Translating the Post-Mortem Brain Multi-Omics Molecular Taxonomy of Alzheimer's Dementia to Living Humans

Abstract

Alzheimer's disease (AD) dementia is characterized by significant molecular and phenotypic heterogeneity, which confounds its mechanistic understanding, diagnosis, and effective treatment. In this study, we harness the most comprehensive dataset of paired ante-mortem blood omics, clinical, psychological, and post-mortem brain multi-omics data and neuroimaging to extensively characterize and translate the molecular taxonomy of AD dementia to living individuals. First, utilizing a comprehensive integration of eight complementary molecular layers from brain multi-omics data (N = 1,189), we identified three distinct molecular AD dementia subtypes exhibiting strong associations with cognitive decline, sex, psychological traits, brain morphology, and characterized by specific cellular and molecular drivers involving immune, vascular, and oligodendrocyte precursor cells. Next, in a significant translational effort, we developed predictive models to convert these advanced brain-derived molecular profiles (AD dementia pseudotimes and subtypes) into blood-, MRI- and psychological traits-based markers. The translation results underscore both the promise of these models and the opportunities for further enhancement. Our findings enhance the understanding of AD heterogeneity, underscore the value of multi-scale molecular approaches for elucidating causal mechanisms, and lay the groundwork for the development of novel therapies in living persons that target multi-level brain molecular subtypes of AD dementia.

Figure 1 |

Schematic approach for identifying, characterizing and translating the brain multi-omics molecular AD dementia taxonomy. A) Multi-modal ante-mortem data including in-vivo blood, longitudinal multi-domain cognition, survey, and neuroimaging data, paired with post-mortem brain data, from the same ROSMAP participants. B) A contrastive machine-learning algorithm was used to integrate the brain’s multi-omics molecular data, obtaining a multidimensional disease space where subjects define biologically-distinctive subtrajectories from no cognitive impairment (NCI) to AD dementia. We first obtained for each participant an AD dementia pseudotime score and subsequently an AD dementia subtype. C) Subsequently, each AD dementia subtype was further characterized by identifying specific cell-type causal drivers and multi-omics features, providing insights into distinct disease mechanisms. D-E) Lastly, we tested whether brain-derived pseudotimes and AD dementia subtypes can be predicted by multi-modal data obtained from the same participants prior to death.

Figure 2 |

Characterization of multi-omics molecular AD dementia pseudotimes. A) AD dementia pseudotime correlation values with multiple cognitive domains and their rates of change over time (all P<0.05, FWE-corrected, based on correlation tests with permutations after adjustment by age, sex, and educational level). Notice the strong associations with episodic memory, global cognition, and their temporal declines. B) Standardized size effects for AD dementia pseudotimes, from the current study and our previous four-omics version from 2022 (10), when concurrently explaining cognitive performance and decline. For each cognitive variable, we conducted linear regression models incorporating both sets of AD dementia pseudotime as predictors along with age, sex, and educational level as covariables. C) AD dementia pseudotime distributions across clinical diagnoses (NCI, MCI, AD; p-value FWE-corrected, based on Ancova tests with age, sex, and educational level as covariables). D-E) AD dementia pseudotime correlation with AD and non-AD specific neuropathologies, respectively (FWE-corrected, based on correlation tests with permutations after adjustment by age, sex, and educational level). F) Distribution of all multi-omics markers’s contributions to AD dementia pseudotimes. G) Omic-specific contributions. H) Top 1% contributing omic markers.

Figure 3 |

Distinctive multi-level AD dementia subtypes identified with post-mortem brain data. A) Three-dimensional visualization of the identified AD dementia subtypes’ subtrajectories from cognitive normality to AD dementia. X and Y axes correspond to the first two dimensions on the reduced multi-omics molecular space, and Z axis to the AD dementia pseudotime. B) Number of participants per AD subtype and background subpopulation (NCI), and corresponding significance obtained with randomization testing (all P<0.001, FWE-corrected). C) Subtype-specific sex proportions, normalizing across all the participants in the cohort. D) Subtype-specific statistical differences in severity of AD-related and non-AD neuropathologies, resulting from ANCOVA tests with subtype as grouping variable (color scale corresponds to F-values, FWE-corrected via permutation tests). E) Across subtypes comparison tests in cognitive performance. Tukey’s honest significant difference criterion was used (with P<0.05 implying significance differences in marginal means; Materials and Methods, Additional Statistical Analysis).

Figure 4 |

Subtypes-specific multi-level molecular alterations. A) Top 30 molecular features distinctively altered for each subtype while preserved in the other subtypes (for complete lists, see Files S1-S3). Contribution values per subtype are normalized by the maximum. B) The raw values displayed above the bars represent the unique counts of distinct contributors for each subtype across specific omics. The first reported percentage for omic type (i) in a given subtype (j) indicates the number of markers from omic type (i) that are uniquely altered for that subtype, normalized by the total number of uniquely altered markers for subtype (j) across all omics. The second percentage for omic type (i) in a subtype (j) reflects subtype-specific percentages of distinctive alterations across all subtypes. Specifically, it corresponds to the number of markers from omic type (i) that are uniquely altered for subtype (j), normalized by the total number of markers of omic type (i) uniquely altered across all subtypes.

Figure 5 |

Cellular-level molecular causal drivers for distinct AD subtypes. A) Top 99-percentile causal drivers associated with each subtype. B) Average cell-type to cell-type causal effects.

Figure 6 |

Blood and MRI based predictions of post-mortem brain molecular AD dementia pseudotimes. A) Top 50 blood markers significantly correlated with pseudotime (all P<0.05, FDR-corrected; see Table S3 for a complete list), adjusting for age, sex, and educational level. Correlations were standardized by comparison with 1000 randomized permutations. B) For each blood data modality, percentage of significantly correlated markers with AD dementia pseudotime. C) DBM maps of spatial associations between AD dementia pseudotime and brain structural properties. Warm and cold colors indicate the magnitude of positive and negative associations of regional brain deformation, respectively. D) Blood and regional MRI 10-folds cross-validated predictions of individual AD dementia pseudotime values. See Materials and Methods, Statistical analyses and DBM analysis.

Figure 7 |

Brain morphology and psychological traits patterns of molecular AD dementia subtypes. A) Structural brain comparisons (all FWE p<0.05; including statistical adjustment for age at death, sex, educational level, postmortem scanning interval and location). Warm and cold colors indicate larger and smaller volumes respectively. B) Association of AD dementia subtypes with negative and positive psychological traits (values were z-transformed for illustration purposes; ANCOVA tests included age, sex and educational level as demographic covariates).

Figure 8 |

Blood, regional MRI and psychological traits-based predictions of post-mortem brain multi-omics AD dementia subtypes. Prediction AUC for every pair of subgroups included in the classification tasks. Notice that the first three columns correspond to subtype-specific vs NCI classifications, while the last three columns correspond to subtype vs subtype classifications.