Geometric brain signatures of Alzheimer's disease progression and subtypes

This study introduces a novel framework that utilizes geometric brain signatures derived from multiple neuroimaging modalities to accurately identify distinct Alzheimer's disease subtypes and progression trajectories, outperforming conventional localized features in stability and biological relevance.

The Gist

The Big Picture: Finding the "Shape" of Alzheimer's

Imagine Alzheimer's disease not just as a list of symptoms, but as a song playing on a radio. For a long time, doctors have tried to understand this song by looking at individual notes (specific brain spots) or by asking the listener how they feel (cognitive tests). But this study suggests that to truly understand the song, we need to look at the entire melody and how the sound waves travel across the whole instrument.

The researchers developed a new way to "listen" to the brain using three different types of brain scans (PET and MRI). Instead of just looking at small, isolated spots, they analyzed the geometric shape of the brain's activity. They found that this approach reveals hidden patterns about how the disease progresses and identifies different "versions" or subtypes of the disease that look the same on the surface but are actually very different underneath.

The Core Concept: The Brain as a Drum

To understand the math behind the study, imagine the brain's surface as a giant, complex drum.

1. The Old Way (Local Spots): Traditional methods look at the drum and say, "There is a dent here, and a scratch there." They treat every spot on the drum as an independent problem.
2. The New Way (Eigenmodes): This study treats the brain like a musical instrument. Just as a drum has specific "vibrational modes" (ways it naturally vibrates when hit), the brain has geometric eigenmodes. These are the fundamental "shapes" or patterns that the brain's surface can naturally take.
   • Low-frequency modes are like the deep, slow rumble of the whole drum (large-scale changes across the whole brain).
   • High-frequency modes are like the rapid, tiny ripples on the surface (small, local changes).

The researchers took the brain scans (showing amyloid plaques, sugar metabolism, and brain thickness) and broke them down into these "vibrational modes." They didn't just look at the dents; they looked at the shape of the vibration.

How They Did It: The Three-Step Recipe

The study used data from two large groups of people (ADNI and OASIS-3) who had various stages of memory issues, from healthy to mild confusion to full Alzheimer's.

1. The Ingredients (The Scans): They used three types of brain scans:
   • AV45 PET: A scan that lights up where "amyloid" (a sticky protein associated with Alzheimer's) is stuck.
   • FDG PET: A scan that shows where the brain is using energy (metabolism).
   • MRI: A scan that measures how thick the brain's outer layer is.
2. The Mixing (The Math): They used a special algorithm (called mcTI) to mix these three scans together. Think of this as blending three different colored paints into a single, rich color that captures the full picture of the disease.
3. The Result (Pseudotime and Subtypes):
   • Pseudotime: The algorithm assigned every person a "progress score" from 0 to 1.
     • 0 is like standing at the start of a road (Healthy).
     • 1 is at the end of the road (Advanced Alzheimer's).
     • This score creates a smooth, continuous timeline of the disease, rather than just jumping between "Healthy," "Mild," and "Severe."
   • Subtypes: The algorithm noticed that people didn't all travel down the road in the exact same way. Some took a "scenic route," others a "highway." These different paths are the subtypes.

What They Found

1. The "Fan" Shape of Disease
When they plotted everyone on a map based on their "progress score," they saw a fan shape.

• Early on (Low scores): People were scattered all over the place. This means that in the early stages, everyone's brain changes in very different, unique ways.
• Later on (High scores): As the disease gets worse, everyone's brain changes start to look more similar. The "fan" closes up. It seems that no matter how different the start was, the end stage of the disease converges into a very specific, stereotyped pattern.

2. Different Roads, Different Genes
The study found that the different "subtypes" (the different paths down the road) were supported by real biology:

• Genetics: People on different paths had different genetic markers (specifically in the APOE gene, a known risk factor for Alzheimer's).
• Biology: One subtype seemed to be driven mostly by protein buildup but didn't show as much loss of brain function initially. Another subtype showed a different pattern of brain shrinkage.
• Stability: The "geometric" method (looking at the vibration modes) was much better at finding these distinct groups than looking at just specific brain spots. It was like using a high-definition camera instead of a blurry one.

3. Why the New Method is Better
The researchers compared their "vibration mode" method against the old "spot-check" method.

• The old method was like trying to understand a storm by looking at individual raindrops.
• The new method was like looking at the shape of the storm clouds.
• The new method was more accurate at predicting who was sick, how sick they were, and which "subtype" of the disease they had. It was also more consistent across different groups of people.

The Bottom Line

This paper doesn't claim to have a new cure or a new drug. Instead, it offers a new map.

By treating the brain's changes as geometric patterns (like musical notes or drum vibrations) rather than just isolated spots, the researchers created a more accurate way to track how Alzheimer's moves through the body. They showed that the disease isn't just one straight line; it has different "lanes" (subtypes) that can be identified early on. This helps explain why some patients respond differently to treatments and why the disease looks different in different people, providing a clearer picture of the journey from health to disease.

Technical Summary

Technical Summary: Geometric Brain Signatures of Alzheimer's Disease Progression and Subtypes

Problem Statement
Alzheimer's disease (AD) presents significant challenges in clinical management due to diagnostic delays, disease heterogeneity, and variable treatment responses. Existing biomarker identification methods often rely on single-modality data or localized features (e.g., specific regions of interest or voxels), which may fail to capture the complex, multiscale nature of neurodegeneration. Furthermore, current progression models often lack the ability to simultaneously infer continuous disease trajectories and identify distinct biological subtypes that reflect heterogeneous pathological mechanisms. The authors propose that conventional localized approaches miss fundamental constraints shaping brain changes and that a geometric perspective, grounded in the brain's physical structure, offers a more robust framework for understanding AD progression.

Methodology
The study introduces a novel framework integrating geometric brain signatures with the multi-omics contrastive trajectory inference (mcTI) algorithm. The methodology proceeds through three primary stages:

1. Geometric Feature Extraction:

   • Data Sources: The framework utilizes multimodal neuroimaging data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Open Access Series of Imaging Studies 3 (OASIS-3), including [18F]-Florbetapir (AV45) PET (amyloid-beta), [18F]-Fludeoxyglucose (FDG) PET (metabolism), and structural MRI (cortical thickness).
   • Eigenmode Decomposition: Instead of using localized region-of-interest (ROI) or vertex-based features, the authors decompose surface-based brain maps into geometric eigenmodes. These modes are derived by solving the eigenvalue problem for the Laplace-Beltrami operator on the cortical mesh (FsAverage template).
   • Spectral Representation: Each brain map is represented as a linear combination of these eigenmodes (standing waves of the cortical geometry). The coefficients ($\beta_j$) of the first 30 modes per hemisphere serve as the "geometric signatures." This approach captures spatial patterns across different scales, from low-frequency global patterns to high-frequency local details.
   • Comparison Baselines: For validation, the study compares these mode-based features against conventional region-based features (68 Desikan-Killiany atlas regions) and vertex-based features reduced via Principal Component Analysis (PCA).

2. Trajectory and Subtype Inference (mcTI):

   • Network Fusion: Features from all three modalities are integrated using Similarity Network Fusion (SNF) to create a patient-dominant fused network.
   • Pseudotime Calculation: A disease progression score ("pseudotime") is computed for each participant as the shortest-path distance from their position in the fused network to the centroid of the cognitively normal (CN) group. Pseudotime is normalized to [0, 1], where 0 represents CN and 1 represents late-stage AD.
   • Subtype Identification: Using an Expectation-Maximization (EM) procedure within the multidimensional scaling (MDS) embedding of the fused network, the algorithm assigns participants to subtypes. The optimal number of subtypes is determined by minimizing the Bayesian Information Criterion (BIC). Stability is assessed via bootstrapping and permutation testing.

3. Validation:

   • The inferred pseudotime and subtypes are validated against clinical diagnoses, cerebrospinal fluid (CSF) biomarkers (A$\beta$, TAU, pTAU), imaging summary scores (SUVR), and cognitive scales (MMSE, MOCA, CDRSB, etc.).
   • Genetic associations (SNPs) are analyzed to determine if subtypes correspond to distinct genetic architectures.

Key Results

• Superior Performance of Geometric Features: Mode-based features significantly outperformed both region-based and vertex-based features. They demonstrated the strongest correlation with clinical diagnosis (Spearman $\rho = 0.255$) and the highest ability to distinguish between CN, MCI, and AD groups (ANOVA $p < 10^{-13}$).
• Biological Validity of Pseudotime: The inferred pseudotime showed monotonic increases across diagnostic groups and strong correlations with established biomarkers. Notably, it correlated significantly with CSF TAU and pTAU despite these not being explicit inputs, suggesting the framework captures downstream pathological cascades.
• Discovery of Distinct Subtypes: The analysis identified four distinct disease subtypes among impaired subjects (plus a CN reference group).
  • Subtype 2 was characterized as a terminal state with high pseudotime, showing a dissociation between amyloid pathology and metabolic/cognitive decline (a "symptomatic floor" effect).
  • Subtypes 1, 3, and 4 exhibited coherent progression patterns with high internal homogeneity ($R^2 > 0.79$).
  • Subtypes displayed unique genetic signatures (e.g., specific SNPs in APOE and APOC1 regions) and distinct biomarker trajectories, confirming they represent biologically distinct pathways rather than arbitrary clusters.
• Multimodal Synergy: Subtype identification relied on a balanced contribution from all three modalities (FDG PET: 35.5%, sMRI: 33.6%, AV45 PET: 30.9%), indicating that no single marker dominates the classification.
• Robustness: The findings were replicated in the independent OASIS-3 cohort, demonstrating the generalizability of the geometric signatures.

Significance and Claims
The paper claims that geometric brain signatures, derived from the fundamental resonant modes of cortical geometry, provide a more holistic and robust representation of AD progression than conventional localized approaches. By shifting the focus from "focal lesions" to "system-wide degradation," the framework captures the intrinsic dynamic responses of the brain constrained by its geometry.

The authors assert that this approach:

1. Unlocks New Markers: It uncovers disease progression trajectories and subtypes that are missed by standard localization methods.
2. Enhances Stability: The identified subtypes and pseudotime estimates are highly stable across datasets and permutation tests.
3. Integrates Biology and Structure: The method successfully links imaging phenotypes with underlying genetic architectures and biological mechanisms, offering a pathway toward more precise diagnostic and prognostic tools for AD.
4. Addresses Heterogeneity: It provides a data-driven mechanism to parse the heterogeneity of AD, potentially explaining why certain treatments fail in specific patient subgroups.

The study concludes that integrating geometric signatures with trajectory inference algorithms offers a promising foundation for understanding disease heterogeneity and developing precision medicine tools for Alzheimer's disease.