An R package for Plasmodium vivax molecular correction: statistical genetic inference of P. vivax
- Relapse
- Recrudescence
- Reinfection
The core function compute_posterior() can be used to obtain per-person posterior probabilities of relapse, recrudescence, and reinfection (recurrence states) using P. vivax genetic data across multiple episodes to update a prior over recurrence states, which is ideally informative (e.g., based on time-to-event information).
Two other important features are more general:
plot_data()can be used to visualise genetic data for molecular correction regardless of the analytical method; for example, Plasmodium falciparum data intended for analysis using a WHO match-counting algorithm.plot_simplex()can be used to visualise per-recurrence posterior probabilities of relapse, recrudescence, and reinfection, or a vector of any three numbers in zero to one that sum to one.
Please be aware of the following points!
Pre-release state:
The R package is still under development and may contain some errors. Moreover, the statistical model around which the Pv3Rs R package is being developed is not yet peer-reviewed and thus liable to modification. The model is documented in the preprint [1]. The model builds upon a prototype documented in [2] (see [1] to better understand how the model underpinning Pv3Rs differs from the prototype).
Prior considerations:
Genetic data are modelled using a Bayesian model, whose prior is ideally informative (in [2] prior estimates 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 used to estimate the probable cause of recurrent P. falciparum malaria by setting the prior probability of relapse to zero. However, the current version of Pv3Rs is suboptimal for P. falciparum molecular correction: 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 probability of relapse is zero a priori (and of recrudescence as relapse when the prior probability of relapse exceeds zero).
Notable assumptions and limitations:
As with any model, Pv3Rs makes various assumptions that limit is capabilities in some settings. They are explained in some detail below and summarised in a table.
Mutually exclusive recurrent states
We model recurrent states (relapse, recrudescence, and reinfection) as mutually exclusive. Pv3Rs was designed with a view towards clinical trials where study participants are actively followed up frequently and where all detected infections are treated to the extent that post-treatment parasitaemia drops below some detectable level before recurrence, if it occurs. In studies where infections persist untreated and/or where events have time to accumulate, the concept of recurrence is ill-defined, and the output of Pv3Rs might not be meaningful.
Complexities of molecular correction that exceed data sampled
We do not model all the complexities around molecular correction. For example, we do not model any population structure (e.g., household effects) or the failure to capture a low-density clone in the body by a blood sample of limited volume [Snounou and Beck, 1998], nor hidden biomass in the spleen and bone marrow (an alternative source of P. vivax recurrence) [Markus, 2019]. As with study design above, we urge the user to interpret Pv3Rs’ output conditional on the input. For example, we expect Pv3Rs to output probable relapse if a person is reinfected by a new mosquito 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) - 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 vignette “Understand half-sibling misspecification”. A descriptive study to explore the extent of half-sibling misspecification is recommended (an example will be provided in an upcoming manuscript).
Observation errors and de novo mutations
We do not model undetected alleles, genotyping errors, or de novo mutations. Recall that 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. Sensitivity analyses that explore the impact of errors and mutations on recurrence state estimates are merited.
Interpreting probable reinfection and recrudescence
When data are not sufficiently informative 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 uniform distribution over graphs (see vignette “Understand graph prior ramifications”). The development of a more biologically-principled generative model on parasite relationships is merited.
| Limitation | Reason |
|---|---|
| Possible misclassification of persistent and/or accumulated states as relapse | Modelling recurrent states as mutually exclusive |
| Possible inconsistency with data on more-and-more markers | Not modelling errors |
| Possible misclassification of relapse as reinfection | Half-sibling misspecification and not modelling errors |
| Possible misclassification of recrudescence | Not modelling errors |
| Possible misclassification of reinfection | Not modelling population structure |
| Strong prior impact on posterior probabilities | Recurrent states are not always identifiable from genetic data alone |
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 counts exceeds eight but there are multiple recurrences, it might be possible to generate recurrent state estimates for recurrences analysing episodes pairwise (this approach was used in [2]).We have not tested the per-marker allele limit of
compute_posterior(). Very high marker cardinalities could lead to very small allele frequencies and thus underflow problems.
Population-level allele frequencies:
In addition to paired data, compute_posterior() requires as input population-level allele frequencies. To avoid bias due to within-host selection of recrudescent parasites, we recommend using only enrolment episodes to estimate population-level allele frequencies. That said, if 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.
Installation
# Install or update latest stable version of devtools from CRAN
install.packages("devtools")
# Install R.rsp, a vignette builder used to build LaTeX vignettes in Pv3Rs
install.packages("R.rsp")
# Install Pv3Rs from GitHub
# I highly recommend doing this in RStudio as RStudio installs pandoc needed to build vignettes.
# If you're working in R outside of RStudio you might need to install pandoc and check its path;
# otherwise set build_vignettes = FALSE
devtools::install_github("aimeertaylor/Pv3Rs", build_vignettes = TRUE)
# 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, e.g.,
?compute_posterior