A Multiomics Approach to Heterogeneity in Alzheimer's Disease

Focused Review and Roadmap

AmanPreet Badhwar; G. Peggy McFall; Shraddha Sapkota; Sandra E. Black; Howard Chertkow; Simon Duchesne; Mario Masellis; Liang Li; Roger A. Dixon; Pierre Bellec


Brain. 2020;143(5):1315-1331. 

In This Article


Complementarity of Omics Biomarkers

Individuals clinically diagnosed with an NDD of ageing (e.g. Alzheimer's disease) exhibit varying loads of neurodegenerative markers (e.g. amyloid-β, tau, α-synuclein, brain atrophy, vascular abnormalities) (Beach et al., 2012; Robinson et al., 2018). Single-domain omics biomarkers can, to an extent, characterize this heterogeneity in vivo. Some data-driven brain atrophy subtypes parallel established clinical diagnoses. For example, the posterior atrophy subtype is evocative of the posterior cortical atrophy Alzheimer's disease variant, and the language atrophy subtype of the logopenic progressive aphasia variant (Ossenkoppele et al., 2015). An active area of research is to determine to what degree the 'bottom-up', fully automated and data-driven subtypes match with established 'top-down' clinical assessments, which usually start with cognitive symptoms and then incorporate specific neuroimaging characteristics, such as left temporoparietal atrophy in logopenic progressive aphasia (Ossenkoppele et al., 2015). The fact that reviewed studies included participants with typical late-onset Alzheimer's disease dementia, and not atypical variants such as posterior cortical atrophy, suggest that specific brain atrophy phenotypes comprise a spectrum of involvement that may overlap a clinical label, but not associate uniquely with one. It is unclear how functional connectivity subtypes tie in with atrophy subtypes, although they both associate with clinical diagnoses and biomarkers and risk factors of Alzheimer's disease (Zhang et al., 2016; Orban et al., 2017b). Although the propagation of functional dysconnectivity parallels the Braak staging of Alzheimer's disease, the mix of connectivity increases and decreases observed in patients may reflect transient compensatory mechanisms as well as neurodegeneration (Badhwar et al., 2017).

To date, we are unaware of studies that have associated data-driven Alzheimer's disease brain subtypes with PRS and/or metabolite panels. However, brain subtypes could be targeted and validated with data-driven metabolomics analyses, with genomics analyses likely contributing precision information to these complementary omics approaches. Our review strongly supports such coordinated multiomics approaches. For example, an Alzheimer's disease PRS composed of genes linked to lipid metabolism and inflammatory response may associate with panels composed of metabolites involved in corresponding pathways. One could further speculate whether the resulting inflammation may cause the diffuse atrophy subtype, and trigger specific functional compensatory mechanisms. Testing these hypotheses will require a cohort that is both deeply phenotyped and captures the entire spectrum of age-related dementias.

Complementarity of Analytical Approaches

There is potential complementary in the information that can be gathered across different analytical approaches, even within a single omics modality.

Different flavours of subtyping methods were used in the imaging papers we reviewed. Some variants, e.g. hierarchical agglomerative clustering using Ward's clustering linkage (Noh et al., 2014; Hwang et al., 2016; Tam et al., 2019) and Louvain clustering (Park et al., 2017), are just different algorithms that produce similar types of cluster representations. These algorithms may still differ by their quantitative performance, but this is hard to establish in the absence of ground truth, in a purely data-driven analysis. A more qualitative difference in some techniques is the use of continuous (rather than discrete) subtypes, with multiple subtype factors contributing to explain the brain map of any particular individual, such as Bayesian latent factor analysis (Zhang et al., 2016). Zhang et al. argued that capturing interindividual variations as a continuous spectrum is important, and that there are no clearly separated, discrete biological clusters. We note that even using traditional, discrete, clustering techniques, it is possible to compute continuous indices of similarity between a subtype map and a given individual, and some groups that applied discrete cluster analyses actually performed statistical analyses using such continuous proximity measures (Tam et al., 2019). Another important analytical variation is to use clinical labels to build discriminant subtypes between clinical cohorts, such as CHIMERA clustering (Dong et al., 2016, 2017). Such an approach will by construction generate subtypes that associate more strongly with clinical variables than unsupervised subtypes based solely on the similarity of brain maps across all subjects, including controls. It can be noted, however, that some studies apply a second stage analysis to combine multiple unsupervised subtype map into a single multivariate signature with high clinical predictive value (Tam et al., 2019). So the distinction between unsupervised versus supervised subtype generation may not be as fundamental as it superficially appears.

Regarding metabolomics studies, five of six of the studies in the section 'Metabolite panels' implemented an orthogonal projection to latent structures-discriminant analysis (OPLS-DA) (Supplementary Table 2). This technique shares similarities with some of the subtyping approaches used in imaging. First, OPLS-DA generates continuous loadings for each factor across subjects, which means that metabolomic data from a particular individual are a combination of multiple latent variables, rather than being associated with one discrete subtype. Second, this technique uses the clinical labels (typically three clinical groups, cognitively normal, MCI and Alzheimer's disease dementia) in its decomposition, attempting to separate clinically relevant factors from (orthogonal) sources of within-group variance. Metabolomic studies typically proceed by further selecting a small panel of metabolites and assess its predictive power, and this panel can be further linked to larger metabolomic pathways, parallel to network analysis in genetics (refer to the 'Metabolomics pathways and networks' section). Emerging analytics in the metabolomic field are thus mixing ideas applied separately in the imaging connectomics and genomics (discussed below) sections of our review.

Regarding the genomics section, the majority of the studies reviewed used PRS derived from omics-based GWAS using both data-driven and hypothesis-driven approaches (Supplementary Table 4 and Supplementary Table 5). These PRSs can be constructed in several ways: (i) all significant SNPs identified by GWAS; (ii) all nominally associated variants based on a particular significance threshold; (iii) combinations of mechanism-related SNPs; and (iv) a variety of interaction analyses. Identifying Alzheimer's disease-associated markers (i.e. genotype-phenotype models) has been an important first step, but, like any single omics approach, GWAS analyses do not directly consider the multigenic and multifactorial mechanisms underlying Alzheimer's disease. Other limitations are low effect size of risk variants and the possible heterogeneity of disease risk variants across populations, with the implication that some rare risk alleles may be under-identified. There is also a need to investigate gene-gene or gene-environment interactions further (Ertekin-Taner, 2011; Bras et al., 2012). Notably, recent machine learning technologies are moving towards analyses of multiple genes, multiple phenotypes, and interactions between them (Gaiteri et al., 2016). Genetic interaction models include epistasis (synergistic or dys-synergistic gene-gene interactions), edgetic (genetic variations of protein-protein interactions), and directed network analysis such as biologically constrained multiscale models that provide information about the genetic variants associated with molecular networks and brain networks. Six studies testing network or multiscale models were reviewed (Supplementary Table 5).

Reproducibility of Subtypes, Panels and Polygenic Risk Score

An important methodological consideration is the reproducibility of subtypes, panels and PRSs. A rigorous evaluation of reproducibility would consist of a series of experiments deriving these subtypes, panels and PRSs in either two independent group samples, or the same sample with different methods. We are not aware of studies conducting such systematic reproducibility experiments. Within the scope of this review, we could only qualitatively comment on the convergence across studies, but it was not possible for us to quantify this convergence (in part because of the limited number of studies) or to identify the factors that influence reproducibility, such as methodological variability, variations due to small sample size, or biological heterogeneity across different recruitment strategies. As these multivariate approaches get more mature and possibly get translated to clinical practice, such systematic reproducibility analyses will be an important area of future work.

Validity of Predictive Power

Another critical methodological consideration in this review is validation of the predictive value of a biomarker. In the machine learning community, the best practice is to estimate the parameters of a predictive model on a training set, and then evaluate the performance of the trained model on data that were not touched during training, called the test set. This principle should in theory guarantee an unbiased estimate of model performance, yet in practice many caveats exist and call for cautious interpretation of some published results. A model with many degrees of freedom can achieve very high accuracy on any training set, but this performance will not generalize to the test set, a phenomenon called overfitting. For a given model complexity, it is however easier to over-fit a small sample size than a large one. If a research group does try many models on the test set and only reports results for the best model, this also leads to overfitting. A recent review of functional MRI biomarkers demonstrated a strong trend across many brain disorders (including Alzheimer's disease), where studies with small sample size tend to report markedly higher accuracy scores than studies with large sample sizes (Figure 4 in Varoquaux, 2018). This strongly suggests that over-fitting in small samples is pervasive in the neuroimaging literature. However, this is not particular to neuroimaging: some of the metabolomics paper reviewed lacked out of sample validation (Wang et al., 2014; Liang et al., 2015) and reported very high area under the curve (AUC) (>0.99), while papers implementing out of sample cross-validation reported more modest effect sizes (Figueira et al., 2016). It is also in general true that accuracy estimates with small sample sizes (n < 100) have a very wide confidence interval, which means that prediction scores reported in studies with a small sample size should be interpreted with caution as the true performance may be very different from the values reported (Figure 1 in Varoquaux, 2018). Both neuroimaging subtypes and metabolomics panels are relatively young, emerging technologies, and some papers reviewed have a relatively small sample size per clinical group, e.g. ~50 (Wang et al., 2014; Figueira et al., 2016). Genomics, by contrast, is a much more mature field, and some studies reviewed tested predictive powers across very large samples (Desikan et al., 2017).

Figure 4.

Proposed roadmap to discovering multiomics Alzheimer's disease biomarkers. COMPASS-ND: The COMPASS-ND cohort is composed of individuals with various types of dementia or cognitive complaints, as well as healthy, cognitively normal individuals. Omics data: Performing dimension reduction for omics data. Featured as examples are some of the results of our review of the Alzheimer's disease literature as presented in the paper. Machine learning, Multiomics biotypes and Prediction: These panels demonstrate how signatures of neurodegeneration derived from the integration of multiomics data using machine learning techniques will better identify individuals on an Alzheimer's disease spectrum trajectory. While our proposed roadmap addresses multiomics biomarkers for Alzheimer's disease, a similar approach can be used for other neurodegenerative diseases of ageing. AD = Alzheimer's disease; CN = cognitively normal; DMN = default mode network; FTD = frontotemporal dementia; G = genomics features; I = imaging features; LBD = Lewy body disease; LIM = limbic network; M = metabolic features; Mixed = mixed aetiology dementia; O = demographic features; SAL = salience network; SCI = subjective cognitive impairment; VCI = vascular cognitive impairment.

Another important consideration is that many types of generalization and test datasets can be implemented, with radically different interpretations. One can test generalization on a different group of subjects, but with data collected using similar methods to the training set and at the same location. This is an important validation step, but remains only a proof-of-concept. Only by testing data collected at different institutions as well as different locations throughout the world, and possibly different data acquisition protocols, can the true predictive performance expected in a 'real world' clinical setting be assessed. Almost none of the imaging or metabolomic studies implemented such large-scale generalization experiments.

A Roadmap for Parsing Heterogeneity in Neurodegeneration

A data-driven characterization of heterogeneity across the NDDs of ageing will require cohorts representative of the spectrum of neurodegeneration.

The cohort assembled by the Canadian Consortium on Neurodegeneration in Aging (CCNA, http://ccna-ccnv.ca/) provides a new opportunity to study the full spectrum of age-related dementia. By 2020, the cohort will include 2310 individuals (aged 50–90) featuring the following cognitive conditions: Alzheimer's disease, vascular, Lewy body, Parkinson's, frontotemporal, and mixed aetiology dementias, as well as subjective cognitive impairment, MCI, vascular MCI, and cognitively normal (Figure 4, COMPASS-ND column). The cohort composition ensures that age-related dementias are more or less equally represented, even for less prevalent dementia types (e.g. frontotemporal dementia). Participants will be deeply phenotyped with extensive clinical, neuropsychological, neuroimaging, biospecimen, and neuropathological assessments.

In Figure 4 we present a roadmap for a multiomics approach to heterogeneity in NDD. We begin with a heterogeneous clinical cohort design (Figure 4, COMPASS-ND column) that enables the discovery of subgroups sharing a common signature across multiple omics domains (biotypes) that are highly predictive of the clinical status and evolution of individual patients. Multiomics biotypes will be complemented by other important variables such as sex, presence of amyloid-β and tau deposits, and vascular abnormalities. Machine learning tools will be applied to identify an optimal combination of different biotypes and explanatory variables that either discriminate different clinical cohorts, or are predictive of future progression of specific symptoms (Figure 4, Machine learning column).

We have three complementary lines of reasoning for including a heterogeneous clinical cohort design (i.e. diverse dementia aetiologies) in our proposed roadmap. First, if we were to just consider the data-driven multiomics biotypes generated using an Alzheimer's disease population (Figure 4, Multiomics biotypes column), having access to diverse dementia aetiologies will allow us to evaluate the uniqueness of each biotype to Alzheimer's disease. Second, as heterogeneity is a feature of Alzheimer's disease as well as other NDDs of ageing (Robinson et al., 2018), our proposed roadmap can be applied to generate multiomics biotypes in other dementia aetiologies. Similar to our review, this will initially require identification of disease-specific indicators from single omic modalities (Figure 4, Omics data column). We envision that for any given NDD of ageing, multiomics biotypes identified will range from pure (but rare) disease-specific biotypes, to biotypes featuring mixed pathologies. Having access to multiomics biotypes from the spectrum of NDDs of ageing will allow better delineation of biotypes, namely, those that are unique to a specific NDD of ageing, and those that show overlap with other NDDs of ageing. Third, recent work has shown that by training machine learning models on large and heterogeneous data it is possible to generalize better to new studies relying on different methodologies and run on slightly different populations (Abraham et al., 2017; Orban et al., 2017a). Such generalizability is critical for a successful translation in clinical practice.

Towards Highly Predictive Multiomics Signatures for Prognosis

Because of the emergent nature of the three omics techniques (neuroimaging-based subtypes, metabolite panels, and polygenic risk score), our review was largely composed of proof-of-concept cross-sectional comparisons of cognitively normal older adults with individuals classified with prodromal or diagnosed Alzheimer's disease, as opposed to the preclinical population. However, the publication of the A/T/N criteria (Jack et al., 2016), along with increasing availability of CSF and PET biomarker data (e.g. amyloid, tau), and longitudinal cohorts provide fertile grounds for additional (and much warranted) research addressing heterogeneity in preclinical and/or at-risk cohorts. Specifically, longitudinal trajectory studies with data-driven neuroimaging subtypes differentially transitioning from cognitively normal to preclinical to prodromal or dementia stages of Alzheimer's disease are needed. In these designs, metabolomics and genomics (PRS or APOE) would probably serve as a variable to add precision to a biomarkers-based prognosis.

Current biomarkers of Alzheimer's disease dementia demonstrate limited predictive power for prognosis in the prodromal phase (Rathore et al., 2017). The best models include a combination of cognitive, structural MRI, fluorodeoxyglucose-PET, and/or amyloid-PET measures (Rathore et al., 2017). A substantial proportion of patients identified as progressors, even by the best model, will remain stable over time. Multiomics signatures will hopefully improve the precision of early prognosis. They will also capture a range of information, ranging from brain networks targeted by the disease, metabolic abnormalities in specific pathways, and distinct genetic backgrounds. The multiomics signature may thus also help elucidate the specific pathophysiological pathways involved, and help refine the A/T/N model. Overall, multiomics biomarkers have the potential to reshape clinical diagnosis, and define new 'bottom-up' cohorts based on markers of underlying pathologies to design and evaluate drugs. Based on this focused but substantial review, we recommend that additional multiomics analyses be performed.