These studies provide a useful resource in identifying HLA Class I alleles that mediate a co-operative additive effect in control of HIV-1 in C-clade infected African cohorts. The extended size of this cohort (>2000 individuals) and adaptation of methodology to identify co-operative additive effects has allowed us to build on previous analyses 
and to identify the impact of individual or paired HLA alleles with greater sensitivity. Importantly, however, in spite of this large cohort size, the analysis remains underpowered given the large number of HLA-pairs and the necessity of a multiple testing correction. These results are therefore likely an underestimate of the true extent of HLA co-operativity, and future studies employing more individuals, or a more restricted set of tests, are likely to reveal further instances of HLA co-operativity. Furthermore, our approach of using most HLA data at two-digit resolution was aimed to maximize statistical power to detect Class I influences on disease control. However, a caveat of this approach is that it limits the detection of possible differences occurring at high-resolution (often a micropolymorphism) level 
; this could be addressed in future by use of larger cohorts.
Effects on disease control were not always seen for both CD4 count and VL. Reasons for this likely include imperfect correlation between CD4 count and VL (r2
0.22, p<0.0001 by linear regression; data not shown), and that the linear models are only idealizations.
Our analysis supports previous evidence that even highly beneficial responses, such as that restricted by HLA-B*57, can be improved upon by addition of other T cell responses 
. The mechanism of this phenomenon has not previously been clearly characterised, but we have here demonstrated that – at least in part - the effect may be explained by the targeting of non-overlapping CD8+ T cell epitopes across the HIV proteome.
The correlation between our ‘sharing score’ (reflecting breadth of epitopes targeted by a pair of alleles) and the probability of a co-operative additive effect mediated by these alleles was only weak (R
−0.08). Any computational method to assess breadth of epitope targeting is a challenge, especially given the density of overlapping CD8+ T cell epitopes in certain regions of the HIV proteome, the bias towards restricting highly targeted epitopes restricted by prevalent Class I alleles, and the complexity of immunodominance patterns. In addition, any single pair of alleles will also be impacted by the other four HLA Class I molecules expressed by a given individual, and the overall disease outcome will be influenced by many factors in addition to HLA genotype. Furthermore, there is no obvious effect size obtainable for the co-operative additive test, and even if there were it would be possible to have large effects for pairs which were not statistically significant. For these two reasons, we chose to measure correlation with the p-value from our test.
These difficulties notwithstanding, these data nevertheless do highlight that two alleles which present different epitopes can each confer a separate benefit (or hazard) to the individual; thus having both of them is better (or worse) than having just one of them and a co-operative additive effect is at play. However, if two alleles present many of the same epitopes (as exemplified by HLA-B*57 and -B*58:01, or HLA-B*42 and -B*81), they are less likely to act together co-operatively – having one of them may be little different from having both. This effect is also underscored by the phenomenon of heterozygote advantage 
, which may be mediated by increased breadth of epitopes presented by HLA class I heterozygotes compared to homozygotes.
As HLA-peptide complexes are ligands not only for T-cell receptors on CD8+ T cells, but also for KIR receptors on NK cells 
, another potential reason for the favourable (or hazardous) interaction of some pairs of HLA alleles is the combined effect of a CD8+ T cell response and an NK-cell response. Homozygosity for KIR ligands may also explain poor disease outcomes in subjects with certain HLA Class I combinations, although many of our pairs involved at least one allele that is not a known KIR ligand.
Characterising interplay between HLA alleles is made difficult by the presence of linkage disequilibrium between alleles. However, our test statistic will not be significant for two alleles simply because they are in linkage disequilibrium, but rather the test can find two alleles to have a co-operative additive effect despite
their being in (incomplete) linkage disequilibrium, albeit with reduced power owing to fewer observations of the alleles acting one without the other. That is, if one observes each allele only in the context of the other, or never together, it is impossible to determine whether nor not they have a co-operative additive effect (hence these pairs removed from analysis; see Methods
section). However, because one needs to observe enough co-occurrences of the alleles, having alleles in incomplete linkage disequilibrium increases the power to detect co-operative additive effects.
In summary, these data highlight the potentially potent interactions between HLA class I alleles to mediate HIV-1 disease control. Even CD8+ T cell responses which are independently associated with strong viraemic suppression and sustained immunological control can be improved upon by the co-expression of certain other favourable HLA class I molecules. This finding underscores the potential benefit of harnessing co-operative effects of multiple CD8+ T cell responses in the development of CD8+ T cell vaccines.