Genotypic resistance testing is gaining widespread use and has become an important tool in the routine clinical management of HIV infection. Genotypic resistance assays have become effective at determining the nucleotide sequence of the HIV-1 reverse transcriptase (RT) and protease enzymes. The introduction of FDA-approved kits is likely to further reduce the amount of variability in the results of sequencing performed in different laboratories.

However, because few clinicians are aware of the clinical significance of all drug-resistance mutations, current genotypic assays are accompanied by some form of interpretation. It is at this stage - the genotypic interpretation - that interlaboratory reproducibility ends. There are many different systems for interpreting HIV-1 genotypes and the same sample may be reported in a number of different ways depending on the interpretation system used. This lack of reproducibility in genotypic assay reporting represents an unfortunate state of affairs. Imagine if the interpretation of a biopsy of a possible malignant tissue was different in different laboratories.

Beginning in 2001, there have been many presentations at meetings on interpreting HIV-1 genotypic sequence data. At the 5^th International Workshop on HIV Drug Resistance and Treatment Strategies, the EuroGuidelines Group for HIV Resistance presented an overview of 19 different interpretation systems that have been either used by laboratories doing genotyping or published on the Web for educational purposes [1]. Each of the interpretation systems was characterized according to (1) whether the system was meant to be applied directly to clinical sequences, (2) organization that created the system and potential for conflict of interest, (3) extent of access to the system, (4) cost to using the system, (5) algorithmic approach, (6) software validation, (7) timeliness, and (8) efforts made towards clinical validation.

The Workshop also had several abstracts comparing the performance of different algorithms. Three types of comparisons were presented: (1) resistance prediction without external validation, (2) resistance prediction with comparison to drug susceptibility data, (3) resistance prediction with comparison to clinical outcome data. None of these comparisons are entirely straightforward. To compare the results of different algorithms it is usually necessary to normalize results into just three categories: susceptible, intermediate, resistant. To correlate algorithmic results with drug susceptibility data, it is necessary to realize that many algorithms have been designed to detect clinically significant mutations that might not cause phenotypic resistance if the mutation is present as a minor variant, is a transitional mutation, or causes resistance to drugs that are difficult to test phenotypically (e.g., didanosine and stavudine). To correlate algorithmic results with clinical outcome data, it is necessary to realize that the accuracy of prediction will be highly dependent on the precise definition of success (i.e., extent of HIV RNA reduction with a salvage regimen, duration of HIV RNA decrease with a salvage regimen) and with the number of effective drugs used as part of salvage therapy.

A comparison of the first type was presented by Shafer et al who examined the level of agreement among three rules-based interpretation systems on 2,200 clinical sequences for a total of 28,600 interpretations (2,200 x 13 drugs evaluated) [2]. Although each of the algorithms were in agreement for 81% of the interpretations, the authors emphasized the need to examine the rules for the mutations present in those sequences that led to discordant results. A comparison of the second type was presented by Schmidt et al, who correlated the predictions of several algorithms on 100 genotypes with phenotypic data on those same isolates [3]. Comparisons of the third type were presented by Lanier et al [4] and Hammer et al [5]. Lanier retrospectively examined data from five clinical trials involving 166 patients in whom abacavir was added as a single agent and compared the predictions of different algorithms with the likelihood of reducing RNA levels by at least 0.5 log10 copies/mL within four weeks. Hammer correlated Virco's VirtualPhenotype^TM and the results of several rules-based methods applied to baseline genotypic data from 99 patients who had virologic responses to therapy at week 24 in an ACTG clinical trial.

The comparisons described in the above studies are logical approaches to characterizing current genotypic interpretation systems. These types of studies should ideally identify the best rules for genotypic interpretation and promote a convergence of the currently divergent algorithms. But to date, none of the comparative studies have been conclusive in identifying either highly predictive rules or algorithms. We anticipate that additional comparative studies of interpretation systems will be presented at the 9^th Conference on Retroviruses and Opportunistic Infections and that the full value of these studies will be realized when they are published in their entirety.

References