(462d) Near Perfect Prediction of HIV-1 Coreceptor Usage Reveals the Interactions Driving Tropism

In recent years, significant advances in the treatment of HIV-1 have been made, and one class of drugs that has contributed to that success is inhibitors that target chemokine receptors CCR5 and CXCR4, collectively known as the HIV-1 coreceptors (1). These therapeutics, including Maraviroc, are able to circumvent the difficulties of thwarting the quickly mutating HIV-1 by targeting host cell coreceptors, and inhibiting a key interaction with the HIV-1 gp120 V3-loop necessary for entry into host cells. The situation is further complicated by HIV-1 tropism, or the ability of the virus to change the coreceptor used, with the transition from a CCR5-specific (R5) virus to a CXCR4-specific (CXCR4) virus often indicating a progression to advanced stages of infection (2). Therefore, tropism determination is performed in conjunction with coreceptor inhibitors to ensure the success of a treatment regimen. Phenotypic methods, such as the Trofile assay, can be costly with a slow turn-around, so genotypic methods based on sequencing the V3-loop and using bioinformatics methods to predict coreceptor usage are also used (3-6).

New structural data regarding the interactions between the coreceptors and their ligands, including two computationally derived structures of CCR5/CXCR4:V3-loop complexes developed by our group (7, 8), have provided molecular level details of potential interactions driving tropism. However, to date, the specific interactions driving tropism have yet to be identified. To this end, we have developed a multifaceted hybrid approach for elucidating the interactions that drive coreceptor usage, using tools from computational biophysics, structural bioinformatics, and machine learning.  Molecular dynamics simulations were used to investigate structural and physicochemical variability of V3-loop:coreceptor complexes. Biophysical insights were converted into structural bioinformatics features using a statistical centroid-centroid force field (9). Nonlinear support vector machines (SVM) were trained to predict coreceptor usage based on the four rules (Net charge, 11/24/25, motif, length)(6), and extracted V3-loop:coreceptor interactions. Finally, a novel non-linear feature selection algorithm was used to narrow down the necessary and sufficient V3-loop:coreceptor interacting pairs.

Of the 624 sets of interactions/rules that were evaluated for predictive utility, the most accurate set achieved a median area under the receiver operator curve (AUC) of 0.981, a median classification accuracy of 0.935, and a median sensitivity of 0.917 at a false positive rate of 5%.  We identify a set of only 10 interactions between the V3-loop residues and CCR5/CXCR4 residues, which in combination with the four rules provide a median AUC accuracy of 0.979. An additional 12 interactions, for a total of 22 interactions, are selected when using interactions only and we still attain an AUC accuracy of 0.974. The identified interactions provide new insights into the factors driving HIV-1 tropism, while the resulting highly accurate and low dimensional SVM model represents a novel method for the prediction of coreceptor usage.

1.      F. S. Younai, Thirty years of the human immunodeficiency virus epidemic and beyond. International Journal of Oral Science 5, 191–199 (2013).

2.      E. A. Berger, P. M. Murphy, J. M. Farber, Chemokine receptors as HIV-1 coreceptors: Roles in viral entry, tropism, and disease. Annual Review of Immunology 17, 657–700 (1999).

3.      J. N. Dybowski, D. Heider, D. Hoffmann, Prediction of Co-Receptor Usage of HIV-1 from Genotype. Plos Computational Biology 6 (2010), doi:10.1371/journal.pcbi.1000743.

4.      O. Sander et al., Structural descriptors of gp120 V3 loop for the prediction of HIV-1 coreceptor usage. Plos Computational Biology 3, 555–564 (2007).

5.      K. Bozek, T. Lengauer, S. Sierra, R. Kaiser, F. S. Domingues, Analysis of Physicochemical and Structural Properties Determining HIV-1 Coreceptor Usage. Plos Computational Biology 9 (2013), doi:10.1371/journal.pcbi.1002977.

6.      C. A. Kieslich, D. Shin, A. Lopez de Victoria, G. Gonzalez-Rivera, D. Morikis, A Predictive Model for HIV Type 1 Coreceptor Selectivity. Aids Research and Human Retroviruses 29, 1386–1394 (2013).

7.      P. Tamamis, C. A. Floudas, Molecular Recognition of CXCR4 by a Dual Tropic HIV-1 gp120 V3 Loop. Biophysical Journal 105, 1502–1514 (2013).

8.      P. Tamamis, C. A. Floudas, Molecular Recognition of CCR5 by an HIV-1 gp120 V3 Loop. Plos One 9 (2014), doi:10.1371/journal.pone.0095767.

9.      R. Rajgaria, S. R. McAllister, C. A. Floudas, Distance dependent centroid to centroid force fields using high resolution decoys. Proteins-Structure Function and Bioinformatics 70, 950–970 (2008).