A frontal cortical world model shapes inference of causal relationships in hippocampus

Technical Summary: A Frontal Cortical World Model Shapes Inference of Causal Relationships in Hippocampus

Problem and Motivation
Causal reasoning requires inferring what leads to what from incomplete and ambiguous observations. While animals and humans can learn cause-effect relationships from single experiences, the nature of these inferences depends on the underlying assumptions or "models" one holds about the environment's causal structure. The authors distinguish between two components: a relational structure, which encodes specific cause-effect relationships, and an abstract world model, which encodes the generative principles governing how these relations are formed and interpreted. Although the hippocampus is known for constructing relational maps in spatial and conceptual domains, and the lateral prefrontal cortex (LPFC) is implicated in abstract rule representation, it remains unclear how these regions interact to support causal inference from sparse data. Specifically, it is unknown whether the hippocampus represents the directional nature of causal relations or merely static associations, and how an abstract world model guides the inference of unobserved causal links.

Methodology
The study employed a novel one-shot causal inference paradigm combined with functional magnetic resonance imaging (fMRI) and transcranial ultrasound stimulation (TUS).

• Task Design: Participants (n=32 for fMRI; n=20 for TUS) performed a task where they inferred whether specific fruits or fruit compounds caused allergic reactions. Crucially, each fruit appeared only once, preventing incremental associative learning and forcing reliance on internal models. The task included four trial types:

  • Backward Blocking (BB): A compound (XY) causes an outcome, followed by one component (X) causing the outcome alone. Participants must infer the status of Y.
  • Unovershadowing (UO): A compound (XY) causes an outcome, but the component (X) does not. Participants must infer Y is causal.
  • Two-cause trials: Two individual fruits independently cause the outcome, designed to shift the participant's world model from a "one-cause" to a "two-cause" hypothesis.
  • No-cause (Control): Neither the compound nor the component causes an outcome.
    Participants provided ratings for unobserved objects before and after observing outcomes, allowing the measurement of "experienced updates" (learning) and "inferred updates" (inference).

• Computational Modeling: A Bayesian normative model was developed to simulate how a prior distribution over causal structures (the world model) combines with observations to update beliefs about unobserved causes.

• Neuroimaging (Experiment 1): fMRI data were analyzed using General Linear Models (GLMs) to identify regions tracking experienced vs. inferred updates. Representational Similarity Analysis (RSA) and repetition suppression analyses were used to examine how object identities and relational structures were represented in the brain.

• Intervention (Experiment 2): Bilateral transcranial ultrasound stimulation (TUS) was used to transiently disrupt the inferior frontal sulcus (IFS) and the hippocampus (HC) in separate sessions. This allowed for causal testing of the dissociable roles of these regions in learning and inference.

Key Results

1. Behavioral Findings:

   • Participants exhibited robust backward blocking and unovershadowing effects, updating their causal ratings of unobserved objects consistent with a Bayesian model.
   • Exposure to "two-cause" trials shifted the participants' world model. Consequently, the magnitude of the backward blocking effect was attenuated in subsequent trials, indicating that the abstract world model dynamically shapes how specific causal evidence is interpreted.

2. Neural Correlates of Learning vs. Inference:

   • Learning (Experienced Update): Activity in the lateral orbitofrontal cortex (LOFC), lateral prefrontal cortex (LPC), dorsal anterior cingulate cortex (dACC), insula, substantia nigra (SN), and ventral tegmental area (VTA) correlated with updates based on direct observation.
   • Inference (Inferred Update): The magnitude of the inferred update specifically correlated with activity in the inferior frontal sulcus (IFS). IFS activity was stronger in trials requiring strong inference (UO) and was modulated by the world model (stronger in BB trials before two-cause exposure).

3. Representation of the World Model:

   • RSA revealed that the right IFS encoded the participant's trial-by-trial belief about the causal structure (the probability of a one-cause vs. two-cause hypothesis). This representation drove the nature of the inference.

4. Reactivation and Relational Mapping:

   • Object Identity: Occipital fusiform gyrus (OccFG) maintained stable representations of object identity. Reactivation of the inferred object in OccFG occurred during the revision phase and was linked to IFS activity.
   • Relational Structure: The hippocampus (HC) did not represent static object identity but rather a dynamic relational structure. In BB trials, the representational distance between the compound and the causal (experienced) object decreased, while the distance to the non-causal (inferred) object increased. In UO trials, this pattern reversed. This geometric updating in the HC was correlated with IFS activity, suggesting the IFS guides the remapping of causal relations in the hippocampus.

5. Causal Dissociation via TUS:

   • HC Disruption: Stimulation of the HC impaired both learning (experienced updates) and inference (inferred updates) but did not alter the pattern of inference. The correlation between learning and inference remained intact, suggesting the HC provides the relational substrate necessary for both processes.
   • IFS Disruption: Stimulation of the IFS left learning relatively intact but selectively impaired inference. Crucially, it abolished the negative correlation between experienced and inferred updates (the standard backward blocking pattern). This indicates the IFS is necessary for implementing the inferential computation itself, likely by encoding the world model that constrains how observations are interpreted.

Significance and Claims
The paper claims to demonstrate that causal inference is implemented through the coordinated activity of dissociable but complementary neural systems:

• The hippocampus instantiates the relational structure, encoding a map of specific cause-effect relationships that is dynamically updated based on new evidence. It supports inference by providing the representational substrate of causal relations but does not perform the inferential computation itself.
• The inferior frontal sulcus (IFS) encodes an abstract world model (generative principles of causal structure). This region guides the inference process by determining how observed evidence is interpreted to infer unobserved relations.

The study concludes that flexible causal reasoning from sparse data relies on the IFS encoding a world model that shapes the updating of relational maps in the hippocampus. This challenges the view that the hippocampus directly performs inference, proposing instead that it maintains the relational knowledge upon which the prefrontal cortex's inferential computations operate.