Assay Harmonization

Assay harmonization is the process of making measurement records comparable before training or evaluating a model. In chem-bio ML, labels with the same column name can come from different protocols, targets, units, and biological contexts.

A measurement should be treated as:

$$y=\mathrm{measure}(M,T,A,c)$$

where $M$ is a molecule, $T$ is a target, $A$ is an assay protocol, and $c$ contains conditions such as organism, construct, endpoint, units, and environment.

Harmonization Fields

Field	Examples	Why It Matters
target identity	protein, isoform, species, construct, mutation	different constructs can change activity
endpoint	$K_d$, $K_i$, ${\mathrm{IC}}_{50}$, percent inhibition, binary active	endpoints are not interchangeable labels
unit and transform	nM, uM, pIC50, Delta G, normalized score	transformations change loss scale and thresholds
censoring	>, <, detection limit, inactive above cutoff	censoring is not exact measurement
replicate policy	mean, median, best, latest, source-prioritized	aggregation changes label noise
assay source	lab, campaign, dataset, protocol, publication	source can leak into train/test splits

Common Conversions

When a concentration $c$ is converted to a logarithmic activity score:

$$\mathrm{p}c=-{\log }_{10}\left(\frac{c}{1\mathrm{M}}\right)$$

This conversion is only meaningful when $c$ is a comparable endpoint and unit. Mixing ${\mathrm{IC}}_{50}$, $K_i$, and $K_d$ can be useful for rough triage but should not be presented as a clean physical target without stating the approximation.

Label Direction Contract

Before training or evaluation, define whether larger means stronger activity:

$$y_{\mathrm{activity}}=-{\log }_{10}\left(\frac{c}{1\mathrm{M}}\right)$$

For concentration-like endpoints, lower raw concentration means stronger binding, but higher transformed p-scale means stronger activity. Mixing raw and transformed values can invert the label.

Raw field	Larger raw value means	Common transformed direction
$K_d$, $K_i$, ${\mathrm{IC}}_{50}$	weaker activity	p-scale larger means stronger
percent inhibition	stronger activity if same concentration and assay	often larger means stronger
binary active	depends on threshold definition	must state active threshold
censored > concentration	weaker than bound only	not an exact value
censored < concentration	stronger than bound only	not an exact value

Conflict Policy

The same molecule-target pair can have multiple labels:

$$\{y_k{\}}_{k=1}^K=\mathrm{measure}(M,T,A_k,c_k)$$

These labels should not be collapsed unless the endpoints, units, constructs, protocols, and censoring rules are compatible.

Conflict	Safer handling
different endpoints	keep endpoint-specific labels or model endpoint as context
different constructs/species	keep target construct explicit
replicate disagreement	aggregate with documented rule or mark uncertain
censored and exact values	keep censoring direction or use interval-aware objective
different assay sources	split or stratify by source when source transfer is claimed

Split Implications

Claim	Split or Control
new molecule in same assay setting	scaffold split with assay held fixed
new assay or campaign	source/campaign holdout
new target family	protein-family split plus assay-source control
cross-dataset transfer	external dataset or source holdout with harmonization audit
target-conditioned activity	keep target, assay, endpoint, unit, and threshold explicit

Checks

• Are endpoint types mixed, such as $K_d$, $K_i$, ${\mathrm{IC}}_{50}$, inhibition percent, or binary labels?
• Are units and transformations recorded?
• Are censored values such as > or < handled explicitly?
• Are replicate measurements aggregated deliberately?
• Are conflicting molecule-target labels resolved or kept assay-specific?
• Does the split hold out assays or campaigns when that is the deployment claim?
• Is the label direction consistent, so larger always means more active or less active?
• Are source-specific missingness and measurement limits documented?
• Is larger-is-better or smaller-is-better made consistent before metric computation?
• Are labels collapsed only when the assay context supports it?

Related

• Assay
• Bioactivity label
• Dataset
• Binding affinity
• Label semantics
• Metadata and provenance
• Leakage
• test split