Negative Set
A negative set defines examples treated as inactive, irrelevant, non-binding, or low-quality. In molecular ML, negative construction is often the benchmark, not a detail.
The key distinction is provenance:
An unmeasured molecule is unknown, not automatically inactive.
Common Sources
| Source | What It Supports | Main Risk |
|---|---|---|
| measured inactive examples | assay-specific inactive classification | assay threshold, censoring, and replicate handling dominate |
| property-matched decoys | virtual-screening enrichment benchmark | decoys may be separable by artifacts or too easy |
| random library molecules | broad retrieval stress test | unrealistic prevalence and trivial negatives |
| assumed negatives from missing labels | weak supervision only | unknowns are mislabeled as true negatives |
| cross-target negatives | target-conditioned selectivity tasks | hidden activity against the target may be unmeasured |
Label Semantics
Negative construction should preserve the distinction:
where is an assay-specific activity threshold. A benchmark should state whether a negative is measured inactive, censored below a threshold, unobserved, or sampled as a decoy.
Evaluation Pattern
| Field | Required Detail |
|---|---|
| negative provenance | measured, decoy, random, assumed, cross-target, or generated |
| prevalence | natural prevalence or rebalanced benchmark prevalence |
| property matching | size, charge, hydrophobicity, topology, and scaffold controls |
| split unit | scaffold, target, assay/source, temporal, or complex-pair split |
| baseline | simple property, fingerprint, similarity, or docking baseline |
| metric | enrichment, ranking, classification, calibration, or decision utility |
Checks
- Are negatives experimentally measured or assumed?
- Are decoys separable by trivial properties such as size, charge, or hydrophobicity?
- Is performance reported separately for measured inactives and decoys?
- Does the active threshold match the assay and task?
- Would a simple property-only or fingerprint baseline solve the benchmark?
- Is the evaluation prevalence the same as the deployment prevalence?
- Are missing labels excluded, modeled as unknown, or incorrectly converted to inactive?