December 17, 2005
Although resistance testing cannot reliably predict the virologic success of HIV treatment regimens, modern computing technology may be able to fill this gap, some researchers say. A study presented at ICAAC 2005 focused on the use of artificial neural networks to achieve this goal.
The Downsides to Resistance Testing
It can be difficult to choose a new antiretroviral regimen for a patient with drug-resistant HIV who has a history of many prior regimens. One reason is that genotypic and phenotypic resistance testing only examine the predominant strains of HIV in the blood. If someone has a history of resistance to prior antiretroviral regimens, there are often stored, or "archived," strains of HIV that may contain other mutations. These archived strains may compromise the effectiveness of a new regimen even though a resistance test would suggest that the regimen should work.
Of course, there are other, well-known reasons for why genotypic and phenotypic resistance tests have their drawbacks. For example, genotypes can be quite difficult to interpret. There are often many relevant mutations that can affect the sensitivity of an antiretroviral medication. Often these mutations interact, and it is difficult to create simple algorithms that will conclusively say whether a medication is likely to be effective. By contrast, phenotypes are simpler to interpret, but they are much more expensive and are less likely to yield a result when a patient's viral load is low. Finally, although both tests are capable of suggesting whether individual antiretroviral medications are likely to be sensitive, neither can evaluate the potential potency of a combination of medications.
The Promise of Artificial Intelligence
Artificial neural networks are being investigated as one method by which clinicians can evaluate the potency of a potential regimen in HIV treatment-experienced patients. At this conference, Brendan Larder, of the HIV Resistance Response Database Initiative in London, presented a study on the use of this technique. The research team involved in this study also included scientists from the British Columbia Centre for Excellence in HIV/AIDS, the U.S. Military HIV Research Program and the U.S. National Institute of Allergy and Infectious Diseases.
An artificial neural network is able to group a wide range of patient-specific data and establish complex relationships that cannot be elucidated by typical analytic techniques normally found in clinical studies. By doing so, the artificial neural network can be used to predict how other patients will respond to new antiretroviral regimens. These investigators have already shown that artificial neural networks can be used to predict virologic responses to a combination regimen based on a genotype.1,2
The investigators "train" artificial neural networks by inputting a large data set of "treatment change episodes" taken from various cohort studies of HIV-infected people. A treatment change episode consists of a variety of data on a particular patient, including the patient's genotype prior to starting a new group of antiretroviral medications, the drugs actually used, the baseline viral load and the resulting change in viral load after starting the new medications. The artificial neural network then uses thousands of these treatment change episodes to effectively train itself into becoming a resistance expert.
As I noted earlier, a genotype does not provide a complete picture of resistance in a patient. Consequently, the artificial neural network may be improved (i.e., made to more accurately predict the change in viral load on a new regimen) if information about archived mutations is included in the data set of treatment change episodes. The point of this study was to test whether the artificial neural network could be improved by incorporating treatment history, prior genotype results or both. The investigators found that the artificial neural network performed better after including treatment history in the treatment change episodes data sets, but was not improved after adding information from prior genotypes. Interestingly, they also found that adding baseline CD4+ cell count data improved the artificial neural network.
What Does the Future Hold for Artificial Neural Networks?
Clinically, an artificial neural network would probably work as an interactive, Web-based tool. The physician would enter a patient's treatment history, baseline CD4+ cell count, baseline viral load and the results of a genotype. The artificial neural network would then recommend the top 10 combinations of antiretroviral medications and the corresponding expected change in viral load for each. The clinician would then use these recommendations to select a new antiretroviral regimen.
Artificial neural networks could also help select an optimized background medication when starting a new drug, such as tipranavir (TPV, Aptivus) or TMC114, that was not included in the treatment change episodes data sets. The artificial neural networks could be updated on a continuous basis by inputting more treatment change episodes that incorporate new agents and new combinations of agents.
The artificial neural networks used in this study are being developed by a nonprofit British organization called the HIV Resistance Response Database Initiative that is planning to make the tool available free of charge. Clinical trials are planned that will test whether better antiretroviral responses are achieved when choosing antiretroviral medications based on artificial neural network results.
For now, these artificial neural networks are still only a research tool. But they have an exciting potential to simplify the complex interpretation of genotypes and help clinicians improve virologic responses among highly treatment-experienced patients.