Global brain efficiency fails to predict general intelligence in large study

Recent research challenges the popular idea that general intelligence relies on the brain’s overall efficiency or “small-world” architecture. A new study demonstrates that broad, whole-brain measures of network organization fail to predict cognitive ability. Instead, the specific connection patterns of individual brain regions drive the relationship between neural architecture and intelligence. This work appears in the journal Cerebral Cortex.

For over a century, scientists have debated the biological origins of general cognitive ability (GCA). This is often referred to as general intelligence. It represents the observation that individuals who perform well on one type of cognitive task tend to perform well on others. Early neuroscience attempted to locate intelligence in specific brain areas. Later, the focus shifted to the connections between these areas.

Modern researchers view the brain as a complex network. They call the map of these neural connections the “connectome.” To make sense of this vast web, scientists use a branch of mathematics known as graph theory. This mathematical approach simplifies the brain into a set of nodes (regions) and edges (connections).

Graph theory allows researchers to calculate specific properties of the network. Some properties describe the entire network at once. For example, “global efficiency” measures how easily information can travel across the whole brain. Other properties function at the local level. These “node-level” measures describe the role of a single brain region within the larger structure.

Previous studies yielded mixed results regarding how these measures relate to intelligence. Some researchers reported that smarter brains are more globally efficient. Others claimed that “small-worldness,” a balance of local clustering and long-range connections, was the key. However, these earlier investigations often suffered from small sample sizes. Many included fewer than one hundred participants. This lack of data led to inconsistent findings and made it difficult to draw firm conclusions.

To resolve these discrepancies, a team of researchers conducted a massive investigation. The study was led by M. Fiona Molloy and Chandra Sripada from the Department of Psychiatry at the University of Michigan. They utilized two of the largest neuroimaging datasets available.

The first dataset came from the Adolescent Brain Cognitive Development (ABCD) study. It included functional magnetic resonance imaging (fMRI) scans from 5,937 children aged 9 to 10. The second dataset served as a replication sample. It came from the Human Connectome Project (HCP) and included scans from 847 adults.

The researchers analyzed “resting state” fMRI data. This method records brain activity while participants are awake but not performing a specific task. It reveals the intrinsic functional architecture of the brain. The team then assessed how well different graph theory measures could predict each participant’s GCA score.

The investigators first examined whole-brain measures. They calculated nine different properties, including global efficiency and small-worldness. They looked at both positive connections (regions activating together) and negative connections (regions where one activates as the other deactivates).

The results were clear. None of the whole-brain graph theory measures showed a meaningful relationship with general intelligence. Even when the researchers combined all global measures into a single predictive model, the association remained negligible. The idea that a smarter brain is simply a more “efficient” or “small-world” network across the board was not supported by the data.

The researchers then turned their attention to node-level measures. These metrics do not assign a single number to the whole brain. Instead, they assign a value to each distinct brain region. This approach asks how specific areas connect to their neighbors.

Here, the findings were quite different. Fifteen out of the sixteen node-level measures evaluated in the ABCD dataset showed a statistically significant relationship with intelligence. The researchers found that properties of individual nodes were robust predictors of GCA.

One specific measure stood out as the strongest predictor. This measure is called “within-module degree.” The brain is organized into communities, or modules, of highly interconnected regions. Within-module degree quantifies how well a specific node communicates with other nodes in its own community.

The study found that higher within-module degree in certain regions predicted higher intelligence. These regions included the temporal poles and the cerebellum. Conversely, in other areas like the striatum and parts of the temporal cortex, a lower within-module degree was associated with higher intelligence. This suggests that for some brain regions, being tightly integrated with their local community supports cognition. For others, a different pattern is optimal.

The researchers validated these results by replicating them in the adult HCP dataset. The patterns observed in the children were largely present in the adults. This successful replication across two independent samples with different age groups provides strong evidence for the reliability of the findings.

The team also compared these graph theory approaches to a “full connectome” model. The full connectome model uses every single connection in the brain—over 87,000 distinct weights—to predict intelligence. This comprehensive model predicted GCA scores with the highest accuracy.

The aggregated node-level graph measures captured about 36 to 39 percent of the predictive power of the full connectome. This indicates that while node-level graph theory simplifies the data immensely, it preserves a substantial portion of the signal relevant to intelligence. It offers a more interpretable summary than looking at thousands of individual wires.

There are limitations to consider. This study focused exclusively on the resting brain. Brain networks reconfigure themselves when people perform active tasks. It is possible that global network properties become more relevant during complex problem-solving.

Additionally, the researchers used a standard atlas to define brain nodes. This atlas groups brain areas based on average functional connectivity. However, individual brain anatomy varies. Future research might benefit from using personalized brain maps for each participant.

Another consideration is the interpretation of “general intelligence” itself. Cognitive ability is intertwined with environmental factors. These include socioeconomic status and educational opportunity. The biological signatures identified here likely reflect a combination of innate neurobiology and life experience.

Despite these caveats, the study provides a definitive update to the neuroscience of intelligence. It moves the field away from simple, whole-brain explanations. The results suggest that the neural basis of intelligence is not a global property like overall efficiency.

Instead, intelligence appears to be supported by a distributed pattern of local connectivity. Specific regions must be integrated into their local networks in precise ways. This nuance was lost in earlier, smaller studies but becomes visible with large-scale data.

The study, “Regional, but not brain-wide, graph theoretic measures are robustly and reproducibly linked to general cognitive ability,” was authored by M. Fiona Molloy, Aman Taxali, Mike Angstadt, Tristan Greathouse, Katherine Toda-Thorne, Katherine L. McCurry, Alexander Weigard, Omid Kardan, Lily Burchell, Maria Dziubinski, Jason Choi, Melanie Vandersluis, Cleanthis Michael, Mary M. Heitzeg, and Chandra Sripada.