GANATCpair: GENERATIVE HARD NEGATIVE SAMPLING FOR IMPROVING DRUG-TARGET INTERACTION PREDICTION AGAINST SNAKE VENOM TARGETS

Abstract

Predicting drug-target interactions (DTIs) remains challenging because experimentally verified non-interaction data are rarely available for bioactive compounds and snake venom targets. This study proposes GANATCpair (Generative Adversarial Network with ATC Pairing), a GAN-based hard-negative sampling strategy for selecting unknown status drug-target pairs. GANATCpair learns latent representations of DTI patterns and uses synthetic vectors as anchors to select informative unknown-status pairs near the positive decision region as putative negatives. The strategy was evaluated using the Snakebite Envenoming Medicines Database and four Yamanishi benchmark datasetsacross SVM, RF, XGBoost, and MLP classifiers. On the Snakebite dataset, GANATCpair improved AUC from 0.626-0.773 under random sampling to 0.916-0.952. The observed differences were supported by paired statistical analysis, and its computation was mainly driven by generative training and latent-space mapping. These findings suggest that GANATCpair strengthens hard-negative sampling and supports computational screening and drug–target prioritisation in snakebite en venoming.