Coordinate Modeling Contract

A coordinate modeling contract records how a model consumes, transforms, predicts, and evaluates 3D coordinates. It is especially important for structure-based modeling, docking, protein structure, molecular conformers, and geometric generative models.

The core question is:

$$X\stackrel{f_{\theta }}{\to }\hat{Y}\text{under}{X}^{\prime }=RX+1t^{\top }$$

where $X\in {\mathbb{R}}^{N\times 3}$ is an input coordinate set, $R$ is a rotation, and $t$ is a translation.

Contract Fields

Field	Question
Coordinate source	Are coordinates experimental, predicted, docked, generated, simulated, or cleaned?
Frame	Are coordinates in a global, centered, receptor, pocket, local residue, ligand, or learned frame?
Unit	Does one point represent atom, residue, bead, grid point, surface point, conformer, or pose candidate?
Ordering	Are atoms, residues, chains, or ligands indexed consistently across preprocessing and evaluation?
Transformation group	Should the model respect translation, rotation, reflection, permutation, or chirality constraints?
Input features	Which features are invariant scalars, equivariant vectors, or frame-dependent metadata?
Output type	Is the model predicting scalar, rank, contact, vector field, coordinate update, conformer, or pose?
Loss	Is the loss defined on coordinates, distances, frames, scores, velocities, or denoising targets?
Metric	Is evaluation RMSD, lDDT-style quality, clash/geometry check, interaction quality, enrichment, or affinity?
Leakage	Does preprocessing use known ligand pose, target-aligned frame, template, or post-hoc alignment unavailable at inference?

Target Symmetry

Output	Expected Behavior	Example
Scalar	invariant	affinity, energy, class probability, ranking score
Pairwise distance/contact	invariant to rigid motion, equivariant to permutation	distance map, contact map, interaction graph
Vector or field	equivariant	force, displacement, velocity, score field
Coordinate update	equivariant	atom update, residue update, ligand pose refinement
Pose or conformer	equivariant up to atom symmetry and rigid motion	docking pose, generated conformer

For scalar prediction:

$$f(RX+1t^{\top })=f(X)$$

For coordinate prediction:

$$F(RX+1t^{\top })=RF(X)+1t^{\top }$$

Evaluation Alignment

Coordinate metrics need a frame and mapping rule.

Metric	Must State
Pose RMSD	receptor frame, atom mapping, symmetry correction, top-1 vs best-of-N
Distance/contact quality	threshold, residue/atom unit, missing residues, chain policy
Geometry plausibility	bond lengths, angles, clashes, chirality, valence, strain
Interaction quality	pocket definition, interaction fingerprint, hydrogen/protonation policy
Affinity/ranking	assay context, split unit, baseline, calibration if probabilities are used

Structure-Based Modeling Risks

• A ligand-defined pocket frame can leak the answer when the ligand pose is unavailable at inference.
• Aligning predicted and reference ligands before RMSD can hide docking failure.
• A model can be equivariant internally but use non-equivariant preprocessing.
• Distance-only features can lose chirality or orientation needed for the task.
• Coordinate denoising loss can improve geometry while not improving ranking or affinity.
• Template-derived structures can make a test split easier than the claimed deployment setting.

Checks

• What coordinate frame does the model receive?
• What transformation group should the input and output respect?
• Is the target invariant, equivariant, permutation-aware, or frame-dependent?
• Does the loss align with the reported coordinate metric?
• Is pose quality separated from scoring, affinity, and virtual screening utility?
• Are atom/residue mappings and symmetry corrections defined?
• Is preprocessing identical across train, validation, test, and inference where required?

Related

• Coordinate frame
• Equivariance
• Invariance
• Distance geometry
• Coordinate update
• Representation contract
• Pose RMSD
• Pose quality
• Template leakage
• Geometry for Structure Modeling