Cross-scale spatially-aware generative modeling of transcriptomic programs underlying neurodegenerative brain organization

Technical Summary: Cross-Scale Spatially-Aware Generative Modeling of Transcriptomic Programs Underlying Neurodegenerative Brain Organization

Problem Statement
Neurodegenerative disorders, particularly Alzheimer's disease (AD), exhibit highly organized, non-random patterns of regional brain vulnerability. While previous imaging-transcriptomic studies have established correlations between gene expression and neuroimaging phenotypes (e.g., cortical thinning), these approaches are predominantly correlative and linear. They often fail to model the latent biological processes or nonlinear cross-scale interactions that explain how microscale transcriptomic organization gives rise to macroscale neurodegenerative structures. Furthermore, existing methods frequently treat brain regions as independent samples, neglecting the spatial smoothness and topological relationships inherent in cortical organization. There is a need for a generative framework that can bridge microscale molecular architecture with macroscale disease vulnerability while respecting biologically plausible spatial constraints.

Methodology
The authors propose a cross-scale spatially-aware generative framework that integrates transcriptomic representation learning with spatially constrained prediction. The pipeline consists of the following components:

1. Data Integration and Preprocessing:

   • Transcriptomics: Regional gene expression profiles were derived from the Allen Human Brain Atlas (AHBA). Data from six neurotypical donors were mapped to the Desikan-Killiany cortical atlas (68 regions). To reduce dimensionality while retaining biological signal, the authors filtered the transcriptome to 910 landmark genes from the LINCS L1000 platform.
   • Neurodegeneration Maps: Cortical thickness measurements were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort (926 cognitively normal controls and 426 AD subjects). Regional neurodegeneration vulnerability ($Y$) was quantified as the continuous difference in mean cortical thickness between AD and control groups ($Y_r = \mu_{NC,r} - \mu_{AD,r}$).
   • Alignment: Both datasets were harmonized to the same 68 cortical regions.

2. Generative Architecture (Variational Autoencoder with Spatial Regularization):
   The model employs a Variational Autoencoder (VAE) architecture modified for neurodegeneration prediction:

   • Encoder: A neural network maps the high-dimensional transcriptomic vector ($x \in \mathbb{R}^{910}$) to a low-dimensional latent distribution ($z \in \mathbb{R}^{64}$), parameterized by mean and variance. These latent variables represent compressed "transcriptomic programs."
   • Decoder: Reconstructs the original transcriptomic profile from the latent embedding to ensure the latent space preserves biologically meaningful molecular structure.
   • Vulnerability Mapper: A separate network projects the latent embeddings ($z$) to predict regional neurodegeneration vulnerability ($\hat{y}$).
   • Spatial Regularization: To enforce biologically plausible spatial organization, a graph-based smoothness constraint is applied. An adjacency matrix representing neighboring cortical regions penalizes large differences in predicted vulnerability between adjacent regions ($L_{spatial} = \sum A_{ij}(\hat{y}_i - \hat{y}_j)^2$).

3. Objective Function:
   The model is optimized using a composite loss function:
   $$L = L_{recon} + \beta L_{KL} + \lambda_1 L_{prediction} + \lambda_2 L_{spatial}$$
   Where $L_{recon}$ is the transcriptomic reconstruction error (MSE), $L_{KL}$ is the Kullback-Leibler divergence for latent regularization, $L_{prediction}$ is the error in predicting cortical thinning, and $L_{spatial}$ enforces spatial smoothness.

Key Contributions

• Generative Modeling of Cross-Scale Biology: Unlike traditional correlation-based imaging transcriptomics, this framework learns latent biological programs that simultaneously reconstruct molecular organization and predict disease vulnerability. This moves beyond descriptive associations to generative representations of how transcriptomic architecture drives degeneration.
• Spatially-Aware Learning: The integration of graph-based spatial regularization ensures that predictions respect the anatomical continuity of the cortex, addressing the limitation of methods that treat brain regions as independent samples.
• Unified Framework: The study provides a computational pipeline that explicitly links microscale gene expression (AHBA) with macroscale structural degeneration (ADNI) through a unified, biologically constrained latent space.

Results

• Predictive Performance: The framework achieved strong predictive accuracy for regional neurodegeneration vulnerability across 68 cortical regions, yielding an $R^2$ score of 0.8604 and a Pearson correlation of $r = 0.9439$ ($p < 0.001$) between predicted and observed vulnerability profiles. The Mean Squared Error (MSE) was 0.001493.
• Spatial Organization: The model successfully recapitulated known AD pathology patterns, identifying high vulnerability in the entorhinal cortex, temporal pole, and limbic systems, while sparing primary sensory regions (e.g., pericalcarine cortex).
• Latent Structure Analysis: Dimensionality reduction (PCA and t-SNE) of the learned latent embeddings revealed structured, non-random trajectories and neighborhoods. These latent representations captured distributed cortical gradients consistent with hierarchical cortical organization (from unimodal to transmodal systems) without explicit supervision regarding anatomical hierarchy.
• Comparison with Baselines: While linear regression and ridge regression achieved comparable $R^2$ values (approx. 0.89), the authors note that the proposed generative approach offers superior biological interpretability by learning compressed latent structures rather than direct feature regression. Standard MLPs without regularization performed poorly ($R^2 = -2.59$), highlighting the necessity of the proposed regularization and generative constraints.

Significance and Claims
The paper claims that this work provides a foundation for spatially-aware generative neurobiology. By demonstrating that biologically constrained generative modeling can bridge microscale transcriptomic organization with macroscale neurodegenerative structure, the study suggests that latent transcriptomic programs are a key determinant of selective cortical vulnerability.

The authors position this framework as a shift from "association studies" to "biologically informed generative modeling." They argue that the learned latent representations offer a biologically interpretable intermediate scale that explains how molecular architecture contributes to systems-level degeneration. The study concludes that such approaches can help close the gap between molecular neuroscience and systems-level neurodegenerative disease modeling, offering a scalable framework for future investigations into cross-scale mechanisms.

Limitations Acknowledged by the Authors
The authors modestly characterize this work as a preliminary proof-of-concept. They acknowledge several limitations:

• Data Source: The AHBA relies on only six postmortem donors, limiting the capture of inter-individual variability.
• Static Nature: Transcriptomic data is static (postmortem) and does not capture the dynamic temporal progression of AD.
• Region-Level Focus: The model predicts group-level spatial susceptibility rather than individual subject trajectories or longitudinal disease staging.
• Simplified Spatial Constraints: The spatial regularization uses a basic neighborhood graph and does not fully incorporate complex structural connectivity or trans-synaptic spread mechanisms.