Light Them Up: Imaging Biomarkers with Radiotracers
New markers candidates such as PAK, drebrin, arc, or calbindin would be more powerful if there was a way to see them in human brain. Henry VanBrocklin of University of California, San Francisco, described how the AD field could begin working toward that goal by using a combination of chemistry and radiological imaging, and he quoted ongoing work in Parkinson disease (PD) as an example. To stimulate early efforts in the AD field, VanBrocklin gave an overview of the development of radiopharmaceutical probes and their application in drug discovery. It is none too soon to start, as developing radiotracers for clinical use can be almost as lengthy and expensive a process as developing a drug.

Two recent changes at the FDA may expedite things, VanBrocklin said. For one, its Critical Path Initiative is part of an effort to get more drugs into the pipeline. Besides identification of biomarkers, its key component is imaging biomarkers or surrogates to assess the value of experimental drugs earlier in the process. In principle, radiotracers for clinical use are subject to the same development process as are new drugs, but recently, the FDA has introduced the “Exploratory IND.'' It lowers the threshold for getting new pharmaceuticals, and also new radiotracers, into humans for study, and represents an effort to cut the time this usually takes from a decade to a year or less, VanBrocklin noted.

Researchers generally use two ways to incorporate radiopharmaceutical probes into the drug development process. They either label the drug itself, with tritium or 14C, to see where drug is and how it gets metabolized. Or they label a receptor system or enzyme that the drug perturbs. In this regard, too, it is important to distinguish between biochemical targets and disease-related targets (see Lieberburg lecture).

Much like development of any other biomarker, radiotracer development involves three steps. The scientist must assure that the tracer hits the target of interest, that it is sensitive enough to track the expected change in target density (see Klunk presentation below for an example where this failed), and to validate it in various animal models of the disease.

For radiotracers to work, the density of the intended target must be adequate so one can see the tracer. If a tracer is to label a low-density receptor, its specificity must be high, and its on-off rate must be within a certain range. It must be selective to avoid nonspecific binding reducing the contrast between background and signal. Blood-brain barrier penetration is vital, and its metabolism must be understood because metabolites can cause background activity.

Neural receptor proteins have been a cornerstone of neuroimaging, and neuropeptides, pre- and postsynaptic neurotransmitter systems, and transporters have seen a variety of tracers made against them, as well. Few of those were ever fully developed and validated, however, VanBrocklin noted. One example relevant to AD is the α7 nicotinic acetylcholine receptor, which a growing number of studies are reporting to be lost in AD-relevant brain areas early in the disease (see Oddo et al., 2005 for a recent demonstration in a triple-transgenic mouse model of AD). Researchers at Johns Hopkins University, working in a joint project with Novartis, recently published the first imaging agent for it (Pomper et al., 2005). Using 11Carbon, its affinity and selectivity for the receptor were good but did not penetrate the brain well. “This is a start, and it is an example for where partnerships between industry and academic PET labs can make great advances,” VanBrocklin said.

Another area that has benefited from neuroimaging with radiolabeled tracers is the study of enzyme activity. The key example in PET is glucose metabolism with 18-fluorodeoxyglucose (FDG), which clears the blood-brain barrier by means of the glucose transporter. This tracer differentiates AD from frontotemporal dementia.

Radiotracer imaging can complement gene therapy, where a therapeutic gene together with a reporter gene gets packaged into a viral vector, introduced into cells, and a radioligand probe picks up expression of the transgenes. For example, the dopamine D2 receptor, a target of antipsychotic drugs, has been used for such an approach. Ideally, the therapeutic and the reporter gene should be the same, and this is the goal of an ongoing study for Parkinson’s conducted jointly by VanBrocklin and collaborators at UCSF and UC Berkeley. The researchers use the enzyme amino acid decarboxylase, which converts L-dopa to dopamine, and detect its expression with 6-fluorometatyrosine (FMT) or fluoro-dopa PET imaging. Such an approach would make it possible to reduce the dose of current L-dopa therapy. This works in a primate model of MPTP lesion on one side of the brain, VanBrocklin said. Both the FMT PET signal and some function returned after injection of the construct into the striatum, and the approach has moved into a phase 1 trial, VanBrocklin added.

Beyond this particular approach, a host of small-animal imaging devices mimicking the human imaging armamentarium has sprung up for use in translational studies in recent years. Micro-PET, microMR, microCT, microSPECT, as well as ultrasound, fluorescence, and bioluminescence are all in place for testing preclinical radioligands in animal models, VanBrocklin said. PET, for example, now works well in mouse, rat, monkey, and human, and thus can support drug discovery all along the way from the bench to the bedside. VanBrocklin finished by saying that imaging agents should be plugged into the development program of a drug as early as possible.

Stream of Consciousness Network for Diagnosis?
Even while researchers have been whipping into shape techniques for animal brain imaging, human brain imaging has made its own advances toward new diagnostic methods. Perhaps the most novel approach comes partly out of Michael Greicius’s group at Stanford University in Palo Alto, California. Greicius is working on a way of exploiting a network of brain activity during mental rest, or free association, for a functional imaging method to predict and monitor AD.

First, some background on fMRI of active and deactivated brain areas. Unlike PET, functional magnetic resonance imaging (fMRI) requires no injection of a radioactive tracer or contrast agent but instead relies on an intrinsic signal contrast known as the blood oxygen level dependent (or BOLD) signal. fMRI exploits different magnetic resonance properties of oxygenated versus deoxygenated hemoglobin, and it can measure brain function because active brain regions contain more of the former.

In a typical fMRI experiment, a person performs a cognitive task in an alternating 30-second task-rest pattern for a few minutes. This generates a “task wave form,” and a statistical program then searches the brain for regions whose BOLD signal time series correlates with this wave. fMRI has relatively low spatial resolution but when overlaid on a high-resolution structural MR scan, the wave signal yields an activation map. This standard scan looks for regions of activation whose BOLD signal increases during the task period and decreases during rest periods.

Deactivation is the opposite. It occurs in regions where the BOLD signal increases during the rest periods and decreases while the person performs the task. In essence, these brain areas get shut off while a person focuses on the task at hand. Deactivation drew interest because it is a consistent phenomenon, whereby the same set of brain areas emerges during rest periods across different tasks and different subjects. Marcus Raichle at Washington University in Saint Louis, Missouri, first put deactivation on the field’s collective radar screen with a study showing that the posterior cingulate cortex, the inferior parietal lobes, and the medial prefrontal cortex were all deactivated across a variety of cognitive tasks (Raichle et al., 2001). Raichle suggested that these regions constitute a network whose activity is suppressed when one has to perform a cognitively demanding task, and he called it the default mode of brain function.

Of course, the resting brain is not truly resting. Brain activity continues even in the absence of a task cued from the outside. How do scientists measure that? In resting-state functional connectivity MRI, scientists do not search the brain with the activation task wave form; instead they pick spontaneous BOLD signal oscillations emanating in a particular region of interest (called the seed region) as the wave form and search the wider brain for regions that correlate with it. Such regions are considered functionally connected. The scientist first isolates that brain area in an activation scan, then the subject is scanned during 5 minutes of rest. To generate the connectivity map, the investigator takes the spontaneous resting-state activity from the seed region (for example, the motor cortex as defined by a preceding finger-tapping task) and searches the brain for regions whose BOLD signal is tightly correlated with it. This provides a resting state network for regions connected to the motor cortex. Subsequent work has convinced the field that these networks truly represent functional connections between brain areas and not mere blood flow artifacts, Greicius noted.

Greicius’s group used this kind of functional connectivity imaging to test Raichle’s hypothesis of the resting-state default mode. They had healthy young adults perform a working memory task to isolate the posterior cingulate as a region of deactivation. Then they used that as a seed region during a second, resting-state scan to look for connected areas. This analysis confirmed that the posterior cingulate indeed formed resting-state connections with all the brain regions Raichle had proposed to be part of the default-mode network.

Different resting-state networks linked to particular functions, such as motor, language, etc., exist. Among them, the default-mode network for memory is the only one that is normally active but needs to be suppressed when a person performs a cognitively demanding task, Greicius said. It has since become clear that most cognitive tasks—working memory, calculation, matching—suppress this network. An intriguing exception is tasks that involve the retrieval of memories; such tasks actually activate this network. Taken together, these findings suggest that the network mediates “stream of consciousness” processing, a silent mode of reminiscing and mulling over recent events in which we spend quite a bit of time every day, but which we suppress when focusing our attention on a specific task.

The network became interesting for AD research when it turned out that the regions involved in this default-mode network overlap remarkably well with regions of decreased metabolism in AD. Along with hippocampal atrophy, the most widely accepted imaging finding in AD is that, when scanned at rest, people with AD have decreased metabolism in the posterior cingulate and inferior parietal lobes on both sides (e.g., Alexander et al., 2002). This raised the question of whether examining resting-state, default-mode activity in these areas might become a diagnostic marker.

Task-activation fMRI is a highly specialized form of imaging that is not used in standard clinical settings. Resting-state fMRI is easier because one need not feed stimuli into the scanner and analyze the person’s responses. As a step toward its use in clinical diagnostics, Greicius developed a simpler, automated means of detecting the network using a statistical method called independent component analysis (ICA). Applied to a publicly available fMRI data set that includes AD patients, the method visualized a robust default-mode network in young and elderly healthy volunteers but a greatly degraded network in the AD group (see Greicius et al., 2004 and, for more detail, ARF related news story). For further free fMRI data sets, Greicius referred researchers to the fMRI Data Center maintained at Dartmouth College.

Group data as in this study are different from what ultimately matters most, that is, a method the neurologist can apply for a patient who comes to the clinic. To this end, the researchers tested whether a measure called the goodness-of-fit score could be useful in distinguishing single AD patients from healthy controls. This score compares the strength of the network measured in a given person to a standard template of the network averaged from a set of healthy controls. Based on it, this first-pass study achieved 85 percent sensitivity and 77 percent specificity, Greicius noted.

Since then, scans with a separate set of 18 patients and 13 age-matched controls have confirmed this first finding. Current work with other patients aims to probe the method’s ability to distinguish AD from other forms of dementia. Early indications from this ongoing, preliminary work are that changes in the posterior cingulate cortex might not only distinguish AD from frontotemporal dementia but also correlate with MMSE scores, Greicius added.

Greicius closed by pointing to a recent collaborative study by Randy Buckner’s group and Klunk and Mathis, which he said convinced him of pursuing this network’s disruption in AD (Buckner et al., 2005). In it, the scientists showed that the default-mode network defined by deactivation, that is, the brain areas that are jointly suppressed while a person focuses on a cognitive task, showed remarkable damage in AD across three separate methods of imaging. Their glucose metabolism was down, as was their volume, and they also comprised the brain areas (posterior cingulate and inferior parietal cortex) that retained the most PIB, that is, had a high amyloid load. For some reason, then, the stream of consciousness network appears to bear the brunt of AD pathology.

In summary, Greicius suggested that resting-state fMRI could be repeated at short intervals in the same person because it is relatively easy to perform and requires neither radiation exposure nor a nearby cyclotron. Ultimately, it could complement the standard structural MRI that is becoming part of the routine workup of patients in memory clinics.

No Bed of Roses: Making PIB Work in Mice
Bill Klunk, of the University of Pittsburgh, Pennsylvania, is best known these days for the success, so far, of the amyloid imaging agent PIB in human studies (Klunk et al., 2004). At the workshop, however, Klunk talked about one of his failures. It is a cautionary tale for any scientist trying to decide how much to stake on mouse studies in a translational biomarker research program.

Klunk and his colleague, radiochemist Chet Mathis, started their search for an amyloid tracer in the late 1980s, first with derivatives of Congo red and then, in the late 1990s, with derivatives of the amyloid-binding dye thioflavin T called BTA compounds. Tinkering with the compounds’ side groups, they gradually improved their affinity to amyloid in the test tube, as well as their ability to enter the brain and leave it again within 10 minutes. They finally settled on a 6-hydroxy derivative called Pittsburgh Compound B.

Klunk and Mathis’s first mouse study worked just fine. They teamed up with Brian Bacskai and Brad Hyman at Massachusetts General Hospital, Charlestown, to validate this compound with multiphoton imaging of live, plaque-ridden Tg2576 mice. This allowed them to watch over the course of 30 minutes how the fluorescent compound labeled the blood vessels, diffused out of the vessel into the brain parenchyma, lit up plaques, and then disappeared from the brain (Bacskai et al., 2003; see image at ARF related news story). Binding studies using nanomolar concentrations of Pittsburgh Compound B with brain homogenates confirmed that the compound selectively recognizes amyloid plaques, not tau.

PIB has since moved into human studies to a stage where centers around the world are beginning to report similar results on its ability to distinguish early AD and MCI from normal aging and related conditions (see ARF related conference story) and where the first reports are appearing on the interplay between PIB retention and CSF biomarkers (Fagan et al., 2005). Amersham and G.E. Healthcare have licensed commercial development, and some natural history studies (Coats and Morris, 2005) have begun using PIB.

But even as the human studies are progressing well, the mouse work with PIB has been a vexing defeat, Klunk said. His group wanted to exploit the powerful technique of micro-PET in transgenic mice to fine-tune PIB and find even better tracers. But the project ran into trouble when PIB sailed through the brain of amyloid-laden PS1/APP transgenic mice as fast as through control brains. This lack of binding also proved true in extensive subsequent ex-vivo studies of various forms of Aβ from mouse brain, including highly sensitive, classic “grind-and-bind” assays with brain homogenate. In all of these tests, PIB stuck poorly to mouse Aβ. The double-transgenic mouse brain had less than one PIB binding site per 1,000 molecules of Aβ. In stark contrast, human AD brain has a PIB binding site for every two molecules of Aβ. Intriguingly, synthetic Aβ aggregated in the test tube is just as invisible to PIB as is Aβ that aggregates in a transgenic mouse brain (Klunk et al., 2005), although all three sources of aggregated Aβ start with the same human Aβ peptide sequence.

The difference lies not in the affinity of binding sites, but in their frequency, Klunk said. The multiphoton experiments worked well because they used 10,000-fold higher PIB concentrations than one uses for micro-PET. At 2 microns, multiphoton microscopy also has almost 1,000-fold better resolution than micro-PET, and this allowed the investigators to focus in on spots where the concentration is the highest, that is, selected plaques, Klunk noted.

Why does PIB “see” only such a tiny subset of synthetic and mouse Aβ? Klunk does not know but suggests that something about the aggregation of Aβ in human and mouse brain is fundamentally different, at least with respect to generating PIB binding sites. There could be cofactors that determine the tertiary structure human Aβ assumes during aggregation. There could be post-translational modifications in humans that don’t occur during faster aggregation in mice, or there could be environmental differences in the brains’ pH and ion composition that account for the difference.

It is unclear at this point what, if anything, this difference between human and mouse implies for AD pathogenesis. Understanding the reasons for the difference might enable researchers to design better animal models of AD and yield new insight into the fundamental process of amyloid aggregation and toxicity itself, Klunk noted. Perhaps differences in Aβ aggregation might even explain why the behavioral deficits in PS1/APP mice are relatively subtle even though their brains are loaded with the human peptide.

For now, transgenic APP/PS mice (and several other transgenic mouse strains that Klunk tested) appear to be of no use for in-vivo micro-PET studies of amyloid deposition, Klunk said. This applies to tracers other than PIB, as well. This experience raises questions about the value of mice for this aspect of AD research, Klunk said. The development of promising PET tracers may stop in its tracks if investigators make human experiments contingent on prior success in mice. “You have to be careful how important you make animal studies along the way,” Klunk said. “I am really glad it worked in humans and not in mice, not the other way round.”—Gabrielle Strobel. See also part 1, part 2, and part 3.