Pv3Rs

An R package for Plasmodium vivax molecular correction via statistical genetic inference of

• Relapse
• Recrudescence
• Reinfection

The core function, compute_posterior(), computes per-person posterior probabilities of relapse, recrudescence, and reinfection (recurrence states) using P. vivax genetic data on two or more episodes. To better understand the core function, in addition to this README, we recommend reading - documentation accessed via ?compute_posterior() - the vignette “Demonstrate Pv3Rs usage” accessed via vignette("demonstrate-usage", "Pv3Rs") - Understand posterior probabilities - the article “Pv3Rs: Plasmodium vivax relapse, recrudescence, and reinfection statistical genetic inference” published in Bioinformatics

Two other important features:

• plot_data() visualises genetic data for molecular correction, regardless of the analytical method (e.g., it can be used to visualise Plasmodium falciparum data intended for analysis using a WHO match-counting algorithm).

• plot_simplex() can be used to visualise per-recurrence probabilities of relapse, recrudescence, and reinfection (or any other probability triplet summing to one).

Please be aware of the following points!

The statistical model is described in the supplement of the Bioinformatics article “Pv3Rs: Plasmodium vivax relapse, recrudescence, and reinfection statistical genetic inference”. It builds on the prototype that features in Taylor & Watson et al. 2019.

Prior considerations:

• Genetic data are modelled using a Bayesian model, whose prior is ideally informative (in Taylor & Watson et al. 2019 priors were generated by a time-to-event model built by James Watson) because the cause of recurrent P. vivax malaria is not always identifiable from genetic data alone: when the data are consistent with recurrent parasites that are relatively unrelated to those in all preceding infections, both reinfection and relapse are plausible; meanwhile, when the data are compatible with recurrent parasites that are clones of those in the preceding infection, both recrudescence and relapse are plausible.

• The main Pv3Rs function, compute_posterior(), could be applied to P. falciparum by setting the prior probability of relapse to zero, but genotyping errors, which are not accounted for under the current Pv3Rs model, are liable to lead to the misclassification of recrudescence as reinfection when the prior probability of relapse is zero (and of recrudescence as relapse when the prior probability of relapse exceeds zero). A post hoc diagnostic to identify misclassified recrudescence is presented in Understand genotyping errors.

Notable assumptions and limitations:

As with any model, Pv3Rs makes various assumptions that limit its capabilities in some settings.

Mutually exclusive recurrent states

Recurrence states are modelled as mutually exclusive, suitable for studies where participants are actively followed up frequently and where all detected infections are treated to the extent that parasitaemia drops below some detectable level before recurrence if it occurs. In studies with untreated or accumulated infections, outputs may not be meaningful.

Unmodelled complexities

We do not model all the complexities around molecular correction. For example, population structure, including household effects; failure to capture low-density clones in a blood sample of limited volume [Snounou & Beck, 1998]; and hidden biomass the spleen and bone marrow [Markus, 2019]. Users must interpret outputs in light of these limitations and in the context of the study and its methods. For example, we expect Pv3Rs to output probable relapse if a person is reinfected by a new mosquito but with parasites that are recently related to those that caused a previous infection, as might happen in household transmission chains.

Sibling misspecification

Relapsing parasites that are siblings of parasites in previous infections can be meiotic, parent-child-like, regular or half siblings, but we model all sibling parasites as regular siblings via the following assumptions:

• Allele inheritance is independent (not true of meiotic siblings, which are complements of one another)
• Sibling relationships are transitive (not true of parent-child-like trios or some half-sibling trios)
• Alleles of a sibling cluster are drawn from at most two parental alleles (not true of half siblings)

In our experience, half sibling misspecification leads to some misclassification of relapses as reinfections; see Understand half-sibling misspecification and Understand posterior probabilities, where half siblings lead to probabilities that behaviour erratically with increasing marker counts. A descriptive study to explore the extent of half-sibling misspecification is recommended.

Genotyping errors and de novo mutations

We do not model undetected alleles, other genotyping errors, or de novo mutations. Recrudescent parasites are modelled as perfect clones under Pv3Rs. As such, the posterior probability of recrudescence is rendered zero by errors and mutations. This becomes more likely when there are data on more markers. Understand genotyping errors explores the impact of errors and mutations on recurrence state probabilities.

Interpreting probable reinfection and recrudescence

When genetic data alone are insufficient to distinguish between recrudescence and relapse (or reinfection and relapse), the posterior probabilities of recrudescence and relapse (or reinfection and relapse) are heavily influenced by our a priori uniform assumption over relationship graphs; see Understand graph-prior ramifications. The development of a more biologically-principled generative model on parasite relationships is merited.

Computational limits:

• Pv3Rs scales to hundreds of markers but not whole-genome sequence (WGS) data.

• We do not recommend running compute_posterior() for data whose total genotype count (sum of per-episode multiplicities of infection) exceeds eight. If the total genotype count exceeds eight but there are multiple recurrences, it might be possible to compute posterior probabilities by analysing episodes pairwise (this approach was used in Taylor & Watson et al. 2019; we’re working currently on an improved version).

• The per-marker allele limit of compute_posterior() is untested. Very high marker cardinalities could lead to very small allele frequencies and thus some underflow problems.

Population-level allele frequencies:

In addition to P. vivax allelic data on two or more episodes, compute_posterior() requires as input population-level allele frequencies. To minimise bias due to within-host selection of recrudescent parasites, we recommend using only enrolment episodes to estimate population-level allele frequencies, and ideally enrolment episodes from study participants selected at random, not only study participants who experience recurrence. That said, if there is strong prior reason to believe most recurrences are either reinfections or relapses, both of which are draws from the mosquito population (albeit a delayed draw in the case of a relapse), assuming there is no systematic within-patient selection (as might occur when infections encounter lingering drug pressure), estimates based on all episodes should be unbiased and more precise than those based on enrolment episodes only.

Read-count data:

Unfortunately, the Pv3Rs model does not exploit data on read counts at present. However, read-count data could be used to compute population-level allele frequencies, assuming they are not biased by experimental artefacts.

Installation

# Install Pv3Rs from CRAN:
install.packages("Pv3Rs")

# Load and attach Pv3Rs
library(Pv3Rs)

# List links to all available documentation
help(package = "Pv3Rs")

# List links to vignettes
vignette(package = "Pv3Rs")

# View function documentation including examples, e.g., 
?compute_posterior

#===============================================================================
# To install the development version of Pv3Rs:
#===============================================================================
# Doing this in RStudio ensures pandoc, required for vignette building, is
# installed. If you're not working in RStudio, you might need to install pandoc 
# and check its path (or set build_vignettes = FALSE)
install.packages("devtools") # Install or update devtools from CRAN
devtools::install_github("aimeertaylor/Pv3Rs", build_vignettes = TRUE)