Treatment history improves the accuracy of neural networks predicting virologic response to HIV therapy

Poster number: 20

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

  1. HIV Resistance Response Database Initiative, London, UK
  2. BC Centre for Excellence in HIV/AIDS Vancouver, Canada
  3. US Military HIV Research Program, Rockville, MD, USA
  4. National Institute of Allergy & Infectious Diseases, Bethesda, MD, USA.

Background

Previous exposure to highly active antiretroviral therapy (HAART) can result in archived resistant virus, undetectable by conventional genotyping, which may undermine the success of new treatment regimens. We assessed the influence of previous treatment on the accuracy of artificial neural networks (ANN) predicting virologic response to HAART.



Methods

A committee of 10 basic ANN models was trained and cross-validated (leave-n-out) using 2,559 treatment change episodes (TCEs), with 71 input variables (55 resistance mutations from current genotype, baseline viral load (VL), follow-up time and 14 drugs). Ten treatment history models were trained using the same procedure but with 4 additional input variables - previous treatment with any of the following: AZT, 3TC, any NNRTI and any PI. The ANNs were tested using the input variables of independent test TCEs, which produced a predicted VL response that was compared to the actual response. The averaged output across the ten ANN models was used when testing each committee.



The importance of the 75 input variables, including previous treatment, was estimated as follows: the whole data set was divided into 12 different groups based on the viral load changes using intervals of 0.5 log10 copies/ml. ANOVA was performed to test the mean differences across groups. P-values for the input variables were obtained and ranked. Statistical significance was accepted if the p-value was <0.05.



Results

Correlations between predicted and actual VL change for basic and treatment history ANN models gave r2 values of 0.30 and 0.45 respectively (p<0.01). The importance of the historical treatment variables was estimated using the ranked p-values. AZT, 3TC, any NNRTI and any PI all had significant impact on VL response with p-values of 0.01, 0.02, 0.01, and 0.00001 respectively, and were ranked in the top 38 to 41 positions among the 75 input variables.



Conclusions

Treatment history data significantly improved the accuracy of ANN and had significant impact upon VL response, confirming that historical exposure to antiretroviral drugs can influence response to a new regimen. This suggests that treatment history may act as a surrogate for minority mutant populations enabling ANN to overcome this shortcoming of current genotyping.