Genome-wide association studies and multiple -omics studies, including proteomics, have revealed that AD pathology is accompanied by perturbations in multiple metabolic and biological pathways that impacts virtually all cell types in the brain (Thompson et al., 2003; Wan et al., 2020a).

Because each protein represents one structural gene, the extensive information represented in proteomics datasets, can be analyzed by appropriate statistical methods to identify structural genes and pathways that are highly correlated with AD, and hence, provide a focus for further research. As is the case for most late-onset neurodegenerative diseases, for Alzheimer’s Disease there are specific aspects of protein dysmetabolism, specifically misfolding and aggregation of specific proteins into abnormal, toxic species that define the neuropathology. Recent work has supported the premise that many biological changes relevant to AD pathophysiology are occurring through mechanisms that are not reflected through changes in mRNA abundance or co-expression (Johnson et al., 2022). This work emphasizes the importance of proteomics analysis and the integration of multiple levels of omics data for understanding the biological mechanisms that underlie development of AD.

Prior multilayer brain proteomic and phosphoproteomic studies have identified molecular networks, including amyloid cascade, inflammation, complement, WNT signaling, TGF-β, BMP signaling, lipid metabolism, iron homeostasis ad membrane transport that are involved in AD progression and contrasted molecular signatures in brain tissue and cerebrospinal fluid proteomic (CSF) with the 5xFAD mouse model (Bai et al., 2020). Reviews of proteomics methods and analysis strategies for unbiased deep profiling of the proteome, specifically differentially expressed proteins and post-translational modifications associated with Alzheimer’s disease have been published (Bai et al., 2021). Multiplexed tandem-mass-tag for ultra-deep proteomics coverage followed by systems biology analysis revealed specific protein signatures for AD across the cortex, CSF and serum that highlighted mitochondrial proteins as involved with the development of AD (9).

For this study, we used human proteomic data from the ROSMAP [Religious Orders Study and the Memory and Aging Project]) study that contains a cohort of 387 individuals well-characterized in terms of sex, race, education, and their state of AD development (Bennett et al., 2012a; Bennett et al., 2012b; Bennett et al., 2018). This dataset contains measurements of 4,913 proteins from the dorsolateral cortex of each individual enrolled in the study.

The mouse proteomic data were obtained from whole brain samples of the CVN-AD AD mouse model. This model faithfully recapitulates the three primary pathologies of human Alzheimer’s disease, amyloid deposits, the accumulation of neurofibrillary tangles, and neuron loss, with minimal genetic manipulation. It is a transgenic model that incorporates human APP bearing the Swedish/Dutch/Iowa (APPSwDI) amyloidogenic mutations under control of the Thy1 promoter (Davis et al., 2004; Wilcock et al., 2008; Colton CA. et al., 2014), on the Nos2 knock-out background. Unlike many other mouse AD models used up until now, the introduced human APP is expressed at a low level, only ∼0.5X the level of endogenous App (Davis et al., 2004). We placed this mutation on the Nos2 knock-out background because inducible nitric oxide synthase has a key role in innate immunity (Bogdan, 2015), and the innate immune response is critical for both the initiation and progression of AD ((Zhang et al., 2013; Kan et al., 2015; Shi and Holtzman, 2018)), However, the expression and activity of human NOS2 are significantly lower than for the mouse Nos2 ((Colton et al., 1996; Mestas and Hughes, 2004)); in order to mimic the human condition we knocked-out Nos2 expression. APPSwDI mice display only amyloid pathology (Davis et al., 2004), and the Nos2 knock-out mice do not exhibit any AD pathology. By contrast, the APPSwDI/Nos2^−/− (CVN-AD) mice develop amyloid plaques and tau pathology, including hyperphosphorylated tau and the accumulation of neurofibrillary tangles, and exhibit neuron loss and learning and memory deficits reminiscent of human AD (Wilcock et al., 2008; Colton C. et al., 2014). Control studies showed that CVN-AD mice exhibit the same pathologies as APPSwDI/huNOS2^Tg, representing CVN-AD engineered to express human NOS2 (Colton CA. et al., 2014). Knocking out the endogenous Nos2 therefore faithfully phenocopies the consequences of the human gene. We have also confirmed the effects of knocking out Nos2 on tau pathology, by crossing another amyloid model, Tg2576 (APPSw), with Nos2 knock-out mice (Colton et al., 2006). Because limited genetic changes, based on well-known and established biology, elicit AD pathology, we chose the CVN-AD model for the studies reported in this paper. The mouse model proteome dataset contains expression measurements of 2014 proteins in 40 samples, and covariate information including mouse model genotype, sex and age.

Regression models have been useful for the analysis of proteomics data. However, the type of regression models must be carefully selected based on their functionalities and advantages in overcoming the challenges of high dimensionality and co-linearity present in the proteomics data. In addition to the problem of high dimensionality, in which the number of proteins (p) far exceeds the number of observations (n), collinearity in the feature space is also a critical issue since expression levels of many related proteins are highly correlated. In this study, we contrast LASSO regression with partial least squares-discriminant analysis (PLS-DA), a variant of Partial Least Squares Regression (PLSR). We compared the mouse and human proteomic analyses at the individual protein and biochemical pathway levels.

Our study focused on statistical approaches to reduce the dimensionality and address the collinearity of “omic” data, specifically proteomic data, in order to compare and contrast across species at the level of individual proteins and biological pathways. Proteomic samples obtained from individuals diagnosed with Alzheimer’s disease and controls were compared with samples from mouse models of AD where the contrast was between mice with a genetic mutation that accelerates the development of AD-related neuropathology and control mice. There were two aims to this study; first to compare statistical methods for addressing the collinearity and high dimensionality of the data for cross-species comparisons, and second to assess the species differences and similarities at the protein and pathway levels. We completed a comprehensive comparison of LASSO regression and Partial Least squares discriminant analysis (PLS-DA) for the analysis of multivariate, high dimensionality datasets with high collinearity. The PLS-DA provided the better statistical performance. The major biological finding of the study was that extracellular matrix proteins and integrin-related pathways were dysregulated in both the human and mouse data. These findings were observable at both the individual protein and pathway levels. The signals in the CVN-AD model that were related to reactive oxygen species (ROS) and innate immunity may reflect adjustments made by the mouse genome to accommodate the loss of Nos2, which is central to both the innate immune response and the generation of ROS species. Likewise, the signal in amyloid clearance could reflect adjustments to the elevated levels of amyloid precursor protein expression in this model, which is estimated to be ca. 1.5X the normal level because it expresses both the human form, at ∼0.5X the mouse level, and mouse App.

The morphogenesis of the CNS and the successful differentiation of all the cell types within it depend on regulatory interactions between the cells and their environments. The extracellular matrix plays an essential role in this communication, and is involved in bidirectional signaling, in-out as well as out-in, from guiding cells and axons during the elaborate processes of developing nerve connections to maintaining tissue homeostasis and regulating cell function, based on ‘nearest neighbors’ signaling. Prior research has suggested the involvement of extracellular matrix (ECM) and integrins in the physiological processes involved in the development of AD. A recent review provided details on the specific ECM proteins that are modulated in the neuropathology of AD (58). Interestingly, the ECM has roles both in regulation of beta amyloid through modulation of amyloid precursor protein (Small et al., 1993; Beyreuther et al., 1996; Ma et al., 2020) and neuroprotection (Cheng et al., 2009; Conejero-Goldberg et al., 2014; Suttkus et al., 2016). ECM substrates fibronectin and vitronectin, but not laminin, promote microglial activation and increased expression of several integrins, cytokines and ECM that are involved in regulation of microglial activity (Milner and Campbell, 2003).

The extracellular matrix (ECM) is comprised of numerous cellular components including proteoglycans, glycosaminoglycans, proteins, proteinases, and cytokines. ECM components are synthesized by both neurons and astrocytes and play an important role in the formation, maintenance, and function of synapses in the central nervous system (CNS) (Dzyubenko et al., 2016). In the CNS, the ECM contains the basement membrane (basal lamina), perineuronal nets, and the neural interstitial matrix (Lau et al., 2013; Mouw et al., 2014; Ma et al., 2020). The ECM is intimately involved in the regulation of beta amyloid. Elastin and heparan sulfate proteoglycans are involved in the upregulation of extracellular Ab. Collagen VI and laminin have been shown to interact with Ab peptides, possibly having an effect on Ab clearance. Our results from the human data show that numerous collagen proteins in addition to amyloid precursor protein are differentially expressed in AD brains in contrast to cognitively normal. Consistent with prior reports, the positive regression coefficients show that concentrations of the collagen proteins were upregulated in the human AD samples relative to cognitively normal (Kalaria and Pax, 1995; van Horssen et al., 2002; Bourasset et al., 2009; Cheng et al., 2009; Tong et al., 2010).

Prior evidence supports the involvement of integrin signaling pathways in the development of AD (58). Studies have suggested the involvement of both the ECM proteins and integrins in modulation of neuroplasticity (Chelyshev et al., 2022), synapse formation (Park and Goda, 2016) and axon regeneration (Pfundstein et al., 2022). It has been suggested that integrins undergo plasticity including clustering through interactions with ECM proteins, modulating ion channels, intracellular calcium and protein kinases signaling, and reorganization of cytoskeletal filaments (Wu and Reddy, 2012). Integrins are also involved in regulation of synapse formation, working with glial signals and neurotransmitter receptor dynamics to regulate synaptic plasticity (Park and Goda, 2016). Integrins also interact with the amyloid precursor protein (APP) (Pfundstein et al., 2022). APP regulates integrin-mediated adhesion and β1-integrins in turn regulate the processing of APP.

This study focused exclusively on analysis of proteomics data. This is in contrast to studies that focus on analysis of mRNA data, either from bulk brain tissue or single cell analysis. There are also studies that have analyzed both proteomics and multilayered omics data. Key findings from one recent study included identification of modules including MAPK/metabolism and matrisome that were associated with AD neuropathology (Johnson et al., 2022). The matrisome module was influenced by the APOE ε4 allele but was not related to the rate of cognitive decline after adjustment for neuropathology (Johnson et al., 2022). The MAPK/metabolism module was strongly associated with the rate of cognitive decline (Johnson et al., 2022). Relevant to our study, the matrisome module consists of a collection of ECM-associated proteins and glycosaminoglycan-binding proteins.

Our study has several strengths. First, two alternative statistical methods for addressing the high collinearity of the proteomic measures were compared with results from each approach reported. Multi-collinearity and dimensionality reduction are common issues for omics studies and this study addressed the question in context of cross-species analysis. For the pathway/signature analysis, well established databases including GO and reactome were used to enable replication studies and other future work. We used a mouse model that reflects the human innate immune response and that leads to age-dependent tau pathology and neuron loss. Careful mapping of mouse to human protein nomenclatures was performed for the proteomics results to allow cross-species comparison. For the mouse model, proteomic determination of both mouse and human APOE concentrations based on specific peptides was performed.

Other statistical methods that address translation between AD mouse models and human data, primarily transcriptomic data, have been published. Some of these approaches share similar a similar statistical basis to our study. Lee et al. presented an approach, “Translatable Components Regression” (Brubaker et al., 2020) that concurrently analyzed transcriptomic data from human brain and AD mouse models to identify pathway-level signatures present in the human data that were predictive of mouse model disease status (Lee et al., 2021). For this approach a principal component analysis (PCA) space for human data is derived and projected in a mouse dataset (Lee et al., 2021). Importantly, this work also utilized linear models to differentiate disease-specific effects from aging and demonstrated that the analysis framework identified cross-species signatures that do not necessarily dominate in at least one of the datasets separately (Lee et al., 2021). Other approaches have focused on cross-species gene set analysis (Miller et al., 2010; Burns et al., 2015), network analysis (Zhang et al., 2013; Mostafavi et al., 2018; Bai et al., 2020; Wang et al., 2020) and meta analyses of co-expression data (Friedman et al., 2018; Wan et al., 2020b). Of particular note are approaches that utilize ultra-deep level proteomics analysis coupled with integrated systems-biology analysis (Bai et al., 2020; Wang et al., 2020; Bai et al., 2021).

Our study also has several limitations. The sample size for the mouse study is relatively small; a larger sample would increase statistical power to detect differences between the CVN-AD and control mice. The comparison between the mouse and human results at both the individual protein and pathway levels are based on single datasets with comparisons to prior literature for the statistically significant proteins and pathways. Future studies could be planned to replicate the mouse, human and combined results using independent datasets. Finally, we used empirical thresholds of 20 proteins for inclusion in the pathway analysis and an FDR significance level of 0.05 for selection of pathways. Alternative analytical approaches and different thresholds may provide additional insight about these datasets.

Future work will assess the impact of age, sex and APOE genotype on the within-species and cross-species comparisons using approaches including the gene set enrichment likelihood ratio test which quantifies gene set enrichment accounting for covariate effects at the gene set level (Bryan et al., 2021), and it will involve additional mouse models of AD that incorporate known genetic risk factors, such as APOE4, and that are created by targeted replacement of the endogenous mouse gene with the hetrospecific isofunctional human homolog, to avoid potential non-specific effects that can blur transgenetic manipulations. A high research priority of the NIH is the development of improved mouse models of Alzheimer’s disease to improve reproducibility, transparency and translatability (https://www.model-ad.org/), and a number of new models have been developed (https://www.alzforum.org/news/research-news/cornucopia-loads-new-mouse-models-available). To identify the most translatable models, comparisons with human data bases, as we have done, will be essential. The intentional incorporation of alterations to the ECM and integrins along the lines discovered here might be a useful approach. In any event, the application of methods we developed here will be helpful in guiding new model development.

In summary, this study addressed several of the critical issues involved in cross-species comparisons of omic data, specifically proteomic data. In addition to providing guidance on alternative statistical approaches to analyze the data, the approach may help inform the development of mouse models that are more relevant to the study of human late-onset Alzheimer’s disease and provide insight about specific biological pathways identified as differentially regulated in individuals with AD and in AD mouse models. The biological results from the cross-species analysis point to specific protein targets that involve the extracellular matrix and integrin pathways. These results can be used to plan future, focused studies on longitudinal changes of these proteins and pathways in context of the development of Alzheimer’s Disease.