Target

A target is the biological object a model or experiment is centered on. In this wiki, a target is usually a protein, pocket, sequence, or genomic region rather than a broad disease program.

A task can be written as:

$$\hat{y}=f_{\theta }(x_{\mathrm{molecule}},x_{\mathrm{target}},c)$$

where $x_{\mathrm{molecule}}$ is a molecule or ligand representation, $x_{\mathrm{target}}$ is target context, and $c$ is assay or experimental context.

When labels are measured, the target should be interpreted through the Target-assay-label contract rather than as a standalone identifier.

Target Context

• Protein sequence.
• Protein family or domain.
• 3D structure or predicted structure.
• Binding pocket.
• Species, isoform, mutation, or construct when public and relevant.
• Assay context when labels come from experiments.

Target Identity Contract

Field	Question	Why It Matters
public identifier	which public target entry is used?	reproducibility
sequence	which amino acid or genomic sequence?	homolog split and mutation effects
family/domain	which family or domain grouping?	transfer claim boundary
structure	apo, holo, predicted, experimental, relaxed?	docking and structure claims
pocket	global target or local binding site?	interaction and pose context
species/isoform	which public biological variant?	label comparability
assay context	which measurement setup?	target label is not context-free

For target-conditioned prediction:

$$p_{\theta }(y\mid m,t,a)$$

is not the same task as molecule-only prediction:

$$p_{\theta }(y\mid m)$$

If a paper describes target-aware modeling, the target representation and target split must be explicit.

Target Split Boundary

Claim	Required Split
new molecules for known targets	molecule/scaffold split
new targets in known family	target-held-out split with family analysis
new protein families	protein-family split
new pockets or structures	pocket/structure-source split
new target-assay settings	target plus assay/source split

Target-level generalization cannot be proven by a row-random split.

Structure-Based Context

When a target is used as a structure:

$$t\to (s,X,P)$$

where $s$ is sequence, $X$ is coordinates, and $P$ is pocket definition. The note should say whether $X$ is experimental, predicted, template-derived, apo, holo, or docked.

Checks

• Is the target a protein, pocket, family, sequence region, or assay-defined object?
• Are target identifiers public and reproducible?
• Are isoform, mutation, and construct differences treated explicitly?
• Are target-level splits separated from molecule-level splits?
• Is the target tied to an assay context, label unit, and split rule?
• Is private collaborator or project context removed?
• Is the target representation available at inference time?
• Are target, pocket, and protein-ligand-complex claims kept separate?

Related

• Protein
• Pocket
• Assay
• Target-assay-label contract
• Protein-ligand interaction
• Protein family split
• Protein-ligand split
• Proteins