Protein

Proteins are biological macromolecules represented as sequences, structures, surfaces, graphs, or dynamics depending on the task.

A protein representation can be sequence-only, structure-only, or fused:

$$r_P=\varphi (s_{1:L},X,c)$$

where $s_{1:L}$ is the residue sequence, $X$ is optional structure or coordinate information, and $c$ is context such as domain, family, pocket, mutation, construct, or assay.

Modeling Views

• Sequence view for language-model and self-supervised learning methods.
• Structure view for folding, docking, and equivariant models.
• Functional view for binding sites, domains, and interactions.
• Target view for assays, docking benchmarks, and virtual screening.
• Dataset view for homolog grouping, sequence identity clustering, and protein-family splits.

Identity Boundary

A protein record is not only a sequence string. Public notes should distinguish:

$$P=(\text{accession},\text{isoform},\text{construct},\text{chain},\text{mutation state},\text{sequence},\text{structure source})$$

The same biological target can appear as a full-length sequence, a domain construct, a mutant, a crystal chain, a predicted model, or an assay-specific construct. These should not be silently collapsed unless the task explicitly allows it.

Structure Availability

For modeling, the most important question is whether structure is available at deployment time:

$$\text{input availability}\in \{\text{sequence only},\text{predicted structure},\text{experimental apo},\text{experimental holo},\text{known complex}\}$$

Using a holo structure or ligand-defined pocket during training can be valid for analysis, but it changes the task if the deployment setting only has sequence or apo structure.

Split and Leakage

Protein-side generalization usually needs sequence or family grouping:

$$\mathrm{cluster}(P_i)\ne \mathrm{cluster}(P_j)\text{for train/test separation}$$

where the cluster can be based on sequence identity, domain family, fold, or target class. For structure-based tasks, also record whether homolog templates or known complexes were available to the model pipeline.

Checks

• Is the task sequence-only, structure-only, or sequence-structure combined?
• Are homologs, protein families, and near-duplicate chains separated across splits?
• Are domains, missing residues, mutations, and cofactors represented explicitly?
• Does the model predict structure, function, interaction, or a downstream assay label?
• Is the split grouped by protein family, sequence identity, target, or assay?
• Is the structure experimental, predicted, modeled, or unavailable at deployment time?
• Does the representation include information that would not be available for the intended task?
• Are isoform, construct, mutation, and chain choices recorded?
• Are sequence embeddings aligned to the same residue indexing as structure coordinates?
• Are templates, MSAs, and bound ligands treated as part of the evidence boundary?

Related

• Entity relation map
• Pocket
• Assay
• Protein representation
• Sequence identity clustering
• Protein domain
• Multiple sequence alignment
• Sequence-structure alignment
• Protein structure prediction
• Protein family split
• Protein modeling
• Protein-ligand complex