A Brain Graph Foundation Model: Pre-Training and Prompt-Tuning for Any Atlas and Disorder

🟩 Clarifying Model Categorization: Time-Series vs. Connectome vs. Graph

Thank you for pointing out the ambiguity in our terminology. We agree that “connectome-based” is a more accurate term than “ROI-based,” as both categories utilize parcellated ROI-level time series. We have revised the manuscript to consistently use “connectome-based” to avoid confusion and better reflect the underlying data representation.

To address this concern, we have significantly clarified the discussion of prior brain foundation models by categorizing them based on their input modality and modeling strategy. Specifically, we have updated Supplementary Table 13 by renaming the column from Model Type to Input Type and replaced “ROI-based” with the more precise term Connectome-based. The revised categorization is as follows:

1. Time-series-based models: e.g., BrainLM, Brain-JEPA. These models take the parcellated ROI-level fMRI time series as input (typically of shape ROI × time) and model temporal dependencies using sequence-based architectures such as Transformers. While expressive, they are computationally intensive due to high input dimensionality and memory demands.

2. Connectome-based or FC-based models (formerly “ROI-based”): e.g., BrainMass, BrainNPT. These models operate on precomputed static functional connectivity (FC) matrices derived from time series (typically ROI × ROI), offering efficiency and interpretability but limited dynamic modeling.

3. Graph-based models (ours): We build a brain graph where ROIs are nodes, node features are global connectivity profiles (i.e., rows from the FC matrix), and edges define sparse topologies via functional/structural priors (e.g., top-k or k-NN). This enables localized message passing and self-supervised objectives tailored for brain graphs, capturing both global and relational structure at low cost.

---

🟦 On Inter-regional Connectivity and Graph Structure

While it is true that our node features are derived from each ROI’s Pearson correlation with all other ROIs (i.e., a row of the FC matrix), we emphasize that the explicit graph structure (edges) plays a distinct and complementary role in our model.

• The node features capture global functional profiles.
• The adjacency matrix (top-k, k-NN) encodes a structural prior, restricting message passing to sparse, biologically relevant neighbors.

Although inter-regional relationships are implicit in the node features, the graph topology:

• Encourages localized and interpretable propagation, and
• Improves generalization by enforcing biologically informed constraints.

This is fundamentally different from flat FC-matrix inputs used in MLPs or CNNs, which only capture global connectivity patterns without explicit structural modeling. In contrast, our graph-based approach combines global functional profiles (from FC-derived node features) with local relational structure (via sparse graph edges), enabling both global and local information propagation through message passing over the graph topology. This dual modeling enhances higher-order representation learning and improves generalization.

---

🟩 Changing ROI to Connectome/FC

We have revised the terminology of ROI throughout the paper to use Connectome/FC-based, which more accurately reflects the nature of the input.

---

🟥 Clarifying Benefits of Functional Connectivity as Input

We thank the reviewer for raising the neuroscience perspective. Functional connectivity (FC) matrices—representing ROI-wise correlations—are a widely adopted representation in neuroscience for capturing stable inter-regional dependencies.

Compared to raw time series, FC-based inputs: • Are computationally efficient • Offer interpretable and stable representations • Are robust to frame-level noise

However, this efficiency comes at the cost of temporal resolution—FC representations inherently lack the ability to model fine-grained spatiotemporal dynamics. In contrast, time-series-based methods can capture both local temporal patterns and global dependencies, but they are computationally expensive due to operating over full-length ROI × T sequences.

Our proposed graph-based model builds upon the efficiency of FC by treating each node’s feature as its FC profile, and introduces an explicit graph structure, which enables not only the modeling of global connectivity patterns, but also the extraction of local relational features through message passing on the graph edges using GCN. Thus, our method strikes a balance between efficiency, interpretability, and expressiveness, making it well-suited for downstream clinical tasks such as diagnosis.

---

🟪 Formatting and Citations

We have revised all inline citations to follow the standard format (e.g., Thomas et al., 2022) and ensured clear visual distinction between citation text and narrative content for improved readability.

---

🟩 Missing Related Works: Prompt Tuning and Meta-Matching

We appreciate the suggestion to include missing related works.

• We have added discussions of prior prompt tuning approaches for brain models ([3], [4]) in the Related Work section. These works provide valuable baselines and motivation for prompt-driven adaptation in neuroimaging.
• We now explicitly compare our approach with Meta-Matching ([5]) in Supplementary Section I.

While both aim to improve generalization under limited labels:

• Meta-Matching performs post-hoc alignment via a linear mapping without episodic training.
• Our method performs full episodic meta-learning, involving inner-loop adaptation and outer-loop meta-optimization.

This enables broader and stronger transfer across diverse tasks, including zero-shot generalization to unseen disorder-atlas combinations.

---

🟦 Choice of 100–300 ROI Schaefer Parcellations

The 100–300 ROI resolutions in the Schaefer atlas provide a strong trade-off between granularity and computational cost. While higher-resolution atlases (e.g., 400+) offer more detailed spatial information, they introduce substantial memory and runtime burdens, especially in large-scale pretraining. Our choice balances anatomical fidelity with scalability and aligns with prior works.

---

🟥 Age Distribution of Pretraining Cohorts

We acknowledge the concern about age distribution. While our dataset is skewed toward younger populations due to inclusion of ABCD, HBN, etc., we also include datasets like OASIS, ADNI, HCP-Aging, and HABS, which cover older adults. In future work, we plan to incorporate large-scale aging cohorts such as UK Biobank to enhance age diversity. We have added a discussion of this limitation in the revised manuscript.

---

🟪 Positional Encoding: Brain Gradient vs. RWSE

We have included Brain Gradient Positioning (BGP) in the ablation study of positional encodings. The results show that BGP and RWSE achieve comparable performance when applied to graph transformers in brain modeling, confirming that Brain Gradient PE is a strong and biologically meaningful encoding for brain networks, indicating both effectively capture brain topological priors.

While BGP performs well, our graph-based model benefits slightly more from graph-specific encodings like RWSE. Nonetheless, Brain Gradient PE remains highly promising and could be further enhanced by adapting it into a graph-aware formulation.

---

🟩 Typo Correction: Overfitting vs. Underfitting

Thank you for catching the error in line 211–213. You are absolutely right—the correct term is overfitting, not underfitting. We have corrected this in the updated manuscript.

---

🟦 Meta-Learning with Multi-Disease Data: Practicality

Our meta-learning framework is flexible:

• If only one disease is available, we can construct pseudo-tasks across sites/sessions/subjects for within-disorder meta-learning.
• If multiple diseases are available (even with few samples), we benefit from task-level transfer, significantly reducing per-disease annotation needs.

Thus, while multi-disease meta-learning enhances generalization, the method remains practical in single-disorder or few-label regimes.

---

🟥 Brain-JEPA Baseline Inclusion

Thank you for the reminder. We have added Brain-JEPA as a baseline in Table 1 (main paper) and Table 13 (supplementary). Brain-JEPA is a more advanced model than BrainLM and shows comparable performance to BrainGFM, validating the strength of high-capacity time-series pretraining.

We fully acknowledge that Brain-JEPA is a strong and representative method of the time-series-based paradigm. Its inclusion highlights the advantages of modeling fine-grained temporal dynamics directly from parcellated ROI time series.

---

🟪 Reproducibility: Code and Error Bars

• We have now uploaded the anonymized code to an anonymous repository in the revised version.
• We have updated Table 1 to include error bars (standard deviations) computed from multiple runs.

🟪 Clarification on Additional Results and PDF Limitations

We thank the reviewer for the detailed review and constructive feedback.
Due to the word limit and the restriction on uploading updated PDFs, we were unable to include all results in the original submission.
We now present them here as additional evidence for further clarification.

---

🟩 Comparison with Meta-Matching in Supplementary Section

We added a comparison with Meta-Matching.
We summarize the key differences in the table below:

---

🟦 Clarification on Brain-JEPA Results and Code Availability

We have evaluated Brain-JEPA on all 10 datasets.
Due to space limits, only results on BrainMass, Brain-JEPA, and BrainGFM are shown on 5 datasets, with standard deviations included.

Table: Performance Comparison (mean ± std) on 5 Datasets (AUC/ACC/SEN/SPE)

The code has been uploaded to an anonymous repository, and we will release it upon paper acceptance.
As per rebuttal guidelines, we are*not allowed to share links, even to anonymous repositories.

---

🟥 Clarification on Prompt Tuning Discussion

To reflect recent work, we added to Section 2.2:

"Recent studies have also introduced brain prompt-tuning methods BrainPrompt and ScaffoldPrompt to adapt pre-trained fMRI models."

In the Appendix, we clarify the distinction:

"While prior works like BrainPrompt and Scaffold Prompt adapt fMRI models using token- or scaffold-style prompts on time-series or ROI-level features, our method introduces a graph-structured prompt designed specifically for brain graphs. These prompts include learnable node and edge parameters aligned with brain graph topology, enabling better structural encoding. Unlike prior methods that tune prompts per task, we adopt a meta-learning approach to generalize across multiple disorder-atlas pairs, supporting few-shot and zero-shot adaptation through a unified and scalable prompt mechanism.”

---

🟧 Clarification on 100–300 ROI Schaefer Parcellations

For each individual disorder, a finer parcellation (i.e., more ROIs) generally results in better performance. However, the optimal parcellation scale is not universal—it varies across disorders.

---

🟩 Clarification on “ANY Atlas” Claim

To support the claim, we conducted:

1. Cross-Atlas Transfer: Train on one atlas (e.g., Schaefer100), evaluate on another (e.g., AAL, Power).
2. Cross-Parcellation Transfer: Train on one scale, evaluate on finer/coarser scales.

Table: Ablation experiments on the ADNI2. The reported improvements (%) indicate the performance gain over models without pre-training.

---

🟨 Response to Concern on "Any Atlas and Disorder" Claim

We appreciate the reviewer’s feedback and acknowledge that the phrase “for Any Atlas and Disorder” in our original title may have been too strong.
While our model is designed to adapt flexibly across a wide range of brain atlases (e.g., Schaefer100–300, AAL, Power) and neurological disorders (e.g., MDD, ASD, AD), we agree that the word "any" may overstate the current empirical coverage.

To address this, we have revised the title to use the term "broad" instead of "any", reflecting a more measured and accurate description of our model’s generalization capacity.

So the new title will be "A Brain Graph Foundation Model: Pre-Training and Prompt-Tuning across Broad Atlases and Disorders".

We hope this clarifies our intent and aligns better with the current evaluation scope.

---

🟧 Clarification on Full-Shot Performance Compared to Brain-JEPA

We appreciate the recognition of our clarification and additional results.
We agree that Brain-JEPA achieves strong performance, particularly in full-shot settings.
However, our model is not solely focused on full-shot accuracy, but rather on introducing a new foundation model paradigm for brain graphs that is:

1. More memory-efficient and computationally lighter than time-series-based models like Brain-JEPA. Our graph-based architecture enables training and inference on large-scale fMRI cohorts with significantly reduced memory usage.
2. Specifically designed to support few-shot and zero-shot generalization, which is crucial for real-world clinical settings where labeled data is scarce.
3. Aligned with the emerging trend of foundation models, our work is the first to explore graph-based foundation pre-training for brain connectomics.

Thus, the core value of our work lies not only in performance, but also in proposing a new, practical and generalizable modeling direction that balances scalability, efficiency, and adaptability.

We hope the reviewer finds this clarification helpful.

🟩 Clarity and Complexity – Integration of Multiple Contributions

We sincerely thank the reviewer for the thoughtful feedback and recognition. We understand the concern regarding the complexity and density of our manuscript. This is primarily because our work presents a comprehensive and integrated framework that unifies both pre-training and downstream fine-tuning, covering not only standard settings but also few-shot and zero-shot scenarios. Rather than addressing each component in isolation, we aim to demonstrate how they collectively contribute to building a Brain Graph Foundation Model — a novel and unified perspective that, to the best of our knowledge, is the first attempt in this direction.

Notably, prior brain foundation models have largely focused on pre-training alone, often overlooking the critical yet practically important downstream tuning strategies, especially in few-shot and zero-shot settings. However, these settings are highly relevant in real-world clinical applications, where labeled data is often scarce or entirely unavailable for new disorders or sites. Our work deliberately addresses this gap by integrating a full-stack solution encompassing pre-training, prompt tuning, meta-learning, and generalization to unseen tasks. As a result, the manuscript may appear information-dense, but we believe this comprehensiveness is essential to convey the full scope and practical utility of our approach.

The extensive experimental design and multilayered methodology reflect our commitment to presenting a complete and practical solution, rather than a partial or fragmented one. Given the scope and ambition of our work, we have carefully organized the supplementary materials to ensure that all technical and implementation details are accessible to interested readers without overwhelming the main text.

---

🟦 Q1: Data Harmonization and Demographic Balance

Thank you for raising this important concern. We would like to clarify that one of the core motivations and advantages of developing a brain foundation model is precisely to address the challenge of data heterogeneity across cohorts. By pre-training on large-scale and diverse datasets that encompass a wide range of acquisition sites, protocols, and subject demographics, BrainGFM is explicitly designed to learn representations that are robust to variations across datasets. This pretraining paradigm naturally performs data harmonization by encouraging the model to generalize across heterogeneous input distributions, rather than requiring explicit harmonization steps prior to training. This is a key distinction from traditional models and a central benefit of the foundation model framework.

In terms of class imbalance, especially for downstream tasks, we carefully constructed balanced evaluation datasets by ensuring demographic (e.g., sex) and clinical label distributions are as uniform as possible. Specifically, we followed the sampling strategy adopted by BrainMass to create downstream classification datasets with balanced class labels and controlled gender ratios. These efforts help to mitigate the potential confounding effects of demographic imbalance on model evaluation.

---

🟥 Q2: Clinical Significance of Metrics and Use Scenarios

Thank you for this valuable and clinically grounded comment. We appreciate your emphasis on the importance of understanding each disorder’s diagnostic context, especially the relevance of sensitivity versus specificity in clinical decision-making. We address these points below by referring to the presented tables and the design of our evaluation pipeline.

In our submission, we already provide detailed context for each brain disorder used in evaluation. Specifically, Table 15 outlines the target diagnosis, age range, and sample sizes for each downstream classification task. The tasks span a diverse set of brain disorders across both neurodevelopmental (e.g., ADHD, ASD) and neurodegenerative/psychiatric conditions (e.g., AD, MDD, PTSD). For each task, we constructed balanced datasets in terms of both class distribution and gender (as described in Table 15 and Table 14), following the protocols of BrainMass and other prior works. This ensures that any performance differences are not confounded by class imbalance or demographic skew.

Regarding the clinical relevance of evaluation metrics, we agree that the importance of sensitivity vs. specificity varies across tasks:

• For Alzheimer’s Disease (AD), specificity is particularly critical in clinical settings because clinicians often suspect cognitive decline, and a high-specificity model helps reduce false positives and overdiagnosis. In our results (Table 1), BrainGFM achieves 84.4% specificity and 85.1% AUC on ADNI2, significantly outperforming non-pretrained baselines (e.g., GCN 73.4% SPE, BrainLM 78.3% AUC but 81.4% SPE). This demonstrates that BrainGFM narrows the gap toward the clinical diagnostic specificity target.

• For early screening tasks such as ASD (ABIDE II) or ADHD (ADHD200), high sensitivity is clinically preferred to avoid missing early-stage cases. For ASD, BrainGFM achieves 70.4% sensitivity, which is competitive given the task complexity and variability across ABIDE sites. For ADHD, we observe 67.3% sensitivity, again higher than all baselines. These results suggest that our model aligns well with the clinical goal of early detection, where missing true positives can delay interventions.

To reflect these differing priorities, we will update Table 1 in the revision to highlight (e.g., bold) the most clinically relevant metric for each task (e.g., specificity for AD, sensitivity for ASD), and include a short summary in the supplementary text indicating the expected clinical performance range for each disorder (where available). This will help readers better interpret how close our model is to the real-world utility level.

Finally, we clarify that BrainGFM is designed to be flexible across both clinical and screening scenarios:

• For disorders like AD or MDD, we envision it as a clinical decision support tool used after initial suspicion, requiring high specificity.
• For neurodevelopmental disorders such as ASD and ADHD, we see it as a pre-clinical screening model, prioritizing sensitivity to identify high-risk individuals early.

This intended use distinction will also be explicitly stated in the revised manuscript.

---

🟪 Paper Formatting Concerns

We thank the reviewer for the careful proofreading. We have corrected all grammatical issues and have thoroughly reviewed the entire manuscript to ensure that even minor errors have been addressed. We greatly appreciate your attention to detail, which helped us further improve the quality and clarity of the paper.

---

🟪 Claims and Evidence – Analysis of GMAE and GCL Components

We thank the reviewer for highlighting the importance of analyzing the individual contributions of GMAE and GCL within our pre-training framework.

To address this, we have conducted comprehensive ablation studies on the ABIDE II dataset using two distinct brain atlases: Schaefer-100 and AAL-116. As shown in Figure 5, we compare the following four settings:

1. Without pre-training
2. With only GCL pre-training
3. With only GMAE pre-training
4. With the sequential combination of GCL and GMAE

The results indicate that GCL slightly outperforms GMAE when applied individually, highlighting its strength in capturing global graph-level representations. In contrast, GMAE focuses on reconstructing local node-level features, offering complementary benefits.

When combined, GCL+GMAE consistently achieves the best performance across all evaluation metrics (AUC, ACC, SEN, SPE), demonstrating the effectiveness of integrating both global and local inductive biases during pre-training. This validates our design choice and confirms the complementary nature of the two pre-training strategies.

We will further emphasize this ablation analysis in the revised manuscript for clarity.

In addition, we provide further theoretical and empirical analysis in the supplementary material:

• Section J provides detailed formulations and mathematical definitions for both GMAE and GCL.
• Section K discusses their application scenarios, strengths, and limitations.

We believe that these additional analyses and experiments adequately address the reviewer’s concerns and further justify our design choices.

🟩 W1: Claims and Evidence – Brain Graph Construction Method

We appreciate the reviewer’s thoughtful critique regarding the brain graph construction strategy.

We would like to clarify that our current method constructs brain graphs based on a Top-k sparsification of the Pearson correlation matrix, which is one of the most widely adopted and standardized paradigms in fMRI-based brain network modeling. Specifically, each node (ROI) is represented by its connectivity profile—computed as Pearson correlations with all other ROIs—and edges are constructed by retaining the top-k strongest connections per node. This strategy effectively captures global connectivity patterns while introducing sparsity, which is crucial for modeling the underlying brain network structure.

This approach builds upon the fundamental assumption in functional connectivity analysis that correlated brain activity implies potential functional interaction. Many prior works (e.g., [relevant citations]) have employed similar top-k strategies or fixed-threshold binarization schemes based on Pearson correlation to ensure robustness and reproducibility in downstream analyses.

While we acknowledge that alternative adjacency construction methods—such as k-NN graphs based on spatial distance, L1/L2-regularized sparse graphs, or learned adjacency matrices—have been explored in the literature, we intentionally selected the Top-k sparsification for its simplicity, interpretability, and reproducibility. This makes our proposed foundation model more generalizable and easier to benchmark.

To further address the reviewer’s concern, we conducted an ablation study comparing our Top-k correlation-based graphs with an alternative k-NN graph construction based on spatial proximity. As detailed in Appendix Section E, Table 3, the k-NN method showed slight performance improvements (~1–2%) in some cases. However, we observed that the choice of node features and pretraining strategy plays a more dominant role in final performance, while the gain from modifying the graph structure is relatively minor.

Table: Comparison of different brain graph construction methods on ABIDE II and HBN (MDD). Our approach uses top-k sparsification of the Pearson correlation matrix to construct brain graphs. We also compare it with KNN-based graph construction. The numbers in the table represent classification accuracy (ACC).

We have added a new section in the revised manuscript to present this analysis and better justify our design choice.

---

🟦 W2: Claims and Evidence – Meta-Learning Task Definition

Thank you for pointing out this important concern regarding the meta-learning formulation in Section 3.3. We appreciate the opportunity to clarify both the definition of meta-learning tasks and the role of ABIDE II in our ablation study.

First, we confirm that the core task in our framework is disease classification. However, each neurological or psychiatric disorder can be represented under different brain atlases, which define how the brain is parcellated into ROIs. These variations result in different feature spaces and connectivity structures, even for the same subject. Therefore, we model each disease+atlas combination as a distinct meta-task, not because we view them as separate tasks clinically, but because each combination represents a different data distribution of the same underlying task. This design allows the model to learn generalizable knowledge not only across different diseases but also across different atlas-specific representations of the same disease.

In this sense, atlas variation is treated as a domain shift, and meta-learning across these shifts helps the model adapt more robustly to unseen configurations. The goal of including atlas diversity is to improve disease classification performance in realistic scenarios where atlas heterogeneity is common across datasets and institutions. Importantly, the “atlas adaptation” is not a separate objective; it serves the disease classification task by encouraging the model to learn atlas-invariant but disease-discriminative features.

Second, regarding the ablation in Table 14 in the supplementary materials, we clarify that ABIDE II is entirely held out from both pre-training and meta-training. It is used only as an evaluation benchmark for comparing different pre-training corpus designs (e.g., using single vs. mixed atlases). Therefore, there is no information leakage or violation of meta-learning principles, as no ABIDE II data is used in either the inner or outer loop of meta-learning, nor in pre-training.

We will revise the manuscript to explicitly state this separation to avoid any confusion.

---

🟥 W3: Claims and Evidence – Zero-Shot Evaluation and Disease Generalization

Thank you for raising this important point regarding the zero-shot generalization setting in our current experiments.

We acknowledge that the disease types in UCLA_CNP and REST-META-MDD also appear in HBN, which is part of our pre-training corpus. However, these overlapping disease types are represented by very limited labeled samples in HBN, and they were not used in supervised training. Therefore, the current zero-shot evaluation is intended to assess the model’s ability to transfer to new datasets and sites under label-free conditions, rather than its capacity to handle entirely novel disease types. We agree that this setting demonstrates label-wise zero-shot transfer, but not strictly disease-wise zero-shot generalization.

To directly address this concern, we conducted an additional evaluation on two uncommon disorders from the HBN dataset—Speech Sound Disorder (SSD) and Adjustment Disorder (AJD)—under full-shot, few-shot, and zero-shot settings. Importantly, the entire HBN dataset was excluded from the pre-training corpus, ensuring that these disorders were never seen during pre-training, even in an unlabeled form. The results, presented in Figure 9 (Supplementary Section D), demonstrate the model’s ability to generalize to truly unseen diseases, particularly under few-shot and zero-shot scenarios.

Dear Reviewer Rfc2,

Thank you again for your detailed and thoughtful review of our paper.

We have carefully addressed each of your concerns in the rebuttal, including clarifications on the brain graph construction method (W1), the meta-learning task definition (W2), and the setup of zero-shot evaluations (W3). We would sincerely appreciate it if you could take a moment to review our responses and let us know if they help clarify the issues you raised. Your further feedback would be extremely helpful to us.

Thank you again for your time.