Bayesian inference of set-point viral load transmission models

Pieter Libin, Laurens Hernalsteen, Kristof Theys, Perpétua Gómes, Ana Abecasis, Ann NowéarXiv 2018 · DOI · arXiv

Abstract

When modelling HIV epidemics, it is important to incorporate set-point viral load and its heritability. As set-point viral load distributions can differ significantly amongst epidemics, it is imperative to account for the observed local variation. This can be done by using a heritability model and fitting it to a local set-point viral load distribution. However, as the fitting procedure needs to take into account the actual transmission dynamics (i.e., social network, sexual behaviour), a complex model is required. Furthermore, in order to use the estimates in subsequent modelling analyses to inform prevention policies, it is important to assess parameter robustness. In order to fit set-point viral load models without the need to capture explicitly the transmission dynamics, we present a new protocol. Firstly, we approximate the transmission network from a phylogeny that was inferred from sequences collected in the local epidemic. Secondly, as this transmission network only comprises a single instance of the transmission network space, and our aim is to assess parameter robustness, we infer the transmission network distribution. Thirdly, we fit the parameters of the selected set-point viral load model on multiple samples from the transmission network distribution using approximate Bayesian inference. Our new protocol enables researchers to fit set-point viral load models in their local context, and diagnose the model parameter's uncertainty. Such parameter estimates are essential to enable subsequent modelling analyses, and thus crucial to improve prevention policies.

The Significance of Set-Point Viral Load

In the study of HIV-1, the "set-point viral load" (SPVL) is a critical clinical marker. It represents the stable amount of virus in a patient's blood during the chronic, asymptomatic phase of infection. SPVL is a primary predictor of both the rate of disease progression to AIDS and the likelihood of onward transmission to others.

The clinical course of HIV infection, highlighting the Set-Point Viral Load (SPVL) during the latency phase.

Crucially, SPVL is partly heritable; the virus carries genetic traits that influence the viral load in the next person it infects. To design effective public health interventions, researchers must model this heritability. However, traditional approaches require granular data on social networks and sexual behavior—data that is often incomplete or entirely unavailable.

A New Bayesian Protocol for Local Epidemics

To bypass the need for explicit social network modeling, researchers have developed a protocol that uses viral genetics as a proxy for transmission dynamics. The method follows a multi-step pipeline:

  1. Genetic Distance Mapping: The protocol calculates genetic distances between virus sequences from different patients. This creates a graph where nodes represent patients and edges represent genetic similarity.
  2. Network Approximation: By applying distance thresholds, the graph is pruned to identify potential transmission partners, forming a Minimum Spanning Tree (MST) as a plausible transmission backbone.
  3. Accounting for Uncertainty: Rather than relying on a single inferred tree, the researchers infer a degree distribution (e.g., Waring or Yule) to sample multiple possible transmission networks, ensuring parameter robustness.
  4. Approximate Bayesian Computation (ABC): Because the likelihood functions for these models are mathematically intractable, the researchers use ABC. They simulate thousands of transmission scenarios and compare the resulting SPVL distributions to real-world data.
Illustration of the transmission modeling process (V-VII) and the comparison of simulated vs. observed SPVL distributions (VIII) using ABC.

Key Insights from the Portuguese Epidemic

The protocol was validated using data from the Portuguese HIV-1 epidemic, focusing on Subtypes B and G. The study revealed several critical insights:

  • The "Shortcut" is Effective: Genetic distance successfully acts as a proxy for transmission networks, allowing for accurate model fitting without behavioral data.
  • Subtype Sensitivity: Different HIV subtypes exhibit distinct evolutionary signatures. Subtype G and Subtype B required different genetic distance thresholds, highlighting the need for local, subtype-specific modeling.
  • Parameter Robustness: By sampling from a distribution of networks, the protocol identifies which model parameters are well-defined by the data and which remain uncertain, providing vital context for policymakers.

Conclusion

This new protocol enables researchers to characterize the "viral fingerprint" of local epidemics with high precision. By leveraging genetic sequencing data—which is increasingly standard in clinical settings—public health officials can better predict epidemic trajectories and refine prevention strategies without the logistical hurdles of mapping complex social networks.

Summary updated

Abstract

When modelling HIV epidemics, it is important to incorporate set-point viral load and its heritability. As set-point viral load distributions can differ significantly amongst epidemics, it is imperative to account for the observed local variation. This can be done by using a heritability model and fitting it to a local set-point viral load distribution. However, as the fitting procedure needs to take into account the actual transmission dynamics (i.e., social network, sexual behaviour), a complex model is required. Furthermore, in order to use the estimates in subsequent modelling analyses to inform prevention policies, it is important to assess parameter robustness. In order to fit set-point viral load models without the need to capture explicitly the transmission dynamics, we present a new protocol. Firstly, we approximate the transmission network from a phylogeny that was inferred from sequences collected in the local epidemic. Secondly, as this transmission network only comprises a single instance of the transmission network space, and our aim is to assess parameter robustness, we infer the transmission network distribution. Thirdly, we fit the parameters of the selected set-point viral load model on multiple samples from the transmission network distribution using approximate Bayesian inference. Our new protocol enables researchers to fit set-point viral load models in their local context, and diagnose the model parameter's uncertainty. Such parameter estimates are essential to enable subsequent modelling analyses, and thus crucial to improve prevention policies.
BibTeX
@misc{libin2018bayesian,
  author = {Pieter Libin and Laurens Hernalsteen and Kristof Theys and Perpétua Gómes and Ana Abecasis and Ann Nowé},
  title = {{Bayesian inference of set-point viral load transmission models}},
  year = {2018},
  eprint = {1811.11042},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1811.11042},
}
Figure