probabilistic learning. Although the role of neuromodulatory systems in learning has been highlighted in several influential previous studies, it has not yet been widely investigated or systematically related to functional neuroimaging data so far. The authors used an efficient approach to address this question by combining group-averaged neurotransmitter maps with functional results from multiple fMRI studies using probabilistic learning tasks with similar structures. This approach provides informative insights into the relationship between the distribution of neuromodulatory systems and cognitive processes from neuroimaging data.

Weaknesses:

One limitation of the study stems from the unavoidable constraints of relying on pre-existing datasets rather than data specifically collected to address the present research question. Because the four fMRI studies differed in their measurements and task structures, the authors defined confidence and surprise on the basis of ideal observer behavior. Thus, "confidence" and "surprise" are not related to individual decision or subjective value, and the PET data is also from group-level data. Thus, it certainly has a limitation in linking with individual learning performance and brain activity. Also, "surprise" in this study does not seem to capture the nature of "surprise" in the learning process, which is a violation of expectation, as it was calculated with improbability. Moreover, the correlation of Study 1-4 for surprise was not consistent and not strong enough to argue for spatial invariance. Thus, these results may not yet be fully conclusive.

Reviewer #2 (Public review):

Summary:

Learning in dynamic, stochastic environments is difficult, and neuromodulatory systems may shape where learning signals appear in the brain. Using fMRI from four probabilistic learning studies and a Bayesian ideal observer model, the authors examined latent variables driving learning, such as confidence and surprise. They found that brain activity related to confidence, and to a lesser degree surprise, is highly spatially invariant across tasks and modalities, suggesting a stable cortical organization. This invariant pattern aligns with PET-derived maps of receptors and transporters, implicating catecholamine and opioid systems, and supporting a neuromodulatory account of adaptive learning with receptor-level hypotheses.

Strengths:

(1) Elegant combination of computational modelling, functional magnetic resonance imaging (fMRI) and positron emission tomography (PET).

(2) The authors describe results of four separate experiments, with very similar results, in effect providing internal replications.

(3) Cross-validated results compared against a meaningful null model.

Weaknesses:

(1) Unclear rationale for using one-sided statistics (e.g., in Figure 3). One-sided tests appear to be invalid, given that the Introduction lacks a preregistered directional hypothesis at an operationalised level. This may have consequences for the following statement in the Discussion: "The associations between receptor architecture and functional topography were substantially weaker for the language network, which is not thought to rely strongly on neuromodulatory systems."

(2) Limited computational modelling. Since learning rates probably differ across subjects, I wonder if they have considered fitting the "volatility" instead of using the generative one. Would that give more meaningful fMRI maps, and better explained variance when correlating these to the PET-based predictors? I was also wondering how their surprise measure relates to "change-point probability" (e.g., Murphy et al., Nat Neurosci, 2021). Finally, I think it would be helpful to show average time courses of surprise and confidence time-locked to state changes.

(3) Lack of GLM validation. It would help to show that the model fits the data well. This is important given the many underlying assumptions (shape of the HRF, linear effects of variables, etc). For example, one could show average insula activity time-locked to state changes, as well as the model-predicted activity, and separately for three strata defined by how surprising the state change was (according to the ideal observer model). Related, the authors use a substantial number of predictors in their GLM, and the language in the Methods is a little casual. It would help to show part of a design matrix, and clearly describe the following: were (occasional) questions and responses modelled by separate stick functions? Which predictors (stimulus, questions, response) varied parametrically with which variables?

Reviewer #3 (Public review):

Summary:

In this unusual paper, Hodapp and Meyniel relate the spatial topography of activity maps for confidence and surprise (from four learning tasks) to the spatial topography of receptor density maps from atlas data. They find that the brain maps for confidence and surprise are largely consistent across four studies using different stimuli/ task demands. They then use a general linear model to predict the spatial pattern of confidence/surprise-related activity from the spatial distribution of receptors (receptor types) for several neuromodulators. Further analyses test which neuromodulators are most important for predicting the functional maps.

Strengths:

The study gives an interesting new perspective on the brain networks for surprise and confidence, indicating that one reason for the involvement of different networks with these computational parameters is the neurochemical sensitivity of tissue within those networks.

Weaknesses:

I felt the paper was light on context.

To what extent are the distributions of receptor types correlated with each other?

What does the spatial topography of receptor density look like for the identified receptors (NET, MOR, 5HT1b)? Could these be displayed alongside the functional networks? I realise these are atlas data, but for me to interpret the result, I'd need to see the map, and I don't want to download the atlas.

To what extent are the correlations with receptor maps network-wide, vs being driven by one big patch of activity in a single region with high receptor density? To me, this would be important - does this study demonstrate that distant regions united in a functional purpose by shared receptor profile (which would in my opinion be more intersting that the alternative, that there is a single region within each network driving the effect).

Finally, I wasn't convinced by the spin test in this particular application. To my mind, permutation tests are valid when the permuted points are interchangeable under the null. The spin test, as used, preserves the distribution and spatial pattern of activity, but assumes that it could equally plausibly be relocated to any angle on the 3D surface (under the null). However, the brain has a lot of structure that is non-uniform across its surface (connectivity patterns and histological boundaries being important ones). The observed data probably follow this structure, but the 'spun' or permuted datasets probably overlay randomly on the connectivity structure (for example), so that one blob of activity has uniform connectivity in the real data, but overlaps the projections of multiple white matter tracts in the permuted data. But then the permuted data would likely be more heterogeneous in terms of both function and histology than the original data. Since connectivity, histology (layer structure) and receptor density are likely correlated, I think it must be impossible to find verticies that differ in one modality whilst being interchangeable in all others, therefore it may not be possible to use permutation logic to make a claim about (say) receptor density independently of connectivity and histology.

I should add I'm not sure how one would carry out a permutation test that respects the underlying brain anatomy here, or whether this is even possible; that is a difficult question.

I would add that I think the observation that the functional networks have different receptor profiles is interesting, even as a qualitative observation, but not convinced the statistical approach can be justified.