DeepBLI: A Transferable Multichannel Model for Detecting β-Lactamase-Inhibitor Interaction.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Pathogens producing beta-lactamase enzymes present a significant challenge in treating antibiotic-resistant infections.
• Current screening methods like high-throughput screening are costly, and structure-based virtual screening has limitations.
• There is an urgent need for novel beta-lactamase inhibitors to combat antimicrobial resistance.

Purpose of the Study:

• To develop a novel multichannel deep neural network (DeepBLI) for effective beta-lactamase inhibitor screening.
• To overcome the limitations of conventional and structure-based screening methods.
• To identify new beta-lactamase-inhibitor interactions and potential drug candidates.

Main Methods:

• Construction of a novel multichannel deep neural network (DeepBLI).
• Pretraining on a label reversal KIBA dataset and fine-tuning on BindingDB data.
• Utilizing convolutional and attention-based encoders, a co-attention module, and fully connected networks for interaction prediction.

Main Results:

• DeepBLI achieved an AUROC of 0.9240 and an AUPRC of 0.9715, outperforming state-of-the-art methods.
• The model successfully identified new beta-lactamase-inhibitor interactions.
• Demonstrated potential in screening inhibitors for metallo-beta-lactamase AIM-1 and repurposing rottlerin.

Conclusions:

• DeepBLI offers an effective computational approach for identifying beta-lactamase inhibitors.
• The model contributes to the development of novel therapeutics against antibiotic-resistant infections.
• DeepBLI shows promise for broad-spectrum inhibitor discovery and drug repurposing.