Protein-ligand Representation Contract

A protein-ligand representation contract records how protein, ligand, pocket, pose, and interaction evidence become model input. It prevents a model note from mixing object identity, geometry, assay labels, and evaluation-only information.

$$r(u)=r(P,L,B,X,c)$$

where $P$ is the protein or target, $L$ is the ligand, $B$ is the pocket or binding site rule, $X$ is coordinate evidence, and $c$ is task context such as assay or structure source.

Required Fields

Field	Record
Protein unit	sequence, chain, isoform, construct, mutation, species, structure source
Ligand unit	molecule identity, salt, stereo, tautomer, protonation, charge, conformer
Pocket rule	known site, predicted site, ligand-defined site, blind grid, residue cutoff
Coordinate source	experimental, predicted, docked, minimized, generated, simulated
Atom/residue mapping	how nodes correspond across reference, generated, and evaluated structures
Pair context	target, assay, endpoint, unit, censoring, source, complex identity
Availability	which features are available at inference time
Split boundary	scaffold, protein family, complex pair, assay/source, template, time

Representation Patterns

Pattern	Input	Main risk
ligand-only	ligand graph, fingerprint, conformer	target context hidden or ignored
protein-only	sequence or structure embedding	ligand-specific claim unsupported
late fusion	separate protein and ligand embeddings	interaction geometry may be weak
pocket-conditioned	ligand plus pocket representation	pocket may use unavailable ligand information
complex graph	protein-ligand graph with edges	edge construction may encode the answer
coordinate model	atom/residue coordinates and updates	frame, chirality, alignment, coordinate-source leakage

Availability Matrix

Feature	Training	Inference	Risk if mismatched
bound ligand pose	sometimes available from complex structures	usually unavailable for new candidates	optimistic pocket or edge construction
assay label	supervised target	unavailable before measurement	label-derived feature leakage
protein sequence	usually available	usually available	isoform or construct mismatch
experimental structure	may be available for benchmark complexes	often missing for new targets	structure-source shift
predicted structure	can be generated uniformly	can be generated uniformly	prediction error becomes part of input
ligand conformer	generated by a protocol	generated by same protocol	train/inference conformer mismatch
docking pose	generated by docking workflow	generated by docking workflow	docking failures and filtering hidden

Claim Boundary

Claim	Representation must prove
ligand property prediction	ligand identity and state are sufficient for the endpoint
target-conditioned activity	protein/target and assay context are explicit
pose quality prediction	pose coordinates, atom mapping, and geometry checks are available
affinity ranking	label semantics, endpoint units, split, and calibration are defined
interaction representation	edges/features are constructible without reference-answer leakage
generation conditioning	condition source is available before generating candidates

Leakage Checks

• Does the pocket require the ground-truth ligand pose?
• Does the graph contain edges computed from the reference complex when the claim is blind prediction?
• Are close ligand analogs or homologous proteins split across train and test?
• Are conformers generated using a protocol that sees evaluation labels or reference poses?
• Are assay labels collapsed across incompatible endpoints, units, or sources?
• Is the representation available under the intended deployment setting?
• Are failed conformer, docking, parsing, or structure-preparation cases included in the denominator?
• Is the same featurization contract used for train, validation, test, and inference?

Related

• Representation Routes
• Protein-ligand complex
• Pocket definition contract
• Protein-ligand split
• Leakage