Accelerated antimicrobial discovery via deep generative models and molecular dynamics simulations.

Area of Science:

• Computational chemistry
• Drug discovery
• Microbiology

Background:

• Designing novel antimicrobials requires exploring extensive chemical libraries for broad-spectrum efficacy and minimal toxicity.
• Existing methods for antimicrobial discovery are often time-consuming and may not yield compounds with desired attributes.

Purpose of the Study:

• To develop and validate an efficient computational method for the de novo design of antimicrobial therapeutics.
• To identify novel antimicrobial peptides with broad-spectrum activity, low toxicity, and reduced potential for resistance development.

Main Methods:

• Utilized a deep generative autoencoder to model molecular latent spaces.
• Employed deep learning classifiers and molecular dynamics simulations for screening generated molecules.
• Synthesized and experimentally tested candidate antimicrobial peptides.

Main Results:

• Identified 20 candidate antimicrobial peptides within 48 days.
• Two peptides demonstrated high potency against Gram-positive and Gram-negative pathogens, including multidrug-resistant Klebsiella pneumoniae.
• The identified peptides exhibited low toxicity in vitro and in vivo (mice) and low propensity to induce resistance in Escherichia coli.
• Live-cell confocal imaging revealed that the peptides act by forming membrane pores.

Conclusions:

• The integrated approach of deep learning and molecular dynamics significantly accelerates the discovery of novel antimicrobials.
• The identified peptides represent promising candidates for broad-spectrum antimicrobial therapeutics.
• This computational strategy can guide the development of next-generation antimicrobials.