a. Expansion of Network to Increase Accuracy
The methods applied in these analyses can be applied to a larger area with the goal of correcting some of the bias in the observed network due to missing nodes. Regions 3-6 and Charlotte TGA (Figure 7) comprise the central area of NC, which is geographically and contextually distinct from the western mountain and eastern coastal areas. The Nexus study (section II.B.) found overlap in the sexual networks between adjoining regions; adding cases from Regions 3 and 4 to the Region 6 cases resulted in large sexual network components that appeared to be smaller, discreet components prior to the addition of the cases and their
partners from other regions, without artificial reduction of the sexual network components due to the constraints of administrative boundaries (Figure 23).
I demonstrated this in our study by including syphilis cases and partners, which increased the size of the network, while resulting in a smaller number of components. The component size 92 in this study (Figure 22) would have been 6 smaller components without the inclusion of the syphilis investigations. Adding cases and partners, including both HIV and
syphilis, from a neighboring region would likely result in a better representation of the true underlying sexual network than what was observed. This may have implications for the model, since I used only the observed Region 6 network to calculate network predictors that are sensitive to network size.
b. Exponential Random Graph Models to Analyze Network Growth Drivers
Markov chain Monte Carlo (MCMC) estimation can be used to create a distribution of randomly-generated networks based upon selected attributes of the observed network, such as number of nodes or edges (see Appendix C section A). Upon comparing the observed network to the distribution of random networks, important properties of the observed network, such as clustering or transitivity above or below what is expected, can be identified based upon divergence from the distribution of randomly-generated networks. Inferences can then be drawn estimating which processes influence tie formation in NC, which may reveal avenues for intervention.
c. Comparison of Centrality Scores and Network Structure Involvement by Infection Recency Behavior changes following diagnosis,77-79 so chronically-infected persons aren’t well-
suited for an assessment of behavior and risk though they are frequently used in publications to describe risk behavior in local populations. An analysis of HIV-positives nodes comparing the centrality scores and risk behaviors of persons estimated to be infected recently (less than 12 months between infection and diagnosis) to persons estimated to be chronically-infected (12 months or more) may provide clues to behaviors and partnership patterns associated with HIV acquisition, and help us determine whether network position differs during the period in which infected was acquired. Comparing recently-infected, chronically-infected, and uninfected individuals may be less biased than networks that are egocentric with respect only to the infected individuals.11 Centrality score mean, across multiple centrality measures, can be
calculated and compared to determine whether there is a significant difference between groups at alpha = 0.05.
d. Geographic or Spatial Analysis Combined with Social Network and Gene Sequence Analyses
The majority of location-based analyses in NC thus far have been geographic rather than spatial. A spatial analysis based on residence would allow location data to be incorporated into models. In addition to being more suited for models and be interpretable on continuous scales, spatial data are more suitable to combine with both network and gene sequence relatedness data; phylogenetic relatedness, spatial closeness, and network distance can all be defined as continuous measures. A previous study of spatial distance and genetic distance used acutely infected individuals with pol gene sequences who were consented into a clinical cohort in NC. The patients were assessed for transmitted drug resistant or drug susceptible virus. Sequenced virus was found to differentiate less within rural areas than within urban areas.51 The application of the combination of these three methods to an entire population of
geographically-defined incident diagnoses is novel and has the potential to significantly improve the understanding of the spread of HIV.
e. Transmitted Drug Resistance
Prevalence of TDRM by sexual network component could be ascertained from a study, which could then be used to guide both clinical and public health practice at a more precise sub- population level. The ability to link drug resistance data extracted from sequences to defined sexual network components would allow assessment not only HIV risk, but also TDRM risk, for individuals within the component.49 Finding a similar mutation profile may result in network
component linkages through an individual believed to be HIV-negative or thought to be anonymous.
Once the assortative factors of the network components were determined, HIV and TDRM risk by type of mutation could be assessed by those factors. Linking the resistance profile to cases who are newly diagnosed or who are failing treatment would allow calculation of a crude prevalence of DRM to be established by drug class within each component – the
prevalence by type of TDRM for the unique set of demographic and risk characteristics which define each component. In another state in the US South, Mississippi, ART-naïve young Black MSM clustered phylogenetically and had TDRM strains, while phylogenetic transmission clusters (TC) containing only older Black MSM did not have TDRM,93 indicating two discrete
sexual networks separated by age and supporting the idea that discrete sexual network components may each have their own circulating drug resistant variants.
One limitation is that a consensus sequence is returned by LabCorp. Minor variants have been shown to incorporate DRM, contributing to the risk of virologic failure and a
resurgence in resistant strains,221-223 which may not be reflected using this sequencing method.
If minority variants coding for DRM are significant in this population then prevalence estimates were underestimated. However, one study that employed ultra-deep pyrosequencing found that half of patients with TDRM only had the mutations in <20% of their virus population.224
f. Predictive Model to Identify HIV-Negative Persons Who Would Benefit Most from PrEP The converse of the model presented in Chapter VII is a model to identify HIV-negative persons who would benefit most from PrEP. A sexual network study that began with 2 acutely- infected patients in NC resulted in a sexual network of 398 persons, nearly half of whom (47%) were of unknown serostatus due to inability to locate or testing refusal. Ninety-two persons in the network were confirmed to be HIV-negative, but 24 of those persons (26%) seroconverted within 3 years.21 HIV-negative partners of new HIV cases and HIV-negative persons involved in
syphilis investigations can be followed through the same 3-year period as the HIV-positive index cases to determine who becomes infected within 3 years. There is significant overlap of syphilis and HIV sexual networks in NC, particularly among MSM,21,96 so the syphilis cases will serve as
an additional HIV-negative population for the baseline period. A predictive model can then be constructed, using the same types of information as the predictive model already developed.