MIT Researchers Unveil 'White-Box' Machine Learning Method Revealing Metabolic Mechanism That Enhances Antibiotic Killing

Most antibiotics work by interfering with core functions, such as DNA replication or the construction of bacterial cell walls. However, these represent only part of how antibiotics exert their effects.

In a new study on the mechanisms of antibiotic action, researchers at MIT developed a novel machine learning approach that uncovered an auxiliary mechanism helping antibiotics kill bacteria. This mechanism involves activating bacterial metabolism to produce nucleotides required for DNA replication.

James Collins, a professor in the Department of Biological Engineering at the MIT Institute for Medical Engineering and Science and the senior author of the study, stated, “Due to drug-induced stress, cells require substantial amounts of energy. These energy demands necessitate a metabolic response that generates toxic metabolites, which contribute to cell death. Leveraging this mechanism can help us identify new drugs that can be used in combination with antibiotics to enhance their bactericidal efficacy.”

Published in Cell on May 9, the paper lists Jason Yang as its lead author. Other co-authors include Sarah Wright, a recent recipient of an MEng award from MIT; Meagan Hamblin, a former researcher at the Broad Institute; Miguel Alcantar, a graduate student at MIT; Allison Lopatkin, an IMES postdoctoral fellow; Douglas McCloskey and Lars Schrubbers from the Novo Nordisk Foundation Center for Biosustainability; Sangeeta Satish and Amir Nili, both recent graduates of Boston University; Bernhard Palsson, Professor of Bioengineering at the University of California, San Diego; and Graham Walker, Professor of Biology at MIT.

For years, Collins and Walker have been investigating the mechanisms of antibiotic action. Their research has shown that antibiotic treatment often induces substantial cellular stress, thereby imposing significant energy demands on bacterial cells. In this new study, Collins and Yang decided to employ a “machine learning approach” to investigate how this process occurs and what its consequences are.

Before initiating computer modeling, the researchers conducted hundreds of experiments in Escherichia coli. In each experiment, they treated the bacteria with one of three antibiotics—ampicillin, ciprofloxacin, or gentamicin—and added one of approximately 200 different metabolites, including a range of amino acids, carbohydrates, and nucleotides (the building blocks of DNA). For each combination of antibiotic and metabolite, they also measured its impact on cell survival.

Yang stated, “We employed various metabolic perturbations to observe their effects on nucleotide metabolism, amino acid metabolism, and other metabolic subnetworks, aiming to fundamentally elucidate which previously undescribed metabolic pathways may be critical to our understanding of how antibiotics kill cells.” Many other researchers use machine learning models to analyze data from biological experiments by training algorithms to generate predictions based on experimental data. However, these models typically operate as “black boxes,” meaning they do not reveal the underlying mechanisms driving their predictions.

To address this issue, the MIT team adopted a novel approach they term “white-box” machine learning. Rather than feeding data directly into machine learning algorithms, they first ran it through a genome-scale computational model of Escherichia coli metabolism, as described by the Palsson Laboratory, to generate an array of “metabolic states” that could be characterized by data. They then fed this array of states into a machine learning algorithm capable of identifying associations between different states and outcomes of antibiotic treatment. Because the experimental conditions for each state were already well understood, the researchers were able to determine which metabolic pathways led to higher levels of cell death.

He further stated, “We demonstrated a network-based simulation of the data, employed machine learning algorithms to construct predictive models for antibiotic lethal phenotypes, and directly mapped the features selected by these predictive models onto established pathways. We have experimentally validated this approach, which is indeed an encouraging development.”

The model has yielded a new finding: nucleotide metabolism, particularly purine metabolism, plays a critical role in the ability of antibiotics to kill bacterial cells. Antibiotic treatment induces cellular stress, leading to impaired utilization of purine nucleotides. Enhancing overall cellular metabolism results in the accumulation of toxic metabolites, thereby causing cell death.

Yang stated, “We believe that to counteract this severe purine depletion, cells automatically activate purine metabolism to address the issue. However, since purine metabolism is itself energetically costly, this further exacerbates the impact of cellular energy imbalance.” The research findings indicate that co-administering certain antibiotics with other drugs that stimulate metabolic activity may enhance the efficacy of some antibiotics. He further added, “If we can shift cells into a more dynamic stress state and induce them to increase metabolic activity, this could represent one approach to potentiating antibiotic effects.”

The researchers stated that the “white-box” modeling approach used in this study could also be applied to investigate how different types of drugs affect diseases such as cancer, diabetes, or neurodegenerative disorders. They are now employing a similar strategy to study how tuberculosis survives antibiotic treatment and develops drug resistance.

(Compiled by: Guo Xiaodong)