Negative examples in protein–protein and drug–target interaction prediction: a critical review of training strategies and evaluation practices

Abstract Models for protein–protein and drug–target interaction prediction usually rely on constructed negative examples because experimentally supported non-interactions are scarce. Unobserved pairs may be sampled at random, filtered by biological rules or replaced with presumed-inactive decoys. These choices affect the patterns a model learns and the difficulty of the benchmark used to assess it. This critical review brings together evidence from interaction prediction, virtual screening and positive-unlabelled learning to examine how negative construction influences model evaluation. We distinguish the evidence supporting a negative label from the rule used to select a pair, and compare the rationale, empirical support and limitations of common selection strategies. Across these studies, more confident negative labels do not necessarily produce more informative tests: selection can make negatives easier to distinguish, amplify degree-related shortcuts or introduce decoy bias. The distinction between training and evaluation helps explain some apparently conflicting findings. A sampling rule that assists learning may give a misleading performance estimate when it also defines the test set. We use this distinction to develop conditional recommendations for selecting negatives, choosing data splits and reporting class ratios and baseline results. A reporting checklist also identifies information that experimental groups can retain to make inactive measurements reusable. We also relate these findings to negative sampling in link prediction, recommendation and knowledge-graph completion. Explicit descriptions of both training and test negatives are needed to interpret reported gains and assess their relevance to the intended biological application.