 |
|
|
| |
Low Sensitivity of Gene-Based Tools in Predicting Coreceptor Use
|
| |
| |
Targeting HIV Entry: 3rd International Workshop
December 7-9, 2007
Washington, DC
Mark Mascolini
Computer tools to predict coreceptor use based on differences in individual HIV envelope gene sequences still lack sensitivity in making correct coreceptor calls, according to results of two separate studies reported at the HIV Entry Workshop. Such a tool would be faster and cheaper than the phenotype-based Trofile assay, but both research teams concluded gene-based methods are not ready for clinical practice.
Stanford University investigators tested four gene-based computer decision trees (or "algorithms") for precision in determining CXCR4 (X4) tropism already established in well-characterized patient isolates by a standard lab method using MT-2 cells [1]. Researchers have used the MT-2 assay for decades, but it is too time-consuming for clinical use.
The Stanford team looked at virus in peripheral blood mononuclear cells collected from people with CD4 counts between 161 and 661 when enrolled in ACTG study 175, a clinical trial from the early 1990s comparing AZT alone, ddI alone, AZT plus ddI, and AZT plus ddC. Study participants had either taken AZT monotherapy or had tried no antiretrovirals when they signed up for the study.
Of the 73 viral samples analyzed, the MT-2 assay identified 29 as syncytium-inducing (SI) virus, inferring X4 coreceptor use, and 44 as non-SI (NSI), inferring R5 use. (Syncytia are cell clusters, and their appearance correlates with advanced HIV infection.) If one assumes these coreceptor assignments are correct, the four gene-based algorithms tested had generally high specificity, but consistently low sensitivity:
- PSSM SI/NSI algorithm: 58.6% sensitivity, 95.4% specificity
- PSSM R5/X4 algorithm: 48.2% sensitivity, 100% specificity
- 11/25 charge rule algorithm: 44.8% sensitivity, 97.7% specificity
- geno2pheno algorithm: 58.6% sensitivity, 95.4% specificity
The algorithms correctly predicted X4 use 8 of 10 times when syncytia took 6 or fewer days to develop in HIV-infected MT-2 cell cultures. But when syncytia took 7 or more days to develop, the algorithms got the X4 call right only 3 of 10 times.
Researchers at the University of California in San Francisco and Berkeley and the University of British Columbia took a different tack in predicting coreceptor preference with genetic data [2]. They started with V3 loop sequences of HIV-1's envelope gene on file in the Los Alamos HIV Sequence Database. Coreceptor use had already been determined for 292 HIV-1 subtype B sequences.
The California-British Columbia team used sequence data from the 292 subtype B viruses to "train" a type of algorithm called a support vector machine (SVM) to predict coreceptor use. The trained SVM did well, identifying coreceptors with a specificity of 93% and a sensitivity of 84% when the researchers used Trofile results as the standard of comparison. Then the investigators used the trained SVM to predict coreceptor use in 951 viral isolates from people with HIV. Specificity remained high, at 95.7%, but sensitivity plunged to 29.9%. Although the algorithm had an 84.1% concordance with the Trofile assay, its positive predictive value was a meager 59.5%.
But this research team reported one ray of optimism in their ongoing work. After discovering the main causes of incorrect coreceptor calls with their algorithm, they set to work revamping this tool so that it can spot minority X4 viral populations more readily and so that it breaks down HIV-1 envelope gene sequences into all possible variations. They hope this retooling "may improve prediction accuracy substantially, making genotype-based coreceptor usage determination an attractive alternative or complement to in vitro phenotypic analysis" with assays like Trofile.
References
1. Huang H, Johnson E, Winters M, et al. Consensus V3 loop sequencing of clinical HIV isolates is a poor predictor of SI phenotype. HIV Entry: 3rd International Workshop. December 7-9, 2007. Washington, DC. Abstract 9.
2. Pillai SK, Good B, Harrigan R, et al. Genotype-based prediction of HIV-1 coreceptor usage via machine learning: clinical applicability. HIV Entry: 3rd International Workshop. December 7-9, 2007. Washington, DC. Abstract LB-1.
|
| |
|
|
|
|
|
