This guest post is by Zamin Iqbal [@ZaminIqbal] and Phelim Bradley [@Phelimb]
Our paper “Rapid antibiotic resistance predictions from genome sequence data for S. aureus and M. tuberculosis” has just appeared on the Biorxiv. We’re excited about it for a number of reasons.
The idea of using a graph of genetic variation as a reference, instead of a linear genome, has been discussed for some while, and in fact a previous biorxiv preprint of ours applying them to the MHC has just come out in Nature Genetics:
In this paper we apply those ideas to bacteria, where we let go of the linear coordinate system in order to handle plasmid-mediated genes. Our idea is simple – we want to see if genomic data can be used to predict antibiotic resistance in bacteria, and we explicitly want to build a general framework that will extend to many species, and handle mixed infections.
The paper does not deal with the issue of discovering mechanisms/mutations/genes which drive drug resistance – we take a set of geno-pheno rules as prerequisite, and then use a graph of resistance mutations and genes on different genetic backgrounds to detect presence of alleles and compare statistical models – is the population clonal susceptible, minor resistant or major resistant? Although it is accepted that minor alleles can sweep to fixation, in general there is neither consensus nor quantitative data on the correlation between allele frequency and in vitro phenotypic resistance or patient outcome (the latter obviously being much harder). At a practical level,in some cases a clinician might avoid a drug if they knew there was a 5%-frequency resistance allele, and in others they might increase the dose. Resistance is of course a quantitative trait, often measured in terms of the minimum concentration of a drug required to stop growth of a fixed inoculum – but commonly a threshold is drawn and samples are classified in a binary fashion.
A paper last year from some of us (http://jcm.asm.org/content/52/4/1182.full) showed that a simple panel of SNPs and genes was enough to predict resistance with high sensitivity and specificity for S. aureus (where SNPs, indels, chromosomal genes and plasmid-mediated genes can all cause resistance) – once you discard all samples with any mixed strains. (Standard process is to take a patient sample and culture “overnight” (12-24 hours), thus removing almost all diversity and samples which show any morphological signs of diversity after culture are discarded or subcultured). By contrast, for M. tuberculosis (which causes TB), known resistance mutations explain a relatively low proportion of phenotypic resistance (~85%) for first-line drugs, and even less for 2nd line (I explain below what 1st/2nd line are). The Mtb population within-host is highly structured and multiple genotypes can evolve in different loci within the body, so it’s important to be able to deal with mixtures. Typical phenotyping relies on several weeks of solid culture (Mtb is slow growing), but mixtures are more able to survive this type of culture than in the case of S. aureus.
We show with simulations that we can use the graph to detect low frequency mutations and genes (no surprise), and that for S. aureus we make no minor calls for our validation set of ~500 blood-cultured samples (no surprise). Each sample is phenotyped with 2 standard lab methods, and where they disagree a higher quality test is used to arbitrate. This consensus allows us to estimate error rates both for our method (called Mykrobe predictor) and for the phenotypic tests. As a result we’re able to show not only that we do comparably with FDA requirements for a diagnostic, but also that we match or beat the common phenotypic tests.
On the other hand for TB, the story is much more complex and interesting. We analyse ~3500 genomes in total, split into ~1900 training samples and ~1600 for validation. For M. tuberculosis, a sample is classed as resistant if after some weeks of culturing under drug pressure, the number of surviving colonies is >1% of the number of colonies from a control strain treated identically – the number 1% is of course arbitrary (set down by Canetti in the 1960s I think), though it has been shown that phenotypic resistance does correlate with worse patient outcome. Sequencing on the other hand is done before the drug pressure, so we are fundamentally testing a different population, and we can’t simply mirror that 1% allele frequency expectation. This is what we use the 1900 training samples for – determining what frequency to set for our minor-resistant model. We ended up using 20%, and also found that there
was an appreciable amount of lower frequency resistance, which did not survive the 6-week drug-pressure susceptibility test, but which might cause resistance in a patient.
Mtb infections can last a long time, and despite their slow growth, the sheer number of bacilli in a host result in a vast in-host diversity. As a result, mono therapies fail, as resistant strains sweep to fixation – standard treatment is therefore with 4 “first-line” drugs, reducing the chance that any strain has enough mutations to resist them all. If the first-line drugs fail, or if the strain is known to be resistant, then it is necessary to fall back to more toxic and less effective second-line drugs. We found, somewhat to our surprise, that
1. Overall, minor alleles contribute very little to phenotypic resistance in first-line drugs, but they do make a significant contribution to second-line drugs, improving predictive power by >15%. This matches previous reports that patient samples had mixed R and S alleles for 2nd line drugs. This could have major public health consequences, as resistance to these drugs needs to be detected to distinguish MDR-TB (resistant to isoniazid, rifampicin) from XDR-TB (isoniazid, rifampicin + second-line), a major concern for the WHO.
2. Interestingly, a noticeable number of rifampicin false-positive calls were due to SNPs which confer resistance but have been shown to slow growth. Since the phenotyping test is intrinsically a measure of relative growth, these strains may be misclassified as susceptible – i.e. these are probably false-susceptible calls due to an artefact of the nature of the test. This has been reported before by the way.
Anyway – please check out the paper for details. We think this large-scale analysis of whether minor alleles contribute to in vitro phenotype, and whether they should be used for prediction is new and interesting both scientifically and in terms of translation. The bigger question is what the consequences are for patient outcome, and how to deal with in-host diversity, and for that we of course need data collection and sharing. We’ve spent a lot of time in the Oxford John Radcliffe Hospital working with clinicians, and trying to determine what information they really need from this kind of predictive test, and we’ve produced both Windows/Mac apps with very simple user-interfaces (drag-the-fastq on, and let it run) for them to use; we’ve also produced an Illumina Basespace app, currently submitted to Illumina for approval, which should enable automated cloud-use.
Our paper also has a whole bunch of work I’ve not mentioned here, where we needed to identify species, and detect contaminants – most interesting when common contaminants can contain the same resistance gene as the species under test.
Our software is up on github here
including some desktop apps and example fastq files so you can test it.
Comments very welcome!
Zam and Phelim
PS By the way, the 4 first-line drugs have different effectiveness in different body compartments – see this interesting paper for the modelling of the consequences: http://biorxiv.org/content/early/2014/12/19/013003.