Rifampicin resistance, predominantly brought about by missense mutations within the rpoB gene, presents a major therapeutic challenge, mainly in tuberculosis and leprosy - where rifampicin is used as backbone treatment, but also in P. aeruginosa and S. aureus infections - where it is reserved as a last-line agent. Current molecular diagnostics solely focus on the rifampicin resistance determining region to identify resistant cases, despite resistance mutations also being reported outside this region. To address this shortcoming, I have used protein structure to develop a resistance predictor with functionality across the whole gene.
The effects of M. tuberculosis variants (203 resistant, 28 susceptible) on protein stability, dynamics and interaction binding affinities were analysed using the mCSM-suite. This identified RNA polymerase complex destabilisation, with consequent reduction of nucleic acid affinity, as a predominant resistance mechanism. Based on these insights, we trained a predictive classifier using the KNN algorithm, and further validated it with a non-redundant M. tuberculosis test set (n=88). Our final tool, SUSPECT-RIF (StrUctural Susceptibility PrEdiCTion for RIFampicin), was built on both protein sequence- and structure-based features, where changes in nucleic acid affinity, wild type deformation energy and the rate of evolution were among the most important features for distinguishing between resistant and susceptible mutations.
SUSPECT-RIF correctly phenotyped 90.9% M. tuberculosis variants (n=319), compared to 42.6% phenotyped by the molecular gold-standard GeneXpert-MTB/RIF. Applying the model to clinical M. leprae, P. aeruginosa and S. aureus variants also showed 97.6%, 100% and 94.1% accuracy respectively, demonstrating clinical applicability to other mycobacterial, Gram-negative and Gram-positive infections. SUSPECT-RIF, is freely available as an interactive website on: http://biosig.unimelb.edu.au/suspect_rif/.