Combination antiretroviral therapy is capable of suppressing HIV replication and preventing disease progression for a number of years. However in most patients, at some point there will be a loss of viral control often involving the development of drug resistance, which will require a change to the drug regimen. The selection of drugs for the next regimen is made from a list of around 20 or so from six classes. In well-resourced settings, the selection of drugs is commonly informed by genotypic antiretroviral resistance testing (GART) indicating to which drugs the virus should still be sensitive. However this is expensive, can miss resistant virus that exists at low levels and requires sophisticated interpretation. There are several interpretation systems in common use but they do not always agree with each other and will only indicate to which individual drugs the virus might be resistant or sensitive, not give a relative indication of the effectiveness of different combinations. Critically, resistance testing is beyond the reach or budget of many healthcare providers around the world, making the individual optimisation of salvage therapy a major challenge.
The HIV Resistance Response Database Initiative (RDI), an independent not-for-profit research group, has developed computational models to predict virological response to a change of antiretroviral therapy following virological failure. The models are trained using real treatment response data from hundreds of physicians, treating hundreds of thousands of patients around the world. Currently the RDI database holds treatment and outcome data on over 250,000 patients, making it one of the largest HIV treatment databases in the world.
Early models used viral loads, CD4 counts, treatment history and resistance mutations from GART to predict response. In dozens of studies over many years these models have regularly demonstrated approximately 80% accuracy when tested with independent data sets and have been evaluated by expert HIV clinicians as being a useful clinical tool.1-3
The RDI has also developed models that do NOT require GART, for use in settings where it is unaffordable.3 The latest versions of these models have shown similar accuracy to modelling with GART in cross study comparisons and have consistently demonstrated significantly superior accuracy to the use of genotypic sensitivity scores derived from GART using rules-based interpretation. 3,4
Most recently, we have developed models that predict the absolute change in viral load over time, following the introduction of a new regimen, for use in settings where the definition of response is greater than <50 copies. The predictions of these models correlate with the actual changes over time with a coefficient of around 0.7.5
The RDI’s models are used to power the RDI’s HIV Treatment Response Prediction System, or HIV-TRePS, which is available free of charge over the Internet. Healthcare professionals can use the system to:
In a pilot study using cases of salvage therapy at a hospital in India, the models identified regimens with a higher probability of response and substantially lower costs than those used in the clinic at for the great majority of cases.6