Target-Assay-Label Contract

A target-assay-label contract states what biological object, measurement process, and label semantics define one supervised chem-bio example. It prevents a dataset row from being treated as a generic molecule label when it is actually target- and assay-conditioned.

The core relation is:

$$y_i=h(m_i,t_i,a_i,c_i)$$

where $m_i$ is a Molecule or Ligand, $t_i$ is the Target, $a_i$ is the Assay, and $c_i$ is measurement context such as endpoint, unit, threshold, dose, species, construct, time, source, and preprocessing.

Contract Fields

Field	Meaning
Molecule or ligand	Standardized chemical identity and representation
Target	Protein, pocket, family, sequence region, or other scoped biological object
Assay	Measurement protocol, readout, endpoint, source, and context
Label	Value, unit, transformation, censoring, threshold, and missing-label policy
Example unit	What one row or complex represents
Split unit	Molecule group, protein family, assay/source group, complex pair, or time
Evaluation	Task, metric, aggregation, and failure mode

Why It Matters

• The same molecule can have different labels for different targets.
• The same molecule-target pair can have incompatible labels across assay protocols.
• A binary active/inactive label can hide endpoint, threshold, unit, and censoring.
• Missing activity is not the same as measured inactivity.
• A split can look clean by row id while leaking molecule, target, assay, or source context.

Split Implication

A target-conditioned task should not be evaluated only as a molecule split unless the claim is molecule interpolation for known targets:

$$s_i=g(m_i,t_i,a_i,{\tau }_i)$$

where $s_i$ is the split assignment and ${\tau }_i$ is optional time or source context. The grouping keys should match the intended generalization claim.

Checks

• Is the label molecule-only, target-conditioned, assay-conditioned, or complex-conditioned?
• Are endpoint, unit, transformation, threshold, and censoring documented?
• Can conflicting measurements be explained by target or assay context?
• Does the split prevent leakage through molecule scaffold, protein family, assay batch, source, or time?
• Does the metric match the label semantics rather than only the model output type?
• Are private target names, collaborator context, internal task names, and unpublished results omitted?

Related

• Entity relation map
• Molecule
• Ligand
• Target
• Assay
• Bioactivity label
• Dataset
• Label semantics
• Missing data
• Censored label
• Weak label
• Dataset split contract
• Assay harmonization
• Protein-ligand split