What’s Cooking in Biotech: Fledgling Products
Discouraging as Bill Klunk’s experience with APP/PS transgenic mice appears, it stands opposite other new data that show, on the contrary, just how useful even the existing models can be when examined with new technologies. Floyd Bloom is a former editor in chief of Science magazine, who in 2000 co-founded the biotech company Neurome, Inc. after leaving the journal. His premise was that the mouse and its genome will be the premier model for molecular neuroscience for some time to come, and transgenic mice for CNS diseases. He believed that knowing which neurons and circuits express a particular gene is a necessary first step toward a functional characterization of the genes one wants to target therapeutically. Because classical chemical neuroanatomy was largely based on rats and cats, the field needed new tools to map and compare gene expression in mouse strains, Bloom reasoned, and he set up Neurome to create high-throughput morphology and histology protocols that can assess ingredients of circuits quickly and comprehensively.

An early project at the company focused on estimating volume changes in brain regions of the PDAPP mouse. Using a 9.4 tesla magnet at CalTech, the scientists obtained digital volumes of specific areas of mouse brain rendered in three dimensions. They observed that between 40 and 90 days of age, the hippocampus in wild-type mice grows significantly, but in PDAPP mice it does not. Something about the expression of the transgene leads to a 14 percent volume reduction already by day 90, and it may set the ground rules for the pathology developing later (see Redwine et al., 2003).

From this initial experience with image analysis, the researchers learned how to align brain sections done in one plane (i.e., coronal) into smooth, virtual atlases in another (i.e., sagittal or horizontal). This enabled them to analyze abnormalities in gene function and view their effects on brain structure in those 3D atlases. The scientists can also perform neurochemical or anatomical experiments on sections, including stereological assessments of the volumes of small components of a given brain region, for example, hippocampus. This showed that the dentate gyrus was the subarea that failed to grow in the PDAPP mice; at 90 days it was 28 percent smaller than in wild-type mice.

To understand why the dentate gyrus languished in this way, the Neurome scientists adapted a DiOlistic technique developed by Jeff Lichtman, then at Washington University in St. Louis, in which a “gene gun” shoots a lipophilic dye on microscopic gold particles into a fixed slice of the desired brain region. This yielded 3D images of fluorescent dentate granule cells in a high-throughput fashion. To make this assay practicable and reproducible, the Neurome scientists developed software for analyzing the dendritic complexity of a granule cell in a short time period and with limited computer memory. This study found that the dendritic trees of dentate granule cells were stunted, and that not all cells were equally affected. Of the six granule cell layers in the dentate gyrus, dendritic length was more reduced in superficial cells than in deep cells, and more reduced in the dorsal blade of the hippocampus than in the ventral blade (Wu et al., 2004). The reason for this difference is unknown, but Bloom noted that work by Fred Gage’s lab has shown the superficial layer of the dentate gyrus to incorporate fewer new neurons from neurogenesis than do the deeper layers.

Morphometric analysis with DiOlistic labeling and a second labeling method based on Golgi impregnations are ongoing at Neurome; the goal is to compare the fate of granule cell dendrites in aging mice to that seen in young PDAPP mice. Early data suggest that PDAPP mice start losing dendritic spines in the most vulnerable layers early in life and that wild-type mice begin showing a similar loss at around 15 months. These could be two independent processes, or they could represent premature aging of the dendritic spines of a particularly vulnerable set of neurons driven by the transgene, Bloom speculated.

These morphometric tools also have made it possible to quantify the distribution of diffuse and compact amyloid in different brain areas of the aging animal in a newly comprehensive way. For example, this study found that diffuse amyloid increases greatly between 12 and 15 months of age in PDAPP mice, suggesting that it could serve as a biomarker for drugs targeting this form of Aβ. “If we had a medication that prevents diffuse amyloid formation, we could test it in that age bracket and get an answer in 3 months,“ said Bloom.

A further finding was that the amyloid load in subareas of the dentate gyrus correlates with their incoming nerve circuitry: PDAPP mice carry a much higher amyloid load in the outer molecular layer and the lateral entorhinal cortex (which projects to the outer molecular layer) than in the middle molecular layer and the medial entorhinal cortex (which projects to the middle molecular layer). Something about that latter circuit allows it to resist the process by which amyloid is laid down, and finding out what it is would hold clues to the underlying disease process, Bloom suggested. There is no explanation for what molecules distinguish these circuits. One candidate for a spatially defined modifier is SUMO2/3, which Barbara Cordell’s group at Scios Inc. in Sunnyvale, California, reported to restrict APP expression (see Li et al., 2003). Bloom’s group noticed that the hard-hit lateral entorhinal cortex at 100 days of age begins to express less SUMO3 than does the medial entorhinal cortex. Incidentally, the lateral entorhinal cortex is an element of the default network discussed by Greicius.

Furthermore, this new system of morphometric analysis reinforced the field’s realization in recent years that the genetic background of APP transgenic mice can greatly influence their phenotype and that data from one strain should not be interpreted in isolation. For example, the MRI volumetric findings in the PDAPP mice were not reproducible in Tg2576 mice, Bloom noted. This strain did, however, lose spines in areas of the hippocampus and cortex that later laid down amyloid, as also found by Greg Cole.

Peter Davies, of Albert Einstein College of Medicine, consults for Applied Neurosolutions on the development of a diagnostic test for AD. This Illinois-based biotechnology company develops an assay based on measuring CSF concentrations of the protein tau phosphorylated on the amino acid threonine 231. At the workshop, Davies evaluated the quality of existing tests for tau; he argued, in essence, that the p-tau231 assay is as good as it can be within the confines of an imperfect clinical diagnosis.

Davies began by noting that tau phosphorylation clearly is increased in AD, and that antibodies visualizing this process light up much more than neurofibrillary tangles. They stain what looks at first glance like intense background but in truth represents evidence of widespread threonine 231 phosphorylation of tau beyond the actual tangles themselves. This biochemical abnormality in tau labels whole cell groups and their processes in large areas of the hippocampus. Threonine 231 and serine 202 are sites on tau that become hyperphosphorylated early in hippocampal pyramidal neurons of very mild AD, well before tangles form there. These markers are useful for detecting early disease, whereas total tau has proved unreliable, Davies said.

Led by Davies’s collaborator Harald Hampel at Ludwig Maximilian University in Munich, Germany, as well as other European investigators, researchers to date have tested the p-tau231 assay in more than 3,000 CSF samples. Most are from patients who came to the clinic for a diagnostic workup (Hampel et al., 2004). AD cases come up positive while controls do not, but the assay also picks up a fraction of patients diagnosed with other dementias. This weakness is often cited. The assay’s sensitivity lies at 90 percent; its specificity ranges from 80 to 100 percent but is lower for vascular dementia and diffuse Lewy body disease. Davies argued that where results diverge from the clinical diagnosis, he suspects that the assay is correct and the neurologist may have made the wrong diagnosis. “Even the world’s best clinicians are not always accurate. I think the patients diagnosed with vascular dementia who were positive in our assay actually have AD,” Davies said.

Overlap in underlying pathologies plays into this issue, as well. A substantial number of vascular dementia cases prove upon autopsy to have had amyloid and tau pathology, as do people clinically diagnosed as having diffuse Lewy body disease. Ironically, people with frontotemporal dementia—the quintessential tauopathy—do not. Their predominant pathology is cell death, and tangles do not accumulate massively as they do in AD brains, Davies said. Accordingly, they come up negative in this assay.

The p-tau231 assay shows no obvious relationship with the MMSE, a crude but widely used cognitive assessment. Why not? Davies pointed to a study that, to date, has measured CSF p-tau231 of 103 mild cognitive impairment (MCI) cases, 163 AD cases, healthy controls, and samples from other neurological diseases. It shows that the signal in MCI already is nearly as high as that of full-blown AD (Buerger et al., 2002), suggesting that by the time a person begins to fall off on the MMSE test, p-tau231 has long accumulated. Davies’s collaborator Mony De Leon is trying to determine exactly when p-tau231 first begins to rise in a cohort of healthy people and MCI patients he is following longitudinally (de Leon et al., 2002). This ongoing study indicates that p-tau231 is fourfold above normal even at the MCI stage when MMSE performance is still fairly high; however, the higher a person’s p-tau231 concentration is at that point, the faster he or she declines on the MMSE in the next few years. “This means we may be able to use this marker to identify patients at a very early stage of disease and predict their progression,” Davies said.

In summary, Davies argued that the clinical diagnosis of AD cannot be an ideal yardstick by which to measure the accuracy of a biochemical assay above the ninetieth percentile, and other investigators agreed. “All biological markers are going to run up against a ceiling effect of the clinical diagnosis. At the present time, our assay has reached this ceiling,” Davies said.

Besides creating circular arguments, this situation makes validating the assay a challenge. Validation on autopsy is difficult in practice because the needed several thousand samples are difficult to obtain in the U.S., where autopsy rates are low and falling. Moreover, people on average live another decade after receiving a diagnosis, and during this time some who were negative at testing would likely develop tau pathology, again muddying the waters. It’s questionable whether such an expensive, long-range study must be done, especially for a combination test of Aβ42 and p-tau231. “We may already have excellent biomarkers; we just do not know it yet,” Davies concluded. (See also ARF related news story on CSF Aβ/p-tau; for new reviews on CSF-based biomarkers, see Formichi et al., 2006; also Wallin et al., 2006).

The well-tested p-tau231 assay illustrates some of the hurdles that a separate, fledgling test introduced at the workshop has yet to clear. It is a proteomic blood test developed in a joint effort between Tony Wyss-Coray at Stanford University and Sandip Ray, who co-founded the startup biotech company Satori with Wyss-Coray.

The process of AD features a vigorous inflammatory response. Astrocytes and microglia become activated and cause the secretion of a large number of proteins, including cytokines, chemokines, growth factors, proteases, and protease inhibitors, which together mediate communication between these cells in the brain. Lymphocytes chime in, as well, especially in vascular forms of the disease. An increasing number of studies indicate that this CNS reaction communicates with the periphery, particularly with peripheral macrophages, lymphocytes, and myeloid cells, Wyss-Coray said. Some of these can travel into the brain, assess its state, and either leave or induce production of factors or even initiate immune responses.

The larger point is that every disease, in every organ, leads to changes in plasma, Wyss-Coray said. The blood is the body’s most complex organ in terms of protein moieties, and Wyss-Coray started his study from the question of whether one can understand a disease process by studying plasma. Scientists have measured individual markers by ELISA, but the low power of this approach has left many studies that report initial discrimination with individual factors without replication by other labs. Mass spectrometry has matured as a tool for mining the proteome, but problems persist there, too, as the method is asked to keep apart many tens to 100,000 different proteins, fragments, and post-translational modifications. Abundant proteins such as albumin tend to overload the system, and efforts to deplete them take down other proteins that may be of interest. This approach is not reproducible enough yet to be clinically useful, Wyss-Coray said. (See also an independent recent attempt to identify a new candidate biomarker set focused around neuroprotective and complement proteins, see Selle et al., 2005.)

The scientists decided on a middle-of-the-road approach between these two extremes. The scientists first picked a set of proteins that might be important in the disease process. “We call this a candidate-based approach by tuning in to the language of cells. How do they communicate when healthy, how when stressed and diseased? Hopefully, we get a disease-specific picture,” Wyss-Coray said.

The scientists gradually whittled down an initial group of 300 proteins from among cytokines, chemokines, growth factors, neurotrophins, hormone-like proteins, acute-phase proteins, complement factors, soluble receptors, proteases, and inhibitors to a set of 12 predictors. They developed an ELISA array of a membrane with monoclonal antibodies specific against these proteins, incubated it with patient plasma samples, and read the signal with chemiluminescence. This yields a picture of relative levels of expression of these 12 factors, which can be quantified.

A first, small study used 48 cases in various stages of AD as well as 50 age-matched controls from seven centers around the world. Of the 17 who have since died, the test had predicted their condition with 100 percent accuracy, Wyss-Coray said. Of the cases that have not yet come to autopsy, the difference in the relative probability of having AD between control and AD groups was large, Wyss-Coray said.

Software developed at Stanford, called significance analysis of microarray (SAM), pulled up 44 proteins whose blood levels differed between AD and controls. Individually, none of these factors predict AD, but together they do, Wyss-Coray said, and a separate procedure of unbiased clustering of these 44 markers based merely on their expression levels reproduced the AD and control groups.

To analyze further whether these plasma differences could predict AD, the scientists turned to another form of analysis called predictive analysis of microarrays (PAM). This algorithm tries to identify a minimal set of markers that can discriminate and predict the proper sample groups without having seen the primary data. In multiple iterations, it adds proteins from within a training data set and calculates their predictive power until it has reached maximal accuracy. This minimal set included 12 proteins, which predicted whether a sample came from AD or control with 97 percent accuracy (i.e., a composite score of 100 percent sensitivity and 94 percent specificity). When the algorithm then applied this information to a different test set it had not seen before, it classified 32 of 33 samples correctly.

The top 12 factors are involved in immune function, energy metabolism, and vascular function. Wyss-Coray proposed that the most abundant changes are consistent with immune and macrophage impairment. There is scarce data on this topic, but a growing trickle of studies is suggesting that mononuclear cells or macrophages isolated from AD patients are impaired in a number of ways and respond poorly to stimulation. (For a current review on serum-based proteomics of neurodegenerative diseases, see Sheta et al., 2006.)

Next on Wyss-Coray’s list is to study related dementias. Initial work on ALS, Parkinson disease, multiple sclerosis, and peripheral neuropathy indicates the AD fingerprint is specific to this disease and does not merely reflect a generic inflammation. The hope is that related dementias will prove to have their own unique pattern of plasma predictors, suggesting that a top 12 set may be found for them, as well.

One important caveat with blood tests is that infections or flu could mask AD in plasma samples. While the scientists have not ruled this out, Wyss-Coray said individual markers clearly change in response to a flu, but a defined set of 12 may not. Confounders such as this imply, however, that an ultimate test for AD may need more than 12 predictors. The present data aim to prove the concept; it is not a commercial test just yet, Wyss-Coray said. —Gabrielle Strobel.

See also part 1, part 2, part 3, and part 4 of this series.