Generative Deep Learning Pipeline Yields Potent Gram-Negative Antibiotics.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Multiresistant bacteria pose a significant global health threat, necessitating new antibiotics.
• Current antibiotic discovery is limited by the biased chemical space of existing libraries.

Purpose of the Study:

• To develop a deep learning (DL) pipeline for *de novo* antibiotic design.
• To generate structurally novel antibiotic candidates beyond current chemical limitations.

Main Methods:

• Utilized a chemical language model trained on diverse drug-like molecules and natural products.
• Applied transfer learning with antibiotic scaffolds to refine generative capabilities.
• Employed predictive modeling and expert curation for compound prioritization and synthesis.

Main Results:

• Identified a lead compound with potent activity against methicillin-resistant *Staphylococcus aureus* (MRSA).
• Synthesized 40 derivatives, yielding 30 active against *S. aureus* and 17 against *Escherichia coli*.
• Compound D8 demonstrated submicromolar potency against MRSA and single-digit micromolar potency against *E. coli*.

Conclusions:

• Validated a DL pipeline for exploring chemical space to generate novel antibiotic candidates.
• Discovered a potent nitrofuran derivative (D8) with a unique mechanism of action (reductive generation of reactive species).
• Generated experimentally validated scaffolds to accelerate future antibiotic development.