Poster presentation at:
The 2nd IAS Conference on HIV Pathogenesis and Treatment
Paris, France; 13-16 July 2003
Neural Network Model Identified Potentially Effective Drug
Combinations for Patients Failing Salvage Therapy
B.A. Larder1, D Wang1, A Revell2 and C. Lane3.
1: The HIV Resistance Response Database Initiative (RDI), Cambridge,
UK; 2: RDI, London, UK; 3: National Institute of Allergy and Infectious
Diseases, Bethesda, USA.
Background: The RDI has developed Neural Network (NN) models
which, when used in conjunction with the RDI clinical database, have
demonstrated accuracy in predicting the trajectory and magnitude of
virological response to antiretroviral therapy from genotype. Here we
examine the potential of NN models to identify potentially effective
alternative treatment regimens for 'salvage' patients, experiencing
continued treatment failure despite having their treatment changed according
to current clinical practice.
Methods: Data from 511 patients from two well-characterized
clinical cohorts were analyzed. All had multiple genotypes and irregular
VL measurements. 747 'treatment change episodes' (TCEs) were obtained:
episodes with a genotype 12 weeks and a VL 8 before treatment change
and follow-up VL within 4-40 weeks. Four NN models were trained (using
690 randomly selected TCEs) and evaluated using the remaining 57 TCEs,
as independent test sets. 108 salvage failure TCEs were then identified
(baseline VL of 4 logs and an increase in VL following treatment change).
The best NN model was then used to predict virological response to 37
common treatment regimens for each case (28,386 individual virtual response
Results: The correlation between the predicted and actual VL
change for the four VL models gave a mean value of 0.6 and mean correct
trajectory prediction rate of 80.7%. The best model ( =0.76) identified
one or more regimens that were predicted would reduce VL in all 108
salvage failure cases. The median actual change in viral load for these
cases was +0.41 logs whereas the median change using the best regimens
from the NN predictions was -2.86 logs.
Conclusions: NN modeling identified potentially beneficial treatment
regimens in all patients who experienced treatment failure, predicting
substantial reductions in viral load. This demonstrates the potential
utility of this approach in supporting treatment decision making in