Systematic bias in malaria parasite relatedness estimation

Abstract

Genetic studies of malaria parasites increasingly feature estimates of relatedness. However, various aspects of malaria parasite relatedness estimation are not fully understood. For example, estimates of relatedness based on whole-genome-sequence (WGS) data often exceed those based on more sparse data types. We explore systematic bias in relatedness estimation using theoretical, numerical and empirical approaches. Specifically, we use a non-ancestral model of pairwise relatedness to derive theoretical results; a simulation model of ancestry to independently verify and expand our theoretical results; and data on parasites sampled from Guyana to explore how theoretical and numerical results translate empirically. We show that allele frequencies encode, locus-by-locus, relatedness averaged over the set of sampled parasites used to compute them. These sample allele frequencies are typically plugged into the models used to estimate pairwise relatedness. Consequently, models of pairwise relatedness are misspecified and pairwise relatedness values are systematically underestimated. However, systematic underestimation can be viewed as population-relatedness calibration, i.e., a way of generating measures of relative relatedness. Systematic underestimation is unavoidable when relatedness is estimated assuming independence between genetic markers. It is mitigated when estimated using WGS data under a hidden Markov model (HMM), which exploits linkage between proximal markers. Estimates of absolute relatedness generated under a HMM using relatively sparse data should be treated with caution because the extent to which underestimation is mitigated is unknowable. That said, analyses dependent on absolute values and high relatedness thresholds are relatively robust. In summary, practitioners have two options: resolve to use relative relatedness estimated under independence or try to estimate absolute relatedness under a HMM. We propose various practical tools to help practitioners evaluate their situation on a case-by-case basis. Author summaryMalaria genomic epidemiology is increasingly recognised as a tool for public health. Relatedness, which captures likeness derived from common ancestry, is a useful concept for malaria parasites. Analyses of malaria parasite relatedness are important for generating results on spatiotemporal scales relevant to disease control. Since shared ancestry is unobservable, relatedness must be estimated under a statistical model. However, not all aspects of malaria parasite estimation are fully understood, including the effects of different data types. In this work, we characterise systematic biases in estimates of malaria parasite relatedness. Our analysis is three-fold: we mathematically interrogate a non-ancestral model of relatedness to derive theoretical results; simulate parasite ancestries from first principles to yield numerical results; and perform an empirical case study of parasites sampled from Guyana. We show that bias may be particularly pronounced when using sparse marker data from inbred parasite populations, which are often found in pre-elimination settings. We chart out a practical roadmap to enable practitioners to assess epidemiological settings on a case-by-case basis. Our findings are relevant to applications in malaria genomic epidemiology that use relatedness directly or indirectly, including molecular surveillance and the genetic-based classification of treatment failure.