Assay

An assay is an experimental measurement that links a biological or chemical setup to an observed label.

In supervised chem-bio ML, an assay is part of the Target-assay-label contract:

$$\text{assay}\to (\text{endpoint},\text{unit},\text{threshold},\text{censoring},\text{context})$$

Why It Matters

• Assay conditions define what a molecular activity label actually means.
• Similar molecules can have labels from incompatible protocols or target contexts.
• Dataset construction should preserve assay identity when possible.
• The same molecule-target pair can have conflicting labels across protocols or sources.

Assay Contract

Field	Meaning	Risk If Missing
target	biological object being measured	labels from different targets collapse together
endpoint	what is measured	binding, function, toxicity, and viability get mixed
readout	experimental signal	proxy signal may not equal desired mechanism
unit	concentration, percent, score, count	regression values become incomparable
threshold	active/inactive boundary	binary labels become source-specific
censoring	greater-than or less-than limits	point regression becomes biased
protocol/source	public assay/source identifier	batch or source leakage is hidden
context	organism, construct, cell line, time, dose	label meaning changes silently

An assay-conditioned label can be written as:

$$y=g(m,t,a,c)$$

where $m$ is the molecule, $t$ is the target, $a$ is the assay, and $c$ is measurement context.

Assay as Model Input

Some papers use assay metadata as a task condition:

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

This is different from removing assay effects during harmonization. If assay identity is an input, the test split must prevent the model from memorizing assay-specific shortcuts.

Split Implications

Split Type	What It Tests
random row split	interpolation over known molecules, targets, and assays
molecule split	transfer to new molecules under known assay context
target split	transfer to new targets or families
assay/source split	transfer to new measurement protocols
time split	future assay records or database releases

Assay/source split is often needed when the paper claims assay-robust or source-robust performance.

Checks

• What target, readout, organism, and protocol produced the label?
• Are labels from different assay types mixed as if they were comparable?
• Is the task predicting assay response, binding, toxicity, or a proxy endpoint?
• Does the train/test split leak through repeated molecules or shared assay batches?
• Are censored or thresholded values handled as such, instead of point labels?
• Is the assay used as metadata, split group, task context, or prediction target?
• Are replicate measurements aggregated, weighted, or kept separate?
• Does the benchmark report performance per assay/source, not only globally?

Related

• Entity relation map
• Target-assay-label contract
• Bioactivity label
• Censored label
• Assay harmonization
• Negative set
• Label semantics
• Dataset shift
• Molecule
• Protein
• Dataset
• Protein-ligand complex