Binding Affinity

Binding affinity measures how strongly a ligand binds to a target. It is related to thermodynamics, but ML datasets often provide assay-dependent proxies such as $K_d$, $K_i$, ${\mathrm{IC}}_{50}$, or activity labels.

A common thermodynamic relation is:

$$\Delta G=RT\ln K_d$$

where $\Delta G$ is binding free energy, $R$ is the gas constant, $T$ is temperature, and $K_d$ is the dissociation constant.

Many datasets use transformed values:

$$p{K}_d=-{\log }_{10}(K_d)$$

Higher $p{K}_d$ means stronger binding when the units are consistent.

Affinity values should be interpreted with the Target-assay-label contract because endpoint, assay, unit, target construct, and censoring change what the label means.

Endpoint Boundary

Endpoint	Meaning	Caveat
$K_d$	equilibrium dissociation constant	closest to direct binding equilibrium when measured cleanly
$K_i$	inhibition constant	depends on assay model and competitor assumptions
${\mathrm{IC}}_{50}$	concentration for half-maximal inhibition	depends on assay conditions, substrate, time, and target context
activity class	thresholded active/inactive label	threshold, source, and missing negatives dominate semantics
enrichment	screening recovery of known actives	ranking utility, not a thermodynamic affinity value

Do not treat these endpoints as interchangeable without an explicit conversion or harmonization rule.

Transformation and Units

If $K_d$ is measured in molar units:

$$\Delta G^{\circ }=RT\ln K_d$$

and

$$p{K}_d=-{\log }_{10}\left(\frac{K_d}{1\mathrm{M}}\right).$$

For $K_d<1\mathrm{M}$, larger $p{K}_d$ usually means stronger binding and more negative $\Delta G^{\circ }$. A note should state units before applying $p{K}_d$, $p{K}_i$, or $pI{C}_{50}$ transformations.

Modeling Boundary

Affinity prediction is usually:

$$\hat{a}=f_{\theta }(P,L,c)$$

where $P$ is the protein or target context, $L$ is the ligand, and $c$ is assay or structural context. If $c$ differs across train and test, the model may learn assay/source effects rather than binding.

Affinity is separate from pose quality:

$$\text{low RMSD}\nRightarrow \text{accurate affinity}$$

A docking workflow can produce a plausible pose and still fail to rank ligands by binding strength.

Checks

• Is the target $K_d$, $K_i$, ${\mathrm{IC}}_{50}$, enrichment, or a binary activity label?
• Are units, assay conditions, and censoring handled explicitly?
• Is the target-assay-label contract preserved from raw measurement to benchmark label?
• Is the model predicting affinity, pose quality, or both?
• Are ligand scaffolds and protein families separated across splits?
• Is target or assay source leakage possible?
• Are censored values such as >, <, or thresholded actives handled as bounds rather than exact labels?
• Does the reported metric match the decision: numeric accuracy, ranking, enrichment, or calibration?

Related

• Scoring function
• Pose quality
• Virtual screening
• Metric
• Label semantics
• Assay harmonization
• Bioactivity label
• Assay
• Target-assay-label contract
• Protein-ligand docking