Topotein: Topological Protein Learning

• Topotein is a protein representation learning framework that models hierarchical structure via combinatorial complexes spanning residues, SSEs, and whole proteins.
• It integrates explicit geometric content with SE(3)-equivariant message passing using the Protein Combinatorial Complex and TCPNet.
• Empirical results demonstrate improved fold classification and robust performance in structure-only tasks compared to residue-centric methods.

Searching arXiv for the literal term “Topotein” and closely related protein representation learning work. Topotein is a topological deep learning framework for protein representation learning that treats hierarchical protein structure as an explicit learning domain rather than as an implicit by-product of residue-level message passing. In the usage established by “Topotein: Topological Deep Learning for Protein Representation Learning” (Wang et al., 4 Sep 2025), the framework consists of a Protein Combinatorial Complex (PCC), which represents proteins across residues, residue interactions, secondary structure elements, and whole proteins, and a Topology-Complete Perceptron Network (TCPNet), which performs SE(3)-equivariant message passing on that hierarchy. The motivating claim is that many sequence-based and graph-based protein encoders remain centered on residues, whereas proteins are naturally organized across multiple scales; Topotein is designed to preserve both this hierarchy and the 3D geometric information needed for structure-aware learning (Wang et al., 4 Sep 2025).

1. Definition and conceptual scope

Topotein is defined as a framework for protein representation learning (PRL) built on combinatorial complexes, with the specific aim of making residues, secondary structure elements (SSEs), and whole-protein organization jointly accessible to learning. The framework is introduced as a response to a representational mismatch: sequence transformers process amino acid strings, and many geometric graph neural networks model proteins as residue graphs, but neither representation explicitly encodes the hierarchical organization by which residues form SSEs and SSEs form folds (Wang et al., 4 Sep 2025).

The paper positions this issue as especially relevant for structurally organized tasks such as fold classification. It notes that protein classification systems such as CATH and SCOP are heavily based on secondary structure organization, and argues that residue-level graphs can overemphasize dense communication within an SSE while inadequately representing relations between neighboring SSEs. A heterogeneous graph with SSE supernodes partly addresses connectivity, but, as described in the paper, compresses geometric detail into coarse supernode features and simple center-of-mass relations. Topotein is therefore not presented as a minor extension of residue graphs, but as a change of learning domain from a graph to a multi-rank topological object (Wang et al., 4 Sep 2025).

A plausible implication is that Topotein is intended less as a universal replacement for all protein encoders than as a structural inductive bias for problems whose difficulty lies in multi-scale organization, especially where SSE arrangement is informative.

2. Protein Combinatorial Complex

The representational core of Topotein is the Protein Combinatorial Complex. The paper first recalls a general combinatorial complex as a triple $(S,\mathcal{X},\mathrm{rk})$, where $S$ is a finite vertex set, $$\mathcal{X}\subseteq \mathcal{P}(S)\setminus\{\emptyset\}$$ is a collection of cells, and $\mathrm{rk}:\mathcal{X}\to\mathbb{Z}_{\ge 0}$ is an order-preserving rank function satisfying

$$x \subseteq y \Longrightarrow \mathrm{rk}(x)\le \mathrm{rk}(y),$$

with all singletons assigned rank $0$ (Wang et al., 4 Sep 2025).

A protein combinatorial complex is then defined as a combinatorial complex $\mathcal{C}=(\mathit{S}, \mathcal{X}, \text{rk})$ whose vertex set contains all residues of a protein, whose cells are sets of residues, and whose rank function maps cells into $R=\{0,1,2,3\}$. The four ranks are fixed:

• rank 0: residues
• rank 1: residue interaction edges
• rank 2: secondary structure elements
• rank 3: the whole protein (Wang et al., 4 Sep 2025)

Rank-1 cells are constructed by connecting each residue to its 16 nearest neighbors. The framework uses directed pairwise edges as 1-cells rather than undirected edges. For a directed edge $(i,j)$, only node $i$ can pass a message to edge $S$0 via $S$1, and edge $S$2 can only pass a message to node $S$3 via $S$4. The paper states that this directionality is introduced to support construction of SO(3)-equivariant edge frames (Wang et al., 4 Sep 2025).

Rank-2 cells are SSEs, constructed by grouping sequentially consecutive residues that share the same SSE label, with a minimum size of three residues. SSE labels are obtained from DSSP, with a 3-way encoding into helix, strand, and coil. Because each residue belongs to at most one SSE, rank-2 cells do not overlap. Rank 3 is a single cell corresponding to the entire protein (Wang et al., 4 Sep 2025).

The PCC also defines incidence and adjacency operators across ranks. The notation includes $S$5 for incidence matrices, $S$6, and $S$7, so the structure supports both direct cross-rank propagation and same-rank propagation mediated by another rank. This is the formal basis on which TCPNet performs hierarchical message passing (Wang et al., 4 Sep 2025).

3. Hierarchical neighborhoods and geometric features

A distinctive feature of PCC is that it augments hierarchy with explicit geometric content at every rank. Because rank-2 cells do not overlap, the paper introduces outer-edge neighborhoods to allow communication between SSEs through residue-contact edges rather than by direct overlap. These neighborhoods are defined as

$S$8

and

$S$9

The stated interpretation is that $$\mathcal{X}\subseteq \mathcal{P}(S)\setminus\{\emptyset\}$$0 maps an SSE to edges that originate inside that SSE and terminate in a different SSE, while $$\mathcal{X}\subseteq \mathcal{P}(S)\setminus\{\emptyset\}$$1 captures the reverse direction (Wang et al., 4 Sep 2025).

This design is meant to avoid a coarse supernode abstraction. Instead of collapsing an inter-SSE relationship into a single edge between two SSE centers, the representation retains the actual residue-level contacts mediating that relationship. The paper repeatedly presents this as central to preserving both fine geometry and part-whole organization (Wang et al., 4 Sep 2025).

The feature design is rank-specific. Rank-0 residue features include scalar channels such as 23-d amino acid one-hot, 21-d 3Di one-hot, 16-d positional encoding, virtual bond and torsion angles $$\mathcal{X}\subseteq \mathcal{P}(S)\setminus\{\emptyset\}$$2, and backbone dihedrals $$\mathcal{X}\subseteq \mathcal{P}(S)\setminus\{\emptyset\}$$3 encoded with sine and cosine; vector channels include displacement vectors to neighboring residues and tetrahedral geometry features. Rank-1 edge features include scalar Euclidean distance and positional encoding of distance, together with the displacement vector between connected residues. Rank-2 SSE features include 3-d SSE type one-hot, SSE size, 20-d positional encoding for start/end residues, consecutive angles between neighboring SSEs, plane torsional angle, eigenvalues, and five eigenvalue-derived shape descriptors: linearity, planarity, scattering, omnivariance, and anisotropy. Rank-3 protein features include protein size, amino acid composition frequencies, SSE type frequencies, SSE size mean and standard deviation, protein eigenvalues, eight shape descriptors from eigenvalues, radius of gyration, contact density measures, three global eigenvectors, and displacement vectors to the ten farthest and ten nearest residues from the protein center of mass (Wang et al., 4 Sep 2025).

A plausible implication is that the PCC is not merely a topological scaffold; it is a deliberately hybrid topological-geometric representation, with topology providing the hierarchy and geometry providing the equivariant signal.

4. Topology-Complete Perceptron Network

The neural component of Topotein is TCPNet, described as an SE(3)-equivariant topological neural network operating on PCCs. Its basic module is the Topology-Complete Perceptron (TCP), a topological generalization of the Geometry-Complete Perceptron. The module consumes scalar features $$\mathcal{X}\subseteq \mathcal{P}(S)\setminus\{\emptyset\}$$4, vector features $$\mathcal{X}\subseteq \mathcal{P}(S)\setminus\{\emptyset\}$$5, and rank-specific localized frames $$\mathcal{X}\subseteq \mathcal{P}(S)\setminus\{\emptyset\}$$6 (Wang et al., 4 Sep 2025).

The TCP module reduces vector channels via

$$\mathcal{X}\subseteq \mathcal{P}(S)\setminus\{\emptyset\}$$7

and

$$\mathcal{X}\subseteq \mathcal{P}(S)\setminus\{\emptyset\}$$8

then combines them with scalar channels through scalarization,

$$\mathcal{X}\subseteq \mathcal{P}(S)\setminus\{\emptyset\}$$9

and produces outputs

$\mathrm{rk}:\mathcal{X}\to\mathbb{Z}_{\ge 0}$0

$\mathrm{rk}:\mathcal{X}\to\mathbb{Z}_{\ge 0}$1

$\mathrm{rk}:\mathcal{X}\to\mathbb{Z}_{\ge 0}$2

Thus scalar and vector channels are kept separate, but scalar channels gate vector outputs (Wang et al., 4 Sep 2025).

The architecture uses edge-centric scalarization as its preferred frame construction. For an edge $\mathrm{rk}:\mathcal{X}\to\mathbb{Z}_{\ge 0}$3, the frame is

$\mathrm{rk}:\mathcal{X}\to\mathbb{Z}_{\ge 0}$4

and edge scalarization is

$\mathrm{rk}:\mathcal{X}\to\mathbb{Z}_{\ge 0}$5

For nodes, scalarization averages projections over incident edge frames, and for SSEs it averages over outer-edge neighborhoods. For the protein level, the paper uses a PCA-based frame

$\mathrm{rk}:\mathcal{X}\to\mathbb{Z}_{\ge 0}$6

with sign ambiguity resolved using the farthest residue from the protein center as an anchor (Wang et al., 4 Sep 2025).

The paper emphasizes that TCPNet should be SE(3)-equivariant, so global translations and rotations of a protein do not alter predictions except in the appropriate equivariant manner for vector channels. It also explicitly states that reflection should not be treated as a symmetry, since proteins are chiral (Wang et al., 4 Sep 2025).

5. Message passing, training configuration, and evaluation tasks

TCPNet is organized as an embedding module followed by $\mathrm{rk}:\mathcal{X}\to\mathbb{Z}_{\ge 0}$7 interaction layers; the reported experiments use 6 layers. Raw features at each rank are normalized by a GVP-style layer normalization and separately embedded:

$\mathrm{rk}:\mathcal{X}\to\mathbb{Z}_{\ge 0}$8

The interaction scheme is described in four stages: edge-level message computation, SSE-level integration, residue-level refinement, and protein-level representation formation (Wang et al., 4 Sep 2025).

At the edge level, each interaction edge aggregates source residue features, target residue features, edge features, and SSE context:

$\mathrm{rk}:\mathcal{X}\to\mathbb{Z}_{\ge 0}$9

At the SSE level, each SSE integrates its previous representation, member residues, internal edges, and messages from outer edges:

$$x \subseteq y \Longrightarrow \mathrm{rk}(x)\le \mathrm{rk}(y),$$0

Residues are then refined using local contact information and parent-SSE messages:

$$x \subseteq y \Longrightarrow \mathrm{rk}(x)\le \mathrm{rk}(y),$$1

and protein-level aggregation combines updated residue-level, SSE-level, and protein-level channels:

$$x \subseteq y \Longrightarrow \mathrm{rk}(x)\le \mathrm{rk}(y),$$2

Rank-specific representations are updated residually with layer normalization (Wang et al., 4 Sep 2025).

The framework is evaluated on four tasks spanning different structural levels: inverse folding on CATH 4.4, fold classification on SCOP 1.75, cellular component prediction on the GO dataset from Gligorijevic et al., and antibody developability on SabDab. The reported setup uses Adam, learning rate $$x \subseteq y \Longrightarrow \mathrm{rk}(x)\le \mathrm{rk}(y),$$3, ReduceOnPlateau, early stopping, A100 GPUs, a 3-layer MLP decoder with 512 hidden size, and SiLU activations in the backbone. An auxiliary 3Di sequence denoising task is added for graph-level tasks by permuting half of the residues’ 3Di labels and masking the other half, then training a two-layer MLP to reconstruct original 3Di types (Wang et al., 4 Sep 2025).

The paper explicitly notes that, due to computational constraints, it did not run extensive hyperparameter tuning or multiple trials. This caveat is important for interpreting the reported margins (Wang et al., 4 Sep 2025).

6. Empirical performance, strengths, and limitations

The main empirical claim is that TCPNet consistently improves over GCPNet and shows its clearest advantage on fold classification, which the paper presents as the flagship test of hierarchical structural reasoning. On inverse folding, TCPNet reports perplexity $$x \subseteq y \Longrightarrow \mathrm{rk}(x)\le \mathrm{rk}(y),$$4 and accuracy $$x \subseteq y \Longrightarrow \mathrm{rk}(x)\le \mathrm{rk}(y),$$5, improving on GCPNet’s $$x \subseteq y \Longrightarrow \mathrm{rk}(x)\le \mathrm{rk}(y),$$6 and $$x \subseteq y \Longrightarrow \mathrm{rk}(x)\le \mathrm{rk}(y),$$7, though GVP-GNN remains best with $$x \subseteq y \Longrightarrow \mathrm{rk}(x)\le \mathrm{rk}(y),$$8 and $$x \subseteq y \Longrightarrow \mathrm{rk}(x)\le \mathrm{rk}(y),$$9. On fold classification, TCPNet is best on all three splits: Fold $0$0, Superfamily $0$1, and Family $0$2, compared with GCPNet’s $0$3, GVP-GNN’s $0$4, and EGNN’s $0$5 (Wang et al., 4 Sep 2025).

On cellular component prediction, TCPNet attains $0$6 in the structure-only / structure+sequence settings, compared with GCPNet’s $0$7, GVP-GNN’s $0$8, and EGNN’s $0$9. On antibody developability, TCPNet attains AUPRC $\mathcal{C}=(\mathit{S}, \mathcal{X}, \text{rk})$0, compared with GCPNet’s $\mathcal{C}=(\mathit{S}, \mathcal{X}, \text{rk})$1, GVP-GNN’s $\mathcal{C}=(\mathit{S}, \mathcal{X}, \text{rk})$2, and EGNN’s $\mathcal{C}=(\mathit{S}, \mathcal{X}, \text{rk})$3. The paper therefore characterizes TCPNet as especially strong in structure-only settings, while not universally best when sequence is also included (Wang et al., 4 Sep 2025).

The comparison with other topological variants is also used as a design argument. GVP-TNN, which adds SSE-level message passing to GVP-GNN, shows mixed results; ETNN, adapted from prior work on combinatorial complexes for small molecules, often underperforms EGNN. This is presented as evidence that topological enhancement must be deeply integrated into the architecture rather than superficially appended (Wang et al., 4 Sep 2025).

The paper’s stated strengths include a clear representational innovation, preservation of hierarchy with geometry, multi-rank equivariant learning, strong fold-classification performance, robust structure-only results, and informative negative controls. Its stated limitations include reliance on DSSP secondary-structure annotations, limited theoretical formalization beyond architectural rationale, moderate gains on several tasks, the fact that inverse folding is not state-of-the-art, likely computational complexity from multi-rank scalar/vector processing, limited tuning, and some notation or rendering issues in the equations (Wang et al., 4 Sep 2025).

A plausible implication is that Topotein is most compelling where the biological signal is distributed across secondary-structure organization and fold-level architecture, and less decisive when the target depends primarily on highly local residue environments.

7. Relation to other uses of similar terminology

The term Topotein has appeared ambiguously in broader discussion, but the literal method name in the protein-learning context is the framework introduced in (Wang et al., 4 Sep 2025). This should be distinguished from unrelated terms and methods with similar orthography. For example, TopoTxR is a topology-derived biomarker and topology-guided attention framework for predicting treatment response from breast DCE-MRI, built from persistent homology and 3D CNNs rather than protein representation learning (Wang et al., 2021). Other similarly prefixed terms, such as TopoTEM or toponium, belong to electron microscopy and high-energy physics respectively, and are conceptually unrelated to protein combinatorial complexes and TCPNet.

Within protein modeling, a nearby but distinct line of work is represented by TopoBind, which combines ESM-2 sequence embeddings with handcrafted structural and topological descriptors, including persistent homology, for antibody–antigen binding free-energy prediction (Yu et al., 27 Aug 2025). TopoBind demonstrates that topology-aware engineered features can complement sequence representations, but it does not introduce a multi-rank topological learning domain comparable to PCC, nor an SE(3)-equivariant topological network comparable to TCPNet. This suggests a useful conceptual distinction: TopoBind is a multimodal feature-engineering framework, whereas Topotein is a topological representation-learning framework centered on a learned hierarchical domain (Yu et al., 27 Aug 2025).

In that sense, Topotein occupies a specific place in the emerging literature: it applies topological deep learning to proteins not by summarizing topology into features, but by making protein hierarchy itself the substrate of equivariant message passing (Wang et al., 4 Sep 2025).