- Nature Neuroscience
- 15, 171–17 (2012) doi:10.1038/nn.3031
- Published online
- 26 January 2012
“Brain regions exhibit specialization for different functions, but such functions are constrained by anatomical connections to other brain regions. A study now finds that, by measuring these connections, we can predict complex functional responses before the subject has even performed the task.
Most people associate London’s Baker Street with Sherlock Holmes, but a look inside its Tube station reveals a peculiar sight. In one section, people are rushing around with suitcases and stony faces, but barely 50 yards away they are sauntering about with friends, cameras and shopping bags. This incongruity may seem odd until you examine the Tube map: passengers can either take the pink line to the financial district or the brown line to the tourist meccas of Soho and Oxford Street. The different connections at Baker Street dictate the type of activity at different corners of the station.
An analogous situation occurs in our brains. Gray matter areas are densely interconnected, and the details of neuronal activity in any one region depend on its connections to distant regions1. Incoming connections dictate the type of information to be processed, and outgoing connections dictate the influence this processing can have on other brain regions. Recent advances in magnetic resonance imaging allow us to measure correlates of these anatomical connections in the living human brain using a technique called diffusion tractography2. In this issue of Nature Neuroscience, Saygin et al.3 show that it is possible to predict complex functional responses using only these anatomical connections.
Given that connections impose functional constraints, it is perhaps not surprising that the pattern of each region’s connections is unique1. By mapping these connectional fingerprints, it has been possible to find places where connectivity exhibits sharp changes4. These points align closely with the boundaries between functional5 and cytoarchitectonic6 regions, suggesting that we may be able to measure these regions, à la Brodmann, but now non-invasively and in vivo.
Saygin et al.3 take this idea substantially further (Fig. 1). They attempted to reconstruct the entire spatial topography of a functional response using only information about connections. As a test case, they chose one of the most famous neural responses in human neuroscience: the face-selective visual response. In monkeys, there are face-selective cells throughout much of the temporal lobe7. Similarly, when a human subject views an image, a complex pattern of blood oxygen level–dependent activity can be seen in the ventral temporal cortex that distinguishes faces from other stimuli8. Crucially, however, when this response is averaged across different subjects, much of the detail is obscured, leaving a band of activity in fusiform gyrus, centered around the notorious fusiform face area9. Saygin et al.3 argue that differences in the detailed response patterns across individuals are a result of differences in anatomical connectivity. If we only knew what connections made for face-selective tissue, we could measure these connections in a new subject and predict his or her face-selective response.
This argument, then, appears to fall at the first hurdle. How can we know what connections lead to face-selective tissue? We might get some inspiration by thinking back to Baker Street, where the suits connect to the business centers and the cameras to the tourist zones, but face processing is a far more complex problem to dissect into inputs and outputs. Saygin et al.3 had a much cleverer solution. They learned face-selective connectivity patterns from previous examples. In essence, they learned a model of the relationship between connections and function.
The authors measured face selectivity in a cohort of subjects and, subsequently, collected diffusion-weighted magnetic resonance images to infer probabilistic connections. In each imaging voxel in the inferior temporal cortex, they inferred connection probabilities to a set of predefined target regions and tried to establish a relationship between these connections and face selectivity. For example, face-selective voxels might, on average, have strong connections to specific prefrontal and parietal regions, but nonselective voxels might have denser connections to regions in the cingulate or somatosensory cortices. To test whether this is the case, Saygin et al.3 used the set of connection probabilities as predictors in a linear model fitted to the map of face selectivity. Each target brain area received a regression weight that revealed the extent to which its connections predict face selectivity in inferotemporal voxels. These weights are of course exactly what we have been looking for: a model for how brain connections influence face selectivity. They tell us which Underground connections are associated with bankers and which with tourists.
The important question is, given a new subject, can we use this model to predict the functional responses? More precisely, is function better predicted by the connections than by spatial or geometrical features of the tissue? Connection strengths decay exponentially with distance10, making these geometric controls crucial. The authors performed a number of tests to confirm that this is indeed the case. For example, the connectivity weights for a new individual predicted functional details that were lost in the spatial group average, but could be seen in the individual’s response. Such details could not be predicted by linear models that only consider geometric features of target regions. Individual differences in function are indeed predicted by individual differences in connectivity.
This demonstration is perhaps the clearest yet in a line of argument that has widespread implications for how we think about neural processing at a systems level. For over half a century, researchers have debated between two extreme models of neural processing. Functional specialization suggests that individual brain regions have responsibilities for precise aspects of neural processing. Functional integration argues that information processing occurs through complex interactions of many brain regions. This debate has been reinvigorated following recent advances in brain imaging9. It is perhaps appropriate that the current study focuses on the inferotemporal cortex, the center of the most heated debate11, 12, 13, 14, but what is most notable about this line of research is that it uses the logic of functional integration to predict functional specialization, providing a clear bridge between these two extreme positions.
This study also raises important questions about how we design experiments and interrogate our neural data. Single-unit physiologists place their electrodes according to anatomical landmarks. Imagers align functional responses across subjects on the basis of brain geometry alone. Surgeons implant stimulating electrodes on the basis of coordinates from group activation studies or successful implants in other individuals. In all of these cases, it is now clear that measurements of brain connectivity would account for an important source of variability15. Saygin et al.3 bring this concept into sharp focus for brain imaging data. By mapping onto a purely spatial template, we lose a great deal of detail that is present in individual responses, and we are left to interpret only the spatial peaks that are consistent across subjects. Saygin et al.3 show that key details are retained if we choose a template that is not spatial, but is instead connectional.
These results come at a time when systems-level brain connectivity is at the forefront of many neuroscientists’ minds. Major funding efforts in both the US (http://www.humanconnectome.org/) and Europe (http://www.brain-connect.eu/) aim to make substantial improvements in in vivo techniques for measuring regional brain connections, and in understanding their effect on neural processing. Saygin et al.’s findings3 are a clear demonstration of the importance of such endeavors and should pave the way for studies investigating the intricate interaction of structure and function across a wide range of neural processes.”