Brain high-throughput multi-omics data reveal molecular heterogeneity in Alzheimer’s disease

Unbiased data-driven omic approaches are revealing the molecular heterogeneity of Alzheimer disease. Here, we used machine learning approaches to integrate high-throughput transcriptomic, proteomic, metabolomic, and lipidomic profiles with clinical and neuropathological data from multiple human AD cohorts. We discovered 4 unique multimodal molecular profiles, one of them showing signs of poor cognitive function, a faster pace of disease progression, shorter survival with the disease, severe neurodegeneration and astrogliosis, and reduced levels of metabolomic profiles. We found this molecular profile to be present in multiple affected cortical regions associated with higher Braak tau scores and significant dysregulation of synapse-related genes, endocytosis, phagosome, and mTOR signaling pathways altered in AD early and late stages. AD cross-omics data integration with transcriptomic data from an SNCA mouse model revealed an overlapping signature. Furthermore, we leveraged single-nuclei RNA-seq data to identify distinct cell-types that most likely mediate molecular profiles. Lastly, we identified that the multimodal clusters uncovered cerebrospinal fluid biomarkers poised to monitor AD progression and possibly cognition. Our cross-omics analyses provide novel critical molecular insights into AD.


Major comments:
-A cri6cal point is that the subtyping was based on direct clustering on the omics data (the features), without controlling in any way for the par6cipants' disease stages.As it has been shown (e.g., Young et al, 2018, Nat.Comms) two relevant "axes" of variability/heterogeneity coexist in every diseased popula6on, one cause by the differences in diseases stages and other by disease subvariants.Trying to detect any of these "axes" independently without considering the other is a common error in many studies focused on progression and heterogeneity.The iden6fied subtypes here are very likely contaminated with variability from having subjects at different stages of disease advance.It may be the case that they are primarily capturing subjects' proximity in the disease's long 6meline, and not similarity in disease mechanisms or subvariants.The authors should consider this important point, i.e., how to disentangle between real underlying subtypes and data clusters poten6ally confounded by similari6es in disease severity.
We appreciate the reviewer for bringing up this significant maZer.We fully agree that it is important to control for disease stages and that genotypes should be considered.However, our study's main goal is not to generate subtypes of the disease, our study aimed at integra6ng mul6modal molecular data with demographic, clinical and neuropathological data to understand the heterogeneity observed in mul6ple modali6es of omics data from postmortem brains.We did not stra6fy par6cipants by their cogni6ve status (e.g., CDR at death) nor pathological stage (e.g., Braak), as we aimed to evaluate whether subtle molecular differences among sporadic AD omics can provide hints for a granular staging of disease (quan6ta6ve molecular neuropathology).We want to emphasize that we inten6onally avoid referring to these clusters throughout the manuscript as "subtypes" since we are focused on the study of the heterogeneity among sporadic ADs (non-carriers of pathological muta6ons) and iden6fy emergent molecular proper6es.For this reason, we chose to call them "profiles" instead.We agree with the reviewer that disease staging and subtyping are highly valuable, but it is outside the scope of our study.This will encompass the inclusion of data from mul6ple omic layers and brain regions for the same cohort of donors.Instead, our study iden6fied and replicated molecular profiles in mul6ple cohorts that ascertained different cor6cal regions, but we agree that molecular subtyping would be highly valuable.We hope this work emphasizes the need for parallel simultaneous ascertainment of mul6ple omics in mul6ple brain region.In addi6on, we iden6fied that molecular markers featuring some of the clusters could be used to track changes in living pa6ents using CSF data, opening the new biomarkers anchored in the biology of AD iden6fied in postmortem brain 6ssue.Furthermore, we recognize the importance of disease staging and subtyping, thus, we employed the method (SuStaIn), suggested by the reviewer, which proved to uncover data-driven disease phenotypes with dis6nct temporal progression paZerns to analyze proteomic data from the cerebrospinal fluid.We selected 8 CSF synap6c proteins associated with an increased risk of demen6a progression and showing different magnitudes of change in earlier AD stages compared to later stages using more tradi6onal neuropathological staging.We applied the SuStaIn method using the CSF proteomic data for these proteins and iden6fied two "subtypes" of AD that were determined to fit our data beZer.These 8 CSF biomarkers showed different progression paZerns across the two subtypes.For instance, among these biomarkers, IGF1 tend to become abnormal in late stages in Subtype1 but predicted to be abnormal in early stages in Subtype2.These observa6ons may suggest that IGF1 might be a good candidate as an early CSF biomarker for AD.In addi6on, NRXN3 (a synap6c cell-cell adhesion molecule) occurs at early stage in Subtype2 but later in Subtype1 which may suggest an earlier increased AD risk for pa6ents in Subtype2 than the ones in Subtype1.In contrast, PLXNC1 start to become abnormal in early stages in Subtype1 and then became stable before star6ng showing abnormality again at later stages, whereas in Subtype2 it is only predicted to be abnormal at later stages.These results show that these CSF protein levels iden6fied dis6nct temporal progression paZerns in AD and detected molecular heterogeneity not only among AD cases but also with the presence of control subjects with CDR = 0. We included these results in the revised manuscript sec6on "Cross-omics integra6on iden6fied CSF synap6c biomarkers for the molecular staging of AD" (page 28-29, lines 673-692): "….Furthermore, we inves:gated whether these 11 CSF proteins could also iden:fy molecular differences among AD pa:ents and their temporal progression paGerns using the Subtype and Stage Inference (SuStaIn) machine learning technique [135,136] .PLXNC1 was shown to be associated with neuronal loss [137] and therefore, these observa:ons may indicate that PLXNC1 might be a good candidate as a biomarker for monitoring the synapse and neuronal loss.Altogether, these results show that these CSF protein levels iden:fied dis:nct temporal progression paGerns in AD and detected molecular heterogeneity among AD cases and in the presence of control subjects with CDR = 0…." Disease subvariants and gene6c heterogeneity: We would like to men6on that we iden6fied similari6es and differences in the analy6cal approaches to the study from Young and collaborators.Alzheimer's Disease has both an Autosomal Dominant (ADAD), and complex gene6c architecture.Indeed, carriers of pathogenic muta6ons in APP, PSEN1, and PSEN2, which we refer to as ADAD, tend to have a dis6nc6ve clinical manifesta6on and a younger age at the onset of cogni6ve decline compared to sporadic non-carriers, which is referred to as sporadic AD.Our group is highly invested in determining whether the gene6c factors iden6fied in sporadic AD play any role in ADAD 1 .We have compared the metabolomic 2 , cellular popula6on structure 3 , and transcrip6onal differences at the single cell-type resolu6on 4 between ADAD and sporadic noncarriers.All our studies indicate that the parietal cortex tends to show a unique and more pronounced aberrant molecular profile than the one we observe in sporadic AD.That is why we did not include different genotypes (ADAD) in this study.Instead, we used unsupervised machine learning approaches to determine whether the brain omics show alterna6ve profiles as emergent proper6es.We mainly aimed at capturing the heterogeneity in postmortem sporadic AD brains, like the study of FTD heterogeneity among c9orf72 carriers reported by Young et al., 2018 and referenced by the reviewer.We want to men6on that we tested whether any of these profiles/clusters are associated with APOE e4 allele, or a polygenic risk score that captures genome-wide risk in sporadic cases.To our surprise and similar to Neff et al., 2021 5 , we found no significant associa6on.These observa6ons were highlighted in several places in the original manuscript, including the Results sub-sec6on "A mul6modal profile of AD brains is associated with poor cogni6ve func6on and molecular aZributes" (page 9, line 196): "….The Knight-C4 cluster was not associated with either AD polygenic risk score (S2 We also men6oned these observa6ons in the "Discussion" sec6on (page 28, line 669): "….Similar to previous studies, we did not find a significant associa:on of APOE e4 or PRS with any molecular profile….".
-Also notable, the study misses relevant references, in par6cular those that have already presented extended mul6-omics integra6on in AD, proposing both staging and subtyping across mul6ple popula6ons.The authors should accurately describe the field and compare main results with previous.See for instance:Yasser Iturria-Medina, Quadri Adewale, Ahmed F. Khan, Simon Ducharme, Pedro Rosa-Neto, Kieran O'Donnell, Vladislav A. Petyuk, Serge Gauthier, Philip L De Jager, John Breitner, David A. BenneZ, 2022.Unified Epigenomic, Transcriptomic, Proteomic, and Metabolomic Taxonomy of Alzheimer's Disease Progression and Heterogeneity.Science Advances.Vol 8, Issue 46.DOI: 10.1126/sciadv.abo6764 We thank the reviewer for bringing this point to our aZen6on.Although we have included and discussed mul6ple previous studies addressing pathological and molecular heterogeneity in AD (references 1-9 and 23-32 in the original manuscript) and par6cularly those that present extended mul6-omics integra6on in AD (Ref.23-32), we missed the study from Iturria-Medina et al.We have included this study and addi6onal studies in the revised manuscript and extended our "Introduc6on" and "Discussion" sec6ons to describe previous studies, including this study, in more detail and discuss how our study is different.
The following text was added to the introduc6on (page 4-5, lines 80-102): "…Previous studies have found molecular features associated with different AD clinical and pathological profiles [2,3,10].For instance, CSF proteins for data-driven clustering found an associa:on between hyperplas:city and increased BACE1 levels with clinical differences in AD pa:ents [2].Similarly, a system biology approach studying transcriptomics from mul:ple cor:cal regions iden:fied AD molecular subtypes associated with mul:ple dysregulated pathways including suscep:bility to tau-mediated neurodegenera:on, amyloid-β neuroinflamma:on, synap:c signaling, immune ac:vity, mitochondria organiza:on, and myelina:on [3].Despite the substan:al effort in these studies to iden:fy molecular associa:ons with clinical and pathological AD features, broader conclusions are limited by its nature, as single-omic analyses typically capture changes in a single component of the biological cascade.Addi:onal studies have leveraged cross-omics data integra:on approaches to study the altera:ons of molecular and cellular pathways underlying AD pathophysiology [35,36] providing more evidence that singleomic analyses do not capture most pathways involved in AD pathology.Moreover, CSF mul:-omics molecular signatures differen:ally related to AD pathology have also been iden:fied [24].Recently, integra:ng AD brain and blood mul:-omics data with clinical and pathological data iden:fied three molecular subtypes and inferred their respec:ve molecular trajectories that diverge from the neuropathologically free control brains [10].Furthermore, the molecular altera:ons underlying AD progression and heterogeneity are more clearly understood when studied in mul:ple brain regions with different burdens of AD pathologies.For instance, the parietal cortex, an understudied brain region affected in later stages of AD [16,37,38], can beGer capture the more ini:al molecular changes in AD e:ology compared to other severely affected regions (e.g.DLPFC), which will help to iden:fy molecular dysregula:ons underlying AD progression before a higher burden of Aβ plaque and tangle occurs.…." We also extended our "Discussion" sec6on to include the following (page 32, lines 769-780): "….Our results are consistent with previous studies that report molecular correlates with AD clinical data and neuropathological staging [3,10].We show that AD heterogeneity cannot be en:rely explained by neuropathological variables (e.g.b-amyloid or tau accumula:on), APOE ℇ4 allele carrier status, or even by differences in age or sex.Most of the previous studies have included a single-omic layer [2,3] capturing molecular changes in a single component of the biological cascade, or have analyzed a single AD cohort and a unique brain region (lacking independent replica:on) [10] as we believe that molecular changes and heterogeneity in AD are beGer understood when studied in mul:ple brain regions.Our study leverages mul:ple AD cohorts (Knight ADRC, MSBB, ROSMAP) and mul:ple brain regions (parietal cortex, DLPFC, PHG) offering us substan:al insights into AD such as our findings of the significant enrichment of AD molecular profiles in synap:c genes and pathways that allow us to iden:fy CSF synap:c biomarkers…..".
-Also, the manuscript should make clear from the introduc6on the added value in the context of the previous work (no, not all previous AD stra6fica6ons have been restricted to a single omic, as claimed in the Introduc6on).As a sugges6on, the authors may accentuate their use of single-nucleus RNA for subtypes-subtypes comparisons.
We agree with the reviewer that this was not described enough in the introduc6on.We have extended our "Introduc6on" (page 4-5, lines 80-102) and "Discussion" (page 32, lines 769-780) sec6ons as men6oned in the previous response to describe more the previous work and highlighted the added value of our work in the context of the previous work with the texts shown in the previous response.

-
Across the manuscript's text, it is unclear and confusing if the snRNA-seq data was used or not for the subtypes discovery, or only the bulk transcriptomics.Even in the abstract, it reads like if it was considered, which would add value to the analyses in terms of novelty, but on Figure 1 it appears on the other/right side, for post-clustering comparisons.I strongly encourage the authors to make this the clearest possible (it is not clear also in the methods).And, if not considered, to clean all possible insinua6ons that it was (for example, when claiming in the abstract "to integrate high-throughput bulk and single-nucleus transcriptomic, proteomic, …").
We apologize for the lack of clarity regarding the use of snRNA-seq data in the previous version of the manuscript.Single-nuclei RNA-seq data is s6ll not available for all of the cohorts and was not used in the discovery of cross-omics profiles.Furthermore, the exis6ng single-cell data available is not as comprehensive as transcriptomics data from 6ssue homogenates (bulk RNAseq) in terms of subject inclusion or sample size.Therefore, we used bulk RNA-seq data to integrate with other omics to study AD molecular profiles.The single-cell data was used to determine the cell-types expressing those genes that we characteris6c in mul6-omics studies.We have included more details in this revised version of manuscript and provided more clarifica6ons and descrip6on of how the snRNA-seq data was used.We have updated the sentence in the Abstract (page 2, line 29) and removed single-nucleus word as follows: "…to integrate high-throughput transcriptomic, proteomic, metabolomic, and lipidomic profiles…".
In addi6on, we extended the Abstract to explain the uses of snRNA-seq data and elucidate that snRNA-seq data was combined with the iden6fied molecular profiles as follows (page 2, lines 38-39): "…Furthermore, we leveraged single-nuclei RNA-seq data to iden:fy dis:nct cell-types that most likely mediate molecular profiles….".
Finally, we updated the 6tle of a results sec6on to reflect this more clearly (page 22, line 511):

"Single-nucleus transcriptomics suggests neurons and alterna:ve glial cells mediate crossomics profiles of AD"
Reviewer #2: The study Brain cross-omics integra6on in Alzheimer's disease by Eteleeb and colleagues is an important study and approach in the analysis and integra6on of molecular and clinical data in Alzheimer's disease.This topic is now of increasing interest given the significance of the disease, powerful computa6onal methods, and the many available exis6ng and new data sources.
We thank the reviewer for the construc6ve and suppor6ve feedback.We appreciate the reviewer's recogni6on of the importance of our work for the field and we extend our gra6tude for the suppor6ve comments on the structure of the manuscript.
The introduc6on is well wriZen and surveys the significance of single cell genomics and AD biomarkers including molecular and trajectory analysis.Cross-omics as proposed in this study approaches reveal how complex biomolecular profiles change in associa6on with a disease and the rela6onships and correla6ons between dis6nct classes of biological molecules.The authors employ a Bayesian integra6ve clustering method to integrate three omics datasets then applying mul6ple associa6on analyses and including 144 associa6on with clinical and neuropathological aZributes (e.g., sex, age of death, age at onset, post-mortem interval, CDR, Braak amyloid stage, Braak neurofibrillary tangle stage), survival, and differen6al expression analyses to characterize AD molecular profiles.
We thank the reviewer for this elegant summary of the work and their support.
The figures are in general well done and Figure 1 presents a clear outline of the approach and methodology.The authors discovered four unique mul6modal molecular profiles, one showing signs of poor cogni6ve func6on, a faster pace of disease progression, shorter survival with the disease, severe neurodegenera6on and astrogliosis, and reduced levels of metabolomic profiles.This profile shows similar cellular and molecular profiles in mul6ple affected cor6cal regions associated with higher Braak tau scores and significant dysregula6on of synapse-related genes and endocytosis, phagosome, signaling pathways altered in AD early and late stages.These are interes6ng associa6ons illustra6ng the power of this type of analysis.While the analysis methods presented are largely standard, there are done carefully and sta6s6cally valid.

Reviewer #3:
The manuscript "Brain cross-omics integra6on in Alzheimer's disease" by Eteleeb et al leveraged a large number of omics measurements to iden6fy AD subtypes, followed by their detailed molecular analysis.The topic of AD subtyping has been gaining momentum lately.Perhaps this is afforded because of the accumula6on of the omics data.The presented study is on par published on the same topic and is worthy contribu6on to the overall forum discussing the subtypes of the AD.There are number of issues that I'd like to see addressed prior the publica6on.
We thank the reviewer for the recognizing the importance of our work and we address their concerns below.

1.
Please discuss the choice of the parietal cortex for studying the AD.What was the ra6onale for selec6ng this brain region?What are the pros and cons compare to a more conven6onal one -DLPFC?
We thank the reviewer for bringing this point to our aZen6on.The parietal cortex is an understudied brain region that is affected in later stages of AD 4,6,7 .Thus, at the 6me of death of most AD pa6ents, the parietal cortex can capture the more ini6al molecular changes in AD e6ology, compared to other severely affected regions that are usually studied in other cohorts, for example the DLPFC for ROSMAP.Furthermore, 6ssue Microarray data from our collaborator, Dr. Sutherland, Director of the New South Wales -Brain Tissue Resource Center (NSW-BTRC), shows that the parietal cortex has a lower burden of plaques and tangles than the DLPFC, but higher than other regions, like the inferior temporal cortex.We include this confiden6al data for evalua6on by the reviewers: Thus, we used data from parietal cortex in our discovery stage, to study the aberrant and dysregulated molecules in the presence of AD pathological lesions, before a higher burden of Ab plaques and NFTs occurs along extended periods of 6mes, and possibly addi6onal compensatory effects confound ini6al changes associated with e6ology.We replicated our ini6al findings in two more brain regions, the DLPFC in ROSMAP cohort and parahippocampal gyrus (BA36) from the MSBB cohort.We believe that studying mul6ple regions with different burdens of Ab and plaques will allow us to beZer understand the molecular changes associated with the progression of the disease, but at the present 6me, we are constrained to generate more data systema6cally in mul6ple brain regions from the same donors.
Since we believe this is an important point that should be included in the revised manuscript, we have extended our "Discussion" sec6on to include the above illustra6on of the choice of using this region (page 29, lines 699-705): "… In the discovery stage, we used data from the parietal cortex, an understudied brain region affected in later stages of AD [16,37,38].The parietal cortex was chosen to study dysregulated molecules in the presence of AD pathological lesions before a higher burden of Ab plaques and NFTs occurs along extended periods of time, and possibly additional compensatory effects confound initial changes associated with etiology.In addition, the parietal cortex typically captures the more initial molecular changes in AD etiology compared to other severely affected regions (e.g., DLPFC), making it a suitable region for our study…."

2.
The main finding of the manuscript are the 4 clusters or subtypes of AD.Since the code was not available, I could only review the analysis based on the descrip6on in the methods sec6on.Use of BIC and deviance criteria for selec6on of the number of clusters is a good choice.However, one discrepancy needs a liZle bit more effort to explain.The supplementary figure 1 states that the op6mum is at number of clusters 3 (if I am reading the X axis right).The legend to the same figure states that K=3, thus the samples are divided into 4 clusters.This is confusing.Please double check the number of clusters and clarify the ra6onale and the arithme6c in the figure and the legend.
We thank the reviewer for bringing this point to our aZen6on and we apologize for the confusion on details on selec6ng the number of clusters.iClusterBayes 8 fits mul6ple clustering solu6ons based on a number of clusters (k, aka eigne features) and then uses Bayesian informa6on criteria (BIC) and deviance ra6o to select the op6mal model.When BIC value reaches the minimum and the deviance ra6o reaches a plateau at a specific K value, the op6mal clustering solu6on with that K value is selected which indicates that the model fits the data best when the samples are divided into K + 1 clusters and therefore the number of clusters will be k + 1.In the previous version of the manuscript, we chose to follow the nomenclature used by the developers of the method, and this cri6cal informa6on was provided in Supplementary Figure 1 legend.However, to avoid any confusion and enhance the readability of the manuscript, we have updated Supplementary Figure 1 and its legend as shown below: Rebu,al Figure 2. The deviance ra4o (top) and the Bayesian informa4on criterion-BIC (bo,om) at each K (number of eigen features).The op6mal solu6on is observed with K =3 when the BIC is the minimum and the deviance ra6o is the maximum dividing our samples into four clusters (k+1).
We have uploaded our analysis code used to integrate all cross-omics data for the three cohorts as well as all downstream analyses to our GitHub repository which can be found at hZps://github.com/HarariLab/Cross-omics-Integra6on.

3.
The discussion sec6on is completely missing other studies that inves6gated the number of subtypes of AD.Please discuss your results within the context of other findings.
We thank the reviewer for this sugges6on.We recognize that we had included mul6ple previous studies addressing pathological and molecular subtypes in AD in the introduc6on sec6on in the previous version of the manuscript (references 1-9), and now provided a more comprehensive descrip6on in this new version.We recognize that our findings should also be contextualized in the discussion.Thus, we have extended our "Discussion" in the revised manuscript to contextualize our results within the previous studies that inves6gate AD molecular heterogeneity using mul6-omics data.The following paragraph was added to the revised manuscript (page 32, lines 769-780): "….Our results are consistent with previous studies that report molecular correlates with AD clinical data and neuropathological staging [3,10].We show that AD heterogeneity cannot be entirely explained by neuropathological variables (e.g.b-amyloid or tau accumulation), APOE ℇ4 allele carrier status, or even by differences in age or sex.Most of the previous studies have included a single-omic layer [2,3] capturing molecular changes in a single component of the biological cascade, or have analyzed a single AD cohort and a unique brain region (lacking independent replication) [10] as we believe that molecular changes and heterogeneity in AD are better understood when studied in multiple brain regions.Our study leverages multiple AD cohorts (Knight ADRC, MSBB, ROSMAP) and multiple brain regions (parietal cortex, DLPFC, PHG) offering us substantial insights into AD such as our findings of the significant enrichment of AD molecular profiles in synaptic genes and pathways that allow us to identify CSF synaptic biomarkers….".

4.
Legend of Figure 1 (page 6, line 128) states "genes=60,754".Human genome is limited to 20K or so genes.It isn't clear what authors referring to.Are these collec6ve measurements of the canonical mRNA transcripts across 3 cohorts or number of transcripts including splice isoforms?I'll leave this up to authors discre6on if they should switch to the term mRNA transcript rather than gene.
We thank the reviewer for highligh6ng this point and we apologize for the lack of clarity.The 60,754 represents all features in the transcriptome including protein-coding genes, long noncoding RNAs, small RNAs, pseudogene, etc.It is a comprehensive gene annota6on data.We have clarified this in the revised manuscript and used the term "features" to represent all entries/bio-features in the en6re transcriptome.

5.
The term "cross-omics" triggers a minor ques6on.Did the authors decide to use the term "cross-omic" to set themselves apart from the rest of the studies that use the term "mul6-omic"?Moreover, once sentence in the discussion sec6on uses both terms in one sentence (page 28, line 664).If there is no good ra6onale for using a dis6nct term, please consider switching to "mul6-omic".Though I leave this issue to author's discre6on too.
We thank the reviewer for raising this interes6ng point.We fully agree with the reviewer that there is no a clearly established difference between mul6-omics and cross-omics yet.Indeed, these terms are some6mes interchanged.
It is our understanding that mul6-omics approaches focus on the simultaneous profiling of mul6omics or other modali6es from the same cells 9,10 .In fact, this term is currently being used to describe plaworms and chemistries that allow the genera6on of gene expression and ATAC-seq in parallel in spa6al resolved transcriptomics and similarly 10X Genomics mul6ome kits.
Nowadays there is a tendency to refer as cross-omics to those approaches that integrate and analyze mul6ple omics data sets from different bio-samples.Indeed, this term has been employed in mul6ple other manuscripts and areas 11 , including T2D 12 , Parkinson Disease 13 , aggrega6on of summary data 14 and White MaZer hyperintensi6es 15 .
We are adop6ng this new term, as we believe it reflects more accurately our work.

6.
Please state how exactly code will be shared.I really hope it won't be "available upon request".Though ideally, I'd prefer the code available at the 6me of the review.

Figure 1 (
Confiden4al): Neuropathological characteriza4on of global AD pathology in the post-mortem human brain 4ssue.(a) Tissue microarray (TMA) was constructed by drilling 2mm 6ssue plugs from FFPE blocks of grey and adjacent white maZer across the brain and immunostained for Ab, total tau and NeuN.(b) Areal frac6ons of amyloid-plaques (Ab) and tau/neurofibrillary pathology respec6vely across the brains of four sporadic AD pa6ents: two males (74 and 80 yo) and two females (78 and 90 yo).Plaques are most dense in the prefrontal cortex (boZom let, highlighted in red color) and lowest in the cerebellum.Neurofibrillary pathology was most dense in prefrontal cortex (boZom right, highlighted in red color) and lowest in the cerebellum.The parietal cortex presents intermediate Ab and neurofibrillary pathology.