Traditional antibiotics primarily target essential bacterial functions like DNA replication or cell wall synthesis. However, these mechanisms only scratch the surface of how antibiotics combat bacterial infections.
In groundbreaking research, MIT scientists have pioneered an innovative machine learning approach that reveals an additional mechanism behind how certain antibiotics eliminate bacteria. This supplementary mechanism involves stimulating the bacterial metabolism of nucleotides, which are crucial for DNA replication within bacterial cells.
“When bacteria experience drug-induced stress, they face enormous energy demands. Meeting these energy requirements triggers a metabolic response, and some metabolic byproducts become toxic, ultimately contributing to bacterial cell death,” explains James Collins, the Termeer Professor of Medical Engineering and Science at MIT's Institute for Medical Engineering and Science (IMES) and Department of Biological Engineering, and the study's senior author. Collins also serves as the faculty co-lead of the Abdul Latif Jameel Clinic for Machine Learning in Health.
Researchers suggest that exploiting this mechanism could pave the way for developing new drugs that work in conjunction with antibiotics to boost their effectiveness.
Jason Yang, an IMES research scientist, leads the paper published in the May 9 issue of Cell. Additional contributors include Sarah Wright, a recent MIT MEng graduate; Meagan Hamblin, former Broad Institute research technician; Miguel Alcantar, MIT graduate student; Allison Lopatkin, IMES postdoc; Douglas McCloskey and Lars Schrubbers from the Novo Nordisk Foundation Center for Biosustainability; Sangeeta Satish and Amir Nili, recent Boston University graduates; Bernhard Palsson, UC San Diego bioengineering professor; and Graham Walker, MIT biology professor.
Transparent AI modeling approach
Collins and Walker have extensively studied antibiotic mechanisms for years. Their previous work demonstrated that antibiotic treatment creates significant cellular stress, imposing substantial energy demands on bacterial cells. In this new study, Collins and Yang employed machine learning to investigate this process and its consequences.
Prior to computer modeling, the team conducted hundreds of experiments using E. coli. They exposed the bacteria to one of three antibiotics—ampicillin, ciprofloxacin, or gentamicin—combined with approximately 200 different metabolites, including amino acids, carbohydrates, and nucleotides (DNA building blocks). For each combination, they measured the effects on cell survival.
“We applied diverse metabolic perturbations to understand how disrupting nucleotide metabolism, amino acid metabolism, and other metabolic subnetworks affects antibiotic efficacy,” Yang states. “Our goal was to fundamentally identify previously unrecognized metabolic pathways crucial for understanding antibiotic lethality.”
While many researchers have utilized machine learning models to analyze biological experimental data, these models typically function as “black-boxes,” obscuring the mechanisms underlying their predictions.
To overcome this limitation, the MIT team developed a novel approach they term “white-box” machine learning. Rather than directly feeding data into a machine-learning algorithm, they first processed it through a genome-scale computer model of E. coli metabolism, previously characterized by Palsson's laboratory. This generated various “metabolic states” described by the data. They then input these states into a machine-learning algorithm, which identified connections between different states and antibiotic treatment outcomes.
Since the researchers knew the experimental conditions producing each state, they could determine which metabolic pathways were responsible for increased cell death.
“Our approach demonstrates that by having network simulations first interpret the data and then having machine learning build a predictive model for antibiotic lethality phenotypes, the elements selected by the model directly map onto pathways we've experimentally validated—which is very exciting,” Yang explains.
Markus Covert, Stanford University associate professor of bioengineering, notes that this study represents significant progress in showing how machine learning can uncover biological mechanisms connecting inputs and outputs.
“Biology, particularly for medical applications, fundamentally revolves around mechanism,” says Covert, who wasn't involved in the research. “You want to find something druggable. For typical biologists, discovering connections between inputs and outputs without understanding why they're linked hasn't been particularly meaningful.”
Metabolic stress implications
This model revealed that nucleotide metabolism, particularly purine metabolism like adenine, plays a crucial role in antibiotics' ability to eliminate bacterial cells. Antibiotic treatment induces cellular stress, depleting purine nucleotides. The cells' attempts to increase production of these DNA-essential nucleotides elevate overall metabolism, leading to harmful metabolic byproduct accumulation that can kill the cells.
“We now believe that in response to severe purine depletion, cells activate purine metabolism to compensate. However, purine metabolism itself is energetically expensive, amplifying the energy imbalance cells already face,” Yang elaborates.
These findings suggest that enhancing certain antibiotics' effects might be possible by administering them alongside drugs that stimulate metabolic activity. “If we can push cells into a more energetically stressed state and induce increased metabolic activity, we might be able to potentiate antibiotics,” Yang proposes.
The researchers believe this “white-box” modeling approach could also prove valuable for studying how different drug types affect diseases such as cancer, diabetes, or neurodegenerative conditions. They're currently applying a similar approach to study how tuberculosis survives antibiotic treatment and develops drug resistance.
The research received funding from the Defense Threat Reduction Agency, the National Institutes of Health, the Novo Nordisk Foundation, the Paul G. Allen Frontiers Group, the Broad Institute of MIT and Harvard, and the Wyss Institute for Biologically Inspired Engineering.
