Poster presentation at:
CROI 2006, Denver, CO, USA, 5-8 February 2006.

Neural Networks Are More Accurate Predictors of Virologic Response to Antiretroviral Therapy Than Rules-Based Genotype Interpretation Systems

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

1: The BC Centre for Excellence in HIV/AIDS Vancouver, BC, Canada (BC-CfE); 2: The HIV Resistance Response Database Initiative (RDI), London, UK; 3: US Military HIV Research Program, Rockville, MD, USA; 4: National Institute of Allergy and Infectious Diseases (NIAID), Bethesda, MD, USA.

Background
HIV genotyping with rules-based interpretation is used to help predict the virologic responses from antiretroviral treatment (HAART) options. The RDI has shown that artificial neural networks (ANN) can predict virologic responses to HAART based upon genotypes. Here we assess directly the accuracy of these approaches by comparing ANN predictions with genotypic sensitivity scores (GSS) from rules-based interpretation systems in terms of correlations with actual virologic responses ( VL).

Method
A committee of five ANN models were trained to predict VL from genotype (55 resistance mutations), the drugs in the regimen and baseline viral load using data from 2,983 clinical treatment change episodes (TCEs) from the RDI database. Three independent test sets were used (each of 25 randomly selected TCEs): two of clinical cohort data (RDI database and Royal Free Hospital) and one from the PhenGen trial of genotyping vs phenotyping. Accuracy was tested by providing the ANN with baseline data from the test TCEs and correlating the models' averaged predictions of VL with the actual VL values.

Total and normalised GSS were derived for the test TCEs using Stanford (HIVdb 4.1.2), ANRS (v2004.09) and Rega (v6.2) rules-based interpretation systems. GSS and Stanford mutation scores (SMS), based on assumed impact on susceptibility to each drug, were also correlated with actual VL for the test TCEs.

Results
Correlations between ANN predictions and actual VL values for the RDI test set gave an averaged value of 0.80. SMS correlated with an of 0.22 and GSS 0.19-0.31. The PhenGen test data yielded values of 0.32 for the ANN, compared with 0.16 for SMS and 0.10-0.24 for GSS. Royal Free data gave values of 0.39 for ANN compared with 0.00 for SMS and 0.00-0.04 for GSS. Correlations between the rules systems were substantial ( = 0.41-0.93), highlighting their similarity. Correlations between the rules systems and the ANN were mostly fairly weak and inconsistent in direction ( = 0.00-0.39).

Conclusions
ANN were considerably more accurate predictors of virologic response than rules-based systems with all three test sets, explaining up to 80% of the variance in VL vs 31%. The ANN were more accurate predicting virologic responses for patients from clinics that have provided data used in training the ANN than for those from ‘unseen' clinics. The results suggest that ANN may have considerable utility as treatment decision-making tools.