Once Alzheimer’s disease pathology takes hold, it inexorably spreads through the brain. Frontotemporal dementias are just as relentless, but they march along different anatomical corridors. What’s behind these patterns? Can brain scans trace the underlying spread of pathologic proteins? Could scientists even foretell where and when pathology will surface just by looking at the connectome of a healthy person?
In the March 22 Neuron, two research groups independently published similar models that predict the progression of neurodegenerative diseases. It’s complex science. Authors William Seeley, University of California, San Francisco, and Ashish Raj, Cornell University, New York, talked curious readers through it in an Alzforum Webinar. Marc Diamond, Washington University, St. Louis Missouri; Jorge Sepulcre, Massachusetts General Hospital/Harvard Medical School; and Lary Walker, Emory University, Atlanta, Georgia, joined them for a panel discussion.
Based on functional and structural connectivity in the brain, the models predict that the transneural spread of misfolded proteins through neural networks determines the distinct pattern of degeneration that marks each disease. Diamond and Walker study transmission of misfolded proteins in cell and animal models, and Sepulcre is a brain imaging expert focusing on graph theoretical approaches.
The Alzforum editors gratefully acknowledge Cell Press for making the papers and an accompanying editorial freely available for download to Alzforum readers until 15 June 2012.
- Preview by Warren et al., 2012
- Raj et al., 2012
- Zhou et al., 2012
- Read a series of questions and panelists' answers.—Posted 16 April 2012
- View/Listen to the Webinar
Click on this image to launch the recording.
- Ashish Raj's Presentation
By Tom Fagan and Gabrielle Strobel
What drives the spread of Alzheimer’s and frontotemporal dementia pathology? Over the last two decades, advances in structural, functional, and amyloid imaging have allowed researchers to peer deep into the brain and address this question. Many scientists now believe that pathology spreads not by proximity, that is, from one neuron to its neighbors, but through connectivity, that is, from one neuron to a distant neuron along fiber tracts and across synapses. Multiple lines of evidence support this hypothesis. In AD, the default-mode network, a series of interconnected brain regions that hum when the mind is not focused on a particular task, is also the network most likely to deposit Aβ (see ARF related news story on Buckner et al., 2005). Cortical hubs, the regions most interconnected with other regions of the brain, are hot spots of Aβ accumulation (see ARF related news story). In animal models, too, the most active brain regions are most likely to deposit Aβ pathology (see ARF related news story), as are regions that are most interconnected (see ARF related news story).
Why are interconnected, active regions of the brain more prone? Perhaps toxic proteins, such as Aβ and tau, spread through cells that are synaptically connected. Researchers found, for example, that if they restrict expression of human APP transgenes to the entorhinal cortex (ERC) in mice, dysfunction ensues in the hippocampus (see ARF related news story). That suggested that toxic protein spread along the perforant path from the ERC to the hippocampus. Earlier this year, two groups using a similar strategy of restricting expression of human tau to the ERC found that tau pathology eventually reached the hippocampus (see ARF related news story). That fit with earlier work from Diamond’s lab showing that tau’s propensity to misfold can propagate from cell to cell, even though tau primarily resides inside cells (see ARF related news story). Long ago, Walker and others reported that specks of Aβ planted in the brain can seed plaques (see Kane et al., 2000), and seeds of tau can do the same (see ARF related news story). All this suggests that, in AD and other neurodegenerative diseases, pathology of toxic proteins spreads by templated protein misfolding.
The new models of dementia progression build on this idea. Seeley’s group compared the vulnerable regions in five different neurodegenerative diseases to maps of brain connectivity recorded with functional MRI in healthy people. The researchers found normal patterns of nodal connectivity that closely resembled atrophy patterns in neurodegeneration. The scientists asked what model of disease best explained those matches. They considered four:
- Nodal stress, which is based on deleterious metabolic activity.
- Trophic failure, whereby dwindling neurotrophic support renders cells susceptible.
- Shared vulnerability, which predicts that all affected neurons start off with some intrinsic predisposition to the disease in question.
- Transneuronal spread, which posits that the greatest degeneration occurs closest to the node where disease started.
Nodes were more prone to disease when they were linked by shorter or stronger functional connections. In other words, the transneuronal spread model best fit disease progression.
While Seeley’s group started with the data and asked what models best fit, Raj and colleagues came at it from the opposite direction. They built a biophysical model and fit to it MRI data from healthy, AD, and FTD brains. These scientists propose that toxic proteins diffuse through the brain in a classic mechanism whereby concentration gradients drive their random dispersion. This, they posit, follows brain connectivity networks. Their so-called network diffusion model precisely recapitulated the pattern of atrophy progression of AD and behavioral variant frontotemporal dementia. It might even be able to predict what brain regions will atrophy in the future based on baseline MRI morphometrics, Raj and colleagues write.
How can these new models help researchers? Can they aid in diagnosis or prognosis? The authors claim the models could even infer population-wide prevalence rates of certain types of dementia whose variable clinical phenotype often leads to wrong diagnoses. What has to happen between the current data and the future promise?
Q: I am wondering whether Ashish used the age-matched normal controls to repeat this great work?—Liang Wang, Postdoc, Princeton University
AR: We used age-matched controls to investigate the correlation between the first eigenmode and normal aging. The result was shown in Figure S2. However, all connectivity data were derived from young healthy volunteers.
Q: Raj et al. should model "incidence" instead of prevalence because incidence (new cases per person-time) reflects the rate of disease occurrence, whereas prevalence is only a measure of the proportion of individuals in a particular population who are "affected" at a point in time. Strictly, then, prevalence is dependent on incidence rate and disease duration/survival. Thus, prevalence could vary by availability of healthcare, which influences survival with disease.—Bud Kukull, Professor of Epidemiology, University of Washington, Seattle
AR: Thank you for the suggestion. When we were researching published prevalence and incidence rates, we found it difficult to obtain reliable and adequate number of incidence data. We agree prevalence rates have other confounders, but so do incidence rates. In order to minimize the effect of healthcare systems and demographics, we mainly included studies done in advanced Western nations whose healthcare delivery systems can be considered comparable.
Q: Access to some of the mathematical aspects presented by the second speaker, beyond an equation, would be helpful. There is certainly much interest in the approach, and there are potentially exciting developments; however, we do need to be careful not to overinterpret findings. For example, R-squared values indicating that 8-20 percent of the variance explained by the observed data may be "significant," yet also begs the question of what explains the other 80-90 percent.—Bud Kukull, Professor of Epidemiology, University of Washington, Seattle
AR: At some point after current research/publications in the pipeline are cleared, we intend to make our methods available to the public. Regarding R-squared values: In our field, we do not expect perfect correlations, due to measurement noise, post-processing artifacts, natural inter-subject variability, and yes, model approximations. The main tool we have to assess whether a model is informative is by assessing significance (p) values, which were significant in our data. I agree that we should not overinterpret the findings due to moderate R values.
Q: The statistical R2 values are too small to make a reliable conclusion.—Kailash Thakur, Modeller at Land Care Research, Phoenix Arizona
AR: See previous response. Again, I take your point, but our numbers are comparable to many, many other publications in the field.
Q: How do you initiate the model (x0)? You must need some initial pathology to drive the diffusion?—Andrew Reid, Postdoc, McGill University, Montreal, Canada
AR: Good question, Andrew. You may have noticed that I have never specified or talked in any great detail about the initial configuration x0. We discussed the reason during the presentation: It is going to be determined by the pathology, the cocktail of misfolded proteins, and their preferred areas of attack. My network diffusion model does not say anything informative about how/where the disease starts. It merely claims that, once disease has taken hold in some starting configuration, the succeeding progression within the network, and the eventual atrophy patterns, will be determined roughly by the eigenmodes and eigenvalues. One could extend the argument and say that the starting configuration will fully determine which eigenmode (hence, which type of dementia) the person is going to get.
Q: You say that higher-order eigenmodes/values are less reliable. Why does AD emerge as the first eigenmode (i.e., the most reliable)? Is it dependent on the way the model was set up? That is, could another way of setting up the model yield other eigenmodes as lower order?—Iris Oren, Postdoc, University College London
AR: I believe the answer is persistence: There is a solid argument about why the most persistent eigenmode should cause the most damage, hence, may be attributed to the most common/widespread dementia. This happens to be Alzheimer's disease. Isn't it fascinating that this correspondence between first eigenmode and AD came about completely from a mathematical model of healthy network diffusion? Please note that this argument has nothing to do with reliability of the eigenmode. It just so happens that higher eigenmodes are increasingly more localized, and increasingly capture noise present in the network. This is a well-known feature of the spectrum of a graph Laplacian, and not a unique result of our methodology.
Q: Have you looked at the nature of connections between regions (i.e., glutamatergic/cholinergic/dopaminergic)? If not, do you intend to? Do you think that this might provide information to your model?—Iris Oren, Postdoc, University College London
AR: This is a nice idea, Iris. As I have discussed in my talk, the minutiae of the network diffusion model (and its rate constant β) would necessarily depend on both neuropathology and the biochemistry of regional variations. But I strongly believe that the overall model would continue to hold, for the simple reason that diffusive processes are reliable and reasonable stand-ins for a large variety of dispersive phenomena in physics, biology, and neuroscience. Further details of the kind you note will likely fine-tune the model, but are unlikely to up-end it.
WS: We have not pursued the neurochemical influences on network organization, but I agree this could prove interesting.
Q: Why is it necessary to invoke a diffusion of pathological agents for correlated pathology among spatially remote regions? Would we see the same consequence of correlated pathology if two regions were no longer communicating? Region A dies, and, eventually, region B dies through lack of afferent activity.—Alan Evans, Professor, McGill University Montreal, Canada
WS: Although correlated pathology does occur in neurodegenerative disease, this was not what we studied. We looked at how correlated activity in the healthy brain (functional connectivity) predicts region-wide atrophy in the diseased brain. In essence, we sought to determine what kind of healthy network nodes prove most vulnerable. The model you describe (Region A’s death begets region B’s death due to lack of afferent activity) evokes the “trophic failure” hypothesis that we tested. In our view, this model predicts that eccentric nodes with sparse connections (low total flow) lack redundant trophic inputs and should prove most vulnerable once disease invades the network. Furthermore, the model predicts that nodes lacking highly connected neighbors (low clustering coefficient) may find themselves lacking in afferent input. Neither of these predictions was supported by our data. There may be other formulations of the trophic failure model, however, and we would be eager to entertain alternative viewpoints.
Q: Can your data or model be used to predict which patients will progress more rapidly than others?—Emily Rogalski, Assistant Professor, Northwestern University
AR: Yes, if the model is correct we should be able to "play out" people's future atrophy and see how quickly they progress. However, note that the model contains a rate constant β, which is unknown a priori, and its value could be different in individuals as well as in different diseases and pathologies. For instance, it is possible that Aβ and tau may have different rates of diffusion in the same network. This rate constant will need to be determined first.
WS: Not yet, but that is exactly where we would like to take this line of investigation.
Q: Do you foresee the efforts to map the human connectome shedding light any time soon on the tendency of Alzheimer's, frontotemporal dementia, Parkinson's, and other forms of neurodegeneration to attack specific areas of the brain while sparing others?—Tom Valeo, Wwriter, Neurology Today
WS: Maybe. We hope to pursue this important question in future studies. It could be that, while all of the diseases spread in a similar fashion, their points of origin differ in some meaningful way that we can identify by examining the network characteristics of the “epicenters” to understand what makes each disease’s epicenter different from all of the others. The other, perhaps more likely, possibility is that there are defining characteristics of each epicenter that the MRI scanner cannot “see,” requiring a cellular-molecular level approach. We are also pursuing this line of investigation.
Q: Braak has suggested that locus ceruleus might be the first place of spreading of pathology, putting the noradrenergic system prior to cholinergic? Have we been misled by the cholinergic theory? Do the nodes you describe correspond to Marcel Mesulam's node concept of projection hierarchy, such as cortical primary cortex to cortical tertiary cortex, where nodes in the more complex cortex are polymodular in functional character? Is this similar to what you see?—Nenad Bodganovic, Medical Director, Pfizer, London
WS: The network hierarchies proposed by Mesulam can indeed be examined, at least in part, using human network mapping strategies. One of the prevailing limitations of these methods, however, is that they cannot distinguish monosynaptic “connections” from those that require multiple synapses to produce the correlated functional signals we measure.
Q: Your paper provides evidence favoring one hypothesis over another. Have you been able since then to formulate a more quantitative test of the hypotheses?—Amy Kuceyeski, Postdoc, Weill Cornell
WS: We entered all graph metrics we studied into regression analyses that determined which metric explained the most variance in atrophy severity. This approach to hypothesis testing was highly quantitative. Nonetheless, the proposed relationships between the mechanistic models and the graph metrics were based on several assumptions. To test the proposed models directly will require an experimental system that works at the cellular level, but such a system is not feasible in living human beings. Therefore, we accepted the limitations of our experimental design and judged that the knowledge gained might help build a dialogue between cellular level and network neuroimaging researchers. Convergent evidence from multiple lines of inquiry and levels of analysis is always the most compelling.
Q: Interesting model...but it does not take into account the role of glia. These disorders are associated with chronic inflammation. How is the contribution of chronic inflammation or glia to the spread of these diseases factored into these theoretical models? Could local initial inflammation play a role in the spread?—Maria Figueiredo-Pereira, Professor, City University New York
AR: Our model only addresses the question of proteopathic transneuronal transmission. None of the other etiologies is considered or modeled. Dr Seeley's paper suggests that, of all possible explanatory hypotheses regarding dementia patterns, the transneuronal spread mechanism appears to provide the best fit. My own belief is that, regardless of etiology and pathology, once the disease has taken hold, subsequent progression along the network can be well captured by network diffusion. The fact that our model predictions appear to be matched by observed patterns of dementia suggests that it is not necessary to invoke glial, inflammatory, and vascular stress processes in order to explain the observed patterns. This does not mean that these processes do not happen in dementia, but simply that these processes may not play a critical role in the spatial expression of the disease.
WS: You are correct that glial responses and neuroinflammation are not factored into these models, and it is quite likely that glia play a role in shaping the severity and perhaps topology of neurodegeneration. One interesting possibility is that glia take up the prion-like proteins as they are released from dead or dying neurons and prevent these proteins from spreading out like a wave over contiguous brain structures.