IC-GNN: Modeling Individualized Brain Community Organization for Disease Prediction
Authors: Yang Zhang, Xiao Zhou, Jonathan Warrell, Avram Holmes, Xuan Zhang, Mark Gerstein, for the Alzheimer's Disease Neuroimaging Initiative
Organizations: Program in Computational Biology and Bioinformatics, Yale University, New Haven, CT, 06520, USA · Department of Psychiatry, Robert Wood Johnson Medical School, and Brain Health Institute, Rutgers University, Piscataway, NJ, 08854, USA · School of Design, Pratt Institute, New York, NY, 11205, USA · Department of Molecular Biophysics and Biochemistry, Yale University, New Haven, CT, 06520, USA · Department of Computer Science, Yale University, New Haven, CT, 06520, USA
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.
Figures & tables
Figure 1: An example of individual differences in functional community organization. The same outlined region of interest (RH_TempPar_5 in the Schaefer atlas, located in the right posterior superior temporal sulcus) belongs to communities comprising different sets of regions in Subject A (blue) and Subject B (purple).
Figure 2: The overall framework of IC-GNN.
Category
Model
ABIDE
ADNI
UK Biobank
ACC
AUROC
ACC
AUROC
ACC
AUROC
ML methods
SVM
63.49±2.42
71.42±5.34
71.22±4.38
82.54±2.21
61.70±0.53
59.06±0.28
Random Forest
61.73±3.14
65.85±4.96
65.64±6.96
70.25±4.40
57.01±1.11
51.69±0.28
General GNNs
GCN ( Kipf and Welling, 2017 )
58.60±3.76
61.46±5.61
69.52±8.51
74.89±2.90
60.35±0.40
57.04±0.46
GAT ( Veličković et al., 2018 )
58.52±3.43
62.97±5.98
73.02±7.76
77.59±3.33
60.83±0.20
57.12±0.11
GIN ( Xu et al., 2019 )
52.34±5.98
56.54±4.34
66.36±5.17
71.98±2.99
59.99±0.54
55.52±0.62
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.
Comparison
Hierarchy
Community
Edge
AUROC
ACC
Full model
Yes
Individual
Learned
75.41±5.73
69.40±4.23
Hierarchy removal
No
–
Learned
71.68±6.29
65.32±5.89
Community control
Yes
Random
Learned
73.79±3.67
67.49±2.91
Yes
Consensus
Learned
74.65±4.49
68.17±2.58
Edge control
Yes
Individual
Open
74.56±4.90
68.17±3.54
Yes
Individual
Uniform
75.16±4.23
67.14±2.85
Table 2: Ablation results of IC-GNN on ABIDE. Values are percentages (mean ± standard deviation) over the five test folds. The best results are in bold.
Figure 3: Differences in community organization between AD and CN participants. For each ROI, we compare how often it shares a community with ROIs of a partner network between AD and CN participants: control with default (top), somatomotor with somatomotor (middle), and salience B with default B (bottom). Colors show the t statistic of the AD vs. CN difference from a regression analysis adjusted for covariates. Red indicates that the ROI is grouped with its partner network more often in AD, and blue means less often. The outlines mark these networks.
Figure 4: Community organization across CN, MCI, and AD. The maps show the Yeo-7 networks and the group-consensus communities in ADNI, each colored by its majority Yeo-7 network. The right panel shows the alignment with Yeo-17 relative to CN, defined as the average fraction of an ROI’s community members that belong to the same Yeo-17 network as the ROI, with 95% confidence intervals.
Figure 5: Regional evidence differences between patients and controls. Brain maps show each ROI’s contribution to the difference in model logits (without the intercept) between patients and controls, expressed as a percentage of the total difference. Positive values indicate higher regional evidence in patients, while negative values indicate higher evidence in controls. Results are shown for ASD versus typically developing controls (ABIDE) and AD versus cognitively normal controls (ADNI). Two coronal slices show subcortical regions. Black outlines mark significant regional differences in ASD. The bars sum the contributions of the ROIs in each functional network, and the ticks indicate each network’s proportion of the 419 ROIs. Results are based on held-out test data and averaged across training seeds.
Model
AUROC
Accuracy
BAcc
Recall
Specificity
Precision
F1
Conventional ML
SVM
71.42±5.34
63.49±2.42
63.06±3.65
62.30±18.28
63.82±12.64
60.50±5.24
60.00±9.93
Random Forest
65.85±4.96
61.73±3.14
60.56±3.41
55.88±19.61
65.24±13.23
58.80±3.38
55.76±11.65
General-purpose GNNs
GCN
61.46±5.61
58.60±3.76
57.64±5.18
52.68±22.55
62.60±13.86
55.10±8.77
51.15±15.30
GAT
62.97±5.98
58.52±3.43
57.75±4.23
59.99±17.04
55.50±14.02
55.24±8.70
55.63±10.77
Table S1: Benchmark results on ABIDE. Values are percentages (mean ± SD over the five test folds).
Model
AUROC
Accuracy
BAcc
Recall
Specificity
Precision
F1
Conventional ML
SVM
82.54±2.21
71.22±4.38
75.50±2.73
82.14±6.47
68.87±6.16
36.72±3.48
50.51±2.59
Random Forest
70.25±4.40
65.64±6.96
65.40±2.53
65.02±16.10
65.79±12.03
29.82±2.73
39.79±1.89
General-purpose GNNs
GCN
74.89±2.90
69.52±8.51
66.82±2.66
62.56±12.58
71.08±12.72
34.01±7.06
42.63±4.48
GAT
77.59±3.33
73.02±7.76
68.87±5.50
62.39±18.89
75.35±12.66
37.12±4.81
44.53±5.80
Table S2: Benchmark results on ADNI. Values are percentages (mean ± SD over the five test folds).
Model
AUROC
Accuracy
BAcc
Recall
Specificity
Precision
F1
Conventional ML
SVM
59.06±0.28
61.70±0.53
55.98±0.23
46.37±0.57
65.58±0.77
25.43±0.27
32.84±0.22
Random Forest
51.69±0.28
57.01±1.11
50.85±0.31
40.51±1.23
61.19±1.69
20.90±0.27
27.56±0.21
General-purpose GNNs
GCN
57.04±0.46
60.35±0.40
54.26±0.32
44.04±1.06
64.48±0.69
23.88±0.25
30.96±0.43
GAT
57.12±0.11
60.83±0.20
54.97±0.51
45.13±1.35
64.81±0.43
24.49±0.39
31.75±0.65
Table S3: Benchmark results on UK Biobank. Values are percentages (mean ± SD over the five folds).
Program in Computational Biology and Bioinformatics, Yale University, New Haven, CT 06520 USA · Department of Psychiatry, Robert Wood Johnson Medical School, and the Brain Health Institute, Rutgers University, Piscataway, NJ 08854 USA · School of Design, Pratt Institute, New York, NY 11205 USA
Department of Biomedical Engineering, Michigan State University, USA. · Department of Electrical and Computer Engineering, Michigan State University, USA.