Our results showed that CTL responses against Gag (as opposed to responses against other proteins) were fundamental to the control of viral replication in a cohort of HIV-1 subtype C infected individuals in KwaZulu-Natal, South Africa, and also suggested that control of viremia was associated with the reduced variability of the Gag protein. Analysis of co-variation in HIV-1 Gag subtype C showed intricate patterns of dependent mutations associated with CTL-driven polymorphisms.
Our co-variation results were derived with a method that can overcome traditional challenges for co-variation studies. In particular, it sets apart co-variation due to selection pressure from co-variation due to shared ancestry by modeling multiple interactions simultaneously. Moreover, it is conditioned on the HLA types observed in each subject, allowing us to identify pathways of particular interest to CTL escape. Nonetheless, there are limitations to the identification of co-variation. For example, associations between two AA that are in different epitopes restricted by the same HLA allele may be a consequence of immunodominance patterns, i.e., the preferential or time-ordered targeting of epitopes. A second issue is the noise in the data; yet, we note that the fuzziness of the network is inherent to the fast pace of HIV evolution and the cross-sectional nature of our analysis. It may also point at biological differences between permanent and transient interactions.
The circle map of dependencies illustrated the topological clustering of AA interactions in Gag, underlining the preponderance of ‘neighboring’ interactions. The distance between co-varying residues suggests two types of interactions: i) structural interactions with combined replacements between spatially proximal AA to maintain the local structure stability, and ii) functional interactions between pairs of distal AA to connect distant functional determinants.
It has been shown in protein-protein interaction studies that essential genes have on average more partners, i.e., higher hubness 
. By analogy, one might expect that hubs in Gag might have a greater impact on viral fitness, since a disturbance at a highly connected residue would probably be more detrimental to the virus than a change at a lowly connected site. Accordingly, we found that HLA-associated sites had conspicuously high numbers of associations and some CTL escape mutations at those HLA-restricted sites are known to have a fitness cost 
. By co-analyzing HLA-imprinting and AA-to-AA interactions in Gag, we described examples of evolutionary constraints that are forced upon some residues and elucidated some mutational pathways associated with HLA-restricted polymorphisms, for example, linked to the B*57-restricted epitopes KF11 and IW10.
A key outcome of our analysis is our ability to stratify the dispensability of particular sites, based on their dependent mutational pathways, and thereby to better understand the relationship between AA mutability and viral fitness. Hence, our CTL-driven mutational dynamics maps reflect the lethality profile of the protein, and we have found some relationships between the presence of rare residues at co-varying sites and lower viral loads for the infected individuals. This pertains to recent vaccine strategies that have proposed to focus on conserved elements of HIV-1 
. The sites that remain conserved despite the influence of selective forces are of specific interest because they correspond to the ‘Achille's heel’ of the virus and are attractive anti-viral or vaccine targets. We previously proposed a CTL-based ‘Conserved Element’ (CE)-vaccine that would be composed of conserved elements of the HIV-1 proteome 
, based on the rationale that ‘an efficacious vaccine must elicit responses toward HIV-1 segments that cannot mutate without severely compromising viral viability, and must not elicit responses against variable, immunodominant decoys’. Here, we present a strategy to expand ‘conserved elements’ while still blocking CTL escape paths by taking into account CTL-driven mutational pathways. Combining results from replicative fitness competition assays and from analyses of the effect of mutations on viremia to co-variation data provides ways to extend CE beyond their minimal length (which is a consequence of their rigid degree of conservation). For example, the vaccine insert could include one or both residues of a co-varying interaction in order i) to link residues that are known to incur a high fitness cost when mutated or ii) to avoid including pairs of residues that can compensate the fitness cost associated with a mutation. In addition, certain HLA-associated hubs under constraints with multiple connections underline sites that may represent vulnerable spots for HIV and that can be incorporated in a vaccine construct.