Is HDL-c plasma concentration a possible marker of HIV replication? A cross-sectional analysis in untreated HIV-infected individuals accessing HIV care in Italy

Aims HIV infection is associated with dyslipidemia and an increased risk for cardiovascular diseases. HIV Nef protein downregulates the generation of nascent HDL. The interplay between HIV-RNA, HDL-c level and CD4/CD8 ratio in naïve HIV patients remains to be elucidated. Methods We included untreated persons living with HIV (PLWH) of the ICONA Foundation Study cohort if they also had ≥2 viral load (VL) measurements prior to ART initiation. We performed unadjusted correlation and linear regression analyses evaluating the effect of VLset on HDL-C. Vlset and CD4/CD8 ratio were fit in the log10 scale, while HDL-c, was fitted in the untransformed raw scale. Results We included 3,980 untreated PLWH. Fifty-eighty (1.5%) were aviremic. We observed a negative correlation between HDL-c and VLset (Pearson R2=0.03), from fitting an unadjusted linear regression model -8.5 mg/dl (95% CI: -15,9 --0,84 p<0.03). There was a dose-response relationship between HDL-c levels and VLset, however, this association was somewhat attenuated after further controlling for gender. Despite a positive correlation between HDL-c and CD4/CD8 ratio, the HDL-c plasma concentration does not satisfy the criteria for a strong surrogate marker. Conclusions Our data show that HDL-c plasma concentration is significantly lower per higher level of VLSet although this was in part explained by gender. Further analyses should be promoted to better understand the molecular mechanisms that underline the relationship between HIV replication, HDL-c formation, and diseases progression.

Lipidomics techniques have also allowed the characterization of the lipidome of enveloped viruses. By this 101 way, HIV lipid envelope has been observed to be different from the producer cell plasma membrane, 102 suggesting that viruses bud from specialized membrane subdomains, which are enriched in particular lipids 103 [24]. 104 The evidence summarised above, supports the notion that plasmatic HDL-c is a should be biochemical marker 105 which is likely to be related to HIV viral budding and inflammation. With this analysis, we aimed to 106 corroborate, in the setting of real-life untreated HIV-infection, the association between VLset and lipids (such 107 as total cholesterol and HDL-c plasma concentrations), and whether VLset mediated HDL-c changes might 108 also correlate with immunological parameters of HIV progressions, such as CD4/CD8 ratio. In this retrospective cross-sectional study, we included untreated HIV-infected people enrolled in the ICONA 113 Foundation cohort. The main aim was to evaluate the association between HDL-c plasma concentration and 114 VL set-point in absence of ART; a secondary objective was to evaluate the association between HDL-c levels 115 and markers of HIV disease progression like CD4/CD8 ratio. We included people for whom ≥2 viral load (VL) 116 measurements prior to ART initiation were available. The viral set point (VLset) was defined as the mean of the first two VL and the date of the 2 nd value chosen as the index date for this cross-sectional analysis. 118 Participants with an estimated VLset <50 copies/mL were labelled as 'aviremic' and the remaining group as 119 'viremic'. People who had started statin therapy prior to the index data and those without a value of HDL 120 over 3 months of the index data were excluded. All laboratory markers test results were included in the 121 correlation analyses if measured within 6 months of the index Vlset date. In order to control for potential confounding factors, a multivariable analysis was conducted for total 138 cholesterol and HDL-c for which an univariable difference between groups was detected. In particular, the 139 association between VLSet (main exposure) and HDL-c (primary outcome) and total cholesterol (secondary 140 outcome) was evaluated by fitting a linear regression model after controlling for a minimal set confounders 141 chosen a priori including gender, age, CD4/CD8 ratio, HCV status (detection of HCVAb), and AIDS diagnosis.

142
Total cholesterol was essentially chosen as a negative control. This list of measured potential confounders was put together using both axiomatic knowledge and literature review. In order to further assess the 144 robustness of the results against potential unmeasured confounding bias, the e-value was calculated and 145 compared to the magnitude of the mean difference seen for predictors showing the strongest association 146 with the outcome (25).

147
HDL-c and total cholesterol, which both showed a symmetric distribution, were fitted in the untransformed 148 raw scale. VLSet instead was fitted in three ways: i) comparing people with ≤50 copies/mL (aviremic) vs. >50 149 copies/mL (viremic); ii) using the log 10 scale and iii) after splitting the study population in groups using pre-150 specified HIV-RNA clinical cut-offs to evaluate a potential dose-response effect.

152
In addition, a refined model has been hypothesised for a third outcome: the CD4/CD8 ratio. In this model, on 153 the basis of the results of the main analysis, BMI was the only confounder of the association between VLSet 154 and CD4/CD8 ratio, while HDL-C was a mediator, i.e. some of the total effect of VLSet on CD4/CD8 is assumed 155 to be explained by a variation in HDL-C. This was visually described using a direct acyclic graph (DAG, Figure   156 1). A mediation analysis was formally performed using the 'medeff' command in Stata 15. All other results 157 were obtained from using SAS version 9.4 (Carey, USA).

160
Study population 161 The clinical and demographic characteristics of HIV positive patients enrolled in the study are shown in Table   162 1. We included 3,980 HIV ART-naive individuals, 58 patients (1.5%) spontaneously aviremic and 3,922 (98.5%) 163 viremic patients, respectively. As shown in Table 1 Table 2). In contrast, there was no evidence for 191 a difference in total cholesterol between the viremic and aviremic group from the unadjusted linear 192 regression with total cholesterol as outcome: -14.5 mg/dl (95% CI:-38.6; +9.36), p=0.23 (Table 4).

194
Role of potential confounding factors 195 The relationship between VLset and HDL-c and TC was re-evaluated after controlling for potential 196 confounders using linear regression adjustment. When VLSet was fitted as a binary exposure (aviremic vs. 197 viremic) it was associated with HDL-C levels independently of age, AIDS diagnosis and HCVAb status.

198
However, after controlling for gender this effect was somewhat attenuated (Table 2). This is because females 199 are known to have a lower VLSet [26] and also a higher HDL-C. Interestingly, confounding was less strong in 200 the analysis in which VLSet was fitted as continuous in the log 10 scale which has greater statistical power.

201
Also, difference could still be seen when comparing aviremic patients with those with very high levels of HIV-202 RNA (>10,000 copies/mL), even after controlling for gender (Table 3). Table 3   Similarly, to move the confidence interval to include the null, an unmeasured confounder that was associated 209 with the outcome and the exposure each by a difference of at least 8.1 logs could do so, but weaker 210 confounding could not. To put this in prospective, the difference associated with the measured factors 211 showing the strongest association was 9.3 logs for gender.

213
In contrast, the model with TC as outcome showed an association with VLSet only when the latter was fitted 214 as continuous in the log 10 scale ( Table 4). The analysis show that other factors such as age, AIDS and HCVAb

Mediation analysis
We further evaluated the total direct effect of VLSet on CD4/CD8 ratio by decomposing the effect in the 220 direct effect of VLSet on CD4/CD8 ratio and the indirect effect through the causal pathway of HDL-C ( Figure   221 1). This analysis indicated that indeed some of the total effect of VLSet on CD4/CD8 is significantly mediated 222 by a variation in HDL-c induced by HIV-RNA. Although significant, this indirect effect is estimated to be only 223 a small percentage of the total effect (Table 5). There was also evidence that the indirect effect was larger, 224 although still small in absolute terms, in people with lower levels of HDL-c which was estimated after formally 225 testing for interaction (data not shown). Especially when dealing with observational data it is important to question whether the findings might be 246 due to bias and these other results, which act as negative controls, are somewhat in support of the evidence. 247 Overall, our results appear to confirm the presence of a link between HIV replication and lipid metabolism.

248
In particular, we speculate that the inverse correlation seen between HIV viremia and HDL-c in our "in vivo" 249 study, is a result of the fact Nef HIV protein was able, through active viral replication, to reduce HDL-c 250 production by impairing ABCA-1, generating cholesterol accumulation within macrophages, promoting their   In addition, our analysis of the possible causal effect of VLSet on HDL-C is based on the assumption of no 283 unmeasured confounding and correct specification of our model (e.g. one of the underlying assumption of 284 our model is that BMI is a predictor of outcome but not a cause of variation in VLSet, etc. see Figure 1).

285
However unmeasured confounding can never be ruled out in real-world data. For example, HCV-RNA at ART 286 initiation which is not available in the database for the majority of our participants is a potential key 287 unmeasured confounding factor. Nevertheless, many important measured confounders have been 288 accounted for and our sensitivity analysis (through calculation of the e-value) shows that results are fairly 289 robust to potential unmeasured confounding bias. Similar considerations apply also to the second part of our 290 analysis, aiming to estimate the indirect and direct effect of VLSet on CD4/CD8 ratio and even more so as one 291 key assumption in mediation analysis is that there is no mediator-outcome unmeasured confounding.       Tables   471  472  Table 1    Mediation analysis on the assumptions described in the DAG (Figure 1).