Protein-Ligand Split

A protein-ligand split defines how complexes, proteins, ligands, assays, and structures are separated across train, validation, and test sets. It is the core benchmark contract for structure-based modeling.

For a complex $c_i=(P_i,L_i,A_i,X_i)$, where $P_i$ is a protein or pocket, $L_i$ is a ligand, $A_i$ is assay or label context, and $X_i$ is structure or pose information, a split function is:

$$s(c_i)\in \{\mathrm{train},\mathrm{val},\mathrm{test}\}$$

The split should match the generalization claim.

Split Axes

Split axis	Held-out unit	Claim tested	Common leakage risk
Ligand-side	scaffold, analog cluster, similarity group	new chemotypes for known targets	close analogs across split
Protein-side	sequence family, structure family, target class	new targets or homolog-shift	homologs across split
Complex-side	protein-ligand pair or binding mode	new pairings	same pocket and close ligand seen separately
Assay/source	assay, campaign, lab, dataset source	cross-assay or cross-source robustness	protocol identity encoded in labels
Temporal	publication or measurement time	future-data generalization	retrospective curation choices

Grouped Split Constraint

Ligand grouping requires:

$$g_L(L_i)=g_L(L_j)\Rightarrow s(c_i)=s(c_j)$$

Protein grouping requires:

$$g_P(P_i)=g_P(P_j)\Rightarrow s(c_i)=s(c_j)$$

where $g_L$ may be scaffold or molecular similarity grouping, and $g_P$ may be sequence identity, protein family, structure family, or target class.

For strict protein-ligand generalization, both constraints may be required. For a narrow interpolation benchmark, the split can be looser, but the claim must also be narrower.

Benchmark Questions

• Pose prediction: is the held-out complex separated from close ligand analogs, homologous targets, and template structures?
• Affinity prediction: are assay context, target family, ligand scaffold, and label semantics separated enough for the claim?
• Virtual screening: are decoys realistic and are ligand-only shortcuts controlled?
• Generative design: are generated molecules evaluated against held-out targets without reusing target-derived pose information?
• Interaction modeling: are binding sites, ligand scaffolds, and interaction fingerprints leaking across split?

Checks

• What is the example unit: molecule, protein, complex, pose, assay row, or target-ligand pair?
• What is the split unit for proteins and for ligands?
• Are standardized molecules used before scaffold or similarity grouping?
• Are protein chains, isoforms, fragments, and homologs grouped before splitting?
• Are bound complexes, template databases, close analogs, or known poses excluded from test-time leakage?
• Are pose quality, affinity, enrichment, and generation claims evaluated under compatible split policies?
• Is the split contract reproducible from public fields rather than private curation notes?

Related

• Template leakage
• Scaffold split
• Protein family split
• Leakage
• Dataset split contract
• Molecular standardization
• Sequence identity clustering
• Pose quality
• Binding affinity
• Virtual screening