Synthetic negative examples scramble rankings of protein interaction models
Because laboratory tests rarely document proteins that fail to interact, computational biologists evaluate interaction models against randomly paired non-interacting proteins, an approach that systematically warps performance metrics. Testing four sequence-based algorithms across six different ways of constructing negative examples showed that altering only the test negatives changed the area under the precision-recall curve by up to 0.26, reshuffling which architecture ranked first. When negative pairs were closely matched in representation space using protein language models, all language model architectures dropped to near-chance accuracy, and carefully constructed training negatives underperformed uniform random sampling in 108 of 120 external comparisons.