Molecular Property Prediction

Molecular property prediction maps a molecule representation to a scalar, class, distribution, or ranked score. It is a core task for ligand modeling, screening, ADMET-style prediction, and molecular representation learning.

A generic supervised setup is:

$$\hat{y}=f_{\theta }(m)$$

where $m$ is a molecule represented as a string, graph, fingerprint, descriptor vector, conformer, or ensemble.

The empirical objective is:

$$\hat{\theta }=\arg \underset{\theta }{\min }\frac{1}{n}\sum _{i=1}^n\mathcal{L}(f_{\theta }(m_i),y_i)$$

Input Choices

• SMILES or SELFIES-style sequence.
• Molecular graph.
• Fixed fingerprint.
• Physicochemical descriptors.
• 3D conformer or conformer ensemble.
• Protein-conditioned molecule representation for target-specific prediction.

Task Contract

Molecular property prediction should state what one example means:

$$e_i=(m_i,c_i,y_i,s_i)$$

where $m_i$ is the molecular object, $c_i$ is optional context such as target or assay, $y_i$ is the label, and $s_i$ is source/provenance metadata.

Field	Question
molecule identity	salt-stripped parent, charged form, tautomer, stereoisomer, or exact submitted structure?
representation	SMILES, graph, fingerprint, descriptor, conformer, ensemble, embedding?
context	molecule-only, target-conditioned, assay-conditioned, or environment-conditioned?
label	class, scalar, censored value, ranking signal, uncertainty target?
split unit	molecule, scaffold, assay, target, time, source, or molecule-target pair?
metric	classification, regression, ranking, calibration, or applicability-domain metric?

For target-conditioned activity, the input is not only a molecule:

$$\hat{y}=f_{\theta }(m,t,a)$$

where $t$ is target context and $a$ is assay or measurement context. Ignoring $t$ or $a$ can turn a conditional task into a noisy molecule-only task.

Label Types

Label Type	Example	Risk
physicochemical property	logP, solubility, charge proxy	preprocessing and unit mismatch
bioactivity class	active/inactive threshold	threshold and prevalence dependence
continuous activity	IC50, Ki, Kd, pIC50	assay condition and censoring
ADMET endpoint	toxicity, permeability, clearance	protocol and species dependence
ranking label	relative preference or screen hit order	candidate-pool dependence
predicted label	teacher model output	inherited bias and circular evaluation

The same molecule can have multiple valid labels under different assay conditions. Do not merge them without a label semantics rule.

Split and Leakage

Random row splits are often too weak. For a model claim $C$, choose the split unit stronger than the intended generalization:

$$\text{split unit}⪰\text{claim unit}$$

Claim	Stronger Split
new analogs near training chemistry	random or molecule split may be enough
new scaffolds	scaffold split
new targets	target or protein-family split
new assays	assay/source split
prospective use	time split
protein-ligand generalization	pair split plus ligand and protein grouping

Evaluation Risks

• Random splits can overestimate scaffold generalization.
• Assay labels can be noisy, censored, and protocol-dependent.
• Activity cliffs make local similarity unreliable.
• Negative sets may mix true inactivity with untested molecules.
• Standardization choices can change duplicates and labels.

Checks

• What does the target label mean and how was it measured?
• Is the task classification, regression, ranking, or uncertainty estimation?
• Are scaffold, time, target, or assay splits needed?
• If the label is target-conditioned, is the protein-ligand split explicit?
• Are stereochemistry, salts, tautomers, and protonation handled consistently?
• Does performance hold outside the training applicability domain?
• Is the molecule identity key defined before deduplication?
• Are censored, repeated, or conflicting labels handled explicitly?
• Does the metric match property prediction, ranking, or decision use?

Related

• Property prediction
• Classification
• Regression
• Ranking
• Molecular graph
• Molecular fingerprint
• Molecular identity
• Chemical state contract
• Conformer
• Bioactivity label
• Label semantics
• Censored label
• Scaffold split
• Protein-ligand split
• Activity cliff
• Applicability domain