Toward Autonomous Antibiotic Discovery

César de la Fuente‐NúñezmSystems 2019 · DOI

TL;DR

The antibiotic discovery pipeline is running dry, coinciding with a surge in multidrug-resistant bacteria. This paper proposes a paradigm shift, leveraging machine learning and computational design to autonomously discover and engineer novel antibiotics, moving beyond the limitations of natural sources.

What Problem Does This Paper Address?

For decades, the discovery of new antibiotics has slowed to a crawl, while the threat of antibiotic-resistant bacteria has escalated into a global health crisis. Historically, nature has been our primary source for these life-saving compounds, with organisms like fungi and bacteria producing a vast array of antimicrobial molecules. However, we appear to have reached a plateau, with newly discovered compounds often resembling existing ones and bacteria quickly developing resistance. This stagnation necessitates exploring new avenues for antibiotic discovery, moving beyond traditional methods of isolating compounds from natural sources. The authors argue that we may have exhausted readily accessible natural sources or lack the technology to fully interrogate complex biological systems.

Methods

The paper outlines two primary machine-driven approaches for antibiotic discovery:

  • Computer-Made Artificial Antibiotics: This involves "teaching" computers the fundamental principles of chemistry, physics, and biology that govern antibiotic production in nature. The goal is to enable machines to autonomously design novel molecular structures with antimicrobial properties, potentially unlike anything found in nature. This approach leverages the vast sequence space of proteins (composed of amino acids) as scaffolds for engineering. A genetic algorithm is described that uses machine-mediated evolution to enhance the antimicrobial activity of host defense peptides (HDPs). This algorithm simulates Darwinian principles of mutation, selection, and recombination, guided by a fitness function that selects for minimal structures capable of interacting with bacterial membranes.
  • Pattern Recognition-Driven Discovery: This method uses pattern recognition-like systems to scan biological systems, specifically protein databases, to identify novel antimicrobial molecules. By decrypting embedded amino acid patterns and leveraging existing knowledge about the molecular characteristics of HDPs, these systems can search for patterns resembling known antimicrobials. The identified "cryptic peptides" are then synthesized and experimentally validated for their antimicrobial activity and safety profiles.

Datasets

Not explicitly detailed in the paper, but the methods imply the use of:

  • Protein Databases: For pattern recognition-driven discovery, protein databases (e.g., from the human body) are explored.
  • Existing HDP Knowledge: Databases and literature on host defense peptides (HDPs) are used to inform pattern recognition and the design of computational models.
  • Experimental Data: For validation, synthesized peptides are tested against panels of foodborne pathogens and in animal models.

Key Insights and Ideas

  • Nature's Limits and the Virtual Frontier: The paper posits that while nature has been an incredible source of antibiotics, we may be reaching its limits. The virtual domain, powered by computation, offers an almost limitless frontier for discovering and designing new molecules.
  • Proteins as Versatile Scaffolds: The immense diversity and sequence space of proteins make them ideal candidates for engineering novel antimicrobial agents. The combinatorial space of even short peptides far exceeds the number of stars in the universe, highlighting the potential for undiscovered functionalities.
  • Machine-Mediated Evolution: Applying evolutionary principles to computational design can accelerate the optimization of antimicrobial properties, leading to potent and effective molecules.
  • Unlocking Cryptic Potential: Biological systems, like the human stomach, may harbor undiscovered antimicrobial peptides that can be identified through advanced pattern recognition and computational analysis.

Key Results

  • The described genetic algorithm enabled the machine-mediated evolution of HDPs, enhancing their antimicrobial activity.
  • Pattern recognition approaches successfully identified cryptic peptides from the human stomach that exhibited efficacy against foodborne pathogens, displayed excellent safety profiles, and were active in an animal model.
  • These computer-designed and identified molecules have the potential to become potent artificial antimicrobials effective both in vitro and in vivo.

Limitations and Future Work

  • Computational Predictions Require Validation: While computational approaches can identify promising candidates, experimental synthesis and rigorous testing are crucial for validation.
  • Need for Multidisciplinary Convergence: The paper emphasizes the need for collaboration across synthetic biology, chemistry, physics, and computer science to advance this field.
  • Future Directions: The authors aim to build the first class of machine-made artificial antibiotics within the next five years, focusing on combating infectious diseases and antibiotic resistance. They envision this multidisciplinary approach leading to groundbreaking insights and novel therapies that can outpace bacterial evolution.

Summary updated

Abstract

Machines hold the potential to replace humans in many societal endeavors, and drug discovery is no exception. Antibiotic innovation has been stalled for decades, which has coincided with an alarming increase in multidrug-resistant bacteria.
BibTeX
@inproceedings{fuentenunez2019toward,
  author = {César de la Fuente‐Núñez},
  title = {{Toward Autonomous Antibiotic Discovery}},
  booktitle = {mSystems},
  year = {2019},
  doi = {10.1128/msystems.00151-19},
}
Figure