There a number of limitations to using the genotype test alone to help predict how a patient will respond to a combination of antiretroviral drugs. There are many different genotype interpretation systems available that do not always agree, the systems only provide categorical results (‘resistant’, ‘sensitive’ or ‘intermediate’) and they do not necessarily relate well to how a patient will respond to a combination of drugs in the clinic.
The RDI was established in 2002 to pioneer a new approach: to develop computational models using the genotype and a wide range of other clinically relevant data collected from thousands of patients treated with HAART all over the world and to use these models to make a quantitative prediction of how an individual patient will respond to different combinations of drugs. The RDI’s goal was to make available a free treatment-response prediction tool over the Internet.
As of October 2010, the RDI has collected data from approximately 70,000 patients from dozens of clinics in more than 15 countries. It is probably the largest database of its kind in the world. The data includes demographic information for the patient, and multiple determinations of the amount of virus in the patient’s bloodstream, CD4 cells counts (a white blood cell critical to the function of the immune system that HIV targets and destroys), genetic code of the patients virus, and details of the drugs that have been used to treat the patient.
The RDI has used these data to conduct extensive research in order to develop the most accurate system possible for the prediction of treatment response. This research involved the development and comparison of different computational modelling methods including artificial neural networks, support vector machines, random forests and logistical regression [1].
The predictions of the RDI’s models have historically correlated well with the actual changes in virus load of patients in the clinic, typically achieving a correlation co-efficient of 0.8 or more [2]. The most recent models have predicted whether a combination treatment will reduce the level of virus in the patient’s bloodstream to undetectable levels with an accuracy of approximately 80%, significantly better than just using a genotype with rules-based interpretation.
HIV and its treatment in resource-limited settings
Most HIV is found in relatively poor countries where healthcare resources are severely limited. For example two-thirds of the 33 million people estimated to be living with HIV in 2009 are in sub-Saharan Africa. In such settings there is a shortage of drugs and often there is not the budget to pay for viral load tests let alone the genotype tests. This makes it very difficult to optimise treatment and select effective new combinations of drugs when a treatment fails.
As clinics in resource-limited settings are often unable to afford genotyping, the RDI has developed models that predict treatment response without the need for a genotype, with only a small loss of accuracy [3].
References
- A comparison of three computational modelling methods for the prediction of virological response to combination HIV therapy. Artificial Intelligence in Medicine (2009) 47:63-74
http://www.hivrdi.org/pdf/RDI-AIIM-Article-Aug09.pdf
- The development of artificial neural networks to predict virological response to combination HIV therapy. Antiviral Therapy (2007); 12:15-24
http://www.intmedpress.com/serveFile.cfm?sUID=1a8ea006-54e4-4f45-a2fa-133d8a37813a
- Modelling response to HIV therapy without a genotype: an argument for viral load monitoring in resource-limited settings. J. Antimicrob. Chemother. (2010) 65(4): 605-607
http://jac.oxfordjournals.org/content/65/4/605.full
