Poster presentation at:
13th International Drug Resistance Workshop
Costa Adeje - Tenerife Sur, Canary Islands, Spain; June 8-12, 2004

Previous Drug Exposure Data Significantly Increase the Accuracy of Artificial Neural Networks in Predicting Virological Response to Combination Therapy

B.A. Larder 1 , D Wang 1 , A Revell 2 , R Harrigan 3 , J Montaner 3 and C. Lane 4 .

1: The HIV Resistance Response Database Initiative (RDI), Cambridge , UK ; 2: RDI Ltd, London , UK ; 3: The BC Centre for Excellence in HIV/AIDS Vancouver , Canada ; 4: National Institute of Allergy and Infectious Diseases (NIAID), Bethesda , USA .

Background: Artificial Neural Networks (ANN) trained using genotype, viral load and drug treatment data, can successfully predict virological response to combination therapy. However, the accuracy of these models may be limited by pre-existing minority populations of resistant virus, not detected by standard genotyping technologies, that can accelerate treatment failure. This study was designed to address whether inclusion of previous AZT exposure data, as a 'surrogate' for minority mutant populations, can increase the accuracy of ANN in predicting response to combination therapy involving d4T, abacavir (ABC) and tenofovir (TDF).

Methods: 716 'treatment change episodes' (TCEs) were selected from five clinical cohorts. These all involved the introduction of d4T, ABC, or TDF and 349 had historical (not immediately prior) AZT exposure. 20 ANN models were developed using mutations, drugs and viral load as input variables, 10 of which also included previous AZT exposure (yes or no) as a binary input variable. Each model was trained using 640 TCE's and tested with the remaining 76. All the training and testing sub-datasets were independent, randomly partitioned and normalised.

Results: Correlations between predicted and actual VL change for the 10 basic models gave a mean r 2 value of 0.64 and mean correct trajectory prediction rate of 87%. However, the 'AZT history' models gave a mean r 2 value of 0.73 and mean correct trajectory prediction rate of 89%. The difference in accuracy between the two groups, as measured by the differences between predicted and actual viral loads, was statistically significant (p<0.05). There were 12 cases with historical AZT but no AZT mutations who experienced virologic failure (<0.5 log reduction) and who were predicted by the basic ANN model to respond but correctly predicted to fail by the 'AZT history' model.

Conclusions: The addition of historical AZT exposure data significantly improved the accuracy of ANN in predicting response to combination therapy including d4T, ABC or TDF, confirming that historical exposure to antiretroviral drugs can influence response to a new regimen. This study suggests that historical treatment information may act as a surrogate for the presence of minority mutant populations and can enable ANN to overcome this potential shortcoming of current genotyping.