Software program aims to rapidly identify drug resistance from bacterial genome sequences

The program, which Oxford's Zamin Iqbal and his colleagues dubbed Mykrobe Predictor, uses de Bruijin graph representations of bacterial diversity to identify both species and resistance profiles of bacterial samples isolated from patients. The researchers used their tool to predict drug resistance in Staphylococcus aureus and Mycobacterium tuberculosis samples, finding that their approach was largely comparable to — and faster than — current gold-standard methods, as they reported today in Nature Communications.

Antibiotic resistance is increasingly becoming a public health issue as drug-resistant strains of bacteria like M. tuberculosis, S. aureus, Klebsiella pneumoniae, Pseudomonas aeruginosa, and others have cropped up, with some strains being resistant to multiple antibiotics.

"Antibiotics that were once lifesavers are in danger of becoming worthless," said Stephen Caddick, director of innovations at the Wellcome Trust, in a statement. "We urgently need new diagnostic strategies that allow us to better target antibiotic use, and thereby safeguard the effectiveness of our existing antibiotics, and any new drugs that are developed in future."

For their program, Iqbal and his colleagues first developed a curated reference of bacterial resistance and susceptible alleles, and assembled a de Bruijin graph on different bacterial backgrounds. This approach, the researchers argued in their paper, is unbiased by either reference choice or assumptions of sample clonality. Further, they said, it could easily be updated as more data on resistance and susceptibility is gathered.

Iqbal and his colleagues tested how well their approach could identify S. aureus and M. tuberculosis from within mixed samples sequenced on the Illumina MiSeq platform, and then gauge their drug resistance.

The researchers reported that their Mykrobe Predictor was able to correctly identify all of the S. aureus samples from within a validation set of 471 S. aureus isolates collected in the UK and 221 coagulase-negative staphylococci.

It did misclassify three non-S. aureus as S. aureus, but upon deeper examination, the researchers concluded that the samples were mislabeled in the National Center for Biotechnology Information's Short Read Archive since both Blast and OneCodex also concluded they were S. aureus.

For the seven drugs with more than 10 resistant samples, Mykrobe missed fewer resistant calls than the British Society for Antimicrobial Chemotherapy disc test and the Phoenix automated microbiology system approaches, the researchers reported. An exception, they added, was ciprofloxacin, which had a false-negative rate of 4.6 percent.

Overall, Iqbal and his colleagues reported that their method has a sensitivity of 99.1 percent and specificity of 99.6 percent across 12 antibiotics for S. aureus.

The researchers likewise found that Mykrobe was able to detect resistance to the first-line antibiotics rifampicin, isoniazid and ethambutol in tuberculosis similarly to the software program KvarQ, and with similar false-positive rates in a validation set of more than 1,600 samples. For instance, Mykrobe has power of 93.7 percent to detect rifampicin resistance while KvarQ has a power of 90.8 percent. The false positive rate for both approaches was 1 percent.

For tuberculosis, Iqbal and his colleagues reported that their method has sensitivity of 82.6 percent and specificity of 98.5 percent. They noted that the lower sensitivity is a function of the more limited understanding of the genetic mechanisms behind resistance in TB.

Iqbal and his colleagues also found the minor alleles helped distinguish multi-drug resistant from extensively drug-resistant TB, though they added that this needs to be tested in a larger dataset.

Mykrobe is a drag-and-drop environment that can be run as Windows or Mac applications, the researchers wrote. There is also a Linux version that could enable a cloud service. They added that it's been run on a laptop, Google Nexus 10 tablet, a Samsung Core Duos phone, and a Raspberry Pi Model B.

A clinical implementation of Mykrobe, they said, could reduce the time to when clinicians would know what drugs their patients are resistant to. They estimated that using a 16-and-a-half hour Illumina MiSeq run, their workflow would give a full set of resistance predictions in about 36 hours, some 12 hours faster than the clinical protocols in place at Oxford University Hospitals.

For the slow-growing tuberculosis, results via Mykrobe would be available in about two weeks, as compared to the five weeks to 17 weeks using standard approaches, they added.

"One of the barriers to making whole genome sequencing a routine part of [UK National Health Service] care is the need for powerful computers and expertise to interpret the masses of complex data," Iqbal said in a statement. "Our software manages data quickly and presents the results to doctors and nurses in ways that are easy to understand, so they can instinctively use them to make better treatment decisions."

Mykrobe also appears to work with other sequencing approaches. Iqbal and his colleagues tested whether Mykrobe could handle reads Oxford Nanopore's MinIon platform. With its small size, they noted that sequencer might be best suited for testing in the field.

After adjusting for higher, per-base error rates, their proof-of-principle test shows that Mykrobe was able to correctly predict that a multidrug resistant S. aureus sample was resistant to penicillin, methicillin, trimethoprim, and other drugs, while noting it was susceptible to tetracycline, vancomycin, and a few others.

Currently, the tool is being piloted at hospitals in three UK cities for three months.

"Our ultimate goal is to be able to provide complete information on a pathogen within 24 hours of culture, linking this information to a national surveillance database to enable more timely and better targeted patient treatment," added co-author Derrick Crook, the director of the National Infection Service of Public Health England.

Source: GenomeWeb