HIV transmission probability parameters

[darcyrao]

SUMMARY: We will revise the VL states, transitions, and transmission risks as described in the final comment on this page and summarized in this Excel sheet.

First comment October 16 I’ve been looking into the transmission probability parameters based on what Cara commented on yesterday and to decide how to model the impact of ART on viral load. Here are some observations and suggestions.

1. I believe the multipliers for transmission risk from Roger’s paper and carried forward to Nick’s and other papers are incorrect.

   • The baseline transmission probability of 0.0006 is cited as coming from Powers et al. 2008 and Boily et al. 2009. Powers reports that estimates are heterogeneous but estimates a per-contact risk of 0.00084 for penile-vaginal intercourse (95% CI 0.0005-0.0017). Boily estimates 0.0004 (95% CI 0.0001-0.0014) for female-to-male and 0.0008 for male-to-female in developed countries. In developing regions the risks were higher at 0.0038 (95% CI 0.013-0.0110) and 0.0030, 95% CI 0.0014-0.0063), respectively. However the developed country estimates were much more heterogeneous. Powers didn’t find differences in transmission for male-to-female and female-to-male, but in Boily this varied by setting.
   • Multipliers are then applied to this baseline transmission probability for the risks by viral load. The reference category here, as shown in the table below, is VL<=1000. This corresponds to viral suppression, which has a much lower risk of transmission. The studies included in Boily and Powers’ reviews included people at all stages of infection. Stratified estimates in Boily et al. by disease stage estimate a per-act risk of 0.00072 during asymptomatic stage, with risks 9.17 times higher during early infection and 7.27 times higher for late infection. Since viral load in asymptomatic infection ranges from ~10,000 to 100,000, with an average ~30,000 (Fraser et al. 2007), I believe the reference category should actually be the 10,000-50,000 category.
     • The multipliers are drawn from Quinn et al. 2000 and Hollingsworth et al. 2008. Hollingsworth estimates that the risk of infection in the acute and late phases is 26 and 7 times higher than in the asymptomatic phase, respectively. But in the table from Roger’s paper, the multiplier of 26 is relative to someone with viral suppression. Quinn et al. estimate that the relative risks for VL 3500-9999, 10,000-49,999, and 50,000+ relative to <3500. Again, in the paper, these are applied relative to the VL<1000 category, which is not quite right. As noted above, Boily et al. 2009 estimate that risks are 9.17 times higher during early infection and 7.27 times higher for late infection relative to asymptomatic.
   • So I think we need to revise this table, as shown at the bottom of this email. I chose the value of 0.2 for VL<1000 somewhat arbitrarily, but to be lower than the value I calculated below for people who take ART with imperfect suppression (see additional comment below). For VL>50,000, I set the value to be similar to the estimates for late-stage infection from Hollingsworth and Boily. For VL 10,000-50,000, I set the RR to 2 as an approximation of the relationships between the risks by VL categories in Quinn et al.

2. The second question is how to model the impact of ART on viral load. The 0.04 multiplier in the table seems to come from HTPN 052 (Cohen et al. 2011), which reported a hazard ratio of 0.04 (95% CI 0.01, 0.27) comparing couples where the HIV-positive partner started ART early vs delayed. But in our model we want to be able to set the level of viral suppression (i.e. to 90% in WHO scenarios). I propose that we assume 0 transmissions from persons with viral suppression (Rodger et al. 2016 and Bavinton et al. 2018), and calculate the multipler to apply for ART based on the level of viral suppression: ART multiplier = 0.0 VS + X (1-VS). For X, I think we could assume that people who are not virally suppressed have a relaive risk of 0.4. This is based on the models by Steve Goodreau et al, (i.e. this) which assume that people who take ART with imperfect adherence have a viral load of 3.5 log 10, and that transmission is multiplied by 2.45^(VL – 4.5), for which they cite Wilson et al 2008. Using that formula, the risk multiplier for a VL of 3.5 log10 is 0.408. For the historical model, I think we could increase viral suppression from ~66% to 90% (see my notes in issue #57 )).

One more comment/question: Is it reasonable to have people progress from acute phase to VL <1000? I didn’t think people dropped below VL 1000 without treatment? https://annals.org/aim/fullarticle/709558/immunopathogenic-mechanisms-hiv-infection

[darcyrao]

Update after talking with Ruanne and Cara:

• For point #2, don't model people who are treated without full suppression. Ruanne says people are either treated or they aren't (there is no partial suppression), so we will just assume that all people on ART are virally suppressed and set the ART multiplier to 0. We will then adjust our coverage for ART to correspond to the proportion with viral suppression.
• For the WHO report, we will just build in discontinuation and adjust the multiplier for transmission to 0 for people on ART. Making the other changes is more substantial and requires more changes. Among them, we would want to change the time people spend in each VL category: currently it assumes people spend ~3.4 years in the VL <=1000 category, which I don't think it right (I actually think maybe we should drop this category altogether). We would then need to re-calibrate which will take some time.

[darcyrao]

12/4/19 email exchange with Jen Ross about her paper and how to interpret the parameters for VL transitions in the context of other studies. The analysis presented in her paper, which our model is based on, assumes that everyone progresses through a state of VL <1000 for ~3 years following acute infection:

---

I found a paper by Jai that also analyzed Partners in Prevention data and reported a median set point viral load (with measures taken 4-18 months after infection) of 4.6 log10, with an IQR of 3.83 to 5.09 (~6,310 - 100,000 copies/mL). While set point VL is highly variable across individuals and populations (see below from Fraser 2007), on average VL in the asymptomatic phase stays above ~10,000. The next plot below, from Hubert et al. 2000, shows the trajectory of VL in a French cohort over the course of infection, with a low of ~3.8 log10 (~6,310 VL copies/mL). Lyles et al. JID 2000 analyzed data from the MACS cohort and show a similar trajectory following infection.

In the van Rooyen paper (JAIDS 2013), which is cited as the source of data for validation of your modeled distribution by VL and CD4 state, I see that 27% of HIV-positive persons who reported no ART use at enrollment had VL<1000. So that would suggest that VL does drop below 1,000. However, I am curious whether there could be reporting errors, or whether some of these individuals may have recently been treated?

[darcyrao]

12/6/19 follow-up with Ruanne, Cara, and Gui after talking with Jen and reviewing more studies.

---

Jen says that Roger did most of the analysis in developing his 2013 model, so she didn’t know all of the details. It does appear that the model is based on data, but I am unclear about exactly how those data were analyzed and developed into parameters. And it seems that there may be some misreporting of ART status in those data, resulting in some people with low VL being classified as not on ART who in fact were treated (Jen says she's analyzed some PHIA data recently that show misclassification of ART status based on self-report compared to plasma drug levels).

Based on the other papers I’ve read and models I’ve seen, it feels strange to me to have everyone progress through a state with VL <1000 for 3 years after acute infection. From Tom Quinn’s analyses, VL<1000 has almost zero transmission risk, and the other papers I’ve included below show that very few people have a VL this low in untreated infection. Most of these natural history studies are in developed settings, so it may be different in sub-Saharan Africa, but I’m not sure it would be so different.

My inclination is to revise the VL states to have an acute phase, asymptomatic/set-point phase, pre-AIDS elevated phase, AIDS, and then a late/final stage. The revised parameters and sources are presented in the attached Excel file, which will be updated and maintained on the Google drive.

A comment about the parameter for the relative risk of transmission in the acute phase: There is a fair amount of uncertainty about the relative risk in this phase, with estimates of 9.2 from Boily et al. 2009 and 26 in Hollingsworth et al. 2008. Monisha seemed to recall some papers providing evidence that 26 is an overestimate, but I haven’t found that. However, I think it’s better to go on the lower side of the estimates for this given the model structure. With the compartmental model structure, we collapse partnerships and all acts that occur within them to the time point in which the partnership is formed. If a partnership is formed when someone is in the acute phase, not all of the acts with that partner will be during acute infection. However, in our model, we assume that they are. There’s not really a way around this, but it makes me inclined to use a lower estimate for transmission in acute infection.

HIV natural history parameters.xlsx

[carajbro]

Slides on effect of late-stage HIV transmission multiplier