Dear HIV-TRePS user / RDI subscriber,
The RDI would be grateful if you could help us raise awareness of HIV-TRePS by introducing colleagues to the system and forwarding the following information about HIV-TRePS to all those in your network you feel could find it useful.
We are writing to you today, World AIDS Day, to tell you about the latest advance with the HIV Treatment Response Prediction system - HIV TRePS. We have updated the two sets of models that make predictions of virological response for cases where a genotype is available and for those without a genotype.
Models that do not require a genotype
These models were trained to predict the probability of virological response to therapy following virological failure using 50,270 treatment change episodes (TCEs) without a genotype from around the world. In order to make HIV-TRePS better suited for use in low-income and rural settings, where clinic visits may be infrequent, the baseline data inclusion criteria were relaxed from 8 weeks to 16 weeks before treatment change for viral loads and from 12 to 24 weeks for CD4 counts. The models were tested with an independent global test set (n=3,000). The models achieved an area under the ROC curve (AUC) value of 0.81 and overall accuracy of 75%. There was no significant difference in accuracy between these models and standard models trained and tested with the subset of data that conformed to the previous, stricter baseline data criteria. These ‘no genotype’ models are also the first to be able to predict responses to regimens including elvitegravir, maraviroc and tipranavir.
Models that require a genotype
In the development of these models we used the relaxed baseline data criteria established above, with genotypes permitted up to 16 weeks prior to therapy change. We tested a new filter for excluding cases that were likely to have been from non-adherent patients. Candidate TCEs where the introduction of a new regimen was followed by virological failure in the clinic were identified. Any that were predicted by our current models to respond and had a GSS score of ≥2 (on all three genotype interpretation systems: ANRS, REGA and Stanford) were excluded due to presumed non-adherence.
Global models were trained with the resulting 17,378 TCEs and tested with 940 independent cases, selected using the same filter. These models were compared to those developed with data using the standard filter (baseline viral load of ≤ 3.0 log10 HIV RNA and an increase in viral load of ≥2.0 following the introduction of a new regimen selected with a recent genotype available). The new models obtained an AUC in independent testing of 0.86 and overall accuracy of 79%, compared to 0.84 and 76% for the standard models.
These two new sets of models have the potential to help optimise therapy particularly in resource-limited settings. We hope this development enhances your use of HIV-TRePS.
A press release covering the introduction of these models is available and a manuscript is being submitted to a peer-review journal for publication:
The RDI team.
Date published: 1st December 2017