IC-GNN: Modeling Individualized Brain Community Organization for Disease Prediction

Yang Zhang, Xiao Zhou, Jonathan Warrell, Avram Holmes, Xuan Zhang, Mark Gerstein, for the Alzheimer's Disease Neuroimaging InitiativearXiv 2026 · arXiv

TL;DR

IC-GNN is a novel graph neural network framework that captures individualized brain community organization to improve disease prediction. By anchoring subject-specific functional communities to a common brain atlas, the model achieves state-of-the-art performance across 30,000+ participants while identifying reproducible, disease-related "fingerprints" in brain connectivity.

What Problem Does This Paper Address?

The human brain is organized into modular "communities"—groups of regions that work together to perform specific functions. Traditionally, neuroscientists use group-level "atlases" to map these communities. However, these atlases assume every brain is organized the same way, ignoring the significant heterogeneity between individuals and how these structures shift during disease.

While researchers can create individualized maps at the voxel level, these require massive amounts of fMRI data (long scan times) and are notoriously difficult to compare across different people because the "boundaries" of regions don't line up. The authors of IC-GNN aim to solve this trade-off: How can we model individualized brain organization in a way that is data-efficient, comparable across subjects, and predictive of disease?

Figure 1: An example of individual differences in functional community organi- zation. The same outlined region of interest (RH_TempPar_5 in the Schaefer atlas, located in the right posterior superior temporal sulcus) belongs to com- munities comprising different sets of regions in Subject A (blue) and Subject B (purple).

Methods

The authors propose IC-GNN (Individualized Community-guided Graph Neural Network). The core innovation is that it uses a standard brain atlas as an "anchor" but allows the functional grouping of those regions to vary for every subject.

The architecture consists of three main novel components:

  1. AP-HPM (Affinity Propagation-based Hierarchical Pooling Module): Unlike standard clustering that requires a pre-defined number of groups, this module uses Affinity Propagation to discover a subject-specific hierarchy of communities directly from their functional connectivity data.
  2. PERM (Probabilistic Edge Refinement Module): fMRI data is notoriously noisy. PERM acts as a learnable filter that prunes "weak" or noisy connections. Crucially, it uses the individualized community information to decide which edges to keep, making the denoising process biologically informed.
  3. Community-Conditioned Readout: To make a prediction (e.g., "Does this person have Alzheimer's?"), the model needs to aggregate information from across the brain. IC-GNN uses a readout mechanism that modulates shared regional weights based on an individual's unique community context. This allows the model to learn that a specific region might be highly predictive only when it belongs to a certain functional neighborhood.
Figure 2: The overall framework of IC-GNN.

Datasets

The model was evaluated on three large-scale public datasets, totaling over 30,000 participants:

  • ABIDE: Focused on Autism Spectrum Disorder (ASD).
  • ADNI: Focused on Alzheimer’s Disease (AD), including data from patients with Mild Cognitive Impairment (MCI).
  • UK Biobank: A massive cohort used here to predict Depression.

All scans were parcellated into a common 419-region space (Schaefer atlas + subcortical regions) to ensure cross-study consistency.

Key Insights and Ideas

  • Brain Fingerprinting: The individualized communities discovered by the model are highly reproducible. In the ABIDE dataset, the model could identify a participant's "fingerprint" across different halves of a single scan with 97.2% accuracy.
  • The "Blurring" of Networks: The model revealed that in diseases like Alzheimer's and Autism, the boundaries between the Control Network (involved in executive function) and the Default Mode Network (involved in internal thought) become blurred. This "network mixing" was a strong predictor of disease state.
  • Hierarchy is Essential: Ablation studies showed that removing the hierarchical structure (moving from ROIs to communities) caused a significant drop in performance. The brain's multiscale nature—how small regions form small groups, which then form larger systems—is critical for capturing disease signals.
  • Beyond Covariates: The authors proved that the diagnostic information captured by these individualized communities goes far beyond what can be explained by simple demographics like age, sex, or head motion.

Key Results

IC-GNN consistently outperformed 15 baseline methods, including standard machine learning (SVM, Random Forest), general-purpose GNNs (GCN, GAT), and specialized brain-network models (BrainGNN, BrainNetTF).

  • ABIDE (Autism): Achieved 75.41% AUROC, significantly higher than the next best brain-specific model (73.84%).
  • ADNI (Alzheimer's): Achieved 86.46% AUROC.
  • UK Biobank (Depression): Maintained leading performance even at a massive scale (60.62% AUROC).
Table 1: Classification performance on ABIDE (ASD vs. controls), ADNI (AD vs. CN), and UK Biobank (depression vs. controls). Values are percentages (mean ± standard deviation) over the five folds. The best result in each column is in bold and the second best is underlined.

Limitations and Future Work

While the atlas-anchored approach is efficient, the authors acknowledge that it still relies on a predefined set of regions. Future iterations might allow for even finer-grained boundary shifts within those regions. Additionally, while the model shows high performance in cross-sectional prediction, the authors suggest that applying IC-GNN to longitudinal data—tracking how a single person's brain communities reorganize over years—could provide even more powerful biomarkers for disease progression.

Summary updated

Abstract

Brain networks exhibit modular organization, where community structures vary substantially across individuals and diseases. Understanding these individual differences is crucial for advancing both basic neuroscience research and precision medicine. However, widely used predefined group-level brain atlases often overlook this heterogeneity. While individualized functional parcellations capture subject-specific variability in community boundaries, they often require lengthy scan times to achieve fine-grained resolution and complicate cross-subject comparisons. To address this trade-off and investigate how individualized community organization captures disease-associated changes across individuals, we propose IC-GNN, an atlas-anchored model that forms distinct, individualized functional communities directly over atlas-defined brain regions. IC-GNN derives discrete, individualized multiscale communities from resting-state functional magnetic resonance imaging (fMRI) data, refines community-guided connections, and adapts shared regional weights to each subject's community context to enable both individualized modeling and group-level comparisons. Across more than 30,000 participants from ABIDE (autism spectrum disorder), ADNI (Alzheimer's disease), and the UK Biobank (depression), IC-GNN achieves the highest performance compared with 15 baseline methods. We further demonstrate that these individualized communities are reproducible within scans and across sessions, and carry predictive power on their own. Furthermore, IC-GNN identifies disease-related shifts in community organization that align with established literature, highlighting a promising approach for modeling individualized community structure for disease prediction and biomarker discovery.
BibTeX
@misc{zhang2026ic,
  author = {Yang Zhang and Xiao Zhou and Jonathan Warrell and Avram Holmes and Xuan Zhang and Mark Gerstein and for the Alzheimer's Disease Neuroimaging Initiative},
  title = {{IC-GNN: Modeling Individualized Brain Community Organization for Disease Prediction}},
  year = {2026},
  eprint = {2611.13070},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.13070},
}
Figure