skip to content

Big Data used to tackle antibiotic resistance

Computer-designed combination treatment takes on antibiotic resistant bacteria.

Dr Daniel Mason, a former postdoc in the Bender Group, has developed a machine learning approach that enables the prediction of more effective antibiotic combination treatments, potentially speeding up the search for new ways to fight antibiotic resistant bugs.

The overuse of antibiotics in livestock, unnecessary prescription in humans and the fact that no new antibiotic classes have been discovered in the past 40 years mean our arsenal of treatments are gradually succumbing to multi-drug resistant organisms.

Daniel, who developed the computer-designed system during his time at the Center for Molecular Informatics, says: “Being able to predict which compounds are likely to produce a synergistic antimicrobial effect when combined together has the potential to greatly reduce time and resources spent on experimental screening for new treatments.”

The work published in the Journal of Medicinal Chemistry shows that once trained using a dataset of 153 pairs of antibiotics screened against E. coli, the method is able to classify synergistic combinations largely correctly when testing on an unseen dataset. This is very significant in practical terms as screening a library of just 50 compound pairs would result in over 1,200 combinations. “The rate of prediction corresponds to a 2.8-fold increase in the discovery of synergistic combinations over that expected by a brute-force screening effort,” says Daniel. The research has resulted in a list of 4,950 combinations of 100 compounds of which 691 are now the starting point for designing future antibiotic combination treatments.

The computer-designed approach was developed in collaboration with the research group of Dr Murat Cokol at Sabanci University in Turkey and Harvard University in the United States and funded by Unilever and the ERC.

Image caption: Using chloramphenicol, 5-fluorouracil and nalidixic acid, this figure symbolizes the description of individual antibiotics by fingerprint features (displayed in different colours and shapes) on the left, and their bitstring representations (which correspond to those features)  in the centre. Based on this input the model is able to model and predict compound synergy and antagonism, which is derived from experimental combination screening data as visualized on the right. Subsequently, the predictive model has been validated in another combination screening experiment to predict synergistic combinations of novel antibiotics.


Daniel J Mason, Prediction of antibiotic interactions using descriptors derived from compound molecular structure, J. Med. Chem., April 6 2017,
DOI: 10.1021/acs.jmedchem.7b00204