Inferring illness causes recruits the animacy semantic network

Reviewer #1 (Public review):

Summary:

In this study, the authors aim to understand the neural basis of implicit causal inference, specifically how people infer causes of illness. They use fMRI to explore whether these inferences rely on content-specific semantic networks or broader, domain-general neurocognitive mechanisms. The study explores two key hypotheses: first, that causal inferences about illness rely on semantic networks specific to living things, such as the 'animacy network,' given that illnesses affect only animate beings; and second, that there might be a common brain network supporting causal inferences across various domains, including illness, mental states, and mechanical failures. By examining these hypotheses, the authors aim to determine whether causal inferences are supported by specialized or generalized neural systems.

The authors observed that inferring illness causes selectively engaged a portion of the precuneus (PC) associated with the semantic representation of animate entities, such as people and animals. They found no cortical areas that responded to causal inferences across different domains, including illness and mechanical failures. Based on these findings, the authors concluded that implicit causal inferences are supported by content-specific semantic networks, rather than a domain-general neural system, indicating that the neural basis of causal inference is closely tied to the semantic representation of the specific content involved.

Strengths:

(1) The inclusion of the four conditions in the design is well thought out, allowing for the examination of the unique contribution of causal inference of illness compared to either a different type of causal inference (mechanical) or non-causal conditions. This design also has the potential to identify regions involved in a shared representation of inference across general domains.

(2) The presence of the three localizers for language, logic, and mentalizing, along with the selection of specific regions of interest (ROIs), such as the precuneus and anterior ventral occipitotemporal cortex (antVOTC), is a strong feature that supports a hypothesis-driven approach (although see below for a critical point related to the ROI selection).

(3) The univariate analysis pipeline is solid and well-developed.

(4) The statistical analyses are a particularly strong aspect of the paper.

Weaknesses:

Based on the current analyses, it is not yet possible to rule out the hypothesis that inferring illness causes relies on neurocognitive mechanisms that support causal inferences irrespective of their content, neither in the precuneus nor in other parts of the brain.

(1) The authors, particularly in the multivariate analyses, do not thoroughly examine the similarity between the two conditions (illness-causal and mechanical-causal), as they are more focused on highlighting the differences between them. For instance, in the searchlight MVPA analysis, an interesting decoding analysis is conducted to identify brain regions that represent illness-causal and mechanical-causal conditions differently, yielding results consistent with the univariate analyses. However, to test for the presence of a shared network, the authors only perform the Causal vs. Non-causal analysis. This analysis is not very informative because it includes all conditions mixed together and does not clarify whether both the illness-causal and mechanical-causal conditions contribute to these results.

(2) To address this limitation, a useful additional step would be to use as ROIs the different regions that emerged in the Causal vs. Non-causal decoding analysis and to conduct four separate decoding analyses within these specific clusters:
(a) Illness-Causal vs. Non-causal - Illness First;
(b) Illness-Causal vs. Non-causal - Mechanical First;
(c) Mechanical-Causal vs. Non-causal - Illness First;
(d) Mechanical-Causal vs. Non-causal - Mechanical First.
This approach would allow the authors to determine whether any of these ROIs can decode both the illness-causal and mechanical-causal conditions against at least one non-causal condition.

(3) Another possible analysis to investigate the existence of a shared network would be to run the searchlight analysis for the mechanical-causal condition versus the two non-causal conditions, as was done for the illness-causal versus non-causal conditions, and then examine the conjunction between the two. Specifically, the goal would be to identify ROIs that show significant decoding accuracy in both analyses.

(4) Along the same lines, for the ROI MVPA analysis, it would be useful not only to include the illness-causal vs. mechanical-causal decoding but also to examine the illness-causal vs. non-causal conditions and the mechanical-causal vs. non-causal conditions. Additionally, it would be beneficial to report these data not just in a table (where only the mean accuracy is shown) but also using dot plots, allowing the readers to see not only the mean values but also the accuracy for each individual subject.

(5) The selection of Regions of Interest (ROIs) is not entirely straightforward:
In the introduction, the authors mention that recent literature identifies the precuneus (PC) as a region that responds preferentially to images and words related to living things across various tasks. While this may be accurate, we can all agree that other regions within the ventral occipital-temporal cortex also exhibit such preferences, particularly areas like the fusiform face area, the occipital face area, and the extrastriate body area. I believe that at least some parts of this network (e.g., the fusiform gyrus) should be included as ROIs in this study. This inclusion would make sense, especially because a complementary portion of the ventral stream known to prefer non-living items (i.e., anterior medial VOTC) has been selected as a control ROI to process information about the mechanical-causal condition. Given the main hypothesis of the study - that causal inferences about illness might depend on content-specific semantic representations in the 'animacy network' - it would be worthwhile to investigate these ROIs alongside the precuneus, as they may also yield interesting results.

(6) Visual representation of results:
In all the figures related to ROI analyses, only mean group values are reported (e.g., Figure 1A, Figure 3, Figure 4A, Supplementary Figure 6, Figure 7, Figure 8). To better capture the complexity of fMRI data and provide readers with a more comprehensive view of the results, it would be beneficial to include a dot plot for a specific time point in each graph. This could be a fixed time point (e.g., a certain number of seconds after stimulus presentation) or the time point showing the maximum difference between the conditions of interest. Adding this would allow for a clearer understanding of how the effect is distributed across the full sample, such as whether it is consistently present in every subject or if there is greater variability across individuals.

(7) Task selection:
(a) To improve the clarity of the paper, it would be helpful to explain the rationale behind the choice of the selected task, specifically addressing: (i) why an implicit inference task was chosen instead of an explicit inference task, and (ii) why the "magic detection" task was used, as it might shift participants' attention more towards coherence, surprise, or unexpected elements rather than the inference process itself.
(b) Additionally, the choice to include a large number of catch trials is unusual, especially since they are modeled as regressors of non-interest in the GLM. It would be beneficial to provide an explanation for this decision.

Reviewer #2 (Public review):

Summary:

In this study, the authors hypothesize that "causal inferences about illness depend on content-specific semantic representations in the animacy network". They test this hypothesis in an fMRI task, by comparing brain activity elicited by participants' exposure to written situations suggesting a plausible cause of illness with brain activity in linguistically equivalent situations suggesting a plausible cause of mechanical failure or damage and non-causal situations. These contrasts identify PC as the main "culprit" in a whole-brain univariate analysis. Then the question arises of whether the content-specificity has to do with inferences about animates in general, or if there are some distinctions between reasoning about people's bodies versus mental states. To answer this question, the authors localize the mentalizing network and study the relation between brain activity elicited by Illness-Causal > Mech-Causal and Mentalizing > Physical stories. They conclude that inferring about the causes of illness partially differentiates from reasoning about people's states of mind. The authors finally test the alternative yet non-mutually exclusive hypothesis that both types of causal inferences (illness and mechanical) depend on shared neural machinery. Good candidates are language and logic, which justifies the use of a language/logic localizer. No evidence of commonalities across causal inferences versus non-causal situations is found.

Strengths:

(1) This study introduces a useful paradigm and well-designed set of stimuli to test for implicit causal inferences.

(2) Another important methodological advance is the addition of physical stories to the original mentalizing protocol.

(3) With these tools, or a variant of these tools, this study has the potential to pave the way for further investigation of naïve biology and causal inference.

Weaknesses:

(1) This study is missing a big-picture question. It is not clear whether the authors investigate the neural correlates of causal reasoning or of naïve biology. If the former, the choice of an orthogonal task, making causal reasoning implicit, is questionable. If the latter, the choice of mechanical and physical controls can be seen as reductive and problematic.

(2) The rationale for focusing mostly on the precuneus is not clear and this choice could almost be seen as a post-hoc hypothesis.

(3) The choice of an orthogonal 'magic detection' task has three problematic consequences in this study:
(a) It differs in nature from the 'mentalizing' task that consists of evaluating a character's beliefs explicitly from the corresponding story, which complicates the study of the relation between both tasks. While the authors do not compare both tasks directly, it is unclear to what extent this intrinsic difference between implicit versus explicit judgments of people's body versus mental states could influence the results.
(b) The extent to which the failure to find shared neural machinery between both types of inferences (illness and mechanical) can be attributed to the implicit character of the task is not clear.
(c) The introduction of a category of non-interest that contains only 36 trials compared to 38 trials for all four categories of interest creates a design imbalance.

(4) Another imbalance is present in the design of this study: the number of trials per category is not the same in each run of the main task. This imbalance does not seem to be accounted for in the 1st-level GLM and renders a bit problematic the subsequent use of MVPA.

(5) The main claim of the authors, encapsulated by the title of the present manuscript, is not tested directly. While the authors included in their protocol independent localizers for mentalizing, language, and logic, they did not include an independent localizer for "animacy". As such, they cannot provide a within-subject evaluation of their claim, which is entirely based on the presence of a partial overlap in PC (which is also involved in a wide range of tasks) with previous results on animacy.

Reviewer #3 (Public review):

Summary:

This study employed an implicit task, showing vignettes to participants while a bold signal was acquired. The aim was to capture automatic causal inferences that emerge during language processing and comprehension. In particular, the authors compared causal inferences about illness with two control conditions, causal inferences about mechanical failures and non-causal phrases related to illnesses. All phrases that were employed described contexts with people, to avoid animacy/inanimate confound in the results. The authors had a specific hypothesis concerning the role of the precuneus (PC) in being sensitive to causal inferences about illnesses.

These findings indicate that implicit causal inferences are facilitated by semantic networks specialized for encoding causal knowledge.

Strengths:

The major strength of the study is the clever design of the stimuli (which are nicely matched for a number of features) which can tease apart the role of the type of causal inference (illness-causal or mechanical-causal) and the use of two localizers (logic/language and mentalizing) to investigate the hypothesis that the language and/or logical reasoning networks preferentially respond to causal inference regardless of the content domain being tested (illnesses or mechanical).

Weaknesses:

I have identified the following main weaknesses:

(1) Precuneus (PC) and Temporo-Parietal junction (TPJ) show very similar patterns of results, and the manuscript is mostly focused on PC (also the abstract). To what extent does the fact that PC and TPJ show similar trends affect the inferences we can derive from the results of the paper? I wonder whether additional analyses (connectivity?) would help provide information about this network.

(2) Results are mainly supported by an univariate ROI approach, and the MVPA ROI approach is performed on a subregion of one of the ROI regions (left precuneus). Results could then have a limited impact on our understanding of brain functioning.

(3) In all figures: there are no measures of dispersion of the data across participants. The reader can only see aggregated (mean) data. E.g., percentage signal changes (PSC) do not report measures of dispersion of the data, nor do we have bold maps showing the overlap of the response across participants. Only in Figure 2, we see the data of 6 selected participants out of 20.

(4) Sometimes acronyms are defined in the text after they appear for the first time.

Author response:

We thank the editors and reviewers for their comments on our manuscript. We found the comments of the reviewers helpful and plan to add new text, analyses, and figures to answer some of the outstanding questions.

In response to the reviewers’ comments, we will clarify the goal of the paper in the introduction: to test the hypothesis that causal knowledge (i.e., an intuitive theory of biology) is embedded in domain-preferring semantic networks (i.e., semantic animacy network). This work links developmental psychology work on intuitive theories and cognitive neuroscience.

As we will emphasize in the revised manuscript, the primary goal of the current paper is to test the claim that semantic networks encode causal knowledge, rather than to rule out the contribution of domain-general reasoning mechanisms to causal inference.

In response to the reviewers’ suggestions, we will add multivariate and univariate whole-cortex analyses that provide further tests for domain-general causality responses. In particular, we will include new figures showing univariate responses to the mechanical inference condition over the non-causal control conditions as well as decoding between these conditions. The reviewers have also asked us to provide individual subject dispersion data. We appreciate this suggestion, and new figures will be added to display this information.

We will also perform additional analysis in the precuneus (PC) to look for shared responses to illness and mechanical inferences. In accordance with our hypotheses, we have shown that the PC responds preferentially to illness inferences. To address the reviewers’ concerns about the selectivity of the PC to illness inferences, we will compare responses to i) illness inferences compared to the noncausal conditions and ii) mechanical inferences compared to the noncausal conditions in the PC to investigate the extent to which a shared response to causal inference across domains emerges in this region.

Critically, we find that the cortical areas that distinguish between causal and non-causal conditions in a ‘domain general manner’ (i.e., for both illness and mechanical inferences) are driven by higher responses to the non-causal condition. Moreover, these responses in prefrontal cortex and elsewhere overlap an RT predictor of neural activity, suggesting that they may reflect difficulty effects.

These results suggest that in the current task, signatures of causal inference are primarily found in domain-preferring semantic networks, rather than in domain-general fronto-parietal reasoning systems. We will provide additional discussion of the argument that the current results do not speak against the role of domain general systems across all types of causal reasoning. Instead, they suggest that the types of implicit causal inferences measured in the current study depend primarily on domain-preferring semantic networks.

The reviewers have asked us to analyze responses to causal inferences about illness in the fusiform face area (FFA). We will perform this analysis. However, we note that univariate and multivariate whole-cortex analyses that are already included in the paper did not identify lateral ventral occipito-temporal cortex as a key region involved in causal inferences about illness. Further, we do not have FFA localizer data in the current participants; therefore, the results cannot be interpreted to reflect activity in functionally defined FFA.

Two reviewers asked us to justify our choice of an implicit magic-detection task, which we will now do more clearly in the manuscript. This task was selected to ensure that participants were attending to the meaning of the vignettes. The goal of the current study was to investigate implicit causal inferences that routinely occur in language comprehension, e.g., when someone is reading a book. Past work has shown that explicitly judging the causality of causal and non-causal stimuli results in differences in response times across conditions (e.g., Kuperberg et al., 2006). In the current study, such judgments would also have introduced a confound between the behavioral decision and the condition of interest: the use of an explicit causal judgment task makes it impossible to know whether any observed neural differences between causal and non-causal conditions are simply due to differences in the selection of task responses. The selection of an orthogonal magic-detection task limits these confounds from complicating our interpretation of the neural data.

One of the reviewers asked us to justify the number of catch trials that we decided to include in our paradigm. Approximately 20% of the vignettes were “magical” vignettes (the same proportion as each of the 4 experimental conditions) to encourage participants to remain attentive throughout the task. Since these catch trials are excluded from analysis, their proportion is unlikely to influence the results of the study. We will clarify this in the manuscript.

A question was raised about the balance of trial numbers across conditions and across runs. To address this, we will include individual comparisons of each causal condition (n=36) with each non-causal condition (n=36; i.e., equal trial counts) where they are not already shown. With regard to runs, each condition is shown either 6 or 7 times per run (maximum difference of 1 trial between conditions), and the number of trials per condition is equal across the whole experiment: each condition is shown 7 times in two of the runs and 6 times four of the runs. This minor design imbalance is typical of fMRI experiments and is unlikely to impact the results. We will clarify this in the manuscript.

We believe that our planned revisions will strengthen the paper and highlight its contributions to our understanding of the neural basis of implicit causal inference.