Enhanced Prediction of Lopinavir Resistance from Genotype by Use of Artificial Neural Networks

Dechao Wang () 0 1 Brendan Larder 0 1 Virco 0 Cambridge Science Park 0 Cambridge 0 United Kingdom 0 0 Received 6 December 2002; accepted 28 March 2003; electronically published 14 August 2003 1 Present affiliation: HIV Resistance Response Database Initiative , London , United Kingdom. Database Initiative, RDI UK Ltd. , 14 Union Square, London N1 7DH , United Kingdom Our objective was to accurately predict, from complex mutation patterns, human immunodeficiency virus type 1 resistance to the protease inhibitor lopinavir, by use of artificial intelligence. Two neural network models were constructed: 1 based on changes at 11 positions in the protease that were previously recognized as being significant for lopinavir resistance and another based on a newly derived set of 28 mutations that were identified by performing category prevalence analysis. Both models were trained, validated, and tested with 1322 clinical samples. A procedure of determining the optimal neural network parameters was proposed to speed up the training processes. The results suggested that the 28-mutation set was a more accurate predictor of lopinavir susceptibility (correlation coefficient, R 2 p 0.88 ). We identified potentially significant new mutations associated with lopinavir resistance and demonstrated the utility of neural network models in predicting phenotypic susceptibility from complex genotypes. - Techniques to assess drug resistance in human immunodeficiency virus type 1 (HIV-1) have been increasingly incorporated into clinical studies of therapeutic regimes since a correlation between the emergence of viral resistance and clinical response was first suggested [1, 2]. With more drugs and therapeutic options becoming available, drug-resistance testing is playing an increasingly important role in the management of HIV-1 infection. At present, 2 general approaches are used for assessing resistance to HIV-1 drugsnamely, phenotyping and genotyping [2]. Phenotyping directly measures the susceptibility of HIV-1 strains to particular drugs, whereas genotyping establishes the absence or presence of specific genetic mutations in HIV-1 that have been previously associated with drug resistance. Although phenotyping is recognized as providing a more direct measurement of resistance than genotyping, it is more complex to perform and takes longer to generate a result [2]. Genotyping has the advantage of being relatively simple and rapid, but the quantitative prediction of drug susceptibility from genotypic data is still challenging. A number of statistical analyses have been used to relate HIV-1 genotype with phenotype [35]. These methods have provided useful information about how susceptibility data correlate with the corresponding genotype. However, difficulties remain in quantitatively relating the phenotype of a specific sample to its genotype and associating specific mutations (or mutation patterns) to resistance. Recently, several studies have associated drug susceptibility with HIV-1 genotypes by means of recursive partitioning algorithms. Although reasonable prediction rates for some drugs have been achieved, the results have been mixed [69]. Moreover, the algorithms were used as a method of classification rather than regression. The use of relational databases has made it possible to obtain a predicted susceptibility of a known genotype by matching archived drug-susceptibility data to resistance patterns. However, a drawback of this approach is that it relies on prior knowledge of resistance patterns. By contrast, artificial neural networks (ANNs) have been successfully applied to addressing complex problems in many fields [1014]. ANNs learn by an iterative process that adjusts the strengths of connections such that the system outputs an appropriate result. Of importance, data processing by these systems does not require assumptions of how outputs relate to inputs or that inputs are independent [10]. Lopinavir is an HIV-1 protease inhibitor that is coadministered with ritonavir to boost patients plasma levels of the drug [15]. In recent studies, mutations in the protease at 11 coding positions were identified as being important determinants of lopinavir resistance [16]. These were at codons 10, 20, 24, 46, 53, 54, 63, 71, 82, 84, and 90 [16]. A total of 23 possible mutations at these 11 codons were taken into consideration when lopinavir resistance was modeled from the genotype [16]. However, it is unlikely that these mutations are the sole determinants of lopinavir resistance. Indeed, a recent study that probed a large database of lopinavir genotypes and phenotypes concluded that additional mutations were associated with resistance [17]. In the present study, ANNs were constructed to examine lopinavir resistance in connection with genetic mutation patterns. Twenty-eight mutations were found to be a more accurate predictor of lopinavir susceptibility, thus enhancing our ability to predict lopinavir susceptibility from the genotype. MATERIALS AND METHODS Plasma samples. The plasma samples used in our study were submitted for resistance testing from HIV-1infected persons from clinical practices in Europe and the United States; they were shipped on dry ice. Plasma samples were stored at 70 C before genotypic and phenotypic analysis. HIV RNA extraction and genotyping. Viral RNA was extracted from 200 mL of patient plasma with the QIAamp viral RNA extraction kit (Quiagen), according to the manufacturers instructions. cDNA encompassing part of the pol gene was produced with use of Expand RT (Boehringer Mannheim) as described elsewhere [18]. A 2.2-kb fragment encoding the protease and reverse transcriptase (RT) regions was then amplified by nested polymerase chain reaction (PCR) with primers and conditions described elsewhere [18]. The PCR products were genotyped by dideoxynucleotide-based sequence analysis. Samples were sequenced with the Big Dye terminator kit (Applied Biosystems) and resolved on an ABI 377 DNA sequencer as described elsewhere [19, 20]. The results of the genotypic analysis are reported as amino acid changes at positions along the RT and protease coding regions compared with the gene sequence of wild-type reference strain HXB2. Phenotypic resistance testing. Phenotypic analysis was done by the recombinant virus assay [21] approach described by Hertogs et al. [18], with the modifications described in Pauwels et al. [22]. In brief, HIV protease and RT coding sequences were amplified from patient-derived viral RNA with HIV-1specific primers. After the homologous recombination of amplicons into a proviral clone from which the protease and RT coding sequences were deleted, the resulting recombinant viruses were harvested, titrated, and used for the testing of in vitro susceptibility to antiviral drugs. The results of this analysis are expressed as fold change values, which reflect the fold increase in the IC50 (micromolar) of a particular drug when it was tested with patient-de (...truncated)