Bioactivity Label

A bioactivity label is a measured or curated value that describes how a molecule, ligand, protein, target, assay, or context behaves in a biological or chemical setup.

It is not just a number. A label should be read as:

$$y=\mathrm{measure}(m,t,a,c)$$

where $m$ is a molecule or ligand, $t$ is a target, $a$ is an assay protocol, and $c$ is measurement context such as endpoint, organism, construct, dose, time, unit, and preprocessing.

This is the label side of the Target-assay-label contract.

Common Forms

• Continuous: $K_d$, $K_i$, ${\mathrm{IC}}_{50}$, ${\mathrm{EC}}_{50}$, percent inhibition, activity score, affinity proxy.
• Transformed: $p{K}_d$, $p{K}_i$, $pI{C}_{50}$, log concentration, standardized score.
• Binary: active/inactive after a threshold.
• Censored: values reported as greater-than or less-than limits.
• Weak or missing: untested, inferred, aggregated, or source-dependent labels.

Why It Matters

• The same molecule-target pair can have different labels under different assay contexts.
• A binary active label may hide endpoint type, threshold, unit, and censoring.
• Regression, classification, ranking, and virtual screening can use the same raw records differently.
• Negative examples may mean measured inactive, not selected, or simply unobserved.

Label Semantics Contract

Field	Question	Example
entity unit	what object receives the label?	molecule, molecule-target pair, complex, assay row
target	is the label target-conditioned?	kinase target, receptor, protein family
assay	which protocol produced it?	binding, functional, cell viability
endpoint	what quantity is measured?	$K_d$, ${\mathrm{IC}}_{50}$, percent inhibition
transformation	raw or transformed?	$pI{C}_{50}=-{\log }_{10}(I{C}_{50})$
threshold	how binary labels are made	active if $pI{C}_{50}\ge \tau$
censoring	inequality or detection limit	$>10\mu M$, $<1nM$
replicate policy	aggregation rule	median, mean, best, latest, source-specific
negative meaning	inactive, decoy, unobserved, or assumed negative	affects ranking and screening

For concentration-style labels:

$$pI{C}_{50}=-{\log }_{10}\left(I{C}_{50}[\mathrm{M}]\right)$$

Unit conversion must happen before the transform.

Task Rewrites

The same raw bioactivity table can define different tasks:

Task	Label Interpretation
regression	predict continuous endpoint after unit/transform policy
classification	predict thresholded active/inactive label
ranking	rank compounds for one target or assay context
multitask learning	predict assay-specific labels with missing entries
virtual screening	enrich measured actives in a candidate pool
target-conditioned prediction	model $p(y\mid m,t,a)$ rather than $p(y\mid m)$

The paper should state which rewrite it uses.

Leakage Risks

Leakage	Why It Happens
molecule duplicate	same compound appears under different salts or strings
target overlap	homologous targets split across train/test
assay/source overlap	assay identity acts as a shortcut
threshold tuning	binary threshold chosen after seeing test behavior
replicate aggregation	test information leaks into global aggregation
negative construction	decoys encode benchmark artifacts

Checks

• What endpoint does the label represent?
• What unit, transformation, threshold, and censoring policy are used?
• Is the label molecule-only, target-conditioned, assay-conditioned, or complex-conditioned?
• Are replicate measurements aggregated, kept separate, or filtered?
• Are conflicting labels resolved globally or kept assay-specific?
• Does the split prevent molecule, target, assay, or source leakage?
• Is the label valid for the intended task, or only for the original assay context?
• Are unmeasured pairs treated differently from measured inactive pairs?
• Are labels aggregated before or after train/test splitting?

Related

• Assay
• Target-assay-label contract
• Molecule
• Ligand
• Target
• Dataset
• Label semantics
• Missing data
• Censored label
• Weak label
• Assay harmonization
• Negative set
• Dataset split contract
• Activity cliff
• Binding affinity
• Molecular property prediction