Elucidating the neuropathological and molecular heterogeneity of amyloid beta and tau in Alzheimer's disease through machine learning and transcriptomic integration.
Zhao K, Xie H, et al. • Alzheimer's & dementia : the journal of the Alzheimer's Association • 2026
A contrastive deep learning framework applied to resting-state fMRI identified A+-specific functional network signatures localized to the right superior temporal and anterior cingulate cortices, linked to attention and memory, with transcriptomic enrichment implicating synaptic dysfunction and glial activity, separate from shared aging-associated network variation.
Key Findings
Methods
A contrastive deep learning framework successfully identified amyloid-positive (A+)-specific functional network dimensions from resting-state fMRI data in older adults.
The study analyzed resting-state fMRI from 289 older adults classified as A+ (n=129) or A- (n=160) based on cerebrospinal fluid biomarkers.
A contrastive graph learning approach was used to separate A+-specific network variation from broader background biological variability.
The framework was designed to distinguish amyloid-associated functional network changes from overlapping aging-associated and non-amyloid biological processes.
Results
A+-specific functional network signatures were spatially localized to the right superior temporal and anterior cingulate cortices.
These A+-specific signatures were distinct from network dimensions shared between A+ and A- individuals.
The localization to right superior temporal and anterior cingulate cortices was linked to attention and memory functions.
This spatial specificity was identified through the contrastive deep learning framework applied to resting-state functional connectivity data.
Results
A+-specific functional network signatures were used to predict individual amyloid beta (Aβ) and phosphorylated tau (p-tau) levels.
The contrastive framework identified network dimensions capable of predicting both Aβ and p-tau biomarker levels at the individual level.
Both CSF-derived Aβ and p-tau were used as reference biomarkers in the classification and prediction framework.
This predictive capacity suggests the identified network dimensions capture biologically meaningful variance related to AD pathology.
Results
Transcriptomic enrichment analysis of A+-specific network signatures implicated synaptic dysfunction and glial activity.
Molecular pathways associated with synaptic dysfunction and glial activity were enriched in regions showing A+-specific functional network alterations.
Transcriptomic integration was used to link imaging-derived network signatures to underlying molecular biology.
This approach connected macroscale functional network changes to cellular and molecular mechanisms relevant to AD pathology.
Results
Functional network dimensions shared between A+ and A- individuals involved language-related regions and aging-associated molecular pathways.
These shared dimensions were distinct from A+-specific signatures and captured broader, non-amyloid-specific biological variation.
Language-related brain regions were prominently involved in the shared network dimensions.
The shared dimensions were transcriptomically enriched for aging-associated molecular pathways rather than AD-specific pathology.
This dissociation demonstrates that the contrastive framework can separate amyloid-associated from aging-associated functional network variation.
Discussion
The study demonstrates neuropathological and molecular heterogeneity in AD-related biomarkers through the identification of both A+-specific and shared network dimensions.
The contrastive framework revealed multiple distinct dimensions of functional network variation with different biological and spatial profiles.
Heterogeneity was observed both at the level of functional network topology and at the transcriptomic level.
The findings provide insight into the heterogeneity of AD-related biomarkers and cognitive dysfunction as stated in the abstract.
The approach helps disentangle the overlapping contributions of amyloid pathology and normal aging to functional brain network alterations.
What This Means
This research used a specialized artificial intelligence technique called contrastive deep learning to analyze brain activity patterns from 289 older adults, some of whom had signs of Alzheimer's disease pathology (amyloid plaques) detected through spinal fluid testing. The AI was designed specifically to find brain network patterns that are unique to people with amyloid buildup, while filtering out patterns that are common to all aging brains. The analysis found that brain network changes specific to amyloid buildup were concentrated in particular regions of the brain — the right superior temporal cortex and the anterior cingulate cortex — areas known to be involved in attention and memory. Separately, brain network changes that occurred in both amyloid-positive and amyloid-negative individuals were linked to language regions and general aging processes.
The researchers then examined which genes were most active in the brain regions showing amyloid-specific network changes. They found that these regions showed molecular signatures related to problems at synapses (the connections between brain cells) and abnormal activity in glial cells (the brain's support cells). This connects large-scale brain network changes visible on MRI to the cellular-level changes known to occur in Alzheimer's disease. The AI framework was also able to use these brain network patterns to predict the actual levels of amyloid and tau proteins measured in individuals' spinal fluid.
This research suggests that it is possible to use brain imaging combined with machine learning to separate the brain effects of Alzheimer's disease pathology from the normal effects of aging — a major challenge in the field. By identifying which brain network changes are truly amyloid-specific versus aging-related, this approach could potentially improve early detection of Alzheimer's disease and help researchers better understand why the disease affects different people in different ways. The integration of brain imaging with genetic expression data also opens a path toward understanding the molecular mechanisms behind the functional brain changes seen in Alzheimer's disease.
Zhao K, Xie H, Jacobs T, Gaggi N, Fortea J, Carlisle N, et al.. (2026). Elucidating the neuropathological and molecular heterogeneity of amyloid beta and tau in Alzheimer's disease through machine learning and transcriptomic integration.. Alzheimer's & dementia : the journal of the Alzheimer's Association. https://doi.org/10.1002/alz.71757