Objects and Entities
Computational biology에서 먼저 정해야 하는 것은 모델이 다루는 대상입니다. 같은 단어라도 protein, ligand, target, assay, structure가 어떤 단위로 정의되는지에 따라 split, leakage, evaluation이 달라집니다.
Entity note의 목적은 용어 정의가 아니라 example unit과 leakage boundary를 고정하는 것입니다. “protein”이라는 단어 하나만으로는 chain, isoform, construct, species, mutation, structure source가 숨겨질 수 있습니다.
Core Objects
| Object | Use when | Main risk |
|---|---|---|
| Target | target-conditioned activity, affinity, selectivity, assay context | target identity가 isoform, construct, mutation, species, assay source를 숨길 수 있음 |
| Protein | sequence, structure, representation, family split | homolog leakage와 residue-index mismatch |
| binding site, docking, pose prediction, structure-based generation | pocket이 ligand-defined이거나 inference에서 unavailable할 수 있음 | |
| Ligand | target-bound molecule, pose, activity, interaction | ligand state와 atom mapping이 protocol마다 달라질 수 있음 |
| Molecule | property prediction, molecular generation, similarity, scaffold split | salt, stereo, tautomer, protonation, duplicate handling |
| Protein-ligand complex | docking, pose quality, interaction prediction, complex graph | pair-level split이 protein 또는 ligand family를 leak할 수 있음 |
| Sequence | protein language model, genome region model, motif task | sequence context와 split unit이 biological claim과 맞지 않을 수 있음 |
| Structure | coordinate, conformer, protein structure, complex | template, frame, coordinate-source leakage |
| Genome | sequence/region/variant-level modeling only | broad omics claim은 명시적으로 열기 전까지 scope 밖 |
여러 object와 relation이 함께 중요하면 Entity relation map을 사용합니다.
Relation Map
| Relation | Meaning | Risk |
|---|---|---|
| protein-target | biological target identity and assay target | isoform/species/construct collapse |
| molecule-ligand | same chemical object with binding context | ligand state differs from parent molecule |
| protein-pocket | binding site or predicted site on structure | ligand-defined pocket leakage |
| ligand-pose | ligand state plus coordinates in a frame | atom mapping and conformer source ambiguity |
| target-assay-label | measured endpoint under assay context | unit, censoring, threshold, source mismatch |
| protein-ligand-complex | pair plus pose/contact/structure context | pair split can leak both axes |
Entity Tuple
Paper note에서는 modeled unit을 단일 단어가 아니라 tuple로 적습니다.
예를 들어 property prediction에서는 row가 molecule-only일 수 있지만, target-conditioned activity에는 molecule, target, assay, endpoint, unit, source가 필요합니다. Protein-ligand pose에는 protein structure, pocket rule, ligand state, atom mapping, pose coordinate, pose source가 필요합니다.
Split Implication
| Claim | Hold out |
|---|---|
| new molecule generalization | scaffold, chemical series, time split |
| new protein generalization | sequence identity, family, fold, target split |
| new protein-ligand pair generalization | ligand/protein 양쪽 axis 또는 explicit pair-level claim |
| new structure generalization | template-aware, homolog-aware, coordinate-source-aware split |
| new assay generalization | assay/source split 또는 endpoint harmonization check |
| new genome region generalization | chromosome, locus, family, annotation-source split |
Example Unit Questions
| If the row is… | Ask |
|---|---|
| molecule | is it parent, salt, mixture, tautomer, or standardized record? |
| activity label | what target, assay, endpoint, unit, censoring, source? |
| protein | which chain, isoform, construct, mutation, species? |
| structure | experimental, predicted, apo, holo, template-derived, minimized? |
| pose | what ligand state, atom mapping, coordinate frame, reference pose? |
Label Objects
Checks
- entity가 biological object, chemical object, assay record, derived feature 중 무엇인가?
- isoform, construct, mutation, chain, ligand state, assay context가 explicit한가?
- one row가 object, pair, pose, measurement, generated sample, candidate list 중 무엇인가?
- label이 있을 때 target, assay, endpoint, unit, threshold, censoring, source가 보존되는가?
- representation이 inference에서 unavailable한 ligand-defined pocket, known pose, template, homolog, future annotation을 쓰는가?
- split unit이 biological claim과 같은가?
- input에 deployment time에 unavailable한 정보가 포함되는가?