Proteochemometric modeling of HIV protease susceptibility
© Lapins et al; licensee BioMed Central Ltd. 2008
Received: 21 December 2007
Accepted: 10 April 2008
Published: 10 April 2008
A major obstacle in treatment of HIV is the ability of the virus to mutate rapidly into drug-resistant variants. A method for predicting the susceptibility of mutated HIV strains to antiviral agents would provide substantial clinical benefit as well as facilitate the development of new candidate drugs. Therefore, we used proteochemometrics to model the susceptibility of HIV to protease inhibitors in current use, utilizing descriptions of the physico-chemical properties of mutated HIV proteases and 3D structural property descriptions for the protease inhibitors. The descriptions were correlated to the susceptibility data of 828 unique HIV protease variants for seven protease inhibitors in current use; the data set comprised 4792 protease-inhibitor combinations.
The model provided excellent predictability (R2 = 0.92, Q2 = 0.87) and identified general and specific features of drug resistance. The model's predictive ability was verified by external prediction in which the susceptibilities to each one of the seven inhibitors were omitted from the data set, one inhibitor at a time, and the data for the six remaining compounds were used to create new models. This analysis showed that the over all predictive ability for the omitted inhibitors was Q2 inhibitors = 0.72.
Our results show that a proteochemometric approach can provide generalized susceptibility predictions for new inhibitors. Our proteochemometric model can directly analyze inhibitor-protease interactions and facilitate treatment selection based on viral genotype. The model is available for public use, and is located at HIV Drug Research Centre.
Despite huge efforts to prevent the spread of HIV, its prevalence continues to increase. Currently over 40 million persons are infected with HIV, and more than 4 million become infected and almost 3 million die from AIDS every year . Intensive treatment with antiretroviral drug combinations has substantially prolonged patient survival. However, the virus is prone to rapid mutation and drug resistant strains emerge, particularly in patients in whom the replication of the virus is only partially suppressed by treatment. The high rate of HIV mutation presents a challenging clinical problem, even a non-treated patient can host many viral variants from which drug resistant strains may evolve once therapy is instituted.
A major pharmacological target in HIV is its protease. The HIV protease is a dimeric protein composed of two identical 99-amino-acid monomers. The protease cleaves the viral Gag-Pol polyprotein, which is a necessary step in the generation of new virus particles. Thus, the HIV protease is essential for the propagation of the virus; nine of the 28 anti-HIV drugs and combination regimens in current use target the HIV protease. However, soon after the introduction of the HIV protease inhibitors it was found that the virus accumulates mutations in the protease, permitting eventual escape from anti-viral therapy. As protease inhibitors differ in their resistance profiles a proper selection of the inhibitor can aid therapy in such cases of drug resistance. The PhenoSense susceptibility test is a widely used bioassay for measuring viral survival during specific drug treatment [2, 3], and this assay is used to develop a proper treatment strategy for individual patients.
A more straightforward and cost-effective method for formulating a therapeutic strategy would be to predict drug susceptibility directly from the HIV genome sequence. Several types of modeling approaches have been developed, variously based on neural networks , support vector machines [5, 6], and other methods [6–8]. A drawback with all of these approaches was that they treated each anti-retroviral drug separately; each inhibitor required a separate model. Accordingly, none of these models can predict the effectiveness of a new drug for mutated proteases. However, such predictions are possible using our proteochemometric approach [9, 10]. Proteochemometrics utilizes the physico-chemical and structural properties of series of ligands and proteins to predict their interaction . Proteochemometrics has been successfully used to model various classes of G-protein coupled receptors [9, 11–17], antibodies , as well as aspartate proteases' ability to cleave their substrates . Here, we show that proteochemometrics can be used to model HIV protease resistance.
Development of a proteochemometric model for drug susceptibility prediction
Performance of proteochemometric models for HIV-1 protease drug susceptibility predictions.
Goodness of fit (R2)
Predictive ability (Q2)
Results of permutation test
P+I, P × I
P+I, P × I, P × P
While all models were statistically valid, Model-2, which included protease-inhibitor cross-terms, performed substantially better than Model-1, which contained only protease and inhibitor descriptors. Adding intra-protease cross-terms (Model-3) provided further improvement. Results from permutation testing also indicated the statistical validity of the models. Thus, for none of the models did the Q2 intercept show a positive value, ensuring that the high original Q2 values were not obtained by pure chance.
As seen in Table 1, adding new descriptor blocks resulted in more positive values for the R2 intercepts (although they remain below the desired level of 0.3), confirming that an increase in the number of X variables often results in better-fitted models in which part of the y data becomes explained by accumulated chance-correlations. Still, the models' predictive ability and interpretability improves because Q2 values increase (in contrast to its intercept for randomized data) and root mean squared errors of prediction (RMSEP) values decrease (Table 1). Thus, according to this analysis, Model-3 is the best performer. The good performance of Model-3 was further demonstrated by its true outer cross-validation; its external predictive ability amounted to P2 = 0.85.
External validation of the drug susceptibility model
External predictions by proteochemometric HIV-1 protease susceptibility models.
These data substantiate the validity of Model-3 and its modeling approach; therefore, we used Model-3 for all subsequent predictions and interpretations.
Use of the drug susceptibility model to analyze the role of individual amino acids in drug resistance
The PLS algorithm we used for model building derives a linear regression equation in which the coefficients reveal the direction and magnitude of the influence of X-variables on the response variable (i.e. protease inhibitor susceptibility). A large absolute value of a coefficient for a protease descriptor (i.e. a coefficient for a z-scale encoding the physico-chemical properties of amino acids at a particular position; see Methods for details) indicates that mutations at this position could induce a large change in drug susceptibility. Alternately, a large absolute value for a coefficient of a cross-term of protease-protease inhibitor descriptor reveals that mutations of the amino acid included in the cross-term can induce large changes in the susceptibilities for some particular inhibitors and not-so-large changes in the susceptibilities for other inhibitors. Finally, a large absolute value of a coefficient for an intra-protease cross-term pinpoints mutations in the protease that regulate drug resistance in a cooperative manner (for a deeper discussion and mathematical derivations relating to this discussion, see ).
Several different mutations (24I, 54V, 73T, 84V, and 90M) reduce susceptibility to all seven inhibitors, although there is variation in the extent of the reduction for specific inhibitors. Moreover, our analysis revealed that the more recently introduced inhibitors lopinavir and atazanavir do not provide increased activity as compared to older protease inhibitors against protease variants bearing these mutations. These findings indicate the need for novel, more adaptive agents that can inhibit proteases harboring these deleterious mutations.
Online prediction of susceptibility resulting from accumulated mutations
Highly resistant forms of HIV protease evolve by accumulating multiple susceptibility-decreasing mutations. The good predictive abilities of Models-1 and 2 indicate that the logarithmically transformed susceptibility data is, to a large extent, a function of the additive independent contributions of each mutation. However, Model-3, which also included intra-protease cross-terms, showed even better predictive ability suggesting that mutations may also interact cooperatively to modify drug resistance.
Complete analysis of the contributions of all possible combinations of amino acid mutations and protease inhibitors to drug resistance is an extensive task and could not be presented easily in a written account such as this. Therefore, we made our model available to the public in form of a Web service, so that users can submit their protease sequence and receive a prediction of drug susceptibility. The use of a Web service makes facilitates integration in other applications and workflows. Access to the Web service is available at HIV Drug Research Centre .
We used proteochemometrics to model susceptibilities of multiple HIV-1 protease variants to seven clinically used protease inhibitors, yielding a model with very good predictability and interpretability. We thoroughly validated the external predictive ability and the statistical significance of the model estimates, and conclude that our model can be reliably applied to the prediction and interpretation of the mechanisms of drug resistance. In fact, our model shows much better goodness-of-fit and predictability (in terms of R2, Q2 and RMSEP) than the hitherto best-performing models reported elsewhere, which, to the best of our knowledge, were obtained by the use of Support Vector Machines applied to each protease inhibitor separately .
Our model uses physico-chemical property (z-scale) descriptors of amino acids rather than encoding the mutations by letter codes or binary indicator variables. This is highly advantageous as many sequence residue positions of the HIV protease are often mutated to amino acids that share similar physico-chemical properties. Our model can evaluate the contribution of each encoded property (e.g. hydrophobicity, steric properties, charge, etc.) to drug susceptibility and perform predictions for mutations to any amino acid, as long as the amino acid's properties fall within the scope of the model. In other words, the information gained from sequence positions with multiple mutations provides for predictions for novel mutations at the same position.
Besides identifying mutations that contribute to general resistance to protease inhibitors, the model also reveals the susceptibilities of particular combinations of protease inhibitors and protease mutants. Thus, the model identifies specific relationships between a particular inhibitor and a particular amino acid(s) in the protease, information that could be useful for analyzing the mechanisms behind inhibitor failure. Moreover, the predictive ability of the model enables the development of targeted treatment based on the genome of a particular viral variant.
It is here appropriate to mention that an alternative way of resistance to protease inhibitors is mutations of the Gag-Pol cleavage sites which lead to enhancement of the processing efficiency of the substrate . Unfortunately the Stanford HIV Drug resistance database does not provide the full genomes of the HIV isolates, and this precludes us to linking the present data-set to complementary mutations in the cleavage sites. However, it seems unlikely that cleavage site mutations play any major role in explaining drug susceptibility in the present case. This is because our model, which was based on protease and inhibitor chemical properties alone, explained over 90% of the variation in the data, where the unexplained variation presumably essentially just represents measurement errors.
In this study we encoded the 3D structures of protease inhibitors by so called GRIND descriptors. These descriptors provide quantitative characterization of the ability of a molecule to form H-bond donor/acceptor and hydrophobic interactions with pharmacophoric groups located at various distance ranges around the molecule. Moreover, we also used a recently-developed GRIND descriptor type (TIP) which describes differences in size and shape of the molecules. These descriptions thus account for all major types of interaction that may contribute to inhibitor binding within the HIV protease, as well as those that can destroy binding (e.g. by steric hindrances). PCA was then applied to compress the descriptions into six orthogonal principal components, and in this case (obtaining six components for seven inhibitors) the PCA did not discard any of the information in the original descriptors. Accordingly the PCAs used herein allow a complete back-tracing to their origin in the original descriptors . An advantage of using GRINDs is that they do not require alignments of molecules and thus are not limited to narrow series of congeneric compounds. On the other hand, difficulties may arise in model interpretations since structural modifications in the molecule often influence the values of multiple GRIND descriptors. A practical approach in the design of improved compounds in such a situation is to predict changes of susceptibility patterns for in-silico modified molecules. For example, such predictions suggest that the loss of susceptibility of nelfinavir to the D30N mutant should be possible to counteract by modifying the 3-hydroxy-2-methylphenyl group of the compound (data not shown). In other words, according to these predictions the resistance to nelfinavir arises due to less favorable interactions of the 3-hydroxy-2-methylphenyl group with the mutated protease compared to its interactions with the wild type protease. Thus our modeling approach may find use to predict susceptibilities to new inhibitors and could potentially be applied in design of new inhibitors. Current susceptibility data is limited to the few clinically used protease inhibitors but proteochemometric modeling could be applied in a more general fashion and aid in the design of new agents with improved ability to withstand the development of resistance.
In summary, proteochemometrics is well suited to the study of HIV protease drug resistance. Our model predicts that relatively few of the more common mutations contribute substantially to a general loss of susceptibility, suggesting that there are limits as to how the virus can escape from the inhibitors. Whether the capacity of the protease to mutate into drug resistant variants is restricted due to inherent biological factors or whether new mutations would appear in response to broader chemical diversity of protease inhibitors remains to be determined. The appearance of new mutations in response to new treatments would require repeated agglomerative modeling of susceptibility data. Analysis of larger datasets (comprising more chemical compounds and more viral variants) would improve the resolution and predictive ability of the proteochemometric model and consequently augment its potential application to drug design and therapy optimization.
Susceptibility data for the seven clinically used HIV-1 protease inhibitors amprenavir, atazanavir, indinavir, lopinavir, nelfinavir, ritonavir, and saquinavir, measured by the PhenoSense assay , were collected from the Stanford HIV Drug Resistance Database . In short, the PhenoSense assay estimates the concentration of the anti-HIV drug that causes 50% inhibition of an HIV isolate's replication in a cell-based assay. The fold-decrease in susceptibility (here abbreviated as FDS) was determined by dividing this concentration by the concentration of the drug causing 50% inhibition of a drug-sensitive reference virus (the wild-type strain NL4-3). Thus, FDS = 1 indicates unchanged susceptibility to a drug, while FDS > 1 indicates decreased susceptibility, that is, increased resistance of the tested isolate as compared to the standard.
We retrieved susceptibility values for 4,794 unique protease-inhibitor pairs (comprising 828 unique protease sequences) from the database. For five of the seven inhibitors, susceptibility data was available for 775 to 824 proteases. For atazanavir and lopinavir, which are more recently approved protease inhibitors, susceptibility data was available for 319 and 513 proteases, respectively.
Numerical descriptions for proteochemometric modeling
Description of proteases
Of the 99 amino acid positions in each protease monomer, 80 were found to be mutated in the data set. Mutated positions were encoded by three z-scale descriptors, z1-z3, of amino acids derived by Sandberg et al. . The three z-scales are based on 26 computed and measured physico-chemical properties of amino acids, and represent hydrophobicity, steric properties, and electronic properties of amino acids, yielding 80 × 3 = 240 protease descriptors.
Description of protease inhibitors
3D structures of organic compounds were generated using Corina software (Molecular Networks GmbH, http://www.molecular-networks.com), and were described by grid independent descriptors (GRINDs)  calculated using Almond 3.1 software (Multivariate Infometric Analysis S.r.l., http://miasrl.com). GRINDs are alignment-independent descriptors that relate to the ability of a molecule to form favorable interactions with independent pharmacophoric groups. Three groups were used: DRY (hydrophobic), O (H-bond acceptor), and N1 (H-bond donor). The overall shape of the molecule was represented by a "TIP-field" using a recently described approach in which the region with repulsion energy of 1 kcal/mol for the N1 group is used to outline the surface of the molecule .
Generation of GRIND descriptors involves several steps: (1) calculation of interaction energies of the molecule with pharmacophoric groups located at grid points surrounding the molecule; (2) calculation of distances between grid points; (3) grid filtering (this is performed by selecting a certain number of grid nodes showing most favorable interactions with the molecule and concomitantly being situated as far as possible from each other); and, (4) computing the products of energy values for all pairs of the selected grid nodes. Finally, the maxima of products falling within specified distance ranges for node pairs obtained using the same probe (DRY, O, N1, and TIP auto-correlograms) and different probes (DRY-O, DRY-N1, DRY-TIP, O-N1, O-TIP, and N1-TIP cross-correlograms) are used as descriptors for the molecules . Thus, the capabilities of the protease inhibitors for hydrophobic, H-bond donor, and H-bond acceptor interactions and the differences in the molecular shapes of the protease inhibitors were encoded by 545 GRIND descriptors. To reduce the number of descriptors and to eliminate their mutual co-linearity, we applied principal component analysis , which transformed all GRINDs into six orthogonal principal components.
Protease-inhibitor and intra-protease cross-terms
Protein-ligand interactions are governed by complex processes that depend on the complementarity of the properties of the interacting entities. In proteochemometrics, this is accounted for by computing protein-ligand cross-terms . In order to account for effects of particular mutations on the susceptibility to particular inhibitors, we computed cross-terms by multiplying mean-centered z-scale descriptors with mean-centered principal components of GRINDs. This yielded 1,440 (240 × 6) protease-inhibitor cross-terms. To account for eventual co-operative coupling of mutations in the protease , we introduced intra-protease cross-terms. These were computed by multiplying mean-centered z-scale descriptors, which gave 240 × 239/2 = 28,680 cross-terms.
Preprocessing of data
Since descriptors were of different origins, they were centered and scaled to unit variance prior to their use. Moreover, to account for differences in the number of descriptors and their formed cross-terms, block scaling was applied. The block scaling was applied onto three descriptor blocks, namely, the block formed from ordinary protease and inhibitor descriptors (P+I block), the block composed of protease-inhibitor cross-terms (P × I block), and the block composed of intra-protease cross-terms (P × P block). Block scaling was performed by systematically varying the standard deviation of P+I block descriptors in one standard deviation intervals and the standard deviation of P × P block descriptors in 0.3 standard deviation intervals until an optimal model was obtained.
The dependent variable (FDS) was logarithmically transformed and mean centered prior to use in the computations.
Correlation by partial least-squares projections to latent structures
The above-derived descriptors and cross-terms were correlated to the susceptibility data by using the partial least-squares projections to latent structures (PLS). In PLS, the independent matrix of X variables (i.e., all descriptors and cross-terms) and a matrix of one or several dependent variables Y (in our case the logarithms of the FDS values comprise a single y vector) are simultaneously projected to latent variables (PLS components), with an additional constraint to maximize the covariance between the projections of X and Y (for an in-depth review of the PLS see ). PLS derives a regression equation for each response y in which the regression coefficients reveal the direction and magnitude of the influence of X-variables on the response.
Validation of modeling
The goodness-of-fit of the PLS models was characterized by the fraction of explained variation of the Y (R2). The predictive ability was characterized by the fraction of the predicted Y-variation (Q2), assessed by cross-validation with seven randomly formed groups, as previously described . The R2 values vary between 0 and 1, where a higher value means a better fit. The Q2 values normally vary between 0 and R2; however, negative values may be encountered, indicating non-predictive models. In PLS, the R2 term increases with each extracted PLS component, while the Q2 value usually reaches a plateau and declines as the model becomes over-fitted. Hence, the predictive ability and not the goodness-of-fit should be used when assessing the optimal number of PLS components.
In addition to the conventional "inner" cross-validation, we performed outer cross-validation in which the entire modeling process (description, scaling, and PLS fit) was performed independently from the excluded data; that is, it corresponded to the modeling practice of the "training set" and "test set" of the data but was performed seven times on random selections of data . The performance of outer cross-validation was assessed by the P2 value, which is calculated in the same way as Q2 of the inner cross-validation.
To assess the statistical significance of the estimated Q2 and R2 values, we employed permutation testing [35, 36]. The susceptibility data was randomly reordered 20 times, and separate models were fitted, correlating X data to each of the permuted y. The results of permutation testing can be displayed by plotting the R2 and Q2 values of these models of partially random data versus the correlation coefficient between the original y and permuted y, and drawing the regression line . The intercepts of the regression lines (that is, when the correlation coefficient is zero) represent the R2 and Q2 of a purely random model. To affirm full statistical significance of the original estimates the desirable limits of intercepts should be R2 intercept <0.3 and Q2 intercept <0.05 .
We also wanted to assess the ability of the proteochemometric model to predict susceptibility to novel inhibitors that were not present in the model in any combinations with the mutated proteases. Therefore, we removed all data for one inhibitor at a time and fitted the model for the remaining six inhibitors. X variables were re-centered and rescaled to unit variance and the y variable was re-centered prior to PLS modeling. The predictions for the excluded inhibitors were calculated from PLS models created on the reduced datasets and assessed by the RMSEP estimate.
This study was supported by the Swedish International Development Cooperation Agency (HIV-2006-019) and the Swedish VR (04X-05957).
- UNAIDS and WHO: AIDS epidemic update. UNAIDS, Geneva December 2006.Google Scholar
- Petropoulos CJ, Parkin NT, Limoli KL, Lie YS, Wrin T, Huang W, Tian H, Smith D, Winslow GA, Capon DJ, Whitcomb JM: A novel phenotypic drug susceptibility assay for human immunodeficiency virus type 1. Antimicrob Agents Chemother 2000, 44: 920–928. 10.1128/AAC.44.4.920-928.2000PubMed CentralView ArticlePubMedGoogle Scholar
- Heilek-Snyder G, Bean P: Role of HIV phenotypic assays in the management of HIV infection. Am Clin Lab 2002, 21: 40–43.PubMedGoogle Scholar
- Draghici S, Potter RB: Predicting HIV drug resistance with neural networks. Bioinformatics 2003, 19: 98–107. 10.1093/bioinformatics/19.1.98View ArticlePubMedGoogle Scholar
- Beerenwinkel N, Daumer M, Oette M, Korn K, Hoffmann D, Kaiser R, Lengauer T, Selbig J, Walter H: Geno2pheno: Estimating phenotypic drug resistance from HIV-1 genotypes. Nucleic Acids Res 2003, 31: 3850–3855. 10.1093/nar/gkg575PubMed CentralView ArticlePubMedGoogle Scholar
- Beerenwinkel N, Sing T, Lengauer T, Rahnenfuhrer J, Roomp K, Savenkov I, Fischer R, Hoffmann D, Selbig J, Korn K, Walter H, Berg T, Braun P, Fätkenheuer G, Oette M, Rockstroh J, Kupfer B, Kaiser R, Däumer M: Computational methods for the design of effective therapies against drug resistant HIV strains. Bioinformatics 2005, 21: 3943–3950. 10.1093/bioinformatics/bti654View ArticlePubMedGoogle Scholar
- Foulkes AS, De GV: Characterizing the Relationship Between HIV-1 Genotype and Phenotype: Prediction-Based Classification. Biometrics 2002, 58: 145–156. 10.1111/j.0006-341X.2002.00145.xView ArticlePubMedGoogle Scholar
- DiRienzo AG, DeGruttola V, Larder B, Hertogs K: Non-parametric methods to predict HIV drug susceptibility phenotype from genotype. Stat Med 2003, 22: 2785–2798. 10.1002/sim.1516View ArticlePubMedGoogle Scholar
- Lapinsh M, Prusis A, Gutcaits A, Lundstedt T, Wikberg JE: Development of proteochemometrics: a novel technology for the analysis of drug-receptor interactions. Biochim Biophys Acta 2001, 1525: 180–190.View ArticlePubMedGoogle Scholar
- Wikberg JE, Lapinsh M, Prusis P: Proteochemometrics: A tool for modelling the molecular interaction space. In Chemogenomics in Drug Discovery – A Medicinal Chemistry Perspective. Edited by: Kubinyi H, et al. Weinheim: Wiley-VCH; 2004:289–309.Google Scholar
- Lapinsh M, Prusis P, Wikberg JE: Proteochemometrics modeling of the interaction of amine G-protein coupled receptors with a diverse set of ligands. Mol Pharmacol 2002, 61: 1465–1475. 10.1124/mol.61.6.1465View ArticlePubMedGoogle Scholar
- Lapinsh M, Prusis P, Mutule I, Mutulis I, Wikberg JE: QSAR and proteochemometric analysis of the interaction of a series of organic compounds with melanocortin receptor subtypes. J Med Chem 2003, 46: 2572–2579. 10.1021/jm020945mView ArticlePubMedGoogle Scholar
- Lapinsh M, Prusis P, Uhlén S, Wikberg JE: Improved approach for proteochemometrics modeling: application to organic compound – amine G protein-coupled receptor interactions. Bioinformatics 2005, 21: 4289–4296. 10.1093/bioinformatics/bti703View ArticlePubMedGoogle Scholar
- Lapinsh M, Veiksina S, Uhlén S, Petrovska R, Mutule I, Mutulis F, Yahorava S, Prusis P, Wikberg JE: Proteochemometric mapping of the interaction of organic compounds with melanocortin receptor subtypes. Mol Pharmacol 2005, 67: 50–59. 10.1124/mol.104.002857View ArticlePubMedGoogle Scholar
- Prusis P, Muceniece R, Andersson P, Post C, Lundstedt T, Wikberg JE: PLS modeling of chimeric MS04/MSH-peptide and MC1/MC3-receptor interactions reveals a novel method for the analysis of ligand-receptor interactions. Biochim Biophys Acta 2001, 1544: 350–357.View ArticlePubMedGoogle Scholar
- Prusis P, Lundstedt T, Wikberg JE: Proteochemometrics analysis of MSH peptide binding to melanocortin receptors. Protein Eng 2002, 15: 305–311. 10.1093/protein/15.4.305View ArticlePubMedGoogle Scholar
- Strombergsson H, Prusis P, Midelfart H, Lapinsh M, Wikberg JE, Komorowski J: Rough set-based proteochemometrics modeling of G-protein-coupled receptor-ligand interactions. Proteins 2006, 63: 24–34. 10.1002/prot.20777View ArticlePubMedGoogle Scholar
- Mandrika I, Prusis P, Yahorava S, Shikhagaie M, Wikberg JE: Proteochemometric modelling of antibody-antigen interactions using SPOT synthesised peptide arrays. Protein Eng Des Sel 2007, 20(6):301–307. 10.1093/protein/gzm022View ArticlePubMedGoogle Scholar
- Kontijevskis A, Prusis P, Petrovska R, Yahorava S, Mutulis F, Mutule I, Komorowski J, Wikberg JE: A look inside HIV resistance through retroviral protease interaction maps. PloS Computational Biology 2007, 3(3):e48. 10.1371/journal.pcbi.0030048PubMed CentralView ArticlePubMedGoogle Scholar
- Prusis P, Uhlén S, Petrovska R, Lapinsh M, Wikberg JE: Prediction of indirect interactions in proteins. BMC Bioinformatics 2006, 7: 167. 10.1186/1471-2105-7-167PubMed CentralView ArticlePubMedGoogle Scholar
- Rhee SY, Fessel WJ, Zolopa AR, Hurley L, Liu T, Taylor J, Nguyen DP, Slome S, Klein D, Horberg M, Flamm J, Follansbee S, Schapiro JM, Shafer RW: HIV-1 Protease and reverse-transcriptase mutations: correlations with antiretroviral therapy in subtype B isolates and implications for drug-resistance surveillance. J Infect Dis 2005, 192: 456–465. 10.1086/431601PubMed CentralView ArticlePubMedGoogle Scholar
- HIV Drug Research Centre[http://www.hivdrc.org]
- Muzammil S, Ross P, Freire E: A major role for a set of non-active site mutations in the development of HIV-1 protease drug resistance. Biochemistry 2003, 42: 631–638. 10.1021/bi027019uView ArticlePubMedGoogle Scholar
- Olsen D, Stahlhut MW, Rutkowski CA, Schock HB, vanOlden AL, Kuo LC: Non-active site changes elicit broad-based cross-resistance of the HIV-1 protease to inhibitors. J Biol Chem 1999, 274: 23699–23701. 10.1074/jbc.274.34.23699View ArticlePubMedGoogle Scholar
- Nijhuis M, van Maarseveen NM, Lastere S, Schipper P, Coakley E, Glass B, Rovenska M, de Jong D, Chappey C, Goedegebuure IW, Heilek-Snyder G, Dulude D, Cammack N, Brakier-Gingras L, Konvalinka J, Parkin N, Kräusslich HG, Brun-Vezinet F, Boucher CA: A novel substrate-based HIV-1 protease inhibitor drug resistance mechanism. PLoS Med 2007, 4: e36. 10.1371/journal.pmed.0040036PubMed CentralView ArticlePubMedGoogle Scholar
- Rhee SY, Gonzales MJ, Kantor R, Betts BJ, Ravela J, Shafer RW: Human immunodeficiency virus reverse transcriptase and protease sequence database. Nucleic Acids Res 2003, 31: 298–303. 10.1093/nar/gkg100PubMed CentralView ArticlePubMedGoogle Scholar
- Sandberg M, Eriksson L, Jonsson J, Sjöström M, Wold S: New chemical descriptors relevant for the design of biologically active peptides. A multivariate characterization of 87 amino acids. J Med Chem 1998, 41: 2481–2491. 10.1021/jm9700575View ArticlePubMedGoogle Scholar
- Pastor M, Cruciani G, McLay I, Pickett S, Clementi S: GRid-INdependent descriptors (GRIND): a novel class of alignment-independent three-dimensional molecular descriptors. J Med Chem 2000, 43: 3233–3243. 10.1021/jm000941mView ArticlePubMedGoogle Scholar
- Fontaine F, Pastor M, Sanz F: Incorporating molecular shape into the alignment-free Grid-INdependent Descriptors. J Med Chem 2004, 47: 2805–2825. 10.1021/jm0311240View ArticlePubMedGoogle Scholar
- Wold S, Esbensen K, Geladi P: Principal component analysis. Chemom Intell Lab 1987, 2: 37–52. 10.1016/0169-7439(87)80084-9View ArticleGoogle Scholar
- Ohtaka H, Schön A, Freire E: Multidrug resistance to HIV-1 protease inhibition requires cooperative coupling between distal mutations. Biochemistry 2003, 42: 13659–13666. 10.1021/bi0350405View ArticlePubMedGoogle Scholar
- Geladi P, Kowalski B: Partial least-squares regression: a tutorial. Anal Chim Acta 1986, 185: 1–17. 10.1016/0003-2670(86)80028-9View ArticleGoogle Scholar
- Wold S: PLS for multivariate linear modeling. In Chemometric Methods in Molecular Design. Edited by: van de Waterbeemd H. Weinheim, Germany VCH; 1995:195–218.Google Scholar
- Freyhult E, Prusis P, Lapinsh M, Wikberg JE, Moulton V, Gustafsson MG: Unbiased descriptor and parameter selection confirms the potential of proteochemometric modelling. BMC Bioinformatics 2005, 6: 50. 10.1186/1471-2105-6-50PubMed CentralView ArticlePubMedGoogle Scholar
- Efron B: Better bootstrap confidence intervals. J Am Stat Assoc 1987, 78: 171–200. 10.2307/2289144View ArticleGoogle Scholar
- Eriksson L: Quantitative structure-activity relationship validation. In Quantitative structure-activity relationships in environmental sciences-VII SETAC, Pensacola Edited by: Schuurmann G, Chen F. 1997, 381–397.Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.