A deep reinforcement learning platform for antibiotic discovery.

Area of Science:

• Computational biology
• Drug discovery
• Infectious diseases

Background:

• Antimicrobial resistance (AMR) poses a significant global health threat, projected to cause 10 million deaths annually by 2050.
• The urgent need for novel antibiotics drives research into innovative drug discovery platforms.

Purpose of the Study:

• To present ApexAmphion, a deep learning framework for the de novo design of peptide antibiotics.
• To demonstrate the framework's ability to rapidly generate potent and broad-spectrum antimicrobial compounds.

Main Methods:

• Coupling a large protein language model with reinforcement learning for antibiotic design.
• Fine-tuning the model on antimicrobial peptide data and optimizing with proximal policy optimization.
• Utilizing a composite reward function integrating minimum inhibitory concentration (MIC) prediction and physicochemical properties.

Main Results:

• In vitro evaluation of 100 designed peptides demonstrated a 100% hit rate with low MIC values (nanomolar range).
• 99% of designed compounds exhibited broad-spectrum activity against clinically relevant bacteria.
• Lead molecules effectively targeted the cytoplasmic membrane for bacterial killing.

Conclusions:

• ApexAmphion offers a scalable and rapid route for de novo peptide antibiotic discovery.
• The framework integrates generation, scoring, and multi-objective optimization within a single pipeline.
• This platform facilitates iterative design for enhanced potency and developability of novel antibiotics.