Poster presentation at:
XVI International HIV Drug Resistance Workshop
12-16 June 2007, Bridgetown, Barbados.
Development of computational models for prediction of virologic response in a prospective clinical study
B.A. Larder1, D Wang1, A Revell1, J Montaner2, R Harrigan2, S Wegner3, HC Lane4.
1: HIV Resistance Response Database Initiative (RDI), London, UK; 2: BC Centre for Excellence in HIV/AIDS, Vancouver, Canada; 3: Uniformed Services University of the Health Sciences, Bethesda, MD, USA; 4: National Institute of Allergy and Infectious Diseases (NIAID), Bethesda, MD, USA.
Virologic response to different antiretroviral combinations is highly complex and dependent on many factors. We have demonstrated that artificial neural networks (ANN) and Random Forests (RF) can predict response from genotype, viral load, CD4 count and treatment history. Here we develop a system combining ANN and RF to predict response to HAART, including darunavir (DRV) and enfuvirtide (ENF), for use in a clinical study.
A single RF and a committee of 10 ANN models were trained to predict virologic response (VL) using 82 input variables (58 mutations, 17 antiretroviral drugs, viral load, CD4 count, treatment history and time to follow-up) from 3,188 treatment change episodes (TCEs) from the RDI database (multiple clinical sources). These included 365 and 215 TCEs involving DRV and ENV respectively. These models were then tested using the input data from an independent test set of 100 TCEs and comparing the RF and ANN predictions of VL, separately and combined (mean), with the actual VL values from the test set. Alternative predictions for TCEs including ENV were made by modelling the backbone and subtracting a further 0.82 log copies/ml from the predicted VL (the mean ENV effect observed in the TORO studies).
Correlations between predicted and actual VL gave values of 0.63 and 0.64 for the RF and ANN respectively. Mean absolute differences between predicted and actual VL were 0.55 and 0.47 log copies/ml respectively. The combined outputs yielded an value of 0.68 and a mean absolute difference score of 0.49. For the 14 test TCEs containing ENF, the results of modelling response to the whole regimens were more accurate (=0.84, mean absolute difference score = 0.58) than those generated by modelling the backbones and applying a constant VL of 0.82 for ENF (=0.79, mean absolute difference score = 0.83).
These results suggest that the combination of outputs from different modelling methods may provide the most accurate predictions of virologic response. The accuracy of the predictions for the new drugs was particularly encouraging. These models are now being used in combination in an international, multicentre clinical trial of the RDI system.
Funded by NCI Contract N01-CO-12400