Medical Advocates for Social Justice
Conference Abstract
from the
1st European HIV Drug Resistance Workshop
Cercle Municipal, Luxembourg

March 06- 08, 2003
 

 
Estimating effective drug combinations against drug-resistant mutants from genotypes [Abstract 3.1]
N Beerenwinkel1, T Lengauer1, M Däumer2, R Kaiser2,
H Walter3, K Korn3, D Hoffmann4, J Selbig5

1 Max Planck Institute for Informatics, Saarbrücken, Germany; 2 Institute of Virology, University of Cologne, Germany; 3 Institute of Clinical and Molecular Virology, German National Reference Center for Retroviruses, University of Erlangen-Nürnberg, Germany; 4 Center of Advanced European Studies and Research (caesar), Bonn, Germany; 5 Max Planck Institute of Molecular Plant Physiology, Golm, Germany		  
Share this Abstract with a Colleague
Background
Selecting a new potent drug combination after failure of antiretroviral therapy (ART) is challenging even in the presence of genotypic information on the viral drug targets. The difficulty is interpreting sequence data with respect to phenotypic resistance and clinical outcomes. Several expert rule sets and some statistical and machine learning approaches provide individual predictions of phenotypic drug resistance to each of the currently 17 approved antiretroviral agents. Here we present a rational approach to selecting an optimal combination of drugs based on the viral genotype.

Methods
To estimate the activity of a regimen against a particular viral strain, we develop a scoring function that takes as inputs the viral pol-gene sequence and the set of drugs making up the combination therapy. The construction of this activity score is based on quantitative phenotype predictions for each drug. A probabilistic model is fit to these predictions that allows for normalizing the predicted values and integration into a score for drug combinations. In order to estimate activity of the regimen on nearby mutants (presumed either to be present in the viral quasi-species or soon to evolve under therapy) we apply a heuristic search through the mutational neighbourhood of the considered sequence. Activity scores are generated for each number of point substitutions up to a certain maximum.

Results
To illustrate the utility of these scores, we analyzed pre-treatment sequences and subsequent regimens from 96 ART-experienced patients. Therapies were categorized as either successful, if virus load dropped by at least two log10 copies/ml approx. 30 days after therapy change (28 cases), or failure otherwise (68 cases). Cross-validation revealed an optimal search depth of three point mutations for this data set. The linear model constructed from these search results classifies therapies with an expected prediction error of 21% (p<0.0001, the probability of observing 20 or fewer errors in 96 predictions (=21%) with probability 28/68 of being incorrect according to the binomial distribution).

Conclusions
The proposed scoring function can predict clinical outcome for a given genotype and regimen. Thus, it may be helpful in designing individualized therapeutic protocols
Main New/Newsworthy Therapeutic Drug Monitoring Data Luxembourg Conference Index      

Estimating effective drug combinations against drug-resistant mutants from genotypes
A Medical Advocates for Social Justice Update