Multilabel classification assigns a set of labels to each instance when several classes may apply simultaneously. In cytochrome P450 inhibition, the question is which of five CYP enzymes a molecule inhibits, including the possibility of none. section 5.4.2
Distinguish multilabel classification from related prediction settings:
- Multiclass classification chooses exactly one class, whereas multilabel classification predicts several labels for the same instance. section 5.4.2
- Multioutput classification predicts multiple output variables and can encompass multiclass and multilabel tasks. section 5.4.2
- Use an independent logistic function for each label rather than softmax, which assumes one correct answer. section 5.4.2
- Use Random Forests directly for multilabel problems, or train independent label models with MultiOutputClassifier when label correlation is not expected. section 5.4.2
Predicting CYP inhibition
Represent each molecule with five binary labels: CYP2C19, CYP2D6, CYP3A4, CYP1A2, and CYP2C9. A prediction can therefore identify any combination of inhibited enzymes instead of forcing the molecule into one class. section 5.4.2
Because labels may be correlated, one model predicting all labels can consolidate shared patterns. Evaluate ranking quality with coverage error and ranking loss, and aggregate per-label F1 with micro- or macro-averaging. section 5.4.2 section 5.5



