Protease inhibitor treatments.
Sequences of 2,244 protease isolates from 1,919 persons met the study selection criteria. Two isolates, one before and one after a protease inhibitor was received, were included from each of 325 persons. The sequences of 1,645 isolates from 1,344 individuals were published previously; the sequences of 599 isolates from 575 individuals have not been published previously. Table groups the isolates in the study according to the protease inhibitor treatments of the persons from whom isolates were obtained. Indinavir, saquinavir, and nelfinavir were each received by >500 persons. Ritonavir was received by 456 persons, ~60% of whom were receiving ritonavir at a low dose as part of a dual protease inhibitor combination. One hundred fifteen persons received amprenavir, which was approved in 1999, and eight persons received lopinavir, which was approved in 2001.
HIV-1 isolates and protease inhibitor exposurea
Protease mutations and their association with treatment.
The median number of protease mutations per isolate increased in proportion to the number of protease inhibitors received, from 4 mutations per isolate in untreated persons to 12 mutations per isolate in persons receiving four or more inhibitors (Fig. ). Table shows the mutation frequencies of the 99 protease positions according to the number of protease inhibitors received. Based on our chi-square analysis, mutations at 45 positions were found to be treatment associated in that mutation frequencies were significantly associated with treatment with at least one protease inhibitor. An additional 17 positions had non-treatment-related polymorphisms; these positions had mutations, but the mutation frequencies were not statistically associated with protease inhibitor treatment. The remaining 37 positions had mutation frequencies of <0.5%, even in isolates exposed to treatment, and were considered invariant.
FIG. 1. Histograms of mutation frequency according to the number of protease inhibitors (PIs) received. The median number of mutations (differences from the consensus B sequence) increased from 4 in untreated persons to 12 in persons receiving ≥4 inhibitors. (more ...)
Mutation frequencies at protease positions 1 to 99 according to the number of protease inhibitors received
The 45 treatment-associated positions included 23 positions previously associated with drug resistance (10, 20, 24, 30, 32, 33, 36, 46, 47, 48, 50, 53, 54, 60, 63, 71, 73, 77, 82, 84, 88, 90, 93) and 22 positions which had not previously been associated with drug resistance (11, 13, 22, 23, 34, 35, 43, 45, 55, 58, 62, 66, 72, 74, 75, 76, 79, 83, 85, 89, 92, 95). Thirteen of the 22 newly described treatment-associated positions (positions 11, 22, 23, 45, 58, 66, 74, 75, 76, 79, 83, 85, 95) showed little or no variation—mutation frequencies of <0.5%—in untreated persons, as shown in Table , column 0. These 13 positions played a significant role in HIV-1 protease variation, with mutations occurring in 92 of 637 (14.4%) persons receiving a single inhibitor and 162 of 603 (26.9%) persons receiving two or more inhibitors. These mutations usually occurred in isolates with one or more primary protease inhibitor resistance mutations (219 of 254 [85.8%]).
Our logistic regression analysis revealed that mutations at 24 positions had statistically significant positive linear relationships between the number of protease inhibitors received and the presence of a mutation (Table ). The positions with the strongest linear relationships were positions 10, 20, 46, 53, 54, 63, 71, 73, 82, 84, and 90. There was a statistically significant negative linear relationship between the number of inhibitors and the presence of a mutation at position 30.
Locations of protease mutations within the enzyme's three-dimensional structure.
The invariant HIV-1 protease positions include the active-site positions (positions 25 to 27); other positions in or near the substrate cleft (positions 28 to 29, 31, and 80 to 81); most of the N- and C-terminal domains, which together with the active site make up the dimer interface; and other positions that appear to be associated with maintaining the enzyme's conformation and flexibility (e.g., 10 conserved glycines, including 3 in the flexible tips of the enzyme flap at positions 49, 51, and 52). The polymorphic positions are found almost entirely in surface loops.
The 23 known drug resistance positions include six substrate cleft residues (positions 30, 32, 48, 50, 82, and 84); four flap tip drug resistance mutations (positions 46, 47, 53, and 54); position 90, which although not in the substrate cleft decreases susceptibility to multiple protease inhibitors; three additional residues which are generally mutated only in treated persons (positions 24, 73, and 88); and nine polymorphic residues (positions 10, 20, 33, 36, 60, 63, 71, 77, and 93). The 22 new drug resistance positions include one substrate cleft residue (position 23), three flap residues (positions 43, 45, 55), one terminal-domain residue (position 95), and 17 residues in the enzyme core. The substrate cleft residues at positions 48 and 50 are also in the protease flap tips.
Correlations between protease mutations.
To identify patterns of drug resistance mutations, we calculated the pairwise binary (phi) correlation coefficients among the 45 treatment-associated and 17 polymorphic protease residues. This analysis was performed separately for the 1,004 isolates from untreated persons and for the 1,240 isolates from treated persons to detect associations that were independent of the treatment status of the individuals from whom the sequenced isolates were obtained. Among the untreated isolates, 23 of the 2,080 possible pairwise correlations were statistically significant, including 19 positive (phi = 0.14 to 0.31) and 4 negative (phi = −0.14 to −0.21) correlations. Among the treated isolates, 115 of the possible 2,080 correlations were statistically significant, including 99 positive (phi = 0.13 to 0.63) and 16 negative (phi = −0.13 to −0.34) correlations.
Table shows the most strongly correlated pairs of positions among the 115 statistically significant correlations in isolates from treated persons. The three most strongly correlated pairs of positions among the treated isolates were 54 and 82 (phi = 0.63), 32 and 47 (phi = 0.51), and 73 and 90 (phi = 0.47). Mutations at two pairs of primary resistance positions had significant positive correlations: positions 84 and 90 and positions 48 and 82. Mutations at positions 82 and 90, although both common, were not significantly correlated with each other. Position 30 was negatively correlated with each of the other primary resistance positions. The positions with the greatest number of positive correlations were positions 10 (16 correlations), 46 (13 correlations), 71 (12 correlations), 90 (10 correlations), 20 (10 correlations), 73 (10 correlations), 82 (9 correlations), 63 (7 correlations), 84 (6 correlations), and 54 (6 correlations).
Most strongly correlated pairs of positions among 115 statistically significant correlations in isolates from treated personsa
Correlations usually involved the most common mutation at each of the two correlated positions (Table ). For example, the strong positive correlation between positions 54 and 82 (phi = 0.63) is in large part due to the strong correlation between I54V, the most common substitution at position 54, and V82A, the most common substitution at position 82 (for I54V and V82A, phi = 0.55). Other combinations of substitutions for these two positions were less commonly observed: I54T and V82A (phi = 0.21) and I54V and V82T (phi = 0.15). In some cases, covariation was dominated very strongly by particular combinations of substitutions. For example the positive correlation between positions 30 and 88 (phi = 0.40) was represented entirely by D30N and N88D (phi = 0.52) rather than by D30N and N88S (phi = −0.05), and the correlation between positions 48 and 54 (phi = 0.29) was represented largely by G48V and I54T (phi = 0.44) rather than G48V and I54V (phi = 0.19).
We can use our measurements of comutation frequencies to construct a graphical model that summarizes the relationships among positions in HIV-1 protease. In this model, we attempt to place positions with high degrees of comutation close together and positions with low or negative degrees of comutation far apart. These relationships are modeled as consistently as possible within the framework of a two-dimensional plot. One computational technique that generates such graphical models is called PCA. We performed PCA on the 45 positions that were associated with protease inhibitor treatment and used the matrix of correlation coefficients as a measure of similarity between positions. The results of our PCA are shown in Fig. . The figure shows that positions 30 and 88 cluster together and are separate from most other positions. It also shows a clustering of positions 54 and 82 and their separation from positions 73, 84, 90, and 93.
FIG. 2. PCA of the 45 positions associated with protease inhibitor treatment. The graph is a two-dimensional projection of the distances among the 45 positions, where the similarity between any two positions is measured by their binary (phi) correlation coefficient (more ...) Correlated mutations and protease residue contacts.
Among the 115 correlated residue pairs, 59 (51%) contained residues that were within 8 Å of each other—many more than the 5.5 pairs predicted when 115 pairs were selected at random. Most of the 59 pairs were close in each of the three structures we examined (liganded, unliganded, and open flap), but four were close only in the open-flap structure from a molecular-dynamics simulation. For example, residues 54 and 82 were separated by 5.4 Å in the open-flap structure but by 8.4 and 8.6 Å in the liganded and unliganded structures, respectively. One of the residue pairs could be explained only by contact between residues on different chains of the protease dimer (residue pair 48 and 82).
Fifty-six (49%) of the 115 correlated pairs were separated by >8 Å. Our Markov chain analysis showed that of these 56 pairs of residues, 16 could be linked by one residue, 21 by two residues, 13 by three residues, and 1 by five residues. However, our permutation analysis, which was designed to determine whether such chains were statistically significant, showed that this amount of chained covariation would be expected by chance in a molecule with 56 correlated, neighboring residues having the size and topology of HIV-1 protease. Therefore, compared with randomly selected residue pairs, the covarying residues we observed were significantly more likely to be within 8 Å of one another but not significantly more likely to be linked by chained covariation.
Figure shows the strongest positive correlations superimposed on the structure of the protease. Most of the strong correlations are in a plane that is adjacent to the substrate cleft and include residues 10, 24, 30, 46, 54, 82, 84, and 90.
FIG. 3. The 50 most highly correlated residues in isolates from treated persons are shown superimposed on the locations of these residues within the folded enzyme. The blue lines represent positively correlated residues (n = 44; phi > 0.2); the (more ...) Clusters of correlated residues.
Pairs of correlated residues can be further grouped into clusters in which all possible pairs within the cluster are mutually correlated. Among the 23 highly correlated pairs found in isolates from untreated persons, there were two mutational clusters, one of three residues and one of four residues. Among the 115 correlated pairs found in isolates from treated persons, there were 30 mutational clusters, ranging in size from three to six residues (Table ).
Clusters of correlated protease positions
Twenty of the 30 clusters in treated isolates contained one or more primary protease mutations, including L90M (12 clusters), V82ATF (6 clusters), and D30N (6 clusters). The substrate cleft mutation I84V was in four of the L90M clusters. The substrate cleft mutation G48V was in one of the V82ATF clusters. Flap tip positions were included in 4 of the 12 clusters containing L90M (position 46, 4 clusters), and each of the 6 clusters containing V82ATF (position 46, 4 clusters; position 53, 1 cluster; position 54, 4 clusters).
Six representative clusters from Table are shown in Fig. . These six clusters occurred in 17% of isolates from all treated persons and 29% of isolates from persons receiving two or more protease inhibitors. Published in vitro susceptibility results for isolates containing each of these six patterns of mutations (and no additional known resistance mutations) reveal that each pattern is associated with reduced susceptibility to each of the protease inhibitors: amprenavir, 2- to 5-fold; indinavir, 10- to 15-fold; lopinavir, 2- to 20-fold; nelfinavir, 10- to 30-fold; saquinavir, 3- to 30-fold; and ritonavir, 3- to 100-fold (25
FIG. 4. Six representative clusters from Table . Each position in a cluster demonstrates statistically significant mutational covariation with each of the other positions within a cluster. (A) Positions 10, 63, 71, 90, and 93; (B) positions 10, (more ...)