Selective convergence and graded divergence of hippocampal and amygdala subregions using functional connectivity

Hippocampus and Amygdala Subregions Exhibit Structured, Graded Co-representation in the Human Cortex

The hippocampus and amygdala are neighboring structures in the medial temporal lobe. Traditionally, they are tasked with the distinct duties of memory and emotion. While it has long been understood that these regions interact to encode emotionally salient memories, the precise way their internal subregions are embedded within the brain's outer layer—the isocortex (the cerebral cortex)—has remained a mystery. Researchers have struggled to determine if these two systems occupy entirely separate cortical territories or if they share a common functional language.

A new study suggests the answer lies in neither total separation nor total overlap, but in a structured continuum. The researchers found that hippocampal and amygdalar subregions are embedded in the isocortex through a graded pattern of "dominance" and "sharedness." Certain cortical areas are specialized for one structure. Others serve as integrated hubs for both.

The Failure of Discrete System Models

For decades, neuroscience has tended to treat the hippocampus and amygdala as discrete modules. One is seen as a "memory system" and the other as an "emotion system." This modular view assumes that stimulating the hippocampus activates a specific set of cortical networks. It assumes stimulating the amygdala activates a different, non-overlapping set.

However, this dichotomy breaks down when looking at actual brain architecture. Anatomical studies in primates show that both structures send projections to the same medial temporal cortices. These include the entorhinal and perirhinal regions. Most human functional MRI (fMRI) studies have relied on simple pairwise correlations. These are often blunt instruments. A simple correlation tells you that two regions fluctuate together. However, it cannot distinguish between a direct connection and a shared dependency mediated by a third area. Without separating direct functional associations from broad, indirect co-fluctuations, we cannot know if these structures are truly competing for cortical real estate.

Mapping the Continuum of Representation

To resolve this, the authors moved away from viewing these structures as monolithic blocks. Instead, they decomposed them into constituent parts. They studied 18 amygdala subnuclei and 30 hippocampal parcels. These parcels were derived from the longitudinal axis of the hippocampus via the HippUnfold algorithm. They then applied two distinct mathematical lenses to the resting-state fMRI data of 722 participants from the Human Connectome Project.

The first lens was Pearson correlation. This captures "broad co-fluctuation," which is the tendency of regions to rise and fall in activity together. This includes both direct and indirect connections. The second was GLASSO (Graphical Lasso), a method of sparse partial correlation. GLASSO acts like a filter that removes the influence of indirect connections. It provides a cleaner estimate of relatively more direct functional associations [Figure 1B, D].

By applying these two estimators, the researchers introduced two novel metrics to quantify cortical organization: 1. Dominance: A signed index measuring which structure has a stronger presence in a cortical parcel. A value of -1 indicates a parcel is exclusively "hippocampus-like." A value of +1 indicates it is "amygdala-like." 2. Sharedness: A metric quantifying the degree of balanced co-representation. High sharedness occurs when both structures provide strong, proportional inputs to the same cortical area.

This dual-metric approach allowed the team to map not just where the structures differ, but where they converge.

Evidence of Structured Interdigitation

The results reveal a sophisticated, spatially organized landscape. Using the GLASSO estimator, which emphasizes focal, direct connections, the authors found that sharedness is concentrated in a specific "core" of the brain. This includes the orbitofrontal and perirhinal cortices [Figure 2A]. In this core, the amygdala and hippocampus are deeply integrated. Conversely, the dominance maps showed clear "preference zones." The visual network was heavily dominated by the hippocampus. The ventral attention and limbic networks showed strong amygdala dominance [Figure 2B].

Perhaps most striking is the organization found at the subregion level. The hippocampus does not interact with the cortex uniformly. Its connectivity shifts systematically along its long axis. The authors report that anterior hippocampal subregions act as a functional interface. These regions show a much higher preference for "shared" cortical territory. They also show stronger intrinsic coupling (internal communication) with the amygdala [Figure 3A, C]. In contrast, the posterior hippocampus is more strictly tied to hippocampus-dominant networks like the default mode.

Within the amygdala, the paralaminar nucleus emerged as a unique outlier. Despite being part of the amygdala, it exhibited a connectivity profile that was remarkably "hippocampus-like." It sits closer to the shared and hippocampal-dominant zones than any other amygdalar nucleus [Figure 3A]. The study notes that the paralaminar nucleus shows a proximity to the hippocampus that stands out from the rest of the amygdala. Finally, the researchers applied diffusion embedding—a technique to find low-dimensional axes of variation. This demonstrated that subregions exist on continuous "gradients." These gradients capture both the broad separation of the two systems and their fine-scale, interdigitated overlap .

Limits of the Functional Lens

While these findings offer a high-resolution map of connectivity, they are not a complete blueprint of brain function. There are several critical caveats to consider.

First, the study relies on resting-state fMRI. This measures spontaneous fluctuations in blood oxygen levels while a participant is at rest. It does not capture how these subregions engage the cortex during active tasks. It does not track processes like emotional memory formation. The "sharedness" observed here may be the baseline state. It does not prove that these regions always work together during behavior.

Second, the dominance and sharedness metrics are "count-based" summaries. They look at the strongest 10% of connections. As the authors note, this is a measure of relative representation. It is not a direct measure of absolute signal strength or causal influence. A subregion with very weak overall connectivity could still appear "dominant." This could happen if its few connections happen to be the strongest in a specific area.

Third, the signal-to-noise ratio (SNR) in subcortical regions is inherently lower in standard fMRI. This applies to the small nuclei of the amygdala and the folds of the hippocampus. This technical limitation means that subtle differences in connectivity might be obscured. The HCP-YA data were not specifically optimized for subcortical resolution or denoising.

The Verdict: A Unified Framework for Affect and Context

The evidence supports a shift away from the "modular" view of the medial temporal lobe. The researchers demonstrated that the hippocampus and amygdala are not two separate systems. Instead, they are woven into the cortical fabric through a highly structured, graded architecture.

The finding that the anterior hippocampus serves as a bridge to shared cortical spaces is vital. This is supported by the link between local intrinsic coupling and extrinsic cortical sharedness [Figure 3C]. This provides a biological mechanism for how the brain integrates emotional salience with contextual information. For those interested in replicating this work, code is reportedly available; see the paper for the canonical link. This study moves us closer to understanding the brain as a continuous, integrated system. Here, memory and emotion are fundamentally inseparable.

Figures from the paper