Representation Routes

Computational Biology에서는 “무슨 모델을 쓸까”보다 “무엇을 어떤 representation으로 바꿀까”가 먼저입니다. 같은 protein-ligand 문제라도 sequence, graph, conformer, pocket, complex graph, coordinate representation은 서로 다른 claim을 만듭니다.

$$r:\text{biological or chemical object}\to \text{model input}$$

Representation route는 model choice보다 앞섭니다. 모델이 아무리 강해도 inference time에 없는 정보, split을 새게 만드는 정보, label을 암시하는 preprocessing을 입력에 넣으면 claim이 무너집니다.

Route Map

Route	Input unit	Good for	Watch
Sequence	amino acid, nucleotide, SMILES token	language modeling, motif, protein family, sequence-only prediction	homolog, duplicate, tokenization leakage
Fingerprint	molecule identity or substructure bits	fast property baseline, retrieval, similarity	stereochemistry, tautomer, protonation loss
Molecular graph	atom and bond graph	property, interaction, generation	standardization, aromaticity, charge policy
Conformer	3D ligand geometry	docking, shape, strain, conformer-sensitive property	conformer generator and energy policy
Pocket	binding-site residues or grid	target-conditioned scoring, docking, retrieval	ligand-defined pocket leakage
Complex graph	protein-ligand pair with edges	interaction, pose, affinity, message passing	pair split and edge construction leakage
Coordinates	atom/residue positions	pose, structure, force, geometry-aware generation	frame, units, alignment, equivariance

Route to Architecture

Representation	Typical architecture	Key inductive bias
Sequence tokens	Transformer, state-space model, RNN	order, context, motif
Molecular graph	GNN, Graph Transformer	locality, bond topology, permutation equivariance
3D coordinates	equivariant GNN, SE(3) Transformer, diffusion/flow	rotation/translation symmetry
Fingerprint	tree model, MLP, retrieval	fast sparse substructure signal
Pocket grid	CNN, voxel model, 3D U-Net	spatial locality on grid
Complex graph	heterogeneous GNN, cross-attention	pair interaction and relation edges

Choosing a Route

Question	Prefer
molecule-only property인가?	fingerprint, molecular graph, conformer
target-conditioned activity인가?	ligand representation + target/pocket context
sequence family generalization인가?	protein sequence with family-aware split
pose quality가 핵심인가?	conformer, pocket, complex coordinates
geometry validity를 주장하는가?	coordinate representation with explicit geometry checks
fast screening baseline이 필요한가?	fingerprint/tree model, similarity, docking score

Representation Contract

각 representation note에는 아래를 남깁니다.

Field	Write
Object	molecule, protein, ligand, pocket, complex, sequence, conformer
Preprocessing	standardization, filtering, missing atoms/residues, protonation, tautomer
Source	experimental, predicted, generated, docked, minimized, database-derived
Axes	token, atom, residue, node, edge, coordinate, candidate
Availability	inference time에 사용할 수 있는 정보인지
Leakage risk	duplicate, homolog, scaffold, ligand-defined pocket, template, assay source

Split Coupling

Representation과 split은 따로 정할 수 없습니다.

Representation	Split risk
Sequence	homolog, duplicate, family leakage
Molecule graph/fingerprint	scaffold, analog series, assay source leakage
Protein-ligand complex	same target, same ligand, same complex family leakage
Pocket/structure	ligand-defined pocket, template, predicted-vs-experimental mismatch
Conformer/pose	generated pose or minimized structure leaking label proxy

Related

• Computational Biology
• Objects and Entities
• Data and Evaluation
• Protein-ligand representation contract
• Model Reading Map