Virtual screening is a critical step in modern drug discovery. By using mathematical models to predict the bioactivity of massive molecular libraries, researchers can prioritize promising compounds that would be impossible to test experimentally. Deep learning models, specifically Deep Graph Neural Networks (GNNs), offer state-of-the-art performance for this task. However, to be effective, they typically require massive datasets to train on—a major hurdle given that existing antibacterial datasets are small and highly imbalanced.

To solve this, CHIMiRA mathematicians collaborated with an interdisciplinary team to apply a machine learning technique called transfer learning. First, they pre-trained a Deep GNN on vast, general molecular datasets, allowing the model to learn fundamental chemical principles. Next, they fine-tuned this model using two small, targeted antibacterial datasets. The resulting model successfully screened databases containing billions of compounds using limited computational power. From this immense pool, they identified a highly targeted list of just 156 compounds for laboratory validation. 

Supported by modest funding from a PhD training grant, the team tested all 156 computationally selected compounds in the laboratory against E. coli, a bacterium with increasingly common resistant strains. The results were remarkable: 54% of the selected compounds demonstrated antibacterial activity. This exceptionally high hit rate yielded several genuinely promising candidates showing antimicrobial activity against a broad spectrum of infections at very low doses.