One such target, HIV-1 integrase, mediates the integration of the reverse-transcribed HIV-1 DNA into the host cell genome and has long been an attractive target for inhibition, as this process is critical within the retroviral life cycle (17). There is the potential to inhibit a number of steps within the HIV-1 life cycle that involve integrase (18); however, the most intensely studied compounds inhibit the strand transfer of the proviral DNA into the host genome (strand transfer inhibitors [STIs]) (12).

HIV-1 integrase comprises a three-domain monomer (5, 6) and a multimeric quaternary structure. The N-terminal domain (codons 1 to 50) is alpha-helical, binds zinc, and is important for protein stability. The peptide chain that links this domain to the catalytic core may be structurally disordered and has not been resolved by X-ray crystallography. The catalytic core domain (codons 51 to 211) forms a two-layer alpha-beta sandwich and contains the active-site triad of residues (Asp 64, Asp 112, and Glu 152) that coordinates magnesium ion binding and the catalysis of the strand transfer process (16). In silico studies have suggested that STIs prevent integrase activity by binding to the active site via the magnesium ion (29). The C-terminal domain (codons 212 to 288) forms an all-beta-sheet SH3 domain that binds DNA in a nonspecific manner (21). The quaternary structure of HIV-1 integrase has not been resolved; however, crystallography and modeling studies have proposed that integrase forms a tetrameric structure (33). This tetrameric model contains two exposed active-site triads and allows for interactions between the three monomeric domains on different chains.

The development of inhibitors that are active against integrase has been slow due to a lack of effective screening methodologies and suitable lead compounds with low toxicity. Nevertheless, two compounds are in advanced stages of clinical evaluation (raltegravir [MK-0518], phase III, and elvitegravir [GS-9317], phase II). As with all other classes of HIV-1 inhibitors, determining and predicting resistance to these compounds will be critical to their effective use in the clinical setting (3, 9). Data from clinical trials and the in vitro passage of HIV-1 in the presence of both lead inhibitors and related compounds has generated an extensive list of mutations that modulate the sensitivity of HIV-1 to these compounds (Table 1). Resistance data from clinical trials has shown that there are two distinct resistance pathways to raltegravir, which are determined by the presence of either the N155S mutation or the Q148K/R/H mutation (13). Depending on which of these mutations develops initially, the accessory mutations that confer high-level resistance will differ. Both of these resistance pathways also confer high-level resistance to elvitegravir (22), even though clinical trial data and in vitro analysis show that alternative mutations also can cause resistance to this inhibitor.

An analysis of the frequency of the occurrence of polymorphisms associated with resistance to HIV-1 integrase inhibitors (Table 1) showed that mutations V72I and V201I were extremely prevalent across all subtypes and occur in 80 and 46% of these 1,165 sequences, respectively (Fig. (Fig.2a).2a). Other mutations, such as V74I/M, T97A, M154I, V165I, and T206S, occur with a frequency of greater than 12% in certain subtypes. Overall, 25 out of 41 integrase resistance mutations were detected within 1,165 inhibitor-naïve sequences spanning all HIV-1 subtypes. The resistance-associated mutations that occurred as natural polymorphisms within the Los Alamos database tended to occur within the most populated subtypes (CRF_01, CRF_02, A, B, C, D, F, and G). It was impossible to say whether the observed relative absence of resistance mutations in subtypes H, J, K, and N was due to the small numbers of sequences from these subtypes, such that the genetic diversity of these subtypes was not represented, or because of some inherent biological constraint. All of the inhibitors and pathways to resistance described previously include mutations that were present within a broad-spectrum inhibitor-naïve population (Fig. (Fig.2b).2b). However, the key mutations ascribed to resistance (where these are known) were rare; thus, key mutations specific for resistance to raltegravir were found only in the subtype B sequence set, which contained one N155H and three G140S/A polymorphisms. Similarly, the key mutations specific for resistance to elvitegravir were either not detected or detected at very low frequency: H51Y, one subtype A sequence; S147G, three subtype CRF_01 and C sequences; and E157Q, eight subtype CRF_02, A, B, C, and D sequences.

In order to utilize HIV-1 sequence data prospectively rather than purely retrospectively, amino acid covariation analysis was performed to determine whether resistance-associated mutations covaried within inhibitor-naïve sequence sets. Covariation analysis has been used extensively to determine sequence-based interactions between HIV-1 drug resistance mutations in the protease and reverse transcriptase enzymes (2, 14, 34) and to infer mechanistic interactions between amino acids in the V3 loop region of gp120 (15). The absence of drug selection pressure on the integrase sequences used within this study conferred an established problem with the determination of amino acid covariation (10); namely, that covariation could be observed as a function of mechanistic interactions between specific amino acids or as a function of the evolutionary relationships between sequences. Without drug selection pressure, some of the observed polymorphisms may confer a selective advantage, but it is unclear what generates that selection pressure. It is conceivable that there are natural inhibitors of integrase or that the inhibition of HIV by inhibitors targeted to other enzymes indirectly alter the selection pressure applied to the integrase coding region. We hypothesize that pairs of mutations in sequences that cluster within a phylogeny are more likely to be mutations that have been fixed within specific clusters by chance. Conversely, pairs of mutations that are present in sequences distributed across a phylogeny may represent multiple mutation and selection events at these positions. In the latter case, amino acid covariation is more likely to represent functional interactions between these residues rather than evolutionary patterns. For this reason, the detection of covariation was weighted toward sequences that did not cluster within a phylogeny.

Covarying mutations at codons 72 and 201 were present in 27 subtype B sequences, and those sequences had an average overall distance (D) of 0.39 (Fig. (Fig.3a).3a). Mutations at codons 154 and 201 co-occurred in 23 sequences, with an overall distance of 0.15, which was lower because all except three of the sequences containing these mutations were present within a single cluster (Fig. (Fig.3b).3b). The distribution of the 154 and 201 mutations within a sample of the random sequence phylogeny (1,000 sequences from a total random population of 10,000) did not show any clustering (Fig. (Fig.3c).3c). To improve the stability of the prediction model, the distance between sequences containing covarying amino acids within the random model (R_D; see Materials and Methods) was estimated using 10 iterations of 10,000 random sequences per subtype. On the basis of this analysis, the R_D was fixed at 0.43 for all covarying pairs of amino acids using sequences from subtypes A, B, and C (average R_D for subtype A, 0.44 [standard deviation, 0.034]; average R_D for subtype B, 0.43 [standard deviation, 0.041]; and average R_D for subtype C, 0.42 [standard deviation, 0.042]).

As the distance between sequences within the data model (D) usually was lower than the random model distance (R_D), Z scores calculated with the incorporation of the distance model (Z_D) were lower than those where this model was not used (Z) (Fig. (Fig.4a).4a). The incorporation of the distance correction factor resulted in the exclusion of 75.3% of the amino acid pairs predicted as covarying significantly using the FDR and a P value of less than 0.001. A comparison between the weighted (Z_D) and unweighted (Z) scores suggested that the weighted score was more sensitive to convergence between sequences and that HIV-1 integrase showed evidence of convergence. Simulations of six sets of 1,000 sequences, constrained such that the distance scaling factor D was fixed within the ranges of 0.1 to 0.3 (most similar, least convergent) to 0.6 to 0.8 (most dissimilar, more convergent), with each bin range increasing by 0.1, were generated, and the relationships between Z_D and Z were plotted. A binomial curve was fitted to each of the scatterplot series, with correlation values ranging between 0.51 (D, 0.1 to 0.3) and 0.84 (D, 0.3 to 0.5). The lowest correlation coefficients were obtained where D was constrained at the highest and lowest ends of the constraint scale. The data from the simulated data was limited to the range of Z and Z_D values observed using the subtype B integrase sequences, and these data and a fitted curve also were presented (r² = 0.63) (Fig. (Fig.4b).4b). As expected, the analysis of weighted and unweighted Z scores, where the distance scaling factor was constrained to a score similar to that used within the random model (D, 0.43), produced a linear relationship. Covariation between sequences constrained to have a high distance scaling factor, indicative of convergence, produced shallow curves that reached a Z score plateau with an increasing Z_D score. While the line generated by observed subtype B data is elevated from the highly convergent simulated sequences, the shape of the curve suggests that the observed data shows a level of sequence convergence at high Z_D scores. The minimum Z_D scores required to give a significant estimated FDR-corrected probability across a range of P values were calculated for each of the six simulated data sets (Fig. (Fig.4c).4c). The data set for which D was constrained between 0.1 and 0.3 resulted in the highest Z_D scores across the range of P values, and the data set for which D was constrained between 0.6 and 0.8 required the lowest Z_D scores to achieve significance. The P value cutoff used to analyze the HIV-1 subtype data sets was 0.001. In all cases, the minimum Z_D score used to assign significance to an instance of covariation in the observed data sets was higher than the minimum Z_D scores generated by the simulations.

Notably, the major mutations associated with pathways to resistance for either raltegravir or elvitegravir (N155H, Q148K/R/H, and E92Q [Table 1]) were not present in the inhibitor-naïve sequence population. However, many of the other mutations associated with these resistance pathways were present, suggesting that the rate at which resistance develops is affected by the baseline genotype. There was no obvious relationship between the subtype and the frequency of occurrence of polymorphisms, other than a slight increase for the more frequently sampled subtypes.

An analysis of pairwise amino acid covariation was performed in an attempt to use sequence analysis prospectively to determine whether resistance-associated polymorphisms covaried and whether there were any previously unidentified positions that covaried with these resistance positions. In order to prevent bias toward the detection of phylogenetically related amino acid pairs, a novel modification of the Jaccard index (26) was implemented. The software generated to perform the covariation analysis takes a multiple-sequence alignment and an optional mutation list as its input and therefore is applicable to a wide range of HIV or other protein sequence data sets. The program was written in Perl and is available on request from the authors.

The covariation analysis performed in this work is a prediction; therefore, in principal, it would have been beneficial to perform a receiver operating characteristic (ROC) curve analysis of the sensitivity and specificity of the modified Jaccard index algorithm. While simulating sequences that encompass a range of Jaccard indices and variable distance values is trivial, defining a set of true-positive and true-negative covarying pairs of amino acids from such a simulated data set is either arbitrary or circular. The frequency with which a pair of simulated covarying amino acids occur together within sequences or the Jaccard index (which is simply derived from the initial frequency) will lead to an arbitrary assignment of true-positive or true-negative covariation for that pair. Any derivation of the ROC using this arbitrary assignment method would be meaningless, especially when the error model was incorporated. The inherent error model within this covariation analysis method (26) compares the Jaccard index derived from a real data set to that derived from a simulated random data set and estimates a probability that a pair of observed amino acids covary to a greater extent than would be expected based on a random model. Using the estimated probability of pairs derived from a simulated data set to assign true-positive or true-negative status to a pair of covarying amino acids would be circular, in that both assignment and ROC curve analyses would be performed using the same score and, consequently, the method would appear to perform perfectly. For these reasons, ROC curve validation was omitted from this analysis and is not generally performed during analyses of covariation. Simulations of the effect of varying the distance score that was incorporated into the probability estimated from the Jaccard index were included, and these suggested that the evolution of the HIV integrase coding region showed signs of convergence due to the repeated observation of pairs of covarying amino acids within separate branches of each subtype phylogeny.