Protein Representation

A protein representation is the form of a protein used by a model: sequence tokens, embeddings, residue graphs, contact maps, coordinates, surfaces, or mixed sequence-structure features.

For a sequence $s_{1:L}$, a representation model maps residues to hidden states:

$$H=f_{\theta }(s_{1:L}),H\in {\mathbb{R}}^{L\times d}$$

A protein-level embedding can be formed by masked pooling:

$$h_{\mathrm{protein}}=\frac{{\sum }_{i=1}^L{m}_i{h}_i}{{\sum }_{i=1}^L{m}_i}$$

where $m_i$ is an attention mask that excludes padding and special tokens.

Sequence Representation

A sequence-only model treats residues as tokens:

$$s_{1:L}\to e_{1:L}\to h_{1:L}$$

where $e_i$ is a residue embedding and $h_i$ is a contextual representation. This is the common setup for protein language models. It captures sequence statistics and evolutionary signals indirectly, but it does not explicitly encode 3D coordinates.

Structure Representation

A structure representation usually contains residue or atom coordinates:

$$X=[x_1,\dots ,x_N],x_i\in {\mathbb{R}}^3$$

If a model consumes coordinates, the transformation behavior must be clear. Scalar predictions such as binding labels should be invariant to rigid transforms:

$$f(RX+t)=f(X)$$

Vector or coordinate outputs should usually be equivariant:

$$f(RX+t)=Rf(X)+t$$

where $R$ is a rotation and $t$ is a translation.

Graph Representation

A residue graph can be written as:

$$G=(V,E),E=\{(i,j):\Vert x_i-x_j{\Vert }_2<\tau \}$$

Edges may come from sequence adjacency, spatial cutoff, k-nearest neighbors, contact maps, or learned attention. The graph policy changes what the model can see.

Fusion Risk

Sequence-structure fusion aligns sequence embeddings with structure residues:

$$(h_i,x_i)\text{for residue }i$$

This alignment is fragile. Missing density, insertion codes, chain breaks, alternate residue numbering, and non-standard residues can silently pair the wrong embedding with the wrong coordinate.

Common Views

• Residue sequence for language-model pretraining.
• MSA profile for evolutionary coupling.
• Residue graph or contact map for structural relationships.
• 3D coordinates for geometry-aware modeling.
• Pocket representation for structure-based interaction tasks.
• Sequence-structure fused representation using aligned embeddings and coordinates.

Checks

• What does one token or node represent?
• Are padding, truncation, chain breaks, and non-standard residues handled?
• Is the representation invariant, equivariant, or frame-dependent?
• Is sequence-structure alignment verified before fusing embeddings with coordinates?
• Is residue indexing preserved after structure cleaning and feature caching?
• Does the representation include full protein, domain, pocket, chain, residue, or atom context?
• Are train-time and inference-time coordinate sources matched, for example crystal structures vs predicted structures?
• Does batching preserve per-protein isolation for pooling, normalization, and attention?

Related

• Residue indexing
• Protein structure cleaning
• Multiple sequence alignment
• Contact map
• Sequence-structure alignment
• Pocket representation
• Equivariance
• Graph construction
• Tokenization
• Protein modeling