At least three expert panels have recommended that HIV-1 resistance testing be done to help select anti-HIV drugs in certain clinical situations. But there are no guidelines on how to interpret these tests and many clinicians are frustrated by their use. Such frustration is a particular problem for genotypic assays which yield a rich but complex form of data which bears little resemblance to the results of typical antimicrobial susceptibility tests.

Fortunately, genotypic interpretation is an ideal application for a computerized expert system because genotypic data consist of machine-readable text (lists of nucleotides) and because rules-based and pattern recognition algorithms can be performed better by computers than by human experts. The days of clinicians having to remember lists of HIV-1 drug resistance mutations will soon come to an end.

Expert systems perform reasoning over representations of human knowledge. Expert systems consist of a computerized knowledge base and an inference engine. A knowledge base with homogeneous data such as genotype-phenotype correlations is amenable to machine learning algorithms. A knowledge base with diverse forms of data such as correlations between genotype and phenotype and between genotype and clinical data is more amenable to rules-based algorithms.

The 5^th International Workshop on HIV Drug Resistance and Treatment Strategies had two sessions devoted to drug resistance interpretation. The abstracts presented included comparisons of methods for interpreting resistance assays, descriptions of methods for correlating genotype and phenotype, and the application of machine-learning and rules-based algorithms to clinical data sets.

Comparisons of methods interpreting resistance assays
J. Schapiro, together with other members of the EuroGuidelines Group for HIV Resistance presented detailed information on 16 interpretation systems and highlighted the extent of variability in the sponsorship, public availability, and methods of these systems (abstract 172) [1]. A. Wensing et al (abstract 133) described the use of interpretation systems for providing drug resistance interpretations for samples in the ENVA-3 panel [2]. Wensing's main finding was that drug resistance interpretations were variable not just because different interpretation systems were used by different groups, but also because the results of the same interpretation system were often interpreted differently in different laboratories. Her abstract argues strongly for the creation of a controlled vocabulary for all terms used in reports of drug resistance interpretation systems.

Description of methods for correlating genotype and phenotype
D. Wang et al (abstract 140) described the use of a 28-mutation neural network model that accurately predicts phenotypic resistance to lopinavir [3]. The neural network was created using 1,322 sequenced isolates for which phenotypic data were available. The neural network was constructed using 1205 samples and tested on 117 samples. The 28-mutation neural network performed slightly better than one relying on the original 11 mutations proposed by D. Kempf et al at last year's Resistance Workshop (r²=0.88 vs r2=0.84; p< 0.001) [4]. N. Beerenwinkel et al developed a decision-tree algorithm (abstract 138) to classify sequences as being susceptible or resistant using a data set of 6000 genotype-phenotype pairs derived from 470 clinical isolates tested in a recombinant virus assay used by the German National Reference Centre for Retroviruses [5].

These two abstracts demonstrated the power of machine learning approaches at correlating genotype to phenotype. However, two limitations were raised during the discussion period. First, as noted earlier, these algorithms cannot take into account known correlations between genotype and clinical outcome because they are programmed to operate on homogeneous genotype-phenotype data. Second, these algorithms are generally incapable of explaining their reasoning. In contrast, rules-based algorithms have the capability of explaining their reasoning in a manner similar to a human expert.

Application of algorithms to clinical data sets
I described a program on the Stanford HIV RT and Protease Sequence Database that allows the results of different algorithms to be compared to identify rates of discordances between algorithms and to identify mutational patterns responsible for those discordances (abstract 134) [6]. The usefulness of this web site was demonstrated by comparing the Stanford HIVDB algorithm, the RCG-DAP algorithm, and the French ANRS algorithm on RT and protease sequences from 2,232 individuals. The abstract concluded that there was 84% concordance in making 29,016 drug resistance interpretations (2,232 x 13 drugs), that four drugs (abacavir, amprenavir, ddI, and d4T) were responsible for nearly 2/3 of the discordances, and that certain commonly occurring mutational patterns causing discordances could be identified.

R. Lanier et al (abstract 137) described the application of 12 algorithms to a clinical data set containing virologic responses to abacavir intensification in patients with persistent viremia on a HAART regimen [7]. D. Costagliola presented an abstract by D. Descamps et al (abstract 136) that developed a genotypic sensitivity score for salvage therapy regimens containing amprenavir based on data from the Narval ANRS 088 trial [8]. S. Hammer et al (abstract 142) examined the predictive value of Virco's VirtualPhenotype^TM and of several rules-based algorithms at predicting the virologic responses observed in the ACTG 372 study [9]. Each of these studies represents an important step in assessing the validity of algorithms and rules for interpreting genotypic resistance data. After the sessions, I spoke to each of the authors emphasizing the importance of making their data publicly available so that other groups can use the clinical data to develop and test drug-resistance interpretations.

During the next year, algorithms will evolve and most likely converge through a process of inter-algorithm comparison and validation using clinical data sets. This is because there is much more concordance among clinical virologists than there is among the various published algorithms. It is also unlikely that algorithms will remain proprietary because there is no precedent for basing important medical decision on proprietary, unpublished data.

References
Abstracts can be accessed by hyperlink after registering for the 5th International Workshop on HIV Drug Resistance & Treatment Strategies Webcast at Mediscover.net

Schapiro JM, et al. Resistance assay interpretation systems vary widely in method and approach. Antiviral Ther. 2001;6(Suppl 1):131. Abstract 172
Wensing AM, Keulin W, Buimer M, Brambilla D, Schurmann R, Boucher CA. Analysis of the world-wide evaluation study on HIV-1 genotype interpretation; ENVA-3. Antiviral Ther. 2001;6(Suppl 1):133. Abstract 133
Wang D, Harrigan R, Larder BA. A 28-mutation neural network model that accurately predicts phenotypic resistance to lopinavir. Antiviral Ther. 2001;6(Suppl 1):131. Abstract 140
Kempf D, et al. Genotypic correlates of reduced in vitro susceptibility to ABT-378 in HIV isolates from patients failing protease inhibitor therapy. Antiretroviral Ther. 2000;5 (Supplement 3):29.
Beerenwinkel N, Schmidt B, Walter H, Kaiser K, Lengauer1 T, Hoffmann D, Korn K, Selbig J. Geno2pheno: a new machine learning approach to predicting phenotypic drug resistance from genotype. Antiviral Ther. 2001;6(Suppl 1):139. Abstract 139
Shafer RW, Gonzales MJ, Brun-Vezinet F. Online comparison of HIV-1 drug resistance algorithms identifies rates and causes of discordant interpretations. Antiviral Ther. 2001;6(Suppl 1):134. Abstract 134
Lanier ER, Scott J, Ait-Khaled M, Stone C2, Melby T, Sturge G, St Clair MH, Steel H, Hetherington S, Pearce G, Spreen W Lafon S. Predicting abacavir antiviral activity using HIV-1 reverse transcriptase genotype: a comparison of 12 algorithms. Antiviral Ther. 2001;6(Suppl 1):137. Abstract 137
Descamps D, Masquelier B, Mamet JP, Calvez V, Ruffault A, Felles F, Goetschel A, Girard PM, Brun-Vézinet F, Costaglio D. A genotypic sensitivity score for amprenavir based on genotype at baseline and virological response. Antiviral Ther. 2001;6(Suppl 1):136. Abstract 136
Hammer S, Peeters M, Harrigan R, Larder BA, and the ACTG 372B/D Study Team. Virtual phenotype is predictive of treatment failure in treatment-experienced patients. Antiviral Ther. 2001;6(Suppl 1):142. Abstract 142