acceptodds
Under review as a conference paper at ICLR 2027

One Shared Graph, Two Mappings: Bidirectional Graph Autoencoder for Brain Structural and Functional Connectivity

Abstract

Structural connectivity (SC) derived from diffusion MRI and functional connectivity (FC) derived from functional MRI provide complementary summaries of brain organization across large-scale networks. Because coherent brain function is supported by anatomically grounded, polysynaptic communication pathways, we model SC and FC as arising from a shared, biologically meaningful scaffold and propose a bidirectional graph autoencoder to estimate a shared latent graph common to both modalities. The method fits jointly trained forward and reverse mappings (SCFC and FCSC) that share a single latent adjacency matrix, which serves as the propagation operator for message passing in both modalities and links the two directions through a common graph over brain regions. On the landmark Human Connectome Project (HCP) and Adolescent Brain Cognitive Development (ABCD) data, the proposed approach improves both bidirectional reconstruction accuracy and graph reliability over existing graph-based and connectome translation methods, yielding roughly 7%-15% lower reconstruction error and about 4%-16% higher top-edge overlap. Moreover, the learned graphs exhibit strong cross-cohort consistency between HCP and ABCD, supporting the utility of the framework for interpretable brain structure-function modeling.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Deep learning, Neuroimaging, Graph Neural Network, Autoencoder

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_mKSCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors present a bidirectional graph autoencoder to estimate a shared latent graph common to both SC and FC modalities. The method uses jointly trained forward and reverse mappings that share a single latent adjacency matrix, which serves as the propagation operator for message passing in both modalities and links the two directions through a common graph over brain regions. The approach is evaluated on the Human Connectome Project (HCP) and Adolescent Brain Cognitive Development (ABCD) datasets. The authors claim that the proposed approach improves both bidirectional reconstruction accuracy and graph reliability over existing graph-based and connectome translation methods, achieving approximately 7%–15% lower reconstruction error and 4%–16% higher top-edge overlap.

Critical strengths

The work is validated in two large cohorts. The manuscript is well written.

Critical weaknesses

Benchmarking against other bidirectional methods: These methods are not included in the literature review, and the authors do not compare their proposed approach against them. All the benchmarking is conducted against non-bidirectional graph-based methods for functional and structural connectivity. This raises concerns about potential plagiarism and significantly reduces the novelty of the idea and its contribution.

Relevant references:

https://ieeexplore.ieee.org/document/11482816 https://www.sciencedirect.com/science/article/abs/pii/S1746809425004112 https://www.sciencedirect.com/science/article/abs/pii/S0957417426005245

To that end, this does not represent a significant contribution to the field.

Questions for authors

No questions.

Reviewer_EJCrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BGAA, a bidirectional graph autoencoder for modeling the relationship between SC and FC in brain networks. The framework learns a shared latent adjacency matrix that serves as a common message-passing operator for both SC-to-FC and FC-to-SC mappings. The shared graph is constructed using anatomical features and learnable edge parameters, while bidirectional reconstruction and graph regularization are jointly optimized. Experiments on the HCP and ABCD datasets evaluate reconstruction accuracy, graph stability, and walk-algebra fit, accompanied by ablation studies and biological analyses. The results demonstrate competitive bidirectional reconstruction performance and suggest that the learned shared graph captures certain consistent structural patterns across cohorts.

Critical strengths
  1. The paper investigates a meaningful research problem by exploring a shared graph representation to jointly model structural and functional brain connectivity. The proposed bidirectional framework provides a reasonable perspective on cross-modal structure–function relationships.

  2. The proposed method integrates shared graph learning with bidirectional reconstruction and is evaluated on two large-scale neuroimaging cohorts. The experiments consider multiple aspects, including reconstruction accuracy, graph stability, and structural consistency, providing relatively comprehensive empirical evidence for the effectiveness of the proposed framework.

  3. The overall structure of the paper is well organized.

Critical weaknesses
  1. Could the authors clarify whether the walk-algebra assumption introduced in Definition 3.1 approximately holds for real SC and FC matrices? How does the proposed reconstruction objective ensure consistency between the actual learning process and this mathematical definition?

  2. What guarantees that the learned shared adjacency matrix represents an identifiable brain communication scaffold, rather than merely one of several possible parameterizations that achieve similar reconstruction accuracy?

  3. What are the fundamental differences between the shared adjacency mechanism and conventional parameter-sharing multi-task learning approaches? Could the authors conduct capacity-controlled comparisons to further validate the independent contribution of the shared graph structure?

  4. Given that the multiple bilinear components in Eq. 11 can be combined into a single matrix, could the authors further explain the practical advantages and necessity of employing multiple bilinear components?

  5. Could the authors clarify whether the permutation analysis is intended to evaluate differences between cognitive performance groups or to examine functional-system enrichment? Furthermore, the authors should explain how the chosen null hypothesis supports the corresponding biological conclusions.

  6. The related work section relies heavily on earlier studies, with much of its coverage concentrated on research published before 2024, and provides limited discussion of recent advances in SC–FC modeling. In particular, the authors do not adequately discuss relevant studies published in 2026.

  7. Could the authors provide a schematic illustration to more clearly and formally convey the research motivation?

Questions for authors

See Weaknesses

Reviewer_aKc4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a bidirectional graph autoencoder for modeling structural and functional brain connectivity. It learns a shared population-level adjacency for SC→FC and FC→SC prediction and evaluates reconstruction, graph stability, and biological organization on the HCP and ABCD datasets.

Critical strengths
  1. The paper addresses an important problem by jointly modeling structural and functional brain connectivity through a shared graph.

  2. The bidirectional architecture provides an intuitive way to encourage the learned adjacency to support both cross-modal mappings.

Critical weaknesses
  1. Better reconstruction with a shared adjacency does not establish that the learned graph uniquely recovers a biological communication scaffold, especially when flexible encoders, decoders, and anatomical priors can also influence the result.

  2. The shared graph is defined through a low-degree walk algebra, but the nonlinear GCN training objective does not directly enforce this definition; the post-hoc walk-algebra fit does not establish recovery of the defined minimal graph.

  3. The reported “test–retest” measure compares graphs learned from disjoint subsets of training subjects rather than repeated scans of the same subjects, so it does not demonstrate conventional scan-level test–retest reliability.

  4. For submission to ICLR, the experimental scope of the paper is too small, which makes it impossible to effectively demonstrate the effectiveness of the method.

Questions for authors
  1. Can different shared graphs achieve similar reconstruction performance, and how robust are the learned edges to changes in initialization, graph regularization, and anatomical features?

  2. What is the precise relationship between the walk-algebra definition and the nonlinear training objective, and what evidence supports recovery of the proposed shared scaffold?

  3. Could the authors clarify that the reported stability measures assess split-sample graph reproducibility, and, if available, evaluate reliability using repeated scans?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.