Artificial intelligence identifies effective drugs for HIV patients
whose treatment is failing
CABO, MEXICO - Thursday 12th June, 2003. New data
presented for the first time today at the 12th International Workshop
on HIV Drug Resistance demonstrated that artificial intelligence (AI)
could find effective treatments for patients whose drug therapy is failing.
The system identified potentially effective drug combinations for patients
who were continuing to fail on therapy, despite having their combinations
of HIV drugs changed by their physicians according to current clinical
practice.
"These patients had high viral loads and were failing because of
drug resistance, despite multiple changes to their treatment and the
use of current resistance tests", commented Professor Julio Montaner
MD, Professor of Medicine and Chair in AIDS Research, at the BC Centre,
University of British Columbia, Canada. "Today's results hold out
the possibility of being able to reverse the process of treatment failure
for such patients, using artificial intelligence to help us identify
the best possible drug combination for the individual."
A 'Neural Network' mathematical model that predicts how patients will
respond to different combinations of drugs with 79% accuracy was fed
the genetic codes of the HIV from these failing patients. In every case
it identified an alternative drug combination that it predicted would
be effective. On average these alternative treatments were predicted
to reduce the amount of virus in the patients' bloodstream by over 99%.
Based on its 79% accuracy, had the system been used to select new drug
combinations for these patients, 110 of the139 patients might have responded
to treatment instead of failing.
"We were very excited when we saw the scale of these results"
commented Dr Brendan Larder, Chair of the RDI's Scientific Core Group.
"This is the first time that we have used this technique specifically
to look for alternative treatments for patients experiencing multiple
treatment failures. While the results of this small study will need
to be confirmed, they certainly hold out great promise."
The study, which was conducted by the HIV Resistance Database Initiative
(RDI), employed the AI technique called Neural Networks to model the
relationships between changes in the genetic code, the genotype, of
HIV and the response of patients to different drug combinations. The
technique mimics the connections in the human brain and can 'learn'
as it analyses. Data collected by the BC Centre for Excellence from
351 patients in Canada, who had received a great deal of HIV drug therapy,
were used for the analysis. These data included the genetic codes (genotypes)
of the patients' HIV, the drugs used to treat them and changes in the
amount of virus in the bloodstream (viral load) following treatment.
The most accurate Neural Network model overall was able to predict the
direction of viral load change to different drug combinations with 79%
accuracy. When this model was then used to look for effective drug combinations
for patients whose actual treatment was failing, based on the genotype
of their virus, it successfully did so in all 139 cases.
The study was originally designed to evaluate the accuracy of Neural
Network models developed using data from uncontrolled clinical cohorts.
Previous results have demonstrated the accuracy of the technique but
used highly regulated data from tightly controlled clinical trials and
the objective was to compare the accuracy of models using much more
variable data. The results showed that the models were at least as accurate,
if not more accurate, than those developed using clinical trial data.
The magnitude of changes in viral load predicted by the most accurate
model in this study were very highly correlated with the actual changes
experienced by the patients, with an R2 value of 0.75. This compares
to an R2 value of 0.55 for the models developed using clinical trial
data.
One of the main reasons why HIV treatment fails is that mutations in
the genetic code of HIV cause resistance to HIV drugs. Tests are already
available to read these genetic changes. The RDI's goal is to enhance
the predictive power of such tests by relating genetic changes in HIV
directly to virological response to drug therapy, using the largest
collection of clinical data of its kind in the world.
Background
RDI Limited is registered in the UK as a not-for-profit company limited
by guarantee and with no share capital. The RDI is headquartered in
London UK, with research based in Cambridge, UK.
The RDI is supported by grants and funds from the BC Centre for Excellence
in HIV/AIDS in Vancouver, Canada (a Founding Academic Member of the
RDI); SAIC-Frederick, for a study in collaboration with the US National
Institute of Allergy and Infectious Diseases; and Gilead Sciences Inc.
Data provided to the RDI does not contain any information relating
to patient identity.
The RDI Scientific Core Group comprises the following scientists and
clinicians:
| 1. Brendan Larder |
Cambridge, UK, Chair |
| 2. Victor DeGruttola |
Harvard School of Public Health, Boston, MA, USA |
| 3. Richard Harrigan |
BC Centre for Excellence in HIV/AIDS, Vancouver, BC, Canada |
| 4. Julio Montaner |
BC Centre for Excellence in HIV/AIDS, Vancouver, BC, Canada |
| 5. Scott Wegner |
US Military HIV/AIDS Program, Washington, DC, USA |
| 6. Maurizio Zazzi |
HIV Monitoring Service, University of Siena, Italy |
###
