Property Prediction

Property prediction maps an entity representation to a scalar, class, distribution, or ranked score. The entity can be a molecule, protein, sequence, structure, material, image, table row, or protein-ligand complex.

The generic task is:

$$\hat{y}=f_{\theta }(r_e,c)$$

where $r_e=\varphi (e)$ is a representation of entity $e$ and $c$ is optional context such as target, assay, environment, source, or time.

For regression:

$$\mathcal{L}=\frac{1}{n}\sum _{i=1}^n\ell (f_{\theta }(r_i,c_i),y_i)$$

For classification:

$$\hat{p}(y=k\mid r,c)=\mathrm{softmax}(W{h}_{\theta }(r,c)+b{)}_k$$

Output Variants

• Scalar regression: affinity, energy, toxicity score, expression level, or physical property.
• Classification: active/inactive, functional class, risk class, or annotation class.
• Distributional prediction: mean and uncertainty, censored range, or calibrated probability.
• Ranking score: prioritization value used for retrieval, screening, or triage.

Molecular Modeling Notes

Molecular property prediction is a special case where the entity is a Molecule and the representation may be SMILES, graph, fingerprint, conformer, or embedding.

Target-conditioned property prediction must preserve the Target-assay-label contract:

$$y=h(m,t,a,c)$$

where $m$ is molecule, $t$ is target, $a$ is assay, and $c$ is measurement context.

Checks

• What entity does one example represent?
• Is the property intrinsic, target-conditioned, assay-conditioned, or context-dependent?
• Is the output a scalar, class, probability, censored value, or ranking score?
• What loss is optimized, and what metric is reported?
• Does the split prevent entity, scaffold, family, source, assay, or time leakage?
• Are missing, weak, censored, and noisy labels handled explicitly?

Related

• Task specification
• Task output space
• Molecular property prediction
• Target-assay-label contract
• Bioactivity label
• Label semantics
• Censored label
• Regression metrics
• Classification metrics
• Calibration