Pith. sign in

REVIEW 4 major objections 6 minor 2 references

Topotein: Topological Deep Learning for Protein Representation Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read TCPNet embeds proteins as residues, interactions, secondary structures, and whole proteins in one SE(3)-equivariant model and reports consistent gains over state-of-the-art geometric GNNs, strongest on fold classification.

desk verdict A real, novel topological representation for proteins, but the empirical case is overstated and rests on single runs. read the letter →

arxiv 2509.03885 v1 pith:J4O5OA6W submitted 2025-09-04 cs.LG

Zhiyu Wang , Arian Jamasb , Mustafa Hajij , Alex Morehead , Luke Braithwaite , Pietro Liò This is my paper · ORCID
classification cs.LG
keywords topologicaldeeplearningproteinrepresentationcombinatorialcomplexesSE(3)-equivariantmessagepassingsecondarystructurefoldclassificationgeometricgraphneuralnetworks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Protein structure is hierarchical, but most learned representations treat a protein as a flat graph of residue nodes. This paper argues that the missing level is secondary structure: the helices and strands whose spatial arrangement defines a protein's fold. It introduces the Protein Combinatorial Complex (PCC), a four-rank object holding residues, residue-residue interactions, secondary-structure elements, and the whole protein, and TCPNet, a network that passes scalar and vector messages across these ranks while remaining equivariant to rotations and translations. On four protein tasks, TCPNet is reported to beat standard geometric graph networks across the board, with the clearest wins on fold classification and the smallest accuracy loss when sequence information is withheld.

What carries the argument

The Protein Combinatorial Complex (PCC): a four-rank cell complex with residues as 0-cells, directed nearest-neighbor edges as 1-cells, DSSP-derived secondary-structure runs as 2-cells, and the protein as a 3-cell. The argument is carried by TCPNet, whose Topology-Complete Perceptron (TCP) module scalarizes vector features against edge-localized frames at every rank, and whose four-step hierarchical message passing—edge messages, SSE updates via outer-edge neighborhoods (which subtract self-edges so SSEs communicate through edges that cross between different SSEs), residue refinement, and global protein readout—moves information up and down the hierarchy while preserving SE(3)-equivariance.

What would settle it

Run TCPNet and its direct GNN counterpart on the same fold-classification benchmark with two versions of each structure: the original DSSP-derived secondary-structure assignments, and corrupted assignments in which a fraction of residues are relabeled or long SSE runs are split into shorter ones. If fold-classification accuracy does not drop materially under corruption, or TCPNet no longer beats the residue-only baseline, then the hierarchy built from SSE labels is not the source of the reported gains.

Watch

Extended reading notes

Core claim

The paper's central claim is that proteins are best represented not as flat residue graphs but as a four-rank combinatorial complex—residues (rank 0), directed residue-residue interaction edges (rank 1), secondary structure elements (rank 2), and the whole protein (rank 3)—and that neural networks can exchange messages across all four ranks while staying equivariant to rotations and translations. The authors define the Protein Combinatorial Complex (PCC), featurize each rank with scalar and vector geometric features, and design TCPNet, whose Topology-Complete Perceptron scalarizes vector features through edge-centered reference frames. A four-step message-passing scheme first computes edge m

Load-bearing premise

The method assumes DSSP secondary-structure labels are accurate, complete, and non-overlapping—each residue has exactly one label and only sequential runs of at least three residues with the same label become structural units—so noisy or fragmented assignments would corrupt the very edges the hierarchy relies on.

Editorial extensions

If this is right

  • Fold classification improves by about 3 percentage points on the hardest Fold split over the second-best model, suggesting that SSE-level message passing captures the structural motifs that define folds.
  • With sequence information removed, TCPNet's cellular-component and antibody-developability metrics drop only 0.6% and 2%, respectively, while graph baselines degrade by up to 10%, making pure-structure representations more practical.
  • Topological enhancement helps only when integrated deeply: TCPNet beats its direct GNN counterpart GCPNet everywhere, while a shallow SSE feature addition gives mixed or worse results.
  • Because the PCC carries scalar and vector features at every rank, the same encoder can be plugged into downstream tasks that need a single protein embedding or per-residue predictions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported structure-only robustness suggests a sequence-free inductive bias: a testable extension is to use TCPNet embeddings as pre-sequence filters or structural priors in inverse folding and de novo antibody design, which the paper only hints at in the conclusion.
  • Because 2-cells are built from sequential runs of identical DSSP labels with minimum length three, short, discontinuous, or overlapping structural elements are invisible to the hierarchy; corrupting or re-assigning SSE labels would reveal how much of the gain depends on exactly these assignments.
  • The outer-edge neighborhoods subtract self-edges; if real secondary-structure elements are occasionally non-contiguous or interleaved, this subtraction may drop biologically meaningful contacts, so an extension is to allow overlapping or probabilistic SSE membership and compare.
  • The paper's comparison places TCPNet above its direct GNN counterpart but below GVP-GNN on inverse folding; an inference is that topological structure helps most at scales above the local residue environment, so hybrid residue/topology pretraining objectives may combine both strengths.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces Topotein, a topological deep learning framework for protein representation learning. It defines a Protein Combinatorial Complex (PCC) that organizes residues, interactions, secondary structure elements (SSEs), and whole proteins as cells of ranks 0--3, and proposes the Topology-Complete Perceptron Network (TCPNet), an SE(3)-equivariant message-passing architecture over PCCs. The main claimed contributions are the hierarchical data structure, the TCPNet architecture, and an empirical evaluation on four tasks: inverse folding, fold classification, cellular component prediction, and antibody developability. The paper compares TCPNet with geometric GNN baselines and with two topological adaptations, GVP-TNN and ETNN, reporting tables of test-set metrics.

Significance. If the central claims were established, the paper would be a useful contribution to protein representation learning: it articulates a principled way to incorporate secondary-structure-level information into equivariant message passing, and it provides an open-source implementation. The hierarchical PCC construction is a sensible and flexible representation that is well motivated by the role of SSE arrangements in protein classification. The architectural design of TCPNet, including outer-edge neighborhoods and edge-centric scalarization, is technically sound as an equivariant mechanism. However, the paper's empirical evidence is not currently sufficient to support the claimed consistent state-of-the-art performance, for the concrete reasons detailed below.

major comments (4)
  1. [Abstract and Section 6.2, Table 1] The abstract claims TCPNet 'consistently outperforms state-of-the-art geometric graph neural networks,' and Section 6.2 claims it 'consistently outperforms its direct GGNN counterpart GCPNet across all tasks.' Table 1 contradicts both statements. GVP-GNN beats TCPNet on inverse folding (perplexity 5.280 vs 5.822; accuracy 0.474 vs 0.441). GCPNet beats TCPNet on structure+sequence cellular component prediction (F1max 0.408 vs 0.398) and on structure+sequence antibody developability (AUPRC 0.878 vs 0.874). The 'consistent' outperformance claim must be corrected, or the evaluation must be rerun and reported in a way that supports the claim.
  2. [Appendix E, Table 1] The entire empirical case rests on single runs. Appendix E states 'we were unable to perform extensive hyperparameter tuning, try various feature combinations, or run multiple trials for each experiment.' No error bars or seed-level results are reported. Several favorable margins are small in absolute terms (e.g., Fold accuracy 0.433 vs 0.384; structure-only cellular component F1max 0.392 vs 0.388) and could easily be within seed noise. This is load-bearing because the central claim is empirical superiority. The authors should provide multiple seeds with standard deviations or at least clearly label the results as preliminary single-run observations and soften the claims accordingly.
  3. [Section 6.1 and Appendix E (auxiliary 3Di task, time limits)] The design choices are described as based on 'preliminary experiments' on the same datasets: the auxiliary 3Di denoising task was added because 'our preliminary results show that this auxiliary task consistently improves all model's performance across our evaluation tasks,' and training time limits were 'carefully chosen through preliminary experiments to allow all models to reach convergence.' This is a selection-on-validation-protocol procedure applied to the same benchmark data, which can bias the comparison in favor of the proposed model. The paper should either report results under a pre-registered protocol or analyze sensitivity to these choices. At minimum, this should be acknowledged as a limitation in the main text, not only in the appendix.
  4. [Section 5.1, Eqs. (2)--(3)] The 2-cells are defined from DSSP labels: each residue has exactly one SSE label and only consecutive runs of at least three residues with the same label become 2-cells. The outer-edge neighborhoods in Eqs. (2)--(3) subtract B^{2->1} to remove self-edges; if DSSP assignments are noisy, or if biologically relevant SSEs are short, discontinuous, or overlapping, the construction will misrepresent the protein and the outer-edge subtraction may drop exactly the inter-SSE edges that matter. The manuscript provides no sensitivity analysis to the DSSP assignment quality, the minimum SSE length, or the k=16 edge cutoff. This is a concrete risk to the main methodological contribution and should be tested (e.g., by perturbing DSSP labels or varying the minimum SSE length).
minor comments (6)
  1. [Tables 1 and 2] The reported cellular component F1max for TCPNet structure+sequence differs between Table 1 (0.398) and Table 2 (0.397). This inconsistency should be resolved or explained.
  2. [Section 5.1] The choice of 16 nearest neighbors for the residue interaction edges is a free parameter, as is the minimum SSE size of three residues. No sensitivity analysis or ablation is provided; at least a sentence justifying these choices would help.
  3. [Eq. (14)] The notation in the GVP layer normalization LN(H) = (LN_s(H_s), H_v / sqrt(||H_v||^2_2 / |H|)) is ambiguous: the denominator should be clarified (what is |H|? the number of vector channels times 3?). This is important for reproducibility.
  4. [Appendix E] The auxiliary 3Di denoising description is terse: 'we randomly permute half of the residues' 3Di types while masking the other half as unknown' leaves unclear whether the model is trained to reconstruct the masked half, the permuted half, or both. Please specify the exact loss target.
  5. [General] Several typos and infelicities appear, e.g., 'vertice set' (Appendix A), 'GVP norms' (caption of Figure 8), and inconsistent use of 'GTNN' (Section 6.3) vs 'TNN.' These should be corrected in revision.
  6. [Section 5.4] Two readout strategies are described (pooling node embeddings vs. using the protein embedding), but no comparison or ablation is reported. Since the protein-level message passing is optional (Figure 8), the reader cannot tell which strategy produced Table 1.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TCPNet's results are measured on held-out benchmarks, not derived from fitted inputs; the internal Table 1 contradictions and single-run limitations are correctness concerns, not circularity.

full rationale

This is an empirical benchmark paper rather than a derivation chain. The methodological components—PCC construction (Section 5.1), the outer-edge neighborhoods (Eqs. 2–3), and the TCP message-passing equations (Eqs. 4–24)—are explicit algebraic definitions and architectural choices, not fitted parameters recycled as predictions. No constant is fitted to a subset of data and then reported as a prediction. The SSE 2-cells are constructed from DSSP labels, and the downstream labels (CATH, SCOP, GO, SabDab) are external; the model does not use the target labels to build its inputs. Self-citations to ProteinWorkshop (Jamasb et al. 2024) and GCPNet (Morehead and Cheng 2024) are framework/baseline citations, and no load-bearing uniqueness theorem or ansatz is imported from the authors' prior work to force the conclusion. The abstract's 'consistently outperforms' claim is contradicted by the paper's own Table 1 (e.g., GVP-GNN inverse-folding accuracy 0.474 vs. TCPNet 0.441; GCPNet structure+sequence cellular-component F1max 0.408 vs. TCPNet 0.398), and Appendix E explicitly states that no multiple trials were run: 'Due to computational constraints, we were unable to perform extensive hyperparameter tuning, try various feature combinations, or run multiple trials for each experiment.' These are serious validity/statistical concerns, but they are not circularity: the reported numbers are measured outcomes rather than consequences of the model's definitions. The auxiliary 3Di denoising task was chosen after preliminary results, which is a model-selection bias, but again it is applied uniformly across baselines and does not make the central comparison true by construction. No circular step can be exhibited from the paper's own equations or citations.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the availability and correctness of SSE labels, the chosen neighborhood sizes, and the training protocol. These are reasonable but not independently justified; the paper does not provide formal proof that the hierarchy is the optimal representation.

free parameters (5)
  • k-NN edge cutoff k = 16
    Residue interaction edges connect each residue to its 16 nearest neighbors (Section 5.1). This choice determines the graph topology and is not justified by an ablation.
  • minimum SSE cell size = 3 residues
    2-cells are only formed from at least three consecutive residues with the same SSE label (Section 5.1); shorter SSEs are absent from the hierarchy.
  • TCP bottleneck parameter lambda = not specified numerically
    Appears in Eq. 5 for vector dimension reduction; no value or sensitivity analysis is given.
  • training time limits = 4h / 4.5h / 10h / 2h per task
    Chosen through preliminary experiments to allow 'near convergence' (Appendix E); could favor models that converge quickly and affect comparisons.
  • auxiliary 3Di denoising mask ratio = 50% of residues permuted, 50% masked
    Set in Appendix E; no ablation is reported despite the claim that it 'consistently improves' all models.
assumptions (4)
  • domain assumption DSSP secondary structure labels are correct and each residue belongs to exactly one SSE.
    Section 5.1 builds 2-cells from 'sequentially consecutive 0-cells that have the same SSE label'; this requires reliable, non-overlapping SSE assignments.
  • domain assumption C-alpha backbone coordinates contain enough geometric information for hierarchical protein representation.
    All featurization uses C-alpha atoms only (Section 5.2); side-chain geometry is ignored.
  • domain assumption SE(3)-equivariance is the correct symmetry; reflection sensitivity is preserved.
    Section 5.3 states this; the edge frames use cross products, which are pseudo-vectors, but the paper claims SE(3) only.
  • standard math PCA eigenvector sign disambiguation yields a consistent global frame.
    Protein-level frames use PCA eigenvectors disambiguated by the farthest residue anchor (Section 5.3); degeneracies for symmetric proteins are not discussed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Topotein: Topological Deep Learning for Protein Representation Learning." pith.science (2026). https://pith.science/paper/J4O5OA6W

@misc{pith2026250903885,
  author       = {Pith},
  title        = {Pith review of: Topotein: Topological Deep Learning for Protein Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J4O5OA6W}},
  note         = {Machine review of arXiv:2509.03885}
}
read the original abstract

Protein representation learning (PRL) is crucial for understanding structure-function relationships, yet current sequence- and graph-based methods fail to capture the hierarchical organization inherent in protein structures. We introduce Topotein, a comprehensive framework that applies topological deep learning to PRL through the novel Protein Combinatorial Complex (PCC) and Topology-Complete Perceptron Network (TCPNet). Our PCC represents proteins at multiple hierarchical levels -- from residues to secondary structures to complete proteins -- while preserving geometric information at each level. TCPNet employs SE(3)-equivariant message passing across these hierarchical structures, enabling more effective capture of multi-scale structural patterns. Through extensive experiments on four PRL tasks, TCPNet consistently outperforms state-of-the-art geometric graph neural networks. Our approach demonstrates particular strength in tasks such as fold classification which require understanding of secondary structure arrangements, validating the importance of hierarchical topological features for protein analysis.

Figures

Figures reproduced from arXiv: 2509.03885 by the authors.

Figure 1
Figure 1. Overview of the Topotein framework. Given PDB protein structures (Berman et al. 2000), we construct Protein Combinatorial Complexes that hierarchically organize residues (rank-0), interactions (rank-1), secondary structures (rank-2), and complete proteins (rank-3). These multi-rank representations can be processed by any topological neural network (e.g. our TCPNet) to generate protein embeddings for downstream predi… view at source ↗
Figure 2
Figure 2. Message passing comparison across protein representation paradigms. The figure illustrates three approaches to protein representation, each showing five consecutive steps of passing information from one secondary structure to another. Note that there should also be edges between nodes and supernodes in the middle row, but they are omitted for clarity. Traditional GNNs represent proteins as graphs where nodes are res… view at source ↗
Figure 3
Figure 3. Domains of Topological Deep Learning. The three main topological domains: hypergraphs (left bottom) with hyperedges connecting multiple nodes; simplicial/cellular complexes (center top) with hierarchical structure and strict boundary requirements; combinatorial complexes (center bottom) unifying both approaches without boundary constraints. The right panel shows relationship types between cells of different ranks: p… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: An example of combinatorial complex. The visualization of this combinatorial complex (A), its constituent cells organized by rank (B), the incidence matrix showing relationships between rank 1 and rank 2 cells (C), and the adjacent matrix showing relationships between …
Figure 5
Figure 5. Figure 5: Topological Neural Networks vs Graph Neural Networks. TNNs generalize GNNs by enabling information flow between any ranks in a topological domain. While GNNs are limited to message passing between nodes (rank 0) and edges (rank 1), TNNs allow flexible multi-rank commun…
Figure 6
Figure 6. Figure 6: Protein Combinatorial Complex. Left: A PCC example showing amino acid nodes (black circles, 0-cell), interaction edges (black lines, 1-cells), secondary structures (blue arrows, 2-cells), and the protein itself (dashed box, 3-cell). Right: Inner edges (purple) directly…
Figure 7
Figure 7. Figure 7: Architecture of Topology-Complete Perceptron (TCP) module. The module processes scalar features hs, vector features hv, and rank-specific frames F (r) i through dual pathways: vector features are reduced via MLPs (Vs, Vd), scalarized using localized frames, concatenate…
Figure 8
Figure 8. Figure 8: Architecture of the Topology-Complete Perceptron Network (TCPNet). The network processes both scalar and vector features through parallel pathways from rank 0 to 3. The network starts with embedding the raw features of each rank with separate TCP modules. Then L intera…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 1 canonical work pages

  1. [2021]

    Resolving the Sign Ambiguity in the Singular Value Decomposition

    DOI: 10.48550/arXiv.2106.12575 (cit. on p. 7). Bro, R. et al. (Feb. 2008). “Resolving the Sign Ambiguity in the Singular Value Decomposition”. In: Journal of Chemometrics 22.2, pp. 135–140. ISSN : 0886-9383, 1099-128X. DOI: 10.1002/cem. 1122 (cit. on pp. 8, 10). Bronstein, Michael M. et al. (May 2, 2021). Geometric Deep Learning: Grids, Groups, Graphs, Ge...

    arXiv doi:10.48550/arxiv.2106.12575 2008

  2. [2025]

    This node-level multi-class classification task predicts amino acid sequences that would fold into given 3D protein structures

    containing 34,652 protein structures (32,652/1,000/1,000 train/validation/test split) filtered at 40% sequence similarity. This node-level multi-class classification task predicts amino acid sequences that would fold into given 3D protein structures. Each backbone position must be classified into one of 23 amino acid classes. The task is evaluated using p...

    work page 1995

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.