Negative sampling labels selected unobserved or unrecorded examples as negative examples. It can expand datasets containing only positive observations, but its central assumption is strong: an unrecorded molecule is treated as a negative case. section 5.4.2
CYP inhibition data
Five CYP inhibitor datasets contained 16,368 molecules, but only 4,805 appeared in every dataset. Rather than assume that missing inhibitor records indicate no inhibition, a full-dataset workflow could label unrecorded molecules as negative examples through negative sampling. section 5.4.2
Use negative sampling with care:
- Separate the technique from evidence of true absence: an unobserved molecule may simply be untested or unrecorded. section 5.4.2
- Combine it with representative train–test splitting when multilabel distributions and label combinations must remain balanced. section 5.4.2



