Chemical State Contract

A chemical state contract records which molecular form is being modeled before features, conformers, docking, labels, or splits are defined. A molecule is not only a graph; it can include salt policy, tautomer, protonation, charge, stereochemistry, and conformer protocol.

The practical object is:

$$m_{\mathrm{state}}=(G,s,\tau ,p,q,c)$$

where $G$ is molecular connectivity, $s$ is stereochemistry, $\tau$ is tautomer state, $p$ is protonation state, $q$ is formal or partial charge policy, and $c$ is conformer or coordinate protocol when 3D geometry is used.

Why It Matters

Two records can share the same topology but differ in binding, activity, docking pose, or featurized representation:

$$G(m_1)=G(m_2)\nRightarrow y(m_1)=y(m_2)$$

This is common for stereoisomers, protonation states, tautomers, salts, and conformers. If a paper does not state the chemical state policy, its reported performance can be hard to interpret or reproduce.

State Fields

Field	Question	Route
Connectivity	What atom-bond graph is treated as the base molecule?	Molecular graph
Salt and fragments	Are salts, counterions, mixtures, or largest fragments kept?	Molecular standardization
Stereochemistry	Are chiral centers and double-bond stereo preserved, flattened, unknown, or enumerated?	Stereochemistry
Tautomer	Is a canonical tautomer chosen, supplied states preserved, or states enumerated?	Tautomer
Protonation	What pH, rule set, receptor context, and hydrogen policy define the state?	Protonation state
Charge	Are formal charge, partial charge, and neutralization policies recorded?	Molecular featurization contract
Conformer	Are 3D coordinates generated, selected, minimized, or observed experimentally?	Conformer

Pipeline Position

Chemical state should be fixed before deduplication, split construction, featurizer caching, conformer generation, and docking:

$$m_{\mathrm{raw}}\stackrel{\text{standardize}}{\to }{m}_{\mathrm{state}}\stackrel{\text{identity key}}{\to }I(m_{\mathrm{state}})\stackrel{\text{split/cache/featurize}}{\to }r$$

If the split key uses topology only but the model consumes state-aware features, near-duplicate states can leak across splits. If the model flattens chemical states that labels distinguish, signal can be destroyed.

2D vs 3D Boundary

For 2D property prediction, the state contract may focus on graph, stereo, tautomer, charge, and label context. For structure-based modeling, the contract must also include conformers, receptor preparation, and pose context:

$$\text{complex state}=(m_{\mathrm{ligand}\mathrm{state}},p_{\mathrm{receptor}},X_{\mathrm{ligand}},X_{\mathrm{pocket}})$$

This separates ligand-only conformers from protein-bound poses and prevents a docking note from silently using deployment-unavailable information.

Paper Reading Checks

• Does the paper define the molecular identity before splitting?
• Are salts, stereo, tautomer, protonation, charge, and conformer policy stated?
• Is chemical state fixed before conformer generation, docking, and featurization?
• Does the benchmark compare methods under the same state policy?
• Are generated molecules evaluated after the same standardization and state assignment as training molecules?
• Are labels aggregated across records that may have different chemical states?
• Is novelty or scaffold split computed before or after state normalization?
• Does a 3D model use ligand-only conformers, experimental structures, predicted structures, or bound poses?

Failure Modes

Failure	Consequence
Splitting raw records before standardization	equivalent compounds can appear in train and test
Flattening stereochemistry	active stereoisomers can collapse with inactive ones
Mixing tautomer/protonation protocols	docking or affinity labels become hard to compare
Caching features without state policy	stale features survive after standardization changes
Evaluating novelty before normalization	generated molecules look falsely novel
Using one conformer without reporting protocol	result depends on an unrecorded geometry choice

Related

• Molecular identity
• Molecular standardization
• Molecular featurization contract
• Stereochemistry
• Tautomer
• Protonation state
• Conformer
• Computational Biology paper intake
• Data and evaluation