acceptodds
Under review as a conference paper at ICLR 2027

TopoScanPE: Topological Positional Encodings for Brain Network Transformers

Abstract

Brain regions can have similar functional connectivity strengths while playing different roles within the connectome. To capture these differences, we introduce *TopoScanPE*, a positional encoding that represents each region of interest (ROI) through the topology of its local neighborhood. For every ROI we extract its local neighborhood from a sparse connectivity graph and apply TopoScan (Uddin et al., 2026), a sliding-window filtration over complementary geometric node functions that summarizes each slice by the topology of its clique complex. A shared encoder maps the resulting TopoScan (Uddin et al., 2026) sequences into a fixed-width code that is concatenated to the ROI token. Experiments on ABIDE and ADNI show that TopoScanPE competitive state-of-the-art brain graph transformers and graph neural networks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Topological Data Analysis, Brain Network Transformers, Brain Disease Prediction, Topological Positional Encoding

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WCKQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TopoScanPE to represent each ROI through the topology of its local neighborhood, encodes the resulting Topo-Scan sequences, and concatenates them with connectivity profiles in BrainNetTF.

Critical strengths
  • The ROI-level encoding provides a sensible way to incorporate local topology into attention-based brain network models.
  • Results on ABIDE are encouraging, with improvements over the reported baselines across all four metrics.
Critical weaknesses
  • The description for some parts of the method is incomplete. The Forman–Ricci node function and heat kernel signature are not explicitly defined. Their formulas, Laplacian convention, and computation within each neighborhood are needed for reproducibility. Distinct quantile values can produce variable-length sequences, whereas the encoder assumes a fixed LL. Please explain how duplicate thresholds, constant functions, and padding are handled.
  • The efficiency analysis is missing. Report the preprocessing cost of ROI-wise topology computation, additional parameters, and training/inference time to substantiate the description of the encoder as lightweight.
Questions for authors
  • What specific pattern has been captured by the proposed PE? How does it align with existing findings for a specific disease?
  • Does the topology encoding remain beneficial when added to a second transformer backbone?
  • Existing work explicitly explores positional and structural encodings for brain networks [1-3]. Why these works are not included for discussion and comparison?

[1] Brain-JEPA: Brain Dynamics Foundation Model with Gradient Positioning and Spatiotemporal Masking. NeurIPS 2024 [2] Contrasformer: A Brain Network Contrastive Transformer for Neurodegenerative Condition Identification. CIKM 2024 [3] A Brain Graph Foundation Model: Pre-Training and Prompt-Tuning across Broad Atlases and Disorders. ICLR 2026

4 comments
AuthorsOpenReview ↗

Response to Reviewer WCKQ (1/4)

Dear Reviewer WCKQ, thank you for the thoughtful review and the questions about interpretation and related work. We added the missing definitions, a cost analysis, an analysis of the patterns the encoding captures and their relation to ASD and AD findings, results on other backbones, and a comparison with Brain-JEPA, Contrasformer and BrainGFM. We hope these additions resolve your remaining doubts and encourage you to reconsider your assessment.

W1. Definitions, duplicate thresholds, constant functions and padding

The description for some parts of the method is incomplete. The Forman–Ricci node function and heat kernel signature are not explicitly defined. Their formulas, Laplacian convention, and computation within each neighborhood are needed for reproducibility. Distinct quantile values can produce variable-length sequences, whereas the encoder assumes a fixed L. Please explain how duplicate thresholds, constant functions, and padding are handled.

Reply: Thank you for pointing this out; these details are indeed needed for reproducibility. We have added the definitions to Appendix A, and Sec. 4.2 and Appendix A now explain the duplicate thresholds and the padding:

All quantities are computed on the induced subgraph Gᵢ = (Vᵢ, Eᵢ) of ROI i's bounded two-hop neighbourhood in the unweighted, symmetrised k-nearest-neighbour graph (k = 8); degrees are taken within Gᵢ.

Forman–Ricci curvature (edge function). We use the combinatorial Forman curvature of an unweighted graph without higher-order cells, F(a, b) = 4 − deg(a) − deg(b) for an edge (a, b) ∈ Eᵢ. Topo-Scan uses F directly on the edges: the thresholds are quantiles of {F(e)} over Eᵢ, and each slice is the subgraph formed by the edges whose curvature falls in the window, together with their endpoints.

Heat kernel signature (node function). We use the symmetric normalised Laplacian of Gᵢ, ℒ = I − D^(−1/2) A D^(−1/2) (rows of vertices without edges equal those of I). With the eigendecomposition ℒ = Σₖ λₖ φₖ φₖᵀ, HKSₜ(v) = Σₖ exp(−t λₖ) φₖ(v)², at t = 1, computed by a full eigendecomposition of each neighbourhood (at most 81 vertices on ABIDE and 37 on ADNI at ρ = 0.4).

Duplicate thresholds and padding. For each filtration function f we take Q = L + m = 22 quantile thresholds of its values (over Eᵢ for Forman–Ricci, over Vᵢ for HKS) and remove duplicates, leaving Q′ ≤ Q distinct thresholds. The windows α_ℓ ≤ f ≤ α_{ℓ+m} give min(Q′ − m, L) slices. Thresholds are never repeated to fill the sequence: the remaining positions are zero columns, (0, 0, 0, 0), so every sequence has length L = 20. Sequences are shorter than L for 7.54% (Forman–Ricci) and 0.26% (HKS) of the ROIs on ABIDE, and for 13.43% and 0.11% on ADNI. We have corrected Eq. (4), which incorrectly implied L = Q − m.

Constant functions. The thresholds are taken from the values of f within each ROI neighbourhood, so f is constant only if it has the same value at every edge (Forman–Ricci) or vertex (HKS) of Gᵢ, e.g. when Gᵢ is a complete graph. This is rare (0.21% of the ROIs on ABIDE, none on ADNI) and needs no separate rule: no threshold is repeated, so all L slices are zero columns. Appendix A now states this.

AuthorsOpenReview ↗

Response to Reviewer WCKQ (2/4)

W2. Efficiency

The efficiency analysis is missing. Report the preprocessing cost of ROI-wise topology computation, additional parameters, and training/inference time to substantiate the description of the encoder as lightweight.

Reply: The signature encoder adds 106,960 parameters (2.4% of the model). Topology is computed once per subject before training, in about 3.6 s per subject (about 1 h for all of ABIDE), and cached. Training and inference are therefore faster than with RWSE or LapPE, which are recomputed in every forward pass.

Cost (ABIDE, BrainNetTF backbone, 200 epochs per run, mean over the five splits; all on one NVIDIA RTX PRO 5000 Blackwell GPU)

EncodingPre-processing (once)Training (min per run)Inference (ms per subject)
RWSE (ALTER's encoding)— (computed in every forward pass)4.771.09
LapPE— (computed in every forward pass)16.244.49
TopoScanPE≈ 3.6 s per subject (≈ 1 h for 1,009 subjects × 200 ROI neighbourhoods)4.350.40

Q1. Which pattern does the encoding capture, and how does it relate to disease findings?

What specific pattern has been captured by the proposed PE? How does it align with existing findings for a specific disease?

Reply: The encoding describes how each ROI's local functional neighbourhood is organised, not how strongly the ROI is connected. Along the Forman–Ricci order of the edges (from edges between hubs to peripheral edges) and the heat-kernel order of the vertices, Topo-Scan records whether the neighbourhood stays connected or splits (β₀), how many cycles are not filled by triangles (β₁), and the slice sizes. It thus distinguishes an ROI whose neighbours form a tight module from one whose neighbours are loosely connected or split into groups, a local measure of segregation versus integration. On the trained models, the learned code predicts clustering (R² = 0.50), betweenness (0.38) and participation (0.37) far better than RWSE (0.07, 0.17, 0.07), and carries little information about connection strength.

Both disorders are commonly described at this level: autism by an atypical balance between local and long-range functional connectivity (Belmonte et al., 2004; Courchesne & Pierce, 2005), and Alzheimer's disease by reduced local clustering and a loss of small-world organisation (Supekar et al., 2008). On ABIDE, 39 of 200 ROIs have signatures that differ between ASD and controls (FDR q < 0.05), mostly in the somatomotor and subcortical/cerebellar networks. In 29 of them the signature is lower in ASD, i.e. fewer vertices, edges, components and cycles per slice; 10, mostly default-mode ROIs, are higher. Half of the 20 ROIs the trained model relies on most are among these 39 (hypergeometric p = 0.001; new Fig. 6, Appendix D). On ADNI (124 participants) no ROI survives FDR correction, so we draw no disease-level conclusion there. We have added this interpretation to Appendix D (Figs. 5 and 6).

AuthorsOpenReview ↗

Response to Reviewer WCKQ (3/4)

Q2. A second transformer backbone

Does the topology encoding remain beneficial when added to a second transformer backbone?

Reply: Yes, for some backbones. We added TopoScanPE, unchanged, to Graphormer, GraphTrans and SAN on ABIDE and ADNI.

Other backbones (test metrics, mean ± std over five splits; base: scores from Table 1; + TopoScanPE: layers and dropout chosen per split on validation loss; Table 3 of the revised paper)

DatasetBackboneVariantAUCACCSENSPE
ABIDEGraphormerbase (Table 1)63.91 ± 4.0561.88 ± 6.8566.30 ± 9.9855.74 ± 11.00
+ TopoScanPE69.81 ± 2.7264.63 ± 3.7867.78 ± 5.3561.27 ± 5.11
GraphTransbase (Table 1)60.13 ± 6.7357.83 ± 4.7165.70 ± 10.3049.77 ± 11.52
+ TopoScanPE57.19 ± 4.0153.79 ± 3.2644.72 ± 11.9664.48 ± 8.27
SANbase (Table 1)71.35 ± 2.1865.34 ± 2.9155.41 ± 9.2968.39 ± 7.50
+ TopoScanPE61.59 ± 6.4957.83 ± 5.4959.02 ± 10.3357.09 ± 8.50
ADNIGraphormerbase (Table 1)60.69 ± 5.2655.75 ± 3.1860.18 ± 11.3647.75 ± 13.53
+ TopoScanPE64.36 ± 5.3262.31 ± 8.5741.82 ± 12.3377.33 ± 17.18
GraphTransbase (Table 1)61.41 ± 3.6558.60 ± 5.4165.57 ± 6.0554.37 ± 3.42
+ TopoScanPE49.45 ± 10.7651.54 ± 7.5425.45 ± 23.9870.67 ± 24.07
SANbase (Table 1)66.11 ± 3.4161.78 ± 4.2253.94 ± 7.5663.63 ± 8.51
+ TopoScanPE61.82 ± 7.8461.54 ± 7.6927.27 ± 19.0786.67 ± 17.38

Against their Table 1 scores, TopoScanPE improves Graphormer on both datasets (AUC +5.90 on ABIDE and +3.67 on ADNI, with higher accuracy). It does not improve GraphTrans (AUC −2.94 on ABIDE, −11.96 on ADNI) or SAN (−9.76 and −4.29), so we describe the transfer to other backbones as partial.

AuthorsOpenReview ↗

Response to Reviewer WCKQ (4/4)

Q3. Existing positional and structural encodings for brain networks

Existing work explicitly explores positional and structural encodings for brain networks [1-3]. Why these works are not included for discussion and comparison?

Reply: Thank you for these references; we have added them to the related work. Brain-JEPA (Dong et al., NeurIPS 2024) positions each ROI by its coordinates on population-level connectivity gradients, and Contrasformer (Xu et al., CIKM 2024) adds a learnable ROI identity embedding; both are the same for every subject. BrainGFM (Wei et al., ICLR 2026) uses RWSE as its structural encoding. TopoScanPE is instead computed per subject from the topology of each ROI's own neighbourhood and can be added to an existing backbone. Under the same backbone and protocol, RWSE reaches 79.92 AUC on ABIDE against 81.96 for TopoScanPE. We also trained the three models themselves on our five ABIDE splits and the five ADNI splits, with their published architectures and learning rates (Brain-JEPA and BrainGFM fine-tuned from their released pre-trained checkpoints) and the same model selection as TopoScanPE:

ABIDE; dropout of each baseline chosen per split on validation loss; TopoScanPE as in Table 1 (test metrics, mean ± std)

ModelAUCACCSENSPE
Brain-JEPA56.82 ± 2.3152.81 ± 3.6055.20 ± 29.7851.96 ± 26.75
Contrasformer69.87 ± 2.9664.24 ± 2.4371.22 ± 5.3556.95 ± 10.78
BrainGFM72.85 ± 2.9867.68 ± 1.7063.54 ± 5.4172.17 ± 5.59
TopoScanPE81.96 ± 1.0475.47 ± 0.8174.99 ± 6.5876.34 ± 5.58

TopoScanPE has the highest AUC, accuracy, sensitivity and specificity.

ADNI; dropout of each baseline chosen per split on validation loss; TopoScanPE as in Table 1 (test metrics, mean ± std)

ModelAUCACCSENSPE
Brain-JEPA51.45 ± 9.6964.62 ± 8.2132.73 ± 9.2788.00 ± 9.80
Contrasformer65.58 ± 4.7561.54 ± 8.4360.00 ± 14.7762.67 ± 20.04
BrainGFM63.64 ± 7.7960.00 ± 10.2050.91 ± 9.2766.67 ± 18.38
TopoScanPE71.15 ± 2.9372.31 ± 1.7258.18 ± 13.7982.67 ± 12.11

On ADNI, TopoScanPE also has the highest AUC and accuracy.

Reviewer_dYs2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TopoScanPE, which applies Topo-Scan to each ROI’s local neighborhood and uses Forman–Ricci curvature and heat kernel signature to generate sequences of Betti numbers and slice sizes. A shared encoder converts these sequences into positional encodings for BrainNetTF. The method performs well on ABIDE and achieves the highest accuracy and specificity on ADNI, although its AUC and sensitivity do not lead.

Critical strengths
  1. The motivation is clear. Incorporating local topological descriptors into ROI tokens allows structural information to influence attention rather than entering only at the final readout.

  2. The ablations are informative. Comparisons with Random PE, individual filtrations, Betti numbers, slice sizes, and removal of the residual projection provide useful evidence about the components.

  3. The ABIDE results are promising, and some evaluation safeguards are explicit. AUC and ACC improve over BQN by 2.11 and 2.94 percentage points, respectively; epoch selection uses validation loss, and feature-standardization statistics are computed from training subjects only.

Critical weaknesses
  1. The advantage over conventional positional and structural encodings is not sufficiently established. Since the main contribution adapts existing Topo-Scan to ROI-level input encoding, direct comparisons with LapPE, RWSE, and simple local structural features under the same BrainNetTF and training protocol are important.

  2. The source of the improvement remains insufficiently isolated. On ABIDE, Random PE already reaches 80.07 AUC, numerically above BQN’s 79.85, while sizes-only reaches 81.07 versus 81.96 for the full model; these results motivate controls that separate topology from the additional encoder and residual projection, whose inclusion in the BrainNetTF baseline is unclear.

  3. Hyperparameter selection raises a test-set selection concern, and the ADNI advantage is mixed. Figure 4 closely matches the reported test results, yet Section 5.3 uses the best performance on these curves to justify selecting (\rho=0.4), without documenting an independent validation-based choice. On ADNI, AUC ranks third, while mean balanced accuracy derived from SEN and SPE is approximately 70.43% versus BQN’s 69.46%, so the higher ACC does not establish a clear overall advantage.

  4. The construction of fixed-length topological sequences requires clarification. The node-level Forman–Ricci function and HKS Laplacian are not explicitly defined, and removing duplicate quantile thresholds may produce variable slice counts despite the fixed (L=20).

Questions for authors
  1. Could the authors compare TopoScanPE with LapPE, RWSE, and simple local structural features under the same BrainNetTF, including matched residual projection and Mixup settings? Please clarify which architectural and training settings are shared with the reported BrainNetTF baseline.

  2. Is Random PE fixed or resampled, and does it pass through the same signature encoder? What are the paired performance differences and uncertainty estimates between the full model, Random PE, and sizes-only?

  3. Were (\rho) and (L) selected without consulting test performance, and which split supplies the curves in Figure 4 and Appendix B? If test results informed selection, please repeat selection using validation data and evaluate on an untouched test set; also clarify whether all methods share the five splits and how any repeated ADNI scans are grouped.

  4. How are node-level Forman–Ricci curvature and HKS defined, including the Laplacian choice? Please explain duplicate thresholds, padding or resampling to a fixed sequence length, and the additional preprocessing and training costs.

5 comments
AuthorsOpenReview ↗

Response to Reviewer dYs2 (1/5)

Dear Reviewer dYs2, thank you for the detailed review and the precise questions. We compared TopoScanPE with standard encodings under an identical backbone and protocol, isolated the encoder and the residual projection, re-trained the baselines on our splits so that every comparison is paired, and repeated the selection of ρ on validation data. All runs share the paper's five splits and the protocol below. We hope the new results address your concerns and encourage you to reconsider your assessment.

Protocol shared by every new run (all encodings, ablations and re-trained baselines)

ComponentSetting used
SplitsThe same five train/validation/test splits per dataset as in the paper, read from file and never redrawn
BackboneBrainNetTF, with the residual projection XW_res of Eq. (11), for TopoScanPE and every other encoding. The re-trained baselines keep their own architectures: BrainNetTF and ALTER are graph transformers, BQN is a quadratic network without message passing or attention
Encoding size16 per ROI, concatenated to the FC row
OptimisationAdam, learning rate 10⁻⁴ → 10⁻⁵ (cosine), weight decay 10⁻⁴, batch 16, 200 epochs
Augmentation and lossMixup; cross-entropy
Model selectionEpoch with the lowest validation loss; test metrics reported at that epoch

W1 / Q1. Comparison with LapPE, RWSE and local structural features under the same BrainNetTF

W1. The advantage over conventional positional and structural encodings is not sufficiently established. Since the main contribution adapts existing Topo-Scan to ROI-level input encoding, direct comparisons with LapPE, RWSE, and simple local structural features under the same BrainNetTF and training protocol are important.

Q1. Could the authors compare TopoScanPE with LapPE, RWSE, and simple local structural features under the same BrainNetTF, including matched residual projection and Mixup settings? Please clarify which architectural and training settings are shared with the reported BrainNetTF baseline.

Reply: Thank you for this suggestion; we ran exactly this comparison. All variants share BrainNetTF, the residual projection, Mixup, optimiser, schedule, splits and model selection; only the encoding changes. We compare TopoScanPE with LapPE and RWSE, with LapPE and RWSE passed through the same signature encoder, and with Topo-Scan using simple local statistics (degree, clustering coefficient, triangle count, closeness) as filtration functions. On both datasets TopoScanPE has a higher AUC than LapPE and RWSE, with or without the encoder. The simpler filtrations are lower, except one ADNI pair (Forman–Ricci + clustering coefficient) that matches its AUC with lower accuracy; curvature carries most of the signal.

AuthorsOpenReview ↗

Response to Reviewer dYs2 (2/5)

Controlled encoding comparison on ABIDE (test metrics at the lowest-validation-loss epoch, mean ± std over the five splits)

EncodingAUCACCSENSPE
TopoScanPE (Forman–Ricci + HKS)81.96 ± 1.0475.47 ± 0.8174.99 ± 6.5876.34 ± 5.58
LapPE80.70 ± 1.1773.79 ± 1.7576.58 ± 3.0070.98 ± 3.23
LapPE + signature encoder79.65 ± 2.1472.51 ± 2.8373.08 ± 6.9872.44 ± 9.26
RWSE (ALTER's encoding)79.92 ± 1.7972.32 ± 2.5072.30 ± 9.0872.71 ± 6.24
RWSE + signature encoder79.91 ± 1.7070.94 ± 1.7972.73 ± 8.5869.50 ± 8.32
Forman–Ricci + degree81.52 ± 1.1474.68 ± 1.3474.53 ± 3.6175.06 ± 4.47
Forman–Ricci + clustering coefficient80.14 ± 1.6272.81 ± 3.2473.98 ± 5.3171.87 ± 5.36
Forman–Ricci + triangle count80.58 ± 1.3673.69 ± 1.0179.50 ± 6.2967.73 ± 7.47
Forman–Ricci + closeness80.89 ± 1.3973.40 ± 2.8277.70 ± 7.7469.25 ± 6.83
FC value (edge) + degree79.34 ± 3.4973.40 ± 2.4577.73 ± 7.1468.65 ± 4.48
FC value (edge) + clustering coefficient78.62 ± 2.4470.74 ± 3.0369.95 ± 9.4771.54 ± 7.21
FC value (edge) + triangle count77.69 ± 4.4370.64 ± 4.9474.25 ± 8.4867.47 ± 13.35
FC value (edge) + closeness80.58 ± 1.3174.09 ± 1.9978.83 ± 5.9169.48 ± 6.17
FC value (edge) + HKS80.60 ± 1.2673.20 ± 2.1377.70 ± 8.2568.71 ± 6.12

Controlled encoding comparison on ADNI (test metrics at the lowest-validation-loss epoch, mean ± std over the five splits)

EncodingAUCACCSENSPE
TopoScanPE (Forman–Ricci + HKS)71.15 ± 2.9372.31 ± 1.7258.18 ± 13.7982.67 ± 12.11
LapPE66.55 ± 3.8860.00 ± 3.9230.91 ± 21.9781.33 ± 21.66
LapPE + signature encoder69.09 ± 2.9261.54 ± 3.4436.36 ± 23.0080.00 ± 17.89
RWSE (ALTER's encoding)68.61 ± 3.7265.38 ± 3.4438.18 ± 21.0585.33 ± 11.47
RWSE + signature encoder64.85 ± 2.9960.00 ± 3.9230.91 ± 15.8581.33 ± 16.55
Forman–Ricci + degree68.61 ± 5.8866.15 ± 2.8840.00 ± 21.9785.33 ± 11.47
Forman–Ricci + clustering coefficient72.36 ± 3.3863.08 ± 1.8841.82 ± 27.3378.67 ± 18.09
Forman–Ricci + triangle count68.00 ± 2.7563.85 ± 5.7649.09 ± 27.9374.67 ± 25.44
Forman–Ricci + closeness69.21 ± 6.1467.69 ± 3.9260.00 ± 21.2073.33 ± 17.38
FC value (edge) + degree67.39 ± 4.6162.31 ± 4.4925.45 ± 23.9889.33 ± 12.36
FC value (edge) + clustering coefficient67.76 ± 3.9560.00 ± 1.8821.82 ± 20.4188.00 ± 12.93
FC value (edge) + triangle count64.97 ± 4.9063.08 ± 4.6243.64 ± 25.3277.33 ± 14.36
FC value (edge) + closeness64.12 ± 7.6663.85 ± 3.9245.45 ± 32.5277.33 ± 18.18
FC value (edge) + HKS67.76 ± 3.9265.38 ± 8.4330.91 ± 24.1290.67 ± 15.55
AuthorsOpenReview ↗

Response to Reviewer dYs2 (3/5)

Shared settings. In the submission, the baseline rows of Table 1 were taken from the published results of BQN (Yang et al., 2025), so they did not share our splits, Mixup or residual projection. We have now re-trained BrainNetTF and ALTER with our protocol: on ABIDE they reach 80.44 and 79.67 AUC, against 81.96 for TopoScanPE. ALTER with and without the residual projection reaches 79.92 and 79.67 on ABIDE and 68.61 and 68.73 on ADNI, so the residual projection alone explains little of the gain.

W2 / Q2. Source of the improvement; Random PE and sizes-only

The source of the improvement remains insufficiently isolated. On ABIDE, Random PE already reaches 80.07 AUC, numerically above BQN's 79.85, while sizes-only reaches 81.07 versus 81.96 for the full model; these results motivate controls that separate topology from the additional encoder and residual projection, whose inclusion in the BrainNetTF baseline is unclear.

Q2. Is Random PE fixed or resampled, and does it pass through the same signature encoder? What are the paired performance differences and uncertainty estimates between the full model, Random PE, and sizes-only?

Reply: Random PE is standard Gaussian noise of the same shape as the Topo-Scan signature, drawn once with a fixed seed and kept fixed for training and evaluation. It passes through the same signature encoder and residual projection, so only the topological information is removed. Sizes-only keeps |V| and |E| of every slice and sets β₀ and β₁ to zero.

Same configuration for all three variants (test metrics, mean ± std over the five splits)

DatasetModelAUCACCSENSPE
ABIDEFull TopoScanPE81.96 ± 1.0475.47 ± 0.8174.99 ± 6.5876.34 ± 5.58
Random PE80.07 ± 1.4072.22 ± 2.0569.51 ± 7.2375.17 ± 3.92
Sizes-only81.07 ± 1.5573.89 ± 1.7473.56 ± 6.0974.43 ± 5.81
ADNIFull TopoScanPE71.15 ± 2.9372.31 ± 1.7258.18 ± 13.7982.67 ± 12.11
Random PE67.39 ± 5.2763.08 ± 5.8347.27 ± 21.7074.67 ± 15.92
Sizes-only69.09 ± 4.8766.15 ± 5.0150.91 ± 33.1577.33 ± 18.62

As you note, Random PE already exceeds BQN’s published AUC on ABIDE (80.07 vs. 79.85). BQN, however, does not use BrainNetTF: it is a quadratic network without message passing or attention, so this comparison differs in architecture as well as in protocol. The matched reference for Random PE is its own backbone, BrainNetTF, re-trained with our protocol: it reaches 80.44 AUC on ABIDE, Random PE does not improve on it (80.07), and TopoScanPE reaches 81.96. We therefore now report the gain against BrainNetTF and ALTER re-trained with the same protocol (W1).

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

This paper proposes TopoScanPE, a topology-based positional encoding for brain network transformers. For each brain region, the method extracts a local neighborhood from the functional connectivity graph and applies Topo-Scan using Forman–Ricci curvature and heat kernel signatures. The resulting topological descriptors are processed by a shared Transformer encoder to generate ROI-level positional encodings, which are incorporated into BrainNetTF for brain disorder classification. Experiments on ABIDE and ADNI compare the method against several baselines, with additional ablation and sensitivity studies.

Critical strengths
  1. The paper addresses a relevant problem by incorporating local topological structure into ROI-level representations.
  2. The proposed encoding is modular and can be integrated into existing brain network Transformer architectures.
  3. The experiments include a broad set of baselines.
Critical weaknesses
  1. On ABIDE, the sizes-only simple variant achieves an AUC close to the full model's. The current ablations do not sufficiently disentangle the contributions of topological information, simple structural statistics, and the additional encoding architecture. Comparisons with standard graph positional encodings (e.g., LapPE and RWSE), simpler structural features (like degree、clustering coefficient、triangle count) variant are missing.
  2. The paper does not explicitly define how Forman–Ricci curvature is converted into a node-level function or which curvature formulation is used. Similarly, the HKS computation lacks a precise definition of the graph Laplacian and its normalization. 
  3. Although TopoScanPE achieves higher accuracy and specificity, its sensitivity is substantially lower than BQN , and its AUC is also lower. This trade-off limits the strength of the claimed performance improvements on ADNI.
  4. ADNI contains only 124 subjects, leaving approximately 25 test subjects per split. With only five random splits and no confidence intervals or paired statistical comparisons, the reliability of the reported improvements over strong baselines remains unclear.
  5. Section 5.3 identifies \rho=0.4 as the best setting based on the reported performance curves, but it is unclear whether \rho were selected using validation data independently of the test results. 
  6. The paper provides no visualization or further analysis of the learned PE representations. As a result, it remains unclear whether these encodings capture the structural distinctions between ROIs that motivate the method.
  7. Although the method is claimed as applicable to different brain network transformers, experiments only demonstrate its integration with BrainNetTF. Its effectiveness across other backbones remains unverified.
Questions for authors
  1. Can the authors provide controlled comparisons with simpler structural positional encodings to better isolate the contribution of Topo-Scan?
  2. Can the authors clarify the exact formulations of the Forman–Ricci and HKS node functions?
  3. Was \rho selected exclusively using validation data, before evaluating test performance?
  4. Can the authors provide paired comparisons or confidence intervals for the improvements over the strongest baselines, particularly on ADNI?
  5. Can the authors provide evidence that the learned positional encodings capture meaningful differences in local brain network topology?
5 comments
AuthorsOpenReview ↗

Response to Reviewer ADsz (1/5)

Comment: Dear Reviewer ADsz, thank you for your careful and constructive review. We added controlled comparisons with standard positional encodings and simpler filtrations, baselines re-trained on our splits with confidence intervals and paired tests, a validation-based selection of ρ, and an analysis of the learned encoding. All new runs share the paper's five splits and training protocol (table below). We hope the new evidence resolves your concerns and encourages you to reconsider your assessment.

Protocol shared by every new run (all encodings, ablations and re-trained baselines)

ComponentSetting used
SplitsThe same five train/validation/test splits per dataset as in the paper, read from file and never redrawn
BackboneBrainNetTF, with the residual projection XW_res of Eq. (11), for TopoScanPE and every other encoding. The re-trained baselines keep their own architectures: BrainNetTF and ALTER are graph transformers, BQN is a quadratic network without message passing or attention
Encoding size16 per ROI, concatenated to the FC row
OptimisationAdam, learning rate 10⁻⁴ → 10⁻⁵ (cosine), weight decay 10⁻⁴, batch 16, 200 epochs
Augmentation and lossMixup; cross-entropy
Model selectionEpoch with the lowest validation loss; test metrics reported at that epoch

W1 / Q1. Disentangling topology, simple structural statistics and the encoder

On ABIDE, the sizes-only simple variant achieves an AUC close to the full model's. The current ablations do not sufficiently disentangle the contributions of topological information, simple structural statistics, and the additional encoding architecture. Comparisons with standard graph positional encodings (e.g., LapPE and RWSE), simpler structural features (like degree, clustering coefficient, triangle count) variant are missing.

Q1. Can the authors provide controlled comparisons with simpler structural positional encodings to better isolate the contribution of Topo-Scan?

Reply: Thank you for suggesting these controls; we have added all of them. Each changes only the encoding: (i) LapPE and RWSE; (ii) LapPE and RWSE passed through TopoScanPE's signature encoder, which separates the information in the encoding from the encoder's capacity; and (iii) Topo-Scan with simple local statistics (degree, clustering coefficient, triangle count, closeness) as filtration functions, paired with Forman–Ricci curvature or an FC-value edge filtration. On both datasets TopoScanPE has a higher AUC than LapPE and RWSE, with or without the encoder, so its gain does not come from the encoder. Among the simpler filtrations, pairs that keep Forman–Ricci curvature come closest, and on ADNI one pair (with the clustering coefficient) matches TopoScanPE in AUC but with lower accuracy; curvature therefore carries most of the signal. The sizes-only variant is not a simple structural statistic: |V| and |E| are counted per slice along the curvature and HKS orders, so they are part of the topological scan.

AuthorsOpenReview ↗

Response to Reviewer ADsz (2/5)

Controlled encoding comparison on ABIDE (test metrics at the lowest-validation-loss epoch, mean ± std over the five splits)

EncodingAUCACCSENSPE
TopoScanPE (Forman–Ricci + HKS)81.96 ± 1.0475.47 ± 0.8174.99 ± 6.5876.34 ± 5.58
LapPE80.70 ± 1.1773.79 ± 1.7576.58 ± 3.0070.98 ± 3.23
LapPE + signature encoder79.65 ± 2.1472.51 ± 2.8373.08 ± 6.9872.44 ± 9.26
RWSE (ALTER's encoding)79.92 ± 1.7972.32 ± 2.5072.30 ± 9.0872.71 ± 6.24
RWSE + signature encoder79.91 ± 1.7070.94 ± 1.7972.73 ± 8.5869.50 ± 8.32
Forman–Ricci + degree81.52 ± 1.1474.68 ± 1.3474.53 ± 3.6175.06 ± 4.47
Forman–Ricci + clustering coefficient80.14 ± 1.6272.81 ± 3.2473.98 ± 5.3171.87 ± 5.36
Forman–Ricci + triangle count80.58 ± 1.3673.69 ± 1.0179.50 ± 6.2967.73 ± 7.47
Forman–Ricci + closeness80.89 ± 1.3973.40 ± 2.8277.70 ± 7.7469.25 ± 6.83
FC value (edge) + degree79.34 ± 3.4973.40 ± 2.4577.73 ± 7.1468.65 ± 4.48
FC value (edge) + clustering coefficient78.62 ± 2.4470.74 ± 3.0369.95 ± 9.4771.54 ± 7.21
FC value (edge) + triangle count77.69 ± 4.4370.64 ± 4.9474.25 ± 8.4867.47 ± 13.35
FC value (edge) + closeness80.58 ± 1.3174.09 ± 1.9978.83 ± 5.9169.48 ± 6.17
FC value (edge) + HKS80.60 ± 1.2673.20 ± 2.1377.70 ± 8.2568.71 ± 6.12

Controlled encoding comparison on ADNI (test metrics at the lowest-validation-loss epoch, mean ± std over the five splits)

EncodingAUCACCSENSPE
TopoScanPE (Forman–Ricci + HKS)71.15 ± 2.9372.31 ± 1.7258.18 ± 13.7982.67 ± 12.11
LapPE66.55 ± 3.8860.00 ± 3.9230.91 ± 21.9781.33 ± 21.66
LapPE + signature encoder69.09 ± 2.9261.54 ± 3.4436.36 ± 23.0080.00 ± 17.89
RWSE (ALTER's encoding)68.61 ± 3.7265.38 ± 3.4438.18 ± 21.0585.33 ± 11.47
RWSE + signature encoder64.85 ± 2.9960.00 ± 3.9230.91 ± 15.8581.33 ± 16.55
Forman–Ricci + degree68.61 ± 5.8866.15 ± 2.8840.00 ± 21.9785.33 ± 11.47
Forman–Ricci + clustering coefficient72.36 ± 3.3863.08 ± 1.8841.82 ± 27.3378.67 ± 18.09
Forman–Ricci + triangle count68.00 ± 2.7563.85 ± 5.7649.09 ± 27.9374.67 ± 25.44
Forman–Ricci + closeness69.21 ± 6.1467.69 ± 3.9260.00 ± 21.2073.33 ± 17.38
FC value (edge) + degree67.39 ± 4.6162.31 ± 4.4925.45 ± 23.9889.33 ± 12.36
FC value (edge) + clustering coefficient67.76 ± 3.9560.00 ± 1.8821.82 ± 20.4188.00 ± 12.93
FC value (edge) + triangle count64.97 ± 4.9063.08 ± 4.6243.64 ± 25.3277.33 ± 14.36
FC value (edge) + closeness64.12 ± 7.6663.85 ± 3.9245.45 ± 32.5277.33 ± 18.18
FC value (edge) + HKS67.76 ± 3.9265.38 ± 8.4330.91 ± 24.1290.67 ± 15.55
AuthorsOpenReview ↗

Response to Reviewer ADsz (3/5)

W2 / Q2. Forman–Ricci curvature and HKS Laplacian

The paper does not explicitly define how Forman–Ricci curvature is converted into a node-level function or which curvature formulation is used. Similarly, the HKS computation lacks a precise definition of the graph Laplacian and its normalization.

Reply: Thank you for pointing this out; both definitions were missing. We have added them to Appendix A. Forman–Ricci curvature is not converted to a node function: Topo-Scan scans it as an edge function.

All quantities are computed on the induced subgraph Gᵢ = (Vᵢ, Eᵢ) of ROI i's bounded two-hop neighbourhood in the unweighted, symmetrised k-nearest-neighbour graph (k = 8); degrees are taken within Gᵢ.

Forman–Ricci curvature (edge function). We use the combinatorial Forman curvature of an unweighted graph without higher-order cells, F(a, b) = 4 − deg(a) − deg(b) for an edge (a, b) ∈ Eᵢ. Topo-Scan uses F directly on the edges: the thresholds are quantiles of {F(e)} over Eᵢ, and each slice is the subgraph formed by the edges whose curvature falls in the window, together with their endpoints.

Heat kernel signature (node function). We use the symmetric normalised Laplacian of Gᵢ, ℒ = I − D^(−1/2) A D^(−1/2) (rows of vertices without edges equal those of I). With the eigendecomposition ℒ = Σₖ λₖ φₖ φₖᵀ, HKSₜ(v) = Σₖ exp(−t λₖ) φₖ(v)², at t = 1, computed by a full eigendecomposition of each neighbourhood (at most 81 vertices on ABIDE and 37 on ADNI at ρ = 0.4).

W3. ADNI: lower sensitivity and AUC than BQN

Although TopoScanPE achieves higher accuracy and specificity, its sensitivity is substantially lower than BQN, and its AUC is also lower. This trade-off limits the strength of the claimed performance improvements on ADNI.

Reply: We agree and have revised the ADNI claims. In Table 1, TopoScanPE ranks third in AUC (71.15 vs. 74.18 for BQN and 71.86 for ALTER), and its higher accuracy comes from higher specificity. Because ACC, SEN and SPE depend on the 0.5 threshold and on the class balance (53 AD, 71 CN), Sec. 5.2 now also reports balanced accuracy (the mean of SEN and SPE): 70.43% for TopoScanPE vs. 69.46% for BQN, which we do not consider a meaningful difference. With the baselines re-trained on our splits (W4), TopoScanPE has the highest sensitivity (58.18 vs. 23.64 for BQN) and balanced accuracy (70.43 vs. 53.82). Choosing the threshold on the 12 validation participants per split instead of 0.5 is unstable and lowers balanced accuracy for every model, so we keep 0.5. Sec. 5.2 now states that on ADNI TopoScanPE is competitive with the strongest baselines, and that the clear improvement is on ABIDE.

Comments on the paper

1 comment

Revised PDF uploaded

We thank all reviewers for their careful reviews and constructive suggestions. We have uploaded a revised version of the paper, with all changes marked in blue. The main updates are:

  • Table 1: added Brain-JEPA, Contrasformer and BrainGFM as baselines.
  • Table 2 (ablation): added LapPE and RWSE, each used directly and passed through our signature encoder.
  • Table 3 (new): TopoScanPE added to other graph transformers (Graphormer, GraphTrans, SAN).
  • Sec. 4.2 and Appendix A: definitions of Forman–Ricci curvature and the heat kernel signature, and the handling of duplicate thresholds, padding and constant functions.
  • Sec. 5.2: revised ADNI claims and balanced accuracy.
  • Sec. 5.3 and Appendix F: selection of ρ on validation loss; Fig. 4 and the L sweeps (Appendix I) are now labelled as test performance.
  • Appendix C: alternative filtration functions.
  • Appendix D: analysis of the learned encoding (R², t-SNE and ROI importance, Figs. 5–6).
  • Appendix E: ADNI baselines re-trained on our splits, with paired DeLong tests.
  • Appendix G: computational cost.
  • Related work: brain-specific positional encodings (Brain-JEPA, Contrasformer, BrainGFM).

We hope the revision addresses your concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.