Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Brain Wiring Knowledge Graph is a high-level abstraction from a physical neuronal wiring diagram with semantic information, helping us better understand brain functions. However, there is currently no approach that simultaneously learns both the physical connectivity and the conceptual semantic connectivity patterns within the connectome. In this paper, we propose using knowledge graphs to integrate physical connectivity and semantic connectivity. We construct knowledge graphs from the connectomes of \textit{Drosophila} and a partial human cortex. Then, we further propose a brain wiring knowledge graph reasoning framework based on Lie Group Embedding for logical neuronal relation inference. By integrating multi-dimensional neuronal data, including synaptic connectivity, spatial localization, functional activity, cellular properties, and morphological characteristics, we construct a heterogeneous brain wiring knowledge graph to capture the intricate relationships between neurons. Link prediction and neuron classification tasks reveal the connection patterns of neurons in brain functions and the distribution patterns of functional regions. Experimental results demonstrate that the proposed method excels in logical reasoning tasks. The learned embeddings of neurons can reveal the taxonomy of complex neuronal functions. Our code is available at \url{https://github.com/zzy2018730/reasoning}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4041_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2025/supp/4041_supp.zip

Link to the Code Repository

https://github.com/zzy2018730/reasoning

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhoZhe_Brain_MICCAI2025,
        author = { Zhou, Zhengyun AND Wan, Guojia AND Liao, Fei AND Hu, Wenbin AND Liao, Minghui AND Qiu, Junchao AND Li, Xinyuan AND Du, Bo},
        title = { { Brain Wiring Knowledge Graph Reasoning: A Region Embedding Approach for Logical Neuronal Relation Inference } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15971},
        month = {September},
        page = {130 -- 139}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a Brain Wiring knowledge graph for relational inference for logical reasoning tasks in state-of-the-art connectomes such as the Hemibrain, MANC, and H01 datasets.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    S1. The paper proposes a method to interrogate connectomes for small connectivity motifs and cell type embeddings.

    S2. The paper uses very recent and extremely challenging data to run experiments.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    W1 Questions about Lie Groups. It remains unclear, why did the authors choose Lie Groups to model logical relationships between neurons? What can Lie Groups express that a simple Cypher query (accessible through Neuprint) can not model? It remains unclear how the proposed knowledge graph is different from the knowledge graph that is directly accessible through Neuprint. Overall, I recommend simplifying the formal definition of the approach and distilling the core concepts to make the approach digestible, given the short manuscript length. To better illustrate the approach, I suggest iterating on figure 1, simplifying the shown math, improving acronyms, nd better explaining why a torus distance function is required for this approach.

    W2 Comparison to State of the Art. Currently, the paper does not compare against state-of-the-art baselines. Tables 1, 2, and Figure 3 do not compare other methods. The authors should demonstrate how their approaches compare to existing work. Additionally, I recommend adding a dedicated paragraph that discusses related work in computational motif analysis in connectomics and neuron cell typing.

    W3 Reproducibility Concerns. While the general approach is well defined in the paper, the experimental setup is not. It is well known that motif analysis is computationally very expensive and, in fact, an NP-complete problem. Thus, I expect some sort of scalability discussion and a statement on how many compute resources the proposed method requires. What’s the maximum size of motifs that can be discovered with reasonable, consumer-level hardware using this approach? Additionally, I would like to see quantitative result for the cell typing experiments with respect to state of the art learning based cell typing methods.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    While the paper adresses a extremly important topic in connectomics, I think it is not ready for publication in it’s current state (see reasons above). My main concern is the lack of comparison to the state of the art.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper proposes a brain wiring knowledge graph (BWKG) reasoning framework that unifies physical synaptic connectivity and semantic neuronal relationships using a Lie group-based region embedding method. The authors construct heterogeneous knowledge graphs from the Drosophila central brain (HemiBrain and MANC) and a partial human cortex (H01). These graphs encode diverse attributes such as synaptic connectivity, spatial location, neuron type, and morphology. The model performs logical reasoning over these knowledge graphs using symbolic query templates (1p, 2p, 2i, ip, 2in), and represents queries as regions on an n-dimensional torus. Through projection, conjunction, and negation operations, the model infers new neuron-level relations and performs classification. Results show strong performance across three datasets, and clustering analyses demonstrate the ability to uncover biologically meaningful neuron groupings.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    This paper presents a novel and theoretically grounded framework for brain wiring knowledge graph reasoning by integrating heterogeneous neuronal data using Lie group-based region embeddings. The proposed model effectively unifies synaptic, spatial, and morphological features from both Drosophila and partial human cortex datasets into a logical reasoning structure using expressive symbolic query templates. Its use of n-dimensional torus geometry for projection, conjunction, and negation operations is both mathematically elegant and well suited for multi-hop relational inference. Empirical results across three datasets demonstrate strong performance in link prediction tasks, and visual analyses such as t-SNE and hierarchical clustering reveal biologically meaningful neuron groupings. The inclusion of publicly available code also supports reproducibility and potential future adoption.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The paper does not include ablation studies to evaluate the contributions of key architectural components, making it difficult to disentangle which aspects of the model are responsible for its performance. In particular, the use of Lie group-based region embeddings is a core innovation, yet there is no comparison with conventional embedding spaces such as Euclidean or hyperspherical geometries, which are standard in knowledge graph reasoning. Furthermore, the individual impact of logical operations—projection, conjunction, and negation—is not analyzed, leaving it unclear whether all components contribute equally or if specific combinations are more effective for certain query types. Incorporating such studies would improve the clarity, robustness, and scientific value of the framework.

    Additionally, while the authors present visualizations of neuron clusters, it could be better if the paper offered more in-depth neuroscientific interpretation or comparisons to existing literature to validate the biological plausibility of the identified neuron groups and inferred connections. Doing so would enhance the model’s relevance for neuroscience applications and further support its claims of biological insight.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper presents a novel and mathematically grounded framework for brain wiring knowledge graph reasoning by integrating diverse neuronal attributes through Lie group-based region embeddings. The proposed method effectively handles logical query templates and demonstrates strong empirical results across three connectome datasets, including Drosophila and partial human cortex graphs. The approach is both conceptually innovative and practically relevant, offering a new perspective on integrating symbolic reasoning with brain structural data. Moreover, the biological clustering visualizations and logical reasoning results are promising, and the availability of source code supports reproducibility.

    However, the paper also has several limitations. It lacks ablation studies and comparative analysis against alternative embedding geometries, making it difficult to isolate the contribution of its core components. The logical operators—projection, conjunction, and negation—are not individually analyzed, and the benefits of the Lie group embedding are not compared with more conventional alternatives. Additionally, while neuron clusters are visualized, the paper could benefit from more thorough neuroscientific interpretation or comparison to prior literature to validate the biological relevance of the findings.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors have adequately addressed my concerns in the rebuttal, particularly with regard to the ablation studies. Overall, I support the acceptance of this paper.



Review #3

  • Please describe the contribution of the paper

    The authors proposed to use a heterogeneous knowledge graph to integrate multi-dimensional neuronal data and a brain wiring knowledge graph reasoning framework based on Lie Group Embedding for neuronal relation inference.

    The authors first converted a connectome like HemiBrain into a knowledge graph that captures the relationship between entities like NeuronID/ROI/types/attributes and inter-connectivities from pre/post synaptic connection. By formalizing the indirect relationships into a knowledge graph, the setup greatly simplifies queries like what type a neuron is or which neuron it connects to.

    The authors then projected entities as embeddings on a n-dim torus and defined conjunction operator with attention, then used various positive and negative query-answer pairs to train the model.

    Eventually the authors reported good performance on complex relationship queries like “Queries neurons that are postsynaptically connected to neuron a, but not to neuron b.”

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The authors formulated the connectomics problem with knowledge graph and group theory and offers refreshing perspectives to connectivity studies. AFAIK, this is the first work that thoroughly studies several cutting-edge connectomics datasets with similar methods
    • The Hits@3 recall performances seems quite impressive, especially for Manc-KG
    • Fig.2 visualization of clustering of neurons was quite convincing in demonstrating the embedding quality
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • 2p was by far the weakest motif benchmarked, which queries the second-degree post-synaptic targets, while the performance degrades understandably due to error propergation, the question is as the graph scales and query complexity goes up, whether the method can still yield reliable results for practical use without human proofreading
    • Overall the method lacks ablation/comparisons against other approaches, e.g. morphology/embedding based cell type classification, or other graph neural network methods.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The authors established knowledge graph as a powerful tool for connectomics data analysis, while also incorporating group theory to reformulate the connectivity relationship. I find the framework fits the usecase quite well and the methodological contributions are quite refreshing for the connectomics field.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors explained the weaknesses in 2p reasoning and even though the system’s performance is limited by 2-hop, it is still relatively competitive, overall I find combining knowledge graph with open-source connectome to be a valuable direction in the field



Author Feedback

Reviewer 1

  • Q1: Comparison with other methods
  • A1:Thank you for your constructive suggestion. Please see A6
  • Q2:Analyze the individual effect of each logical operator
  • A2:Each query type(1p,2p,2i,ip,2in)involves fixed logical operations,so individual operator ablation would alter the query meaning.Their effects are reflected in the performance across query types.
  • Q3:Offer more in-depth neuroscientific interpretation
  • A3:Notably, the neuron groups identified by our model align well with known brain regions such as the mushroom body, lateral horn,and antennal lobe,which are associated with memory formation,olfactory processing and sensory integration,respectively(Aso et al.,eLife,2014;Dolan et al.,eLife,2019)

Reviewer 2

  • Q4:Comparison with other methods.Can the method still perform reliably on 2p queries?
  • A4:Please see A6 for more comparison.While 2p queries are more prone to error due to two-hop reasoning,our Hits@50 results remain stable across datasets,showing the model can still retrieve biologically meaningful candidates

Reviewer 3

  • W1: 1)Why was a Lie group chosen?Why is it necessary to model a torus distance function? 2)What can Lie groups express that Cypher queries cannot? 3)How does your knowledge graph differ from NeuPrint?4)Valuable suggestions

  • A5:Thank you for your valuable comments 1) Since logical reasoning requires closure under operations,it is important to choose an appropriate embedding space.Lie groups,as compact manifolds,constrain logical operators within the manifold,whereas Euclidean space is an open manifold and may lead to embedding divergence.Since the operates on a Lie group manifold, torus distance is suitable. 2) e.g.2p is the abstract of the cypher query “MATCH (a)-[*1..2]-(b) RETURN b”.Our reasoning framework could perform on incomplete node connections that Cypher can`t reach the target node. 3) NeuPrint(NeuPrint,Front.Neuroinform.,2022)is a graph database for retrieving connectomic data using python interface based on NEO4J/Cypher.It only supports explicit relationship retrieval and lack semantic reasoning over incomplete or unknown neurons.We construct a semantic knowledge graph with our framework that enables inference of missing connections and neuron functions and supports incomplete logical query answering. 4) We will revise these.

  • W2:Provide a comparison with existing work.Discuss related work
  • A6:Thank you for your constructive suggestion.According to your suggestion,we compared our approach on the H01 dataset with representative methods that perform well on related tasks with real-valued embeddings using MRR:the Euclidean space-based method On 1p,2p,2i and ip scored 31.67,37.06,80.12,88.87,underperformed compared to our Lie group-based method.To our best knowledge,we first construct logical reasoning knowledge graphs on connectomes and perform motif-based reasoning.There are almost no existing method targets this setting.We will add more discussion on motif analysis and neuron cell typing.
  • W3: 1) No experimental setup. 2) Motif analysis is costly and NP-complete. 3) Show results vs.sota cell typing methods
  • A7: 1) The detailed experimental settings could be seen in our code (BWKGR/example.sh) 2) We agree with you that motif matching is theoretical NP-hard with the growth of motif size.Specifically,the query motif we used are small(<5 triplets)so that the cost is acceptable.Technically,we use Dotmotif(JK Matelsky,et al.2021 Sci.Rep.)that is a highly optimized connectomic motif matching package conducted in CPU to solve this problem.In our case,motif extraction is done before training and does not require GPUs, keeping the computational overhead low. 3) Insightful feedback! Learning based cell typing methods, such as morphology-based, connectivity-based could predict cell types.However,these type information have been in our KGs as semantic relationships,making our reasoning tasks and classification tasks difficult to compare directly.It is insightful to adopt KG reasoning to perform cell typing task.We are glad to



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top