Poster presentation at:
4th International Congress on Drug Therapy in HIV Infection
17th November 2002 - 21st November 2002
Wang D & Larder BA. On Behalf of The HIV Resistance Response Database Initiative (RDI)
The goal of the RDI is to create a sufficiently large relational database to correlate baseline HIV-1 resistance genotypes and drug therapy to virological response. Initially, we wished to develop neural network (NN) models to predict virological failure.
Baseline genotype, viral load (VL) and treatment data, plus week 24 VL from clinical studies is being collected in a customised Oracle database. We plan to collect data from 5000 - 10,000 individuals. Initial NN models were based on complex (12 drugs and 49 mutations) or simplified (6 drugs and 31 mutations) data input parameters. Patient samples from the Vigilance II trial were used for NN training. The VL input was either the continuous variable of absolute VL change from baseline, or a dichotomous variable of virological success or failure. Virological failure was defined as VL>400 copies/ml at week 24.
Training of both NN models using 700 patient cases was surprisingly successful. For the complex NN model, the correlation between the predicted and actual absolute VL change for the training set gave an r@ value of 0.85 (p<0.0001). In independent test and validation sets, this correlation gave an average r@ value of 0.5?0.05 (p<0.0001). The VL trajectory was correctly predicted in 75% (?1.8%) of cases in the test sets by the complex NN model. Of interest, the simplified NN model was slightly less accurate in predicting absolute VL change (average r@ = 0.47?0.05 for test and validation sets) or the VL trajectory (72?1.8% correct for the test and validation sets). Correct prediction of failure based on baseline genotype and treatment was achieved in 100% of cases in the training data set and an average of 82%?1.6% of cases in multiple cross validation data sets. This model was also used successfully to predict 'in silico' VL failure using genotypes and a variety of hypothetical treatment regimens.
Initial development of NN models resulted in successful training with a limited dataset and demonstrated the feasibility of this approach to predict absolute VL change, VL trajectory and virological failure from complex input variables. As the size of the training data set increases, it is anticipated that the accuracy of these predictions will also increase and that this technique will prove to be a useful tool for predicting treatment outcome from genotype.