Protein Modeling Concepts
Protein modeling concept는 sequence, evolutionary signal, structure, geometric constraint가 model input과 output으로 바뀌는 과정을 설명합니다.
Protein note의 기본 단위는 항상 먼저 고정합니다.
단위가 바뀌면 representation, split, metric, leakage risk가 같이 바뀝니다.
Protein Modeling Contract
Protein modeling note는 sequence, structure, context, target을 분리해서 적어야 합니다.
| Part | Meaning | Typical question |
|---|---|---|
| biological unit | residue, domain, chain, family, pocket, complex, assay row? | |
| sequence | isoform, mutation, construct, truncation, residue indexing? | |
| structure or coordinates | experimental, predicted, apo, holo, complex, template-derived? | |
| context | MSA, template, pocket, ligand, species, assay, family? | |
| representation | token embedding, pooled chain embedding, residue graph, pocket graph? | |
| target | residue label, function, contact, coordinate, binding, affinity, design? | |
| metric | likelihood, accuracy, contact precision, RMSD, ranking, calibration? | |
| split boundary | sequence identity, protein family, fold, target, complex, time? |
This contract prevents a common error:
Route Map
| Need | Start | Risk |
|---|---|---|
| choose a protein representation | Protein representation | pooling, tokenization, MSA/template leakage |
| understand sequence-only pretraining | Protein language model, Multiple sequence alignment | likelihood confused with fitness |
| define protein units and groups | Protein domain, Sequence identity clustering | homologs crossing train/test |
| connect sequence to structure | Sequence-structure alignment, Residue indexing | residue mismatches, missing residues, insertion codes |
| prepare structures | Protein structure cleaning | chain selection, alternate locations, unresolved regions |
| model structure | Protein structure prediction, Contact map | benchmark leakage and coordinate-frame assumptions |
| model binding context | Binding site, Pocket representation | ligand-defined pockets and apo/holo mismatch |
Task Boundary
| Task Type | Object Unit | Evidence Boundary |
|---|---|---|
| sequence representation | residue, domain, chain | downstream split, pooling rule, homolog control |
| structure prediction | chain, domain, complex | template/MSA policy, coordinate target, confidence calibration |
| functional prediction | residue, domain, chain, assay row | label semantics, source, family split |
| binding-site modeling | pocket, residue set, local structure | site definition, apo/holo state, deployability |
| protein-ligand modeling | protein, ligand, pocket, complex | protein-family split plus ligand scaffold or complex split |
| design/generation | sequence, backbone, pocket, complex | validity, novelty, diversity, downstream experimental gap |
Representation Contract
For any protein note, state what the model actually sees:
where is the residue sequence, is optional structure or coordinates, and is context such as MSA, template, pocket, species, assay, ligand, or family.
| Input Source | Common Use | Main Public Claim Risk |
|---|---|---|
| raw sequence | protein language models, classification, variant scoring | homolog leakage and family memorization |
| MSA | coevolution, structure prediction | database date and template leakage |
| experimental structure | geometric modeling, pockets, docking | unavailable-at-deployment information |
| predicted structure | structure-aware downstream model | circular evidence and model-bias transfer |
| pocket or ligand context | binding, docking, interaction prediction | ligand-defined pocket leakage |
Claim Separation
Protein papers often move between sequence, structure, function, and binding. Keep the claim at the evidence level.
| Claim | Evidence needed | Does not prove alone |
|---|---|---|
| sequence modeling | held-out sequence likelihood, family split, downstream probe | structure accuracy or biological function |
| representation transfer | fixed representation, downstream task, split, baseline | mechanistic understanding |
| structure prediction | coordinates, contact/distance, template policy, confidence | ligand-ready pocket geometry |
| binding-site prediction | pocket definition, apo/holo policy, localization metric | ligand pose or affinity |
| protein-ligand modeling | pair representation, complex split, ligand scaffold control | protein-only or ligand-only generalization |
| design or generation | validity, novelty, diversity, downstream assay or structural evidence | expression, stability, activity, selectivity |
When a method uses MSA, templates, predicted structures, or ligand context, state whether those inputs are available at deployment time. Otherwise the benchmark may be easier than the intended use case.
Evidence Strength
Do not read all protein metrics as the same type of evidence.
| Evidence | Supports | Does Not Prove Alone |
|---|---|---|
| sequence likelihood | sequence distribution or representation quality | binding, function, expression, foldability |
| contact/distance accuracy | coarse structural constraints | ligand-ready all-atom pocket geometry |
| backbone accuracy | fold-level structure quality | side-chain placement or docking quality |
| pocket localization | site-finding under a stated definition | affinity or pose scoring |
| downstream benchmark score | task-specific performance | broad biological generalization |
Split and Leakage Controls
| Control | Why |
|---|---|
| sequence identity threshold | prevents near-identical proteins from crossing train/test |
| family or domain split | tests transfer across homologous groups |
| template date and database policy | controls structure-prediction leakage |
| MSA source and depth | changes available evolutionary signal |
| chain and construct policy | prevents silent task changes from truncations or isoforms |
| complex-pair split | prevents memorizing a protein-ligand or protein-partner pair |
| pocket definition | ligand-defined pockets may use unavailable test information |
| assay/source split | controls label-source shortcuts in functional or binding tasks |
Core Concepts
| Group | Notes |
|---|---|
| Representation | Protein representation, Protein language model, Multiple sequence alignment |
| Structure | Protein structure cleaning, Protein structure prediction, Contact map |
| Indexing | Residue indexing, Sequence-structure alignment |
| Binding | Binding site, Pocket representation |
Checks
- What is the example unit: residue, domain, chain, family, pocket, complex, or assay row?
- Is the model sequence-only, structure-only, or sequence-structure fused?
- What is available at inference time: MSA, template, predicted structure, experimental structure, ligand, or pocket?
- Is sequence likelihood being treated as evidence for function, stability, or binding without downstream validation?
- Is the pocket representation ligand-defined, predicted, or deployable?
- Are residue indices aligned across sequence tokens, structure residues, and coordinates?
- Are homologs and near-duplicate proteins separated across splits?
- Are missing residues, insertion codes, non-standard residues, and chain IDs handled explicitly?
- Does the evaluation measure geometry, function, interaction, or downstream task transfer?
- Is the claim narrowed to the evidence type: sequence, structure, pocket, binding, or function?