Machine learning reveals a new bacterial weak spot to antibiotics

Biotechnology and biochemistry

Feeding lots of experimental data into an open machine learning algorithm revealed that antibiotics create certain metabolic stress in bacteria. This knowledge could be used to stress bacteria further during treatment in order to enhance the effects of antibiotics. Also, this approach could reveal weak spots in cancer, diabetes, or neurodegenerative diseases.

Antibiotics often disrupt bacteria’s ability to replicate their DNA or in some way destroy the cell’s outer defense system – the cell wall. But in a new study involving extensive lab work and ‘white box’ machine learning, researchers found another mode of action for must antibiotics: They cause cellular stress, which means that the bacteria run low on certain DNA-building blocks.

“We hope to be able to use this knowledge to co-administer drugs that can selectively and more effectively kill bacteria responsible for infection,” says Postdoc Douglas McCloskey from the The Novo Nordisk Foundation Center for Biosustainability, co-author of the new study, which has just been published in Cell.

Cells try to boost production of building blocks

In response to this cellular stress, bacteria boost their production of certain building blocks for DNA called purines. These building blocks are essential for the cell, but purine production is very energy demanding. Hence, trying to boost purine production causes further cellular energy imbalances, which leads to cell death during antibiotic treatment.

"We hope to be able to use this knowledge to co-administer drugs that can selectively and more effectively kill bacteria responsible for infection"
Douglas McCloskey, Postdoc and Co-author

The researchers found this correlation by first doing extensive lab work in E. coli combining antibiotics with one of 200 different supplements including purines that could potentially improve offset the metabolic imbalance caused by antibiotics. For each combination, the researchers measured the effect on survival.

This approach is quite common. But the next step was completely novel. ‘Normally’ one would have used traditional machine learning to analyze the massive amounts of data using a ‘closed’ algorithm – a black box – which doesn’t allow researchers to know how the predictions were generated.

Open algorithm gave interesting results

Instead the researchers put the data into a genome-scale computer model for E. coli developed previously by Professor Bernhard Palsson and his team. This allowed them to study the drugs mechanisms of action. In this way they could go from correlation-based machine learning to causality-based ‘machine reasoning’.

The white-box approach led to some metabolic states within the bacteria. These states were then used as input for a black-box machine learning algorithm, which then revealed that purine synthesis was severely impaired by some antibiotics.

“This way of first using a model to calculate some basic metabolic states and then using machine learning to deduce some correlations between drug actions and cell response is completely new. But we expect this area to grow, because it provides greater interpretability than using a purely machine learning approach,” says Douglas McCloskey.

The researchers also expect that the white-box approach could be useful for studying how drugs affect diseases such as cancer, diabetes and neurogenerative diseases. Their hope is to find more weak spots that can be used to kill the harmful cells off even faster than current drugs allow. Now, the team is using a similar approach to study how the bacteria causing tuberculosis becomes resistant to antibiotic treatment.

The research was funded by the Novo Nordisk Foundation, the Defense Threat Reduction Agency, the National Institutes of Health, the Paul G. Allen Frontiers Group, the Broad Institute of MIT and Harvard, and the Wyss Institute for Biologically Inspired Engineering.

MIT research scientist Jason Yang is head-author of the paper. Authors from the Novo Nordisk Foundation Center for Biosustainability are Postdoc Douglas McCloskey, Senior Analytical Chemist Lars Schrübbers and CEO at the Novo Nordisk Foundation Center for Biosustainability and Galletti Distinguished Professor of Bioengineering at the University of California at San Diego, Bernhard Palsson.

Resources: Cell (2019): A White-Box Machine Learning Approach for Revealing Antibiotic Mechanisms of Action