Two primary mechanisms have been suggested to explain HFCs [14,15]. First, relatively homozygous individuals may be more susceptible to infection because they are inbred. Here, average heterozygosity at the panel of markers being genotyped estimates genome-wide heterozygosity, which in turn estimates the inbreeding coefficient, F. However, several theoretical treatments have come to the conclusion that such a mechanism is unlikely to operate in most real populations [16-18]. The problem is that random mating generates extremely few individuals with sufficiently high F for their heterozygosity to stand out when measured at tens or even hundreds of markers, unless the population is very small or highly polygynous. Humans may offer a further exception in cultures where cousin marriages are actively encouraged [19], potentially increasing the rate of heritable diseases [20,21].

The second mechanism that may generate HFCs involves chance linkage between one or more of the markers and a gene(s) experiencing balancing selection. Balancing selection has often been thought to be rather rare, particularly in humans [22] where the classical example is sickle cell anemia [23] remains one of very few examples. Moreover, while some argue that polymorphism at immune function genes is maintained by overdominant balancing selection [24], there is evidence that this is unlikely to be effective at maintaining more than two alleles [25-27]. Regardless of theory, a number of recent HFC studies report convincing associations between heterozygosity at one particular locus and the measured trait [13,28-31].

For the study of HFCs a number of measures of heterozygosity have been proposed that offer potential benefits over straight heterozygosity, weighting scores variously by allele size (mean d²)[3], allele frequencies (internal relatedness, IR) [6] and the variability of loci scored (HL) [34]. However, in automated high throughput studies, heterozygosity assessment can sometimes be problematic, particularly where time for scrutiny of every trace is limited. Thus, null (non-amplifying) alleles, allele drop-out and, at some loci, high levels of stutter-bands can all contribute to a tendency for a minority of loci to carry misleading genotypes where heterozygotes are called as homozygotes or vice versa. Issues have also been identified with allele binning, in some cases causing single alleles to be split between two length classes [35]. In an attempt to circumvent these problems we spent most empirical effort ensuring that heterozygotes and homozygotes were accurately scored and introduce a variant of the measure Standardised Heterozygosity (SH) [2], designed to be highly conservative. SH controls for missing data by expressing heterozygosity as the ratio of the observed heterozygosity in an individual relative to the expected value at the markers genotyped, assuming Hardy-Weinberg equilibrium. Our measure, Standardised Observed Homozygosity (SOH), follows the same principle but instead of calculating the expected homozygosity from the allele frequencies, we used the observed homozygosity at each locus. In this way, SOH measures the extent to which any given individual is more or less homozygous relative to the level expected if all genotypes were randomized among individuals, negating the requirement for accurate allele frequency estimates and reducing the impact of allele drop out, null alleles and other possible artefacts.

We first asked whether SOH varied significantly among disease categories by conducting a one-way ANOVA. Raw SOH values exhibit a slightly skewed distribution, but this is removed by a simple log transformation (Shapiro-Wilk normality test, W = 0.9941, p = 0.322). Following transformation, SOH revealed highly significant variation among disease classes (F _[5,281] = 6.75, P = 5.89 × 10^-6) (Figure ​(Figure1).1). However, when the control class was excluded, the ANOVA was no longer significant (F _[4,144] = 0.785, P = 0.54), indicating that the main effect is driven by a difference in heterozygosity between cases and controls rather than between disease classes. The direction of the deviation is toward greater homozygosity in cases compared with the controls.

We next asked whether there was evidence of local effects due to chance linkage between one or more markers and a gene(s) experiencing balancing selection. To test this proposition we calculated age-adjusted odds ratios of mortality at each locus in turn (Figure ​(Figure2).2). Most markers show either a non-significant or borderline (at alpha = 0.05) association between homozygosity and risk of mortality. However, nine markers reveal a strong associations with experiment wide significance using full Bonferroni correction (p < 0.00037, see Table ​Table1).1). This is a highly conservative threshold since where multiple markers are expected, the less stringent false discovery rate approach can be justified [36]. Although the spacing between markers is sufficient to ensure they behave as if unlinked, it is possible that multiple markers contribute to the same risk through linkage to related genes. Consequently, we then constructed a multivariable logistic regression model with mortality as the response and age, sex, locality, SOH and homozygosity at each of the nine largest-effect markers as explanatory variables. Sequential removal of terms that did not contribute significantly (likelihood ratio test, p < 0.05) yielded a final model containing age, location and five markers (D12S310, D13S158, D14S275, D16S3103, D16S423). SOH is dropped as a marginally significant term (LHR, p~0.07) whether fitted as a continuous variable or as a factor with five levels. By implication, it seems that genome-wide effects (inbreeding depression) are minimal or absent.

Beginning with the final model derived above, we explored further possible interactions between markers, and also between each marker and age. Among all possible pair-wise combinations of markers, two revealed significant interactions, both of which were retained in the model regardless of the order in which they were added (Table ​(Table2,2, last two columns). No significant interactions with age were detected. Our data contain approximately equal numbers of individuals who died from Gram positive (n = 63), gram negative (n = 79) or both (n = 6) infections and this allowed us to ask whether our markers identify genes that impact differently on diseases caused by bacteria of different classes. We therefore repeated the logistic regression approach above on each bacterial class separately, including dual infections in both analyses. Given the smaller datasets, criteria for initial inclusion in the model were relaxed to an initial OR significant at p < 0.005. The final models are summarized in Table ​Table22 and reveal surprising complexity, with susceptibility to gram positive and gram negative infections associated mostly with non-overlapping genomic locations. Only marker D12S310 is significant in all three models. Marker D9S164 reveals an interaction with age, infants being more likely to die if homozygous (OR = 1.65) and older children less (OR = 0.18). Interactions between marker pairs in the whole dataset suggest that homozygosity at both markers together confers no greater risk than homozygosity at either one alone. However, in the gram negative model the interaction of homozygosity at two markers, D7S486 and D16S423, indicates a significantly synergistic risk of mortality (odds ratio 40.7) where homozygosity at either of the markers alone confers no risk.

Finally, to assess the magnitude of homozygosity effects with respect to the population, we calculated the population attributable risk fraction (PARF) [37] for each marker (Table ​(Table3).3). PARFs indicate the proportional reduction in mortality due to bacterial infection that would result if homozygosity at the locus could be eliminated. However, such direct interpretation in our case is problematic for many reasons, including the fact that the risk is probably driven by specific alleles whose prevalence differs greatly from that of homozygotes in general. None the less, the ORs of the full model indicate that the risks we describe are sizeable, particularly since the markers provide only indirect measures of homozygosity at the genes themselves. Furthermore, given that the population prevalence of homozygosity at the relevant markers is high, the population-wide effects of homozygosity are likely to impact very considerable on the total burden of invasive bacterial disease mortality, with the majority of deaths being genetically determined.

There is currently a debate as to whether the benefits of heterozygosity accrue mainly through genome-wide effects (inbreeding) [38,39] or through individual balanced polymorphisms [13,14]. We found that inbreeding effects are either small or absent in this population. This is perhaps not surprising because, in contrast to some other populations such the Fulani [40] and some Arab communities [19], consanguineous marriages tend to be discouraged, with a preference for marriages between rather than within clans [41]. In contrast, five loci independently contribute significant risk factors, lending strong support to the local effects model. However, it should be remembered that human populations differ greatly in their structure and that, in contrast to most animals populations, some human populations actually favour consanguineous marriages [19,40,42]. In such populations a rather different pattern may well emerge.

To find several balanced polymorphisms in a relatively small study of just 134 markers is surprising, given how few have been identified previously in humans [22]. Two factors may contribute to this discrepancy. First, a large majority of genome scans focus on complex, non-infectious diseases, and these are likely to differ from infectious diseases mechanistically. Most heritable non-infectious diseases involve mutant alleles at one or more loci where function is removed or disrupted, and hence are mostly recessive. In contrast, the efficacy of immune-function genes is widely though to benefit from high diversity, a larger palette of alleles increasing the range of pathogen types that can be recognised, and therefore these loci tend naturally towards heterosis. Second, classical association studies tend to be applied to diseases that are known to run in families [43,44], and hence susceptibility will tend to have an appreciable additive component. As such, patterns where heterozygosity is important will tend to be overlooked because heterozygosity per se tends not to be heritable. Instead there is a strong focus on searching for associations between particular alleles and disease [43,45,46]. It will be interesting to see the extent to which future studies reveal a much higher prevalence of balancing selection, thereby supporting results from many non-human systems.

Our current study is relatively small-scale, with several of the smaller chromosomes being scored for only three or four markers. Consequently, there are large tracts of the genome where further loci could be located with the potential to contribute even further to genetic susceptibility, and implying that the five regions we identify are not the complete set of the loci that could potentially be identified in a larger study. This is surprising because the loci we have uncovered exhibit large individual and combined effect sizes, to the extent that mortality appears highly non-random. Moreover, it should be remembered that the overall risk factor combines both genetic susceptibility and variation in exposure. Unless exposure to pathogens is highly uniform, the impact of genetic factors will be even higher than we report and could rise further if our study has missed further contributory loci.

The effect sizes we report appear much larger than expected. Across the five loci identified as having highest impact, population attributable risk fractions (PARFs) all lie in the range 25–55%. PARFs provide an indication of the proportion of total risk that can be attributed to each genetic factor, given the local prevalence of exposure. Since the calculations assume overlapping effects, these do not sum to one. None the less, our analysis suggests that half or more of the observed deaths would probably not have occurred if the individuals concerned had been heterozygous for these loci, a figure that would surely be even higher if we had been able to genotype SNPs in the genes concerned rather than at linked microsatellites.