Conformer

A conformer is a 3D arrangement of a molecule for a specific topology and stereochemical state. The same molecule can have multiple low-energy conformers.

Coordinates can be represented as:

$$X=[x_1,\dots ,x_N{]}^{\top },x_i\in {\mathbb{R}}^3$$

where $N$ is the number of atoms.

An ensemble of conformers is:

$$\mathcal{C}(M)=\{X^{(1)},\dots ,X^{(K)}\}$$

Generation Contract

A conformer is tied to a protocol:

$$X^{(k)}\sim G_{\pi }(M,s_k)$$

where $G_{\pi }$ is the conformer generator, $\pi$ contains settings such as force field, embedding method, pruning, and minimization, and $s_k$ is a seed or stochastic state.

The contract should state:

• molecular identity used for generation;
• stereochemistry, tautomer, protonation, and charge state;
• number of conformers and pruning rule;
• energy minimization or ranking method;
• coordinate units and software version.

Checks

• Was conformer generation seeded and versioned?
• Are protonation, tautomer, charge, and stereochemistry fixed before conformer generation?
• Is the model trained on crystal structures but evaluated on generated conformers?
• Are metrics stable across conformer ensembles?
• Does the task need ligand-only conformers or protein-bound poses?
• Is the conformer generation protocol part of the featurization contract?
• Is a force field used for ranking or minimization?

Evaluation Boundary

If the model consumes or predicts conformers, evaluation should separate:

• conformer validity: chemically plausible geometry;
• conformer coverage: whether low-energy or bioactive-like states are represented;
• pose correctness: placement relative to a protein pocket;
• property utility: whether conformers improve the downstream prediction.

Conformer vs Pose

A ligand-only conformer is not the same as a bound pose:

$$X_{\mathrm{conf}}\ne X_{\mathrm{pose}}(P)$$

The pose depends on the protein or pocket context $P$. A model trained on bound poses but evaluated on ligand-only conformers is under a representation shift.

Routing

Question	Route
How is the molecule standardized before coordinates?	Molecules
Is this ligand-only geometry or protein-bound placement?	Structure-based modeling
Is a pose generated, scored, or filtered in a pocket?	Docking
Does the model need equivariance or coordinate losses?	Coordinate modeling contract
Is conformer generation used as featurization?	Molecular featurization contract

Failure Modes

• Train structures are experimental or bound poses, while inference uses generated conformers.
• Only the lowest-energy conformer is used even when the bioactive state differs.
• Conformer ensembles are generated with a different tautomer/protonation policy than training.
• Metrics ignore variance across conformers and report one lucky geometry.
• Coordinates are centered or aligned with deployment-unavailable context.

Related

• Molecular featurization contract
• Chemical state contract
• Force field
• Energy minimization
• Computational Biology Scope
• Molecules
• Docking
• Stereochemistry
• Protonation state
• Pose generation
• Pose quality
• Geometric architecture
• Ligand
• Structure