Alzheimer's disease biomarker discovery using in silico literature mining and clinical validation

Alzheimer’s disease (AD) is one of the commonest causes of dementia resulting in a severe loss of intellectual abilities including memory. The main histological features of AD in brain are amyloid plaques and neurofibrillary tangles, due to accumulation, respectively, of amyloid beta (Aβ) peptide and tau protein in insoluble form. The causes of this, and other pathology found in the AD brain, are not known with certainty but are likely to be multifactorial. This multifactorial and only partially understood pathogenesis complicates both drug and biomarker discovery.

The search for biomarkers to aid accurate diagnosis, predict progression and for use in clinical trials has become a major research goal [1, 2]. The most widely used strategy for the discovery of biomarkers is predicated on the identification of potential candidate biomarkers using knowledge of disease processes followed by validation, comparing healthy control to affected subjects [3]. In many respects, the optimal source of human tissue for the investigation of AD biomarkers is cerebrospinal fluid (CSF) and the demonstration of lowered CSF Aβ and raised CSF tau and phosphorylated tau in AD is the prime example of candidate biomarker discovery and validation leading to qualification using an affected case-normal elderly control design [4].

However, all three approaches - candidate discovery, case–control design and the use of CSF have their drawbacks. CSF although ideal in many respects as it bathes the diseased organ, is limited by patient and clinician acceptability, especially for repeated measurements. Affected case versus normal elderly control experimental design is severely limited by the long prodromal phase in AD, meaning many apparently unaffected controls may have substantial occult pathology. Candidate biomarkers are limited by the complex multifactorial pathology in AD meaning that beyond Aβ and tau it becomes difficult to ascertain true candidates. In an effort to overcome the limitations of CSF, we, and others, have pursued, with some early indications of success, markers in plasma that might act as biomarkers of disease [5]. To overcome the problems of occult pathology and to identify biomarkers of potential utility, two broad approaches have been employed; firstly comparing people with prodromal states, for example in mild cognitive impairment (MCI) who progress to AD, to those who do not, and secondly comparing both affected and asymptomatic people by a non-clinical marker of disease such as MRI measures of atrophy or PET measures of amyloid load [6, 7]. This latter approach has been described variously as an endophenotype or extreme phenotype design. To mitigate the limitations of the candidate marker approach many have used a range of data-driven technologies including proteomic and genomic platforms and increasingly these datasets are being aggregated for biomarker discovery and understanding pathogenesis; for example in renal disease where over 200 datasets are mined in a web-based application to identify disease associated proteins [8].

Here we report a novel approach to biomarker discovery through the utilization of data-driven biomarker discovery in silico followed by validation in blood using automated analysis of structural MRI as an endophenotype measure of disease. The in silico study started by generating massive volumes of assertional data represented in the form of an Intelligence Network. Assertions are simple factual representations of statements made in the biomedical literature and other sources. They can be compiled into a semantically consistent form by applying comprehensive vocabularies and lexical matching approaches to yield a navigable database known as an Intelligence Network (IN). For this study, the IN contained assertions relating to proteins expressed in the brain and associated with pathology relevant to AD. These are two important features for an ideal biomarker for AD, assuming that such proteins could be measured in readily accessible fluids such as blood or urine. An in vitro study of two putative markers identified by the in silico screen - Choline Acetyltransferase (ChAt) and urokinase-type Plasminogen Activator Receptor (PLAUR) - provided evidence supporting the validity of the method suggesting that in silico screening for biomarkers in AD and, by extrapolation, other disorders, is a productive approach.

The discovery of biomarkers for AD is an increasingly important task – both for early diagnosis and for use in experimental medicine. However, it is a task complicated by at least three major intrinsic difficulties. First the complexity of AD pathology means that identification of candidate markers, beyond the low hanging fruit of Aβ and tau, is problematical. Second, collection of the optimal peripheral fluid for biomarker identification, CSF, is relatively invasive and unsuitable for repeated measures in elderly people. Third, AD has a prolonged prodrome when apparently normal elderly people harbor considerable pathological load meaning that the conventional case–control design is confounded by pathology in clinically unaffected subjects. Previously we and others have attempted to mitigate the third of these complications by using a design of biomarker discovery where the independent or outcome variable is not clinical diagnosis but an endophenotype of disease such as structural MRI evidence of atrophy [7] or PET evidence of Aβ load [6]. The second of the limitations in biomarkers for AD – the availability of CSF – has prompted many groups to seek markers in other fluids such as plasma. The first of the limitations, the identification of candidates, has been previously attempted by two broad categories of studies; either using candidates based on the researcher’s own understanding of disease or using a data-driven, most often proteomic, approach. Here we combine the use of endophenotypes to complement diagnostic category as an outcome measure, with the use of plasma as a biomarker tissue, with an entirely novel approach to the identification of candidates. This innovation makes use of linguistic and textual analysis to interrogate the entire biomedical knowledge base in the form of all the major publicly available databases to identify candidate markers using a consensus driven set of primary assertions. We accessed the various data sources in 2006, and clearly in a fast moving field data will have changed considerably in the intervening years. Indeed some of the proteins identified in 2006 (such as Transthyretin and Clusterin) had not at that time been considered as biomarkers. However, when the in vitro analysis was carried out, data had been provided from our own proteomics studies that these proteins were in fact putative biomarkers. Thus this time lag has inadvertently provided further substantiation of the proof of concept of the in silico approach that we discuss here.

The textual analysis of publicly available data sources suggested a total of 25 potential candidate biomarkers. Some of these have previously been identified as potential biomarkers in plasma. For example, using MRI measures of atrophy as an outcome endophenotype we identified and confirmed plasma Clusterin [7] and Transthyretin [16] as measure of severity of disease and using PET measures of amyloid identified apoE protein as the primary correlate in plasma [6]. All these studies used gel based proteomics as the discovery tool and the fact that textual analysis identifies the same proteins before these proteomic studies were performed is a strong indicator of the power of the method. Other promising candidates suggested by textual analysis, and where there is published data suggesting that these proteins are altered either in blood or CSF, include CRP [17‐19], Complement factor 1 [20, 21], butyrylcholinesterase [22] and BACE1 [23, 24]. In all but the case of butyrylcholinesterase, this biomarker data was published after the IN lock-down and hence these biomarker utility data are independent of the IN and act as independent proof of concept.

As the IN identified as potential protein biomarkers proteins previously identified in proteomic studies – without this data entering into this particular network – we were encouraged to attempt further validation in plasma. We chose two proteins - neither previously identified as potential plasma biomarkers to our knowledge - and measured these in over 200 subjects most of whom had as part of the European AddNeuroMed project, automated analysis of structural MRI data available. One of these proteins – PLAUR – was significantly decreased in AD relative to controls with MCI being at an intermediate level. Both PLAUR and ChAT showed a correlation, inverse in the case of PLAUR, with imaging evidence of atrophy in control cases and both showed a smaller and non-significant, but in the same direction, correlation in AD cases. We used semi-quantitative immunoblotting as a screening method as in previous studies as this approach, in contrast to ELISA for example, yields information on degradation products and post translational modifications. In fact the data on these two chosen proteins suggested whole protein correlation with disease state suggesting future biomarker replication and qualification studies, beyond the intention of the present investigation, might progress rapidly to fully quantitative methods.

Urokinase plasminogen activator receptor (PLAUR) is a protein involved in many biological functions including cell signaling [25, 26]. By binding urokinase plasminogen activator (uPA), with which it forms an active complex (uPA-PLAUR), it catalyzes the transformation of zymogen plasminogen into the active protein plasmin, a serine protease which degrades fibrin. The receptor is also involved in cell signaling and in chemotaxis, and controls cell adhesion. Increased levels of PLAUR have been previously reported in inflammatory disorders [27] and has been implicated in chemotaxis leading to microglial accumulation in the core of amyloid plaques in brain in transgenic rodent models of AD. Aβ induces PLAUR [28] and PLAUR is increased in microglia cells of human AD brains and in brains treated with amyloid β peptide [29, 30]. The inverse relationship we observe between soluble PLAUR and AD and brain atrophy is noteworthy and might suggest an inverse relationship either between soluble and, functional, membrane bound PLAUR or between central and peripheral PLAUR more generally. An inverse relationship between central amyloid load and peripheral, CSF, amyloid has been previously and extensively noted. The other novel protein association with pathology we identify, ChAt, is a key component of the cholinergic pathway which is severely affected in AD and is the target for the first symptomatic therapies for AD. We observe a relationship between cerebral atrophy and ChAT protein and this may reflect the loss of cholinergic neurons known to occur early in disease process.

In summary we show here that the extraction of data from huge volumes of biological datasets including text based information is possible and that the creation of hypothesis or assertion-driven analysis yields potential biomarkers. As some of these markers have been independently generated using proteomics, and, as here we show at least partial validation of the two markers tested, this finding offers strong support to a text mining approach to biomarker discovery using the ever increasing publically available datasets.