In this independent analysis of the ADNI Plasma Proteome dataset, we have demonstrated the value of a novel multivariate feature selection approach for identifying signatures of plasma analytes that distinguish pre-clinical AD from healthy controls more effectively than a collection of the ‘best’ markers as determined by statistical univariate analysis. The important difference between this type of approach and other more conventional analyses of putative blood biomarkers is that it considers information about individual participants rather than just assessing univariate measures of class central tendency and variance. As a result, a signature set will sometimes contain analytes that do not vary significantly between groups of control and test samples yet still contribute to distinguishing these groups through the contrast between their behaviour and that of other analytes within individuals. This is not taken into consideration by univariate approaches that assess the levels of single analytes in isolation. In addition, particular analytes that do not vary significantly between two large groups can sometimes provide information about a subset of samples with profiles that are not consistent with the majority of the sample pool. It is therefore important to stress that the unitary components of a multivariate signature should not be trialled as stand-alone univariate biomarkers but instead need to be validated in the context of all the analytes comprising that signature, using appropriate classification algorithms.
The analytes that were measured by the ADNI study were not all selected because of specific links with AD, however a number of the analytes comprising the signatures we identified have been shown previously to be altered in blood or CSF from people with MCI or AD (Table S15
). Notably, other studies have found α2-macroglobulin levels to be higher in plasma from AD patients 
and transthyretin and transferrin levels to be lower in serum from AD patients than controls 
, consistent with the findings from the ADNI plasma dataset. While it is difficult to directly compare the longitudinal and meta-feature analyses with the single analyte comparisons previously reported in the literature, several of the longitudinal signature components have also been potentially implicated in AD through previous biomarker studies. These include cystatin C, sortilin and kidney injury molecule 1, which have been identified in an independent analysis of ADNI CSF samples 
, In addition, sortilin shows similarity to the sortilin-related receptor SORL1 that is genetically associated with AD risk 
. It is noteworthy that the directions of the longitudinal effects observed for these analytes in plasma appear consistent with the directions of changes in AD patients relative to controls in the ADNI CSF study discussed above.
It is striking that the majority of the analytes involved in the longitudinal signatures have been highlighted in the literature as important in renal disease in particular and, often in relation to this, in cardiovascular disease or diabetes. Notable examples include cystatin C, kidney injury molecule 1, cancer antigen 19-9, complement factor H and macrophage inflammatory protein 1α. However, while this suggests that differences relating to these conditions may exist between controls and MCI progressors, not all of the group differences were in directions that would normally be associated with pathogenicity and some may instead reflect compensatory mechanisms, complicating interpretation.
While there have been no published studies which have used the ADNI plasma dataset to identify signatures that distinguish controls from MCI progressors, one recently published study by O'Bryant and colleagues used the ADNI dataset to test the classification accuracy of a signature designed to discriminate cognitively normal controls from patients with clinical AD 
. This signature, which was selected based on serum biomarker data from the Texas Alzheimer's Research Consortium, returned a sensitivity of 54% and specificity of 78% when tested on baseline ADNI data. The authors reported that as seen in other studies, accuracy was improved substantially by incorporating demographic and clinical lab data. Similar to our signatures for discriminating controls and AD patients, the signature of O'Bryant and colleagues comprised a total of 11 analytes, however only one of the analytes in this signature (tenascin C) passed the entropy filter used in our analyses. While we have not conducted a detailed analysis of how the multitude of clinical and demographic variables collected by ADNI can be combined with plasma protein data to generate signatures with improved classification accuracy, this is likely to be an important direction for future analyses.
We have previously applied our analysis approach to the plasma proteomic dataset contributed by Ray and colleagues 
, refining their 18-protein biomarker signature for distinguishing controls from AD patients to a 5-protein signature 
. However, as described above, there are inconsistencies between the ADNI dataset and the dataset of Ray and colleagues. This discrepancy may reflect differences in the participant cohort, differences in the sensitivity of the assays used by the two studies, differences in the selection of thresholds to eliminate background noise or other factors. The Luminex technology used in the ADNI study has been well validated and the assay protocols include stringent quality controls, measurement of standards to allow calculation of the absolute concentration of plasma analytes and calculation of the least detectable dose (LDD) to facilitate identification of unreliably low readings. Further assessment of these different techniques is required to explain the discrepancies.
In addition, the study by Ray and colleagues did not assess APOE
genotype. Our analyses show that APOE
genotype has important effects on plasma levels of APOE and possibly other biomarkers. The finding that levels of APOE in the plasma are affected by APOE
genotype is consistent with previous studies, which have demonstrated a gradient of APOE plasma concentrations as a function of APOE
genotype (ε2>ε3>ε4) 
. It is well established that plasma concentrations of total cholesterol and low density lipoprotein cholesterol also differ considerably depending on APOE
and it is feasible that levels of various plasma proteins are regulated in response to changes in cholesterol or APOE levels. In support of this, APOE
genotype has previously been associated with changes in blood levels of apolipoprotein A 
, apolipoprotein B 
and C-reactive protein 
. The effect of APOE
genotype on the putative plasma biomarkers explored here highlights the importance of first considering an individual's APOE
genotype if a plasma biomarker panel is to be used as a diagnostic tool. It may even be necessary to test different biomarker signatures depending on APOE
genotype, particularly in view of the variability in frequencies of different APOE
alleles across different populations 
While the effect of APOE
genotype on plasma APOE concentration observed here was independent of clinical diagnosis, it is nonetheless possible that plasma APOE levels are still relevant to AD pathogenesis. Various studies have demonstrated that APOE
genotype can influence brain Aβ levels (with APOE
-ε4 carriers having greater Aβ deposition than non-carriers) 
but few have investigated how this relates to plasma APOE levels. Consistent with the various studies just mentioned 
, one paper on the relationship between APOE
genotype, brain Aβ deposition and plasma APOE concentration in non-demented individuals 
reported higher Aβ burden in the medial temporal cortex of APOE
-ε4 carriers than non-carriers, as assessed by Pittsburgh Compound B retention 
. However it also reported a positive
correlation between plasma APOE concentration and brain Aβ burden, with higher
plasma APOE levels in APOE
-ε4 carriers (n
10) relative to non-carriers (n
29) as measured by ELISA 
. This is not consistent with several other studies of the relationship between plasma APOE levels and APOE
or with our findings in the ADNI cohort. This indicates the need for further research into the relationship between plasma APOE and events in the brain.
Plasma levels of APOE and other proteins may also provide insights into vasculopathy in particular individuals. This may be informative as co-existing vasculopathy can affect AD onset and progression. In this context it is interesting to note that serum glutamic oxaloacetic transaminase, which we identified in signatures discriminating controls from both MCI progressors and AD patients, has been proposed as a predictive biomarker for functional outcome following ischemic stroke 
In addition, we cannot exclude the possibility that common polymorphisms in genes other than APOE may also affect plasma levels of their corresponding protein or other proteins – this will require further investigation by an integrated analysis of genomic and proteomic data. There could also be effects on levels of plasma proteins due to diet or factors, for example systemic inflammation, which may be affected in various conditions common in older people (e.g. diabetes, heart disease, cancer and arthritis). Such effects may partly account for the large number of inflammatory markers (e.g. interleukins) that were identified in the 18-protein signature of Ray and colleagues but were either not altered or not detectable in the ADNI study.
Another factor that may affect interpretation of the ADNI proteomics data is that, as noted in the ADNI Data Primer, the control samples chosen for proteomic studies were subject to selection bias. Samples selected for inclusion had baseline CSF Aβ42 levels above the median baseline CSF Aβ42 levels of the control cohort. This led to an abnormally low frequency of the APOE-ε4 allele, presumably due to an association between APOE genotype and CSF Aβ42 levels. While this is an appropriate strategy for improving detectability of differences between the controls and the disease groups, it may have other unanticipated effects, such as those involving plasma APOE levels described above. In addition, the disparity in APOE-ε4 frequency is likely to have led to an overestimation of the ability of APOE genotype, alone or in combination with demographic variables such as age and gender, to distinguish the clinical groups.
Differences in the size of the participant groups had a profound influence on the values for sensitivity and specificity determined by the classification algorithms. This probably occurs because most classifiers use a training protocol that involves optimizing classification strategies to achieve maximal values for total prediction accuracy. This leads to a bias towards strategies that correctly classify the group with the larger sample size, as this group will constitute a higher proportion of the total sample and will therefore return higher values for total prediction accuracy when called correctly. This highlights one limitation of the plasma analyte dataset currently available for ADNI, which contains considerably fewer control samples than MCI or AD.
Classification strategies that favor sensitivity over specificity are unlikely to be desirable in a clinical diagnostic setting, where it is important to avoid giving healthy people the false impression that they have a terminal disease with no effective treatment. We anticipate that the clinical applicability of the signatures will improve as data become available for larger numbers of cognitively normal controls that are more representative of the general population, allowing more appropriate classification strategies to be selected. Until further control data become available, an alternative approach might be to manually tune classification strategies, based on approaches derived from receiver operating characteristic (ROC) curve analysis, to make specificity a high priority in addition to sensitivity.
The consideration of meta-features representing pair-wise differences generally led to biomarker signatures with improved prediction accuracy relative to signatures comprising single analytes. This may be due to the identification of two analytes that are mathematically or biologically synergistic with regard to disease prediction capacity, the mitigation of confounding effects that arise from inter-sample biological variability or technical variability, or a combination of these factors. The lower accuracy of signatures comprising pair-wise sums probably arises because calculating the sum of Z-score values of log10-transformed data (comparable to calculating the product of the relative abundance of two analytes) will compound any effects of variability rather than mitigate them.
The meta-feature signatures () might also help identify protein interactions of possible biological relevance, as some of the meta-features selected by our method comprise analytes with related molecular functions. Examples of pairs which are potentially related include the chemokine meta-feature pair macrophage inflammatory protein-1α (CCL3) and pulmonary and activation-regulated chemokine (CCL18) and the matrix metalloproteinase-9 and -10 pair. In any event, the improvement in classification accuracy using meta-features demonstrates that the consideration of meta-features represents a useful tool in the search for biomarkers.
While the various signatures proposed here provided accurate classification when considered in the context of the baseline dataset from which they were identified, some performed poorly when tested on data collected at 12 month follow-up. This was particularly true of signatures designed to discriminate controls from individuals with pre-clinical AD (here used to mean individuals with MCI who later progress to AD), in contrast to the signatures designed to discriminate controls from individuals with existing AD, which performed well at both time points. The reasons for this are uncertain. One possible explanation is that if plasma protein profiles reflect the extent of disease progression, it would be expected that protein profiles of controls and AD patients differ more than those of controls and individuals with MCI. As a result, the AD group, being further separated from controls, might be expected to be relatively more robust than the MCI group against fluctuations in plasma analyte levels for any reason.
The ADNI proteomics data that are currently available only cover two time points 12 months apart, a period that may be of insufficient duration to detect substantial clinical or proteomic differences. It would be informative to be able to test our proposed signatures across a wider range of time points if data become available.
Nonetheless, the most accurate classification results we obtained came from signatures that considered the longitudinal change in analyte levels or meta-feature values over the 12 month period. While assessing changes within individuals over time is less convenient than a single test, the stronger performance of a signature based on longitudinal changes suggests that this is an avenue that should be explored in order to improve predictive accuracy.
It was interesting to note the lack of biomarkers that could reliably distinguish individuals with MCI who have progressed to AD from those who have not yet progressed. This is likely to be at least partly due to a high degree of heterogeneity among the non-progressor group. Some might progress to AD in subsequent years, while others may progress to different neurodegenerative conditions. It is also likely that a number of these individuals will remain with only MCI or even revert to control status. It will be important to determine whether the signatures proposed here have the ability to predict which of these individuals will later convert to AD, and we look forward to the outcomes of follow up clinical evaluations.
In addition to unknown endpoints within the MCI group, accurate biomarker identification might also be affected by incorrect ascertainment of AD or co-existing pathologies. As a definitive diagnosis of AD cannot be made until brain pathology is assessed post-mortem, a number of study participants with a clinical diagnosis of AD may have dementias of other etiologies. The extent of co-existing pathologies such as cerebrovasculopathy is also best assessed post-mortem. It will only be possible to select signatures with optimal accuracy for identifying pre-clinical and clinical AD through the use of retrospective analyses after post-mortem pathology has been assessed.
In addition, heterogeneity within the participant groups may also stem from other disease co-morbidities. For example, as noted above, the signatures obtained when considering the longitudinal change from baseline to 12 months follow-up (for both single analytes and meta-features) highlighted a number of analytes that have been reported to have associations with renal failure, heart failure or the metabolic syndrome and diabetes, all of which have been associated with cognitive impairment or AD risk. These conditions are common in older people and may lead to altered plasma protein profiles and a high degree of analyte variation within the participant groups. The classification accuracy of signatures might therefore be improved by first taking into account co-morbidities.
The analyses presented here highlight that in order to identify reliable biomarkers, several factors need to be taken into account. Biomarkers should be robust against age, gender and genotype. In addition, a reliable biomarker should demonstrate minimal variation unrelated to the disease of interest and measurements should not fluctuate substantially between different time points in healthy individuals. These factors may be affected by some of the limitations of the current dataset. For example, as mentioned above, the ADNI control sample was subject to selection bias and so may not accurately represent the parent population. In addition, the Luminex analyte panel used was not designed specifically for AD and better plasma biomarker signatures might be achievable with the measurement of additional components, whether other proteins or non-protein factors such as Aβ, lipids and metals. For example, for CSF biomarkers, combining novel analytes identified using a Luminex proteomics assay with established CSF biomarkers such as Aβ and tau has been shown to substantially improve accuracy in discriminating controls from AD patients 
. Similar approaches should also be effective with plasma biomarkers. More accurate (although more expensive) diagnostic classification might also be achieved through an integrative approach combining blood and CSF proteomics, genomics and brain imaging.
In conclusion, the findings of this study suggest that sets of plasma analytes can act as useful biomarkers for pre-clinical AD but can be influenced by a number of confounding variables, in particular APOE genotype. More research is required on larger samples which allow stratification by potential co-variates while retaining sufficient power for analyses of subgroups. It is likely that plasma biomarkers of the future will involve sets of analytes rather than individual analytes and that accurate pre-clinical diagnosis might require panels of multiple biomarkers. With technological advances in multiplexing protein assays, financial considerations relating to measuring large biomarker panels are becoming less of a barrier to implementation and more importance will instead be placed on assembling optimal panels rather than minimizing the number of proteins.
Furthermore, if costs continue to come down, it may become feasible to perform routine measurements of panels of plasma analytes in ‘at risk’ individuals and monitor the change over time, as is currently done in clinical biochemistry for various markers of health and disease. In addition to providing a cost-effective and minimally-invasive test capable of diagnosing AD in its pre-clinical stages, these approaches may allow us to identify molecular signposts of disease progression, improving understanding of the disease course and facilitating the monitoring of changes over time and responses to interventions.