Coordinate linkage of HIV evolution reveals regions
of immunological vulnerability
Vincent Dahirela,b,c,1, Karthik Shekhara,b,1, Florencia Pereyraa, Toshiyuki Miurad, Mikita Artyomovc,e, Shiv Talsaniab,f,
Todd M. Allena, Marcus Altfelda, Mary Carringtona,g, Darrell J. Irvinea,h,i, Bruce D. Walkera,h,2
and Arup K. Chakrabortya,b,c,i,2
aRagon Institute of Massachusetts General Hospital, Massachusetts Institute of Technology, and Harvard University, Boston, MA 02129; Departments of
bChemical Engineering,cChemistry, andiBiological Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139;eDepartment of Chemistry,
Moscow State University, Moscow 119991, Russia;fDepartment of Chemical Engineering, Loughborough University, Leicestershire LE11 3TU, United Kingdom;
gCancer and Inflammation Program, Laboratory of Experimental Immunology, SAIC-Frederick, Inc., National Cancer Institute-Frederick, Frederick, MD 21702;
hHoward Hughes Medical Institute, Chevy Chase, MD 20815; anddInstitute for Medical Sciences, University of Tokyo, Tokyo 108-8639, Japan
Edited* by Laurie H. Glimcher, Harvard University, Boston, MA, and approved May 20, 2011 (received for review April 4, 2011)
Cellular immune control of HIV is mediated, in part, by induction of
single amino acid mutations that reduce viral fitness, but compen-
satory mutations limit this effect. Here, we sought to determine if
higher order constraints on viral evolution exist, because some
coordinately linked combinations of mutations may hurt viability.
Immune targeting of multiple sites in such a multidimensionally
conserved region might render the virus particularly vulnerable,
because viable escape pathways would be greatly restricted. We
analyzed available HIV sequences using a method from physics to
reveal distinct groups of amino acids whose mutations are collec-
tively coordinated (“HIV sectors”). From the standpoint of muta-
tions at individual sites, one such group in Gag is as conserved as
other collectively coevolving groups of sites in Gag. However, it
exhibits higher order conservation indicating constraints on the vi-
tions likely destabilize multiprotein structural interactions critical
for viral function. Persons who durably control HIV without medi-
cations preferentially target the sector in Gag predicted to be most
vulnerable. By sequencing circulating viruses from these individu-
als, we find that individual mutations occur with similar frequency
in this sector as in other targeted Gag sectors. However, multiple
mutations within this sector are very rare, indicating previously un-
such regions with higher order evolutionary constraints provides a
novel approach to immunogen design for a vaccine against HIV and
other rapidly mutating viruses.
cytotoxic T-lymphocyte response|elite controllers|random matrix theory
tinues to wreak havoc around the world, particularly in poor
nations. An efficient vaccine is urgently needed, and such a vac-
cine is likely to be one that induces both antibody and cytotoxic
T-lymphocyte (CTL) responses (1–3). The extreme variability of
the HIV envelope has precluded vaccine-induced generation of
broadly neutralizing antibodies (1) that can prevent acquisition,
leading some to focus on CTL-based vaccines (4) to prevent
disease progression and possibly to limit acquisition (5).
CTLs recognize and respond to viral peptides presented by
class I MHC proteins. Single point mutations within and sur-
rounding such HIV epitopes targeted by CTLs can enable escape
from immune pressure, leading to a focus on targeting conserved
regions of the HIV proteome. Conserved regions are defined to
be ones where the frequency of occurrence of mutations at single
sites is small, indicating [according to evolutionary theory (6)]
that the corresponding mutant viral strains are replicatively less
fit. Thus, if such a site is targeted, the outgrowth of a mutant
virus that escapes the immune pressure is less likely (7). The
emergence of a compensatory mutation that restores fitness is a
challenge to this approach (8).
espite the efficacy of life-extending medications, HIV con-
Characterizing the frequency of occurrence of viral strains
based on the effects of mutations at single sites results in a unidi-
mensional measure of conservation, which ignores potential cou-
to structural/functional constraints. A useful multidimensional
in the viral proteome could be identified, such that sites within
a group coevolve in a collectively interdependent manner to in-
groups. If such groups exist, it is possible that a greater proportion
of the combinations of mutations involving sites in the group
are harmful in some groups compared with other groups. Such
a group of sites is multidimensionally more conserved in that
another group (as it coevolves independently). Such regions
should be particularly vulnerable to multiple points of CTL pres-
sure, because escape pathways would be restricted. Targeting
multiple points would promote the emergence of multiple muta-
tions to escape the immune pressure, but multiple mutations in
such a region would be more likely to result in unfit viruses.
Thus, we sought to determine collectively coevolving groups
of residues in HIV proteins. We analyzed publicly available se-
quences from the Los Alamos HIV Sequence Database (http://
www.hiv.lanl.gov/) of clade B HIV proteins derived from patients
with diverse genotypes using random matrix theory (RMT) (9) to
identify groups of sites (not pairs) that evolve collectively, pre-
sumably because they act together to perform a function im-
portant for the virus. RMT has been applied to diverse realms of
physics, to analyze stock price fluctuations (10, 11), and to ana-
lyze sequences of an enzyme (12). Although inspired by Halabi
et al. (12), we outline our analysis of HIV proteins by analogy
with analyses of financial markets.
From price fluctuations of various stocks over time, one can
compute the extent of correlation between changes in price of
a pairofstocks,averagedoverall available timepoints. Thisyields
a matrix of pair correlations. Each element of the matrix describes
correlations between a particular pair of stocks, andmathematical
Author contributions: V.D., K.S., and A.K.C. designed research; V.D., K.S., and A.K.C.
performed research; F.P., T.M., M.C., and B.D.W. contributed new reagents/analytic tools;
V.D., K.S., M. Artyomov, S.T., B.D.W., and A.K.C. analyzed data; and V.D., K.S., T.M.A.,
M. Altfeld, D.J.I., B.D.W., and A.K.C. wrote the paper.
The authors declare no conflict of interest.
*This Direct Submission article had a prearranged editor.
Freely available online through the PNAS open access option.
1V.D. and K.S. contributed equally to this work.
2To whom correspondence may be addressed. E-mail: email@example.com or arupc@
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
| July 12, 2011
| vol. 108
| no. 28www.pnas.org/cgi/doi/10.1073/pnas.1105315108
properties of the entire matrix (Results) reveal groups of stocks
whose price fluctuations are collectively interdependent. These
properties of the matrix are influenced by “noise” because of a fi-
nite sample of times and correlations that reflect changes in
overall stock prices attributable to external forces (e.g., reces-
sions). RMT has been used to “clean” the correlation matrix of
these effects and obtain groups of companies whose economic
activities are intrinsically collectively coupled and essentially in-
was composed of stocks that define known economic “sectors,”
such as financial service companies and automotive companies.
A similar analysis has enabled us to identify collectively
coevolving groups of sites in the HIV proteome (which we term
HIV sectors). Further, we have identified a sector in which
mutations at multiple sites are collectively more constrained, and
thus might be particularly vulnerable to multiple points of im-
mune pressure. We examined immune targeting and sequenced
circulating viruses in those rare persons who are able to control
HIV without medications to determine whether such mechanisms
might contribute to durable immune control. These data support
our predictions. We suggest how this new understanding can be
harnessed to design immunogens for a vaccine against HIV.
Sequence Analysis Identifies Significant HIV Sectors. Each aligned
sequence of an HIV polyprotein is analogous to stock price data
at a particular time point. In a specific sequence, one asks if
a mutation away from the most frequent amino acid (“wild type”)
at a particular position i appears simultaneously with a similar
mutation at another site j, and this is done for all pairs of ami-
no acids. Average values for the frequency of occurrence of
double mutations for each pair of sites in the protein (fijfor sites i
and j) are obtained by repeating this procedure for all sequences
(Methods and SI Appendix 1). Similarly, the frequencies with
which mutations are observed at individual sites (fiand fjfor
sites i and j) are obtained. A pair correlation matrix, C, can be
(defined in the main text) with a given magnitude (abscissa) is shown. In total, 1,600 sequences of Gag polyproteins (500 residues long) were used. Note that
a relatively large number of sequences of HIV proteins are available. (Lower) Distribution of eigenvalues obtained from 1,000 randomly generated matrices of
the same size as C for Gag (described in the main text and SI Appendix 2). (B) Analyzing the eigenvectors corresponding to eigenvalues larger than the highest
eigenvalue for random matrices yields collectively coevolving groups of sites or “sectors” (SI Appendix 3–7). Sites within a sector are grouped together along
the rows and columns so that groups of collectively coevolving sites are vivid as squares along the diagonal of a heat map representing the values of the
correlations (obtained from the “cleaned” correlation matrix as described in SI Appendix 2). (C) Mean frequency of the dominant amino acid at single sites
within each of the five Gag sectors. (D–F) Threshold value defining a significant correlation was chosen to be such that correlations with magnitudes greater
than this value arise with vanishing probability in the randomized matrices (P < 0.02). Changing this threshold value does not change qualitative results (SI
Appendix, Fig. S11). (D) Percentage of significant negative (Cij< −0.03) pair correlations within each sector. (E) Percentage of significant positive (Cij> 0.1) pair
correlations within each sector. (F) Percentage of significant three-site negative (Cijk< −0.01) correlations within each sector (method in SI Appendix 17). (G)
Percentage of significant three-site positive (Cijk> 0.1) correlations within each sector.
Defining collectively coevolving groups of sites (sectors) in Gag. (A, Upper) Number of eigenvalues (ordinate) of the correlation matrix, C, for Gag
Dahirel et al.PNAS
| July 12, 2011
| vol. 108
| no. 28
defined, each element of which, Cij(for sites i and j), reflects
the interdependence of mutations at these two sites. Cij¼
quency of occurrence of a double mutant and that which would be
observed if the two mutations occurred independently; the var-
iances of the distributions of mutations (Viand Vj) at sites i and j,
respectively, normalize the values of Cijfor different pairs of sites.
Another method (12) for computing Cij yielded qualitatively
similar results (SI Appendix, Fig. S6).
Properties of this correlation matrix, termed eigenvectors,
contain information about collective coevolution of sites. Each
eigenvector represents a specific combination of sites, whose
mutations occur in a coupled way but are essentially independent
of mutations in a combination of sites represented by another ei-
genvector. The data on HIV evolution can be represented as de-
individual sites. The association of each site with a particular ei-
An eigenvector is also associated with a number, called an eigen-
value, whose magnitude reflects the contribution of this eigen-
vector to the correlations between sites (SI Appendix, Eq. S4).
Eigenvalues of the correlation matrix, C, for HIV Gag poly-
proteins are shown ordered according to their magnitudes (Fig.
1A, Upper). The corresponding eigenvectors describe groups of
sites whose mutations are collectively linked, presumably by the
requirements of maintaining viral fitness. However, the infor-
mation in some eigenvectors is not significant because of noise
(attributable to a finite sample of sequences) or because it re-
flects phylogeny (13).
RMT theorems can help to determine which eigenvectors re-
flectthe influence of noise because they describe the properties of
correlation matrices derived from independent random variables.
For example, given the length of the Gag polyprotein and the
number of available sequences, RMT states (SI Appendix, Eq. S3)
reflected noise only. However, this theorem assumes that the
length of the polyprotein and the number of sequences are very
over all aligned sequences of a particular HIV protein and recal-
culate C, we obtain a random correlation matrix of the same size
as our sample (SI Appendix 2). The distribution of eigenvalues
obtained by randomizing the sequences 1,000 times is shown in
Fig. 1A, Lower. We see that the RMT theorem is rather accurate
even for finite samples, because very few eigenvalues exceed 2.4
and none exceed 3. The continuous part of the eigenvalue spec-
in the same way as that for the randomized sequences (compare
panels in Fig. 1A for Gag). The eigenvectors corresponding to
eigenvalues <3 correspond to noise.
As noted before (12), and derived more completely in SI Ap-
pendix 3, the contribution of phylogeny alone to a representation
where the contributions from each individual site in the poly-
protein have the same sign. For all HIV polyproteins, the largest
eigenvalue corresponds to such an eigenvector. Because each
eigenvector is independent, we conclude that the eigenvector
corresponding to the largest eigenvalue describes correlations
attributable to phylogeny (13). Eigenvectors corresponding to the
next few largest eigenvalues contain information on collective
correlations between groups of sites that originate largely from
relationships between evolutionary constraints and sequence.
By parsing these eigenvectors, groups of sites (sectors) in HIV
proteins that coevolve together but essentially independent of
others were identified. Many sites do not collectively coevolve
we found five strongly collectively coupled groups of sites (SI
Appendix 6 and Fig. S4), which we term sectors 1–5. As expected,
we also found an additional group of residues that is weakly
and measures the difference between the fre-
coupled collectively, as a result of mutations arising in the context
of specific HLA class I molecules that drive mutations at multiple
sites (SI Appendix 8 and Figs. S4 and S5). The weakly collective
“footprints” [i.e., these mutations arise in persons with the same
HLA type (14)]; thus, we do not consider this quasi-sector further.
Gag sectors 1–5 can be visualized as a heat map reflecting the
correlations between sites (Fig. 1B). As expected, the sites that
comprise a sector are not contiguous along the linear protein
sequence. The positions of the sites along the rows and columns
in Fig. 1B group the sites in a sector together. For example, if
a sector was composed of sites 27, 45, and 53 along the linear
protein sequence, these sites would be adjacent to each other
along a row or column. Each group of sites that comprises
a sector is outlined in red. We observed similar groupings for other
HIV proteins (SI Appendix, Figs. S7–S10). We focus further
analysis on Gag because of data indicating correlations between
Gag-specific responses and viral control (15) and a strong impact
of Gag mutations on viral fitness (16).
Identification of an Immunologically Vulnerable Sector in Gag. From
the standpoint of identifying regions with single sites that are
more conserved, all five Gag sectors are equivalent (Fig. 1C).
However, if one examines the relative occurrence of positive and
negative correlations between multiple mutations at sites that
comprise each collectively evolving sector, differences between
the sectors become apparent.
Negative correlation between sites implies that the multiple
mutant isobserved lessfrequently than ifthe individual mutations
were to arise independently. For example, consider a simple case
being 90% conserved [i.e., the frequency with which single muta-
tions at sites i and j are observed is 10%, (fi= fj= 0.1)]. For Cij
(proportional to fij− fifj) to be negative, the frequency with which
the double mutant is observed (fij) must be less than 1%. This
implies that when multiple sites are negatively correlated, the
multiple mutant is considerably less fit compared with the single
escaping the immune pressure is likely to require more than one
mutation. The greater the proportion of negative correlations in
the sector, the more difficult it is to both escape the CTL pressure
and maintain virus fitness, because multiple mutations are much
less tolerated, likely because of fitness limitations.
Positive correlation between sites implies that the corre-
sponding multiple mutants are observed more frequently than if
the mutations were to occur independently. Thus, positive cor-
relations can be associated with compensatory mutations; in-
deed, known compensatory mutations do appear as positive
correlations (SI Appendix, Table S9). Thus, sectors characterized
by larger numbers of positive correlations would be less effective
targets, because escape from CTLs without substantial loss in
virus fitness is more likely.
For pairs of sites within a sector, one finds that Gag sector 3
contains the largest proportion of negative correlations compared
with other sectors and a relatively small number of positive cor-
relations (Fig. 1 D and E). This is also true for three-site (Fig. 1 F
and G) and higher order correlations. The magnitudes of the
negative pair correlations between sites in sector 3 are as large
as they can be, given the level of single site conservation in this
are expected to be observed very rarely. The eigenmaps in SI Ap-
pendix, Fig. S4 B and C also show that several sites in sector 3 are
collectively (i.e., at higher order) negatively correlated to several
others, which is not true for other sectors. Note that we do not
identify important negative pair correlations by screening all pair
correlations for those that exceed a cutoff (17). Rather, we first
identified groups of sites (sectors) that are significantly collectively
coupled (because RMT theorems help to eliminate noise-induced
| www.pnas.org/cgi/doi/10.1073/pnas.1105315108Dahirel et al.
correlations). We then examined the signs of multibody corre-
lations only within these meaningful collectivelycoupledsectors.
Our results suggest that sector 3 is the most immunologically
vulnerable multidimensionally constrained region in Gag, be-
cause multiple mutations are most constrained in this sector.
Because sequence analysis methods are not exact, we tested this
prediction, and its consequences for HIV infection and vacci-
nation, against existing and new experimental data.
Sector 3 Is Immunologically Vulnerable Because of the Importance
of Assembling Multiprotein Structures Critical for Viral Capsid
Formation. In sector 3, 52 of the 57 sites are contained in the p24
random (Fig. 2A; individual amino acids in SI Appendix, Table
S1). However, hexamers of the p24 protein form the viral capsid
(18), and superimposing sector 3 sites on the structure of the p24
is at interfaces between p24 proteins in a hexamer or at interfaces
of this group of sites originates from constraints important for
assembly of functional multiprotein structures, highlighting the
importance of determining collective correlations.
The structural locations of sector 3 sites suggest the origin of
negative correlations. Whereas one mutation in interface residues
may still allow formation of p24 hexamers and assembly of the
would likely destabilize these protein-protein interfaces. Fitness
Sector 1 sites also reflect structural constraints associated with
supramolecular assembly because they largely comprise the core
of the hexamer (Fig. 2C). However, our analyses suggest that
sector 1 is not as immunologically vulnerable as sector 3. This is
because of the following:
i) The ratio of negative to positive correlations for pairs of
mutations is a factor of 3 greater for sector 3.
ii) At the level of three-site correlations, the ratio of negative
to positive correlations is more than threefold greater for
iii) Sector 3 is characterized by collective higher order negative
correlations between sites that are absent in sector 1 (SI
Appendix, Fig. S4).
These results may be because multiple mutations in sector 3
residues result in disruption of intra- and interhexamer inter-
faces, whereas mutations in the core of the hexamer are less
likely to disrupt virion assembly (21).
Elite Controllers of HIV Preferentially Target Multiple Sites in Sector 3.
Our results suggest that applying CTL pressure at multiple sites
contained in sector 3 is likely to facilitate durable control of viral
load to low levels; multiple mutations that allow immune escape
are more likely to hurt viral fitness significantly, because multidi-
mensional escape pathways are restricted. Can HLA (human
MHCclass I)molecules presentpeptides containing multiple sites
within the identified vulnerable region? Because elite controllers
achieve viral control (16), we first analyzed the locations of sites
contained in peptides targeted in individuals with HLA alleles
associated with spontaneous HIV control (HLA-A*25, B*57,
B*27, B*14, Cw*08) (22). Remarkably, the sector with the largest
controllers is sector 3 (P = 3·10−7; Fig. 3A and SI Appendix, Table
S3). CTL pressure imposed by individuals with these HLA alleles,
however, is not the driver of this collectively coevolving group of
sites for reasons that include:
i) There are clear structural explanations (Fig. 2) for the col-
lective correlations that characterize sector 3 sites, and
structural features are independent of immune pressure.
ii) Negative correlations cannot be induced by immune pres-
sure, because mutations induced in persons with the same
HLA would be more likely to arise than if they occurred
iii) As noted, a quasi-sector describes weak correlations associ-
ated with HLA-associated mutations (SI Appendix, Fig. S5),
and these sites are not a part of sector 3.
Thus, the sector we identify to be most vulnerable by analyses
of viral sequences and supramolecular structures is the most
targeted by controllers.
Examining targeted epitopes lends further insights. For ex-
ample, eight amino acids in the dominant epitopes targeted by
B57-restricted CTLs are in sector 3; this multiplicity is further
augmented by the enhanced cross-reactivity of these CTLs (23).
B14+individuals target three amino acids in this sector. Popu-
lation studies indicate that HLA-B57+/B14+persons are partic-
are shown as purple spheres. (B) Sector 3 sites represented on the structure of the p24 hexamer (PDB code 3GV2) and the structure of the interface between
hexamers (PDB code 2KOD). Sites at interfaces between two p24 proteins belonging to two adjacent hexamers are shown in green, and sites at interfaces
between two p24 molecules within a hexamer are shown in red. The few remaining sites in sector 3 that are not part of these interfaces are shown in purple.
(C) Sector 1 sites are shown in cyan on the structure of the p24 hexamer.
Protein structures reveal the origin of collective correlations. (A) Sector 3 sites represented on the structure of the p24 monomer (PDB ID code 3GV2)
Dahirel et al. PNAS
| July 12, 2011
| vol. 108
| no. 28
ularly efficient controllers of HIV infection. Our results suggest
that this is because viral strains that can escape the multiple
points of immune pressure in this sector are likely to have very
low fitness. Thus, point mutants with lower viral fitness that only
partially escape immune pressure or strains with multiple muta-
tions that are not negatively correlated are likely to be the only
options for the virus. This is consistent with the observed lower
viral load and lower average viral fitness in controllers (16). Note
also that CTLs restricted by HLA molecules associated with
progression do not target sector 3 sites significantly (P = 0.37;
SI Appendix, Table S3).
Virus Sequences from Patients Show That Multiple Mutations in
Sector 3 Are Unlikely. Roughly 10% of sites in sectors 1, 4, and 5
are targeted by controllers, and sector 2 is not targeted (Fig. 3A).
However, because sectors 4 and 5 are composed of only 10 and
14 sites, respectively, only 1 site (and epitope) is targeted by
controllers in these sectors. Because sector 1 comprises 79 sites,
many sites in sector 1 are targeted by controllers. If two regions
of the proteome are equally targeted, all things being equal, the
number and types of mutations observed in these regions should
be similar. Sectors 1 and 3, both in Gag, exhibit similar levels of
single site conservation, but we find that sector 3 is more mul-
tidimensionally constrained. Thus, we predict that even though
sector 3 is targeted more than sector 1, fewer HIV strains with
multiple mutations in sector 3 sites would be viable compared
with those with similar mutations in sector 1.
To test this prediction, we sequenced viruses obtained from
elite controllers because they target both sectors 1 and 3. We
obtained 72 sequences from plasma and 103 including both
plasma and peripheral blood mononuclear cells (PBMCs). The
qualitative results obtained from analyzing either set of se-
quences are the same (Fig. 3B and SI Appendix 18). As per our
predictions, the frequency of single mutations in sectors 1 and 3
is similar but multiple mutations are significantly rarer in sites
that comprise sector 3. Our results suggest that this is because
viral strains that can escape multiple points of immune pressure
in sector 3 are more likely to be associated with negative cor-
relations, and hence have particularly low replicative fitness be-
cause of defective capsid assembly. Thus, they are unlikely to be
viable, and we do not observe them. These results show that the
calculated differences between collective correlations among sites
that comprise sector 3 and other sectors (Fig. 1) result in quali-
tative differences in viral mutations observed in humans that
target sector 3 sites. Note also that almost all the few viral strains
that we observe with multiple mutations in sector 3 are associated
with positive correlations (SI Appendix, Fig. S13). Although only
positive correlations can correspond to compensatory mutations,
positive correlations can also correspond to multiple mutants
that are only moderately less fit compared with single mutants,
unlike negative correlations, which are only associated with
multiple mutants that are very deleterious.
Our results strongly support our hypothesis that applying CTL
pressure at multiple points in a multidimensionally constrained
group of sites can trap the virus, because we find that controllers
target multiple sites in sector 3, but strains with multiple muta-
tions in sector 3 are not viable. Thus, viable strains are still likely
to be subject to immune pressure.
Immunogens That May Induce CTL Responses in a Population That
Hurt HIV. Can such potent responses be induced by vaccination in
persons who are not blessed with HLA alleles that mount the
correct dominant responses? Individuals with other HLA alleles
present subdominant peptides containing sites in the vulnerable
regions we have identified (using the Epitope Tables of the Los
Alamos HIV Molecular Immunology Database, http://www.hiv.
lanl.gov/content/immunology/tables/tables.html). This suggests the
following strategy for design of immunogens that could induce
memory CTL responses targeting the most vulnerable regions of
HIV in a population: Select protein segments that simultaneously
maximize sites in sector 3 (and some sites in sector 1), which are
characterized by negative correlations, and those contained in
subdominant and dominantepitopes presented by HLA molecules
that span a broad section of a population. Such immunogens
should elicit memory CTLs that onlytarget the multidimensionally
conserved regions of HIV, because epitopes that are dominantly
targeted ineffectually are excluded. During natural infection, these
memory CTLs are expected to mount robust responses that hit
HIV where it hurts early, before naive CTLs mount ineffective
dominant responses. Mounting the right responses early is im-
portant for HIV infections (2), because there is a short window of
time before the immune system is compromised.
To illustrate this strategy, we considered a target population,
white Americans with the 25 most frequent haplotypes (∼39% of
this population) (24). We create all groups of 10 known epitopes
(as an example) presented by HLA-A/B molecules that comprise
these haplotypes and select the top 10 groups according to the
two criteria noted above (SI Appendix 12). One top-scoring im-
munogen (p24 residues 160–188 and 240–277) contains at least 1
or 2 targeted epitopes in 97% and 56% of the target population,
respectively (a lower bound, because many epitopes are un-
known). Previous empirical data show that controllers target p24
sites 240–272 (25). Sites 160–188 contain only 1 epitope tar-
elite controllers in each sector and comparison of mutation patterns in
sectors 1 and 3 in viruses derived from elite controllers. (A) Epitope is defined
as dominant if it is the most targeted (in Gag) by individuals with the cor-
responding HLA allele. Concentration is the number of epitope sites in the
sector divided by the total number of residues in the sector. The P value of
association with each sector is in SI Appendix 10. Only sector 3 is significantly
enriched with sites from the dominant epitopes of HLA molecules associated
with control. Sectors 4 and 5 are not targeted at multiple points because
they contain only 10 and 14 sites, respectively. Sectors 1 and 3 are targeted
at multiple points. (B) Comparison of the number of unique viral strains
obtained from a cohort of elite controllers that contains different numbers
of mutations in sectors 1 and 3.
Concentration of sites from the dominant epitopes presented by
| www.pnas.org/cgi/doi/10.1073/pnas.1105315108Dahirel et al.
geted by known controllers but represent an equally vulnerable
region that provides broad coverage.
Immunogens designed in this way to include only the multi-
dimensionally constrained regions of the HIV proteome, which
also contain epitopes presented by diverse HLAs in a population,
are candidates for peptide vaccines using synthetic vectors (26),
or by linking each segment together, they can be analogs of
mosaic immunogens (27) delivered by traditional vectors. Such
an immunogen would target the most vulnerable regions of HIV
rather than the whole proteome. Targeting the latter is more
likely to elicit responses from which HIV can escape via muta-
tions (28), while hindering a focused response directed at regions
of immunological vulnerability we have identified. Our goal of
identifying such regions, and the methods we have used toward
this end, are different from previous efforts to study coevolution
of HIV proteins (e.g., 14, 29–31). Further analysis following the
logic described by us should reveal additional regions of HIV
proteins that are multidimensionally constrained, thereby en-
hancing the list of regions that are candidates for inclusion in
a potent vaccine. A practical vaccine may also require inclusion
of CD4 epitopes and flanking residues needed for antigen pro-
cessing which do not contain protein segments that are likely to
elicit ineffectual memory CTL responses. In vitro experiments
and studies with animal models are required to develop this
new concept, which might also be applied to design efficacious
immunogens against other viruses. Our results also suggest the
design of new small-molecule inhibitors of HIV replication.
HIV Sequences and Similarity Analysis. Multiple sequence alignments of nu-
cleotide sequences of HIV-1 Gag, RT, and Nef were downloaded from the Los
Alamos HIV Sequence Database (http://www.hiv.lanl.gov/) (clade B, one se-
quence per patient). We checked the phylogenetic homogeneity of each set
by performing a principal component analysis of the similarity matrix of
sequences (SI Appendix 5).
Identification of HIV Sectors. The correlation matrix for a given set of
main text and SI Appendix 1–4). The definition of sectors follows the strategy
proposed by Halabi et al. (12), grouping positions with a particular spectral
signature into a sector (SI Appendix 6).
Targeting by Elite Controllers. We defined a set of epitopes dominantly tar-
geted by HLA alleles associated with control (22), using a published dataset of
the frequency of recognition of epitopes in an HLA-specific population (32).
Sequencing Viruses from Elite Controllers. HIV-1 Gag was amplified using
nested RT-PCR from previously frozen PBMCs or plasma from elite con-
trollers as previously described (33). PCR products were purified, and cycle-
sequencing reactions were performed using 60 HIV-1–specific sequencing
primers. Population sequences were obtained using an ABI 3730 PRISM
(Applied Biosystems) automated sequencer. Gag mutations in the sequences
derived from elite controllers were defined with reference to the clade B
ACKNOWLEDGMENTS. We thank D. Barouch and H. Eisen for valuable
discussions. Financial support was provided by the Ragon Institute, a Na-
tional Institutes of Health Director’s Pioneer Award (to A.K.C.), National
Institutes of Health Grants RO130914 (to B.D.W.) and PO1 AI074415 (to M.
Altfeld and T.M.A.), The Howard Hughes Medical Institute (B.D.W.) and
the Mark and Lisa Schwartz Foundation (B.D.W.) This project has been
funded in whole or in part with federal funds from the National Cancer
Institute, National Institutes of Health, under Contract HHSN261200800001E.
1. Walker BD, Burton DR (2008) Toward an AIDS vaccine. Science 320:760–764.
2. McMichael AJ, Borrow P, Tomaras GD, Goonetilleke N, Haynes BF (2010) The immune
response during acute HIV-1 infection: Clues for vaccine development. Nat Rev
3. Amanna IJ, Slifka MK (2011) Contributions of humoral and cellular immunity to
vaccine-induced protection in humans. Virology 411:206–215.
4. Barouch DH, Korber B (2010) HIV-1 vaccine development after STEP. Annu Rev Med
5. Hansen SG, et al. (2009) Effector memory T cell responses are associated with pro-
tection of rhesus monkeys from mucosal simian immunodeficiency virus challenge.
Nat Med 15:293–299.
6. Hartl DL, Clark AG (2007) Principles of Population Genetics (Sinauer, Sunderland, MA),
7. Altfeld M, Allen TM (2006) Hitting HIV where it hurts: An alternative approach to HIV
vaccine design. Trends Immunol 27:504–510.
8. Goulder PJ, Watkins DI (2004) HIV and SIV CTL escape: Implications for vaccine design.
Nat Rev Immunol 4:630–640.
9. Wigner EP (1967) Random matrices in physics. SIAM Rev 9:1–23.
10. Plerou V, et al. (2002) Random matrix approach to cross correlations in financial data.
Phys Rev E Stat Nonlin Soft Matter Phys 65:066126.
11. Laloux L, Cizeau P, Bouchaud JP, Potters M (1999) Noise dressing of financial cor-
relation matrices. Phys Rev Lett 83:1467–1470.
12. Halabi N, Rivoire O, Leibler S, Ranganathan R (2009) Protein sectors: Evolutionary
units of three-dimensional structure. Cell 138:774–786.
13. Bhattacharya T, et al. (2007) Founder effects in the assessment of HIV polymorphisms
and HLA allele associations. Science 315:1583–1586.
14. Brumme ZL, et al. (2009) HLA-associated immune escape pathways in HIV-1 subtype B
Gag, Pol and Nef proteins. PLoS ONE 4:e6687.
15. Kiepiela P, et al. (2007) CD8+ T-cell responses to different HIV proteins have
discordant associations with viral load. Nat Med 13:46–53.
16. Miura T, et al. (2009) HLA-associated viral mutations are common in human
immunodeficiency virus type 1 elite controllers. J Virol 83:3407–3412.
17. Hoffman NG, Schiffer CA, Swanstrom R (2003) Covariation of amino acid positions in
HIV-1 protease. Virology 314:536–548.
18. Ganser-Pornillos BK, Yeager M, Sundquist WI (2008) The structural biology of HIV
assembly. Curr Opin Struct Biol 18:203–217.
19. Pornillos O, et al. (2009) X-ray structures of the hexameric building block of the HIV
capsid. Cell 137:1282–1292.
20. Troyer RM, et al. (2009) Variable fitness impact of HIV-1 escape mutations to cytotoxic
T lymphocyte (CTL) response. PLoS Pathog 5:e1000365.
21. Adamson CS, Jones IM (2004) The molecular basis of HIV capsid assembly—Five years
of progress. Rev Med Virol 14:107–121.
22. Pereyra F, et al.; International HIV Controllers Study (2010) The major genetic
determinants of HIV-1 control affect HLA class I peptide presentation. Science 330:
23. Kosmrlj A, et al. (2010) Effects of thymic selection of the T-cell repertoire on HLA class
I-associated control of HIV infection. Nature 465:350–354.
24. Maiers M, Gragert L, Klitz W (2007) High-resolution HLA alleles and haplotypes in the
United States population. Hum Immunol 68:779–788.
25. Streeck H, et al. (2007) Recognition of a defined region within p24 gag by CD8+ T cells
during primary human immunodeficiency virus type 1 infection in individuals
expressing protective HLA class I alleles. J Virol 81:7725–7731.
26. Melief CJ, van der Burg SH (2008) Immunotherapy of established (pre)malignant
disease by synthetic long peptide vaccines. Nat Rev Cancer 8:351–360.
27. Barouch DH, et al. (2010) Mosaic HIV-1 vaccines expand the breadth and depth of
cellular immune responses in rhesus monkeys. Nat Med 16:319–323.
28. Altfeld M, Goulder PJ (2011) The STEP study provides a hint that vaccine induction of
the right CD8+ T cell responses can facilitate immune control of HIV. J Infect Dis 203:
29. Liu Y, Eyal E, Bahar I (2008) Analysis of correlated mutations in HIV-1 protease using
spectral clustering. Bioinformatics 24:1243–1250.
30. Fares MA, Travers SAA (2006) A novel method for detecting intramolecular
coevolution: adding a further dimension to selective constraints analyses. Genetics
31. Bickel PJ, et al. (1996) Covariability of V3 loop amino acids. AIDS Research and Human
32. Streeck H, et al. (2009) Human immunodeficiency virus type 1-specific CD8+ T-cell
responses during primary infection are major determinants of the viral set point and
loss of CD4+ T cells. J Virol 83:7641–7648.
33. Miura T, et al. (2008) Genetic characterization of human immunodeficiency virus type
1 in elite controllers: Lack of gross genetic defects or common amino acid changes.
J Virol 82:8422–8430.
Dahirel et al. PNAS
| July 12, 2011
| vol. 108
| no. 28