AI Hunter Outperforms Rivals in Search for Novel Antibiotics

The rise of pan-drug-resistant bacteria has created an urgent need for structurally novel antibiotics. While artificial intelligence offers a faster route than traditional screening, it often stumbles over a specific hurdle: class imbalance. In molecular datasets, inactive compounds vastly outnumber the handful of effective antibiotics, causing models to struggle with accuracy.

To solve this, researchers introduced selective similarity-based methods to balance the data. Working with a diverse library of 14,393 compounds—containing only 534 known antibiotics—they trained and tested seven different models. The team utilised techniques like K-means clustering to intelligently reduce the noise from inactive compounds.

The results were striking. A Graph Convolutional Network (GCN) model emerged as the clear winner, achieving a stellar 0.97 ROC-AUC score. Even classical models saw performance boosts when using the new under-sampling datasets. By comparing their findings with fourteen similar studies, the authors confirmed their method outperformed existing approaches by 10% to 18%, promising a more efficient future for drug discovery.