Caution in inferring viral strategies from abundance correlations in marine metagenomes

Abstract

Coutinho et al. 1 reported metagenomics-derived evidence in support of the ‘Piggyback-the-Winner’ (PtW) hypothesis that lysogeny prevalence increases at high microbial abundances. Coutinho et al.1 did not directly estimate lysogenic prevalence, but instead, found that the ratio of virus-to-microbial host abundances decreased as microbial cell abundances increased. This pattern represents potential (albeit indirect) evidence in support of PtW. Here, we show that the bulk of these reported abundance relationships are likely spurious. Instead, we find absence of evidence for positive, sublinear correlations between virus and microbial abundances as estimated in dozens of putative virus-microbe pairs identified by Coutinho et al.1 The absence of correlations between virus and microbial abundances is a counter-indicator for PtW. Altogether, our re-analysis suggests the need for caution in using correlation-based inference to identify viral strategies from metagenomics-derived abundance relationships. To begin, consider the work of Coutinho et al.1, who developed a metagenomics-based approach to characterize the diversity, ecology, host-associations, and strategies of marine phage. In doing so, they introduced a “new method for host prediction based on co-occurrence associations”, in which “virus–virus abundance associations were used for host affiliation”1. As a result, Coutinho et al.1 claimed that observed abundance relationships amongst phage and bacterial hosts in a range of marine habitats are consistent with the recently introduced mechanism of PtW2. The hypothesis underlying PtW is that viruses have increased lysogenic prevalence (and decreased lytic activity) with increasing microbial abundances. This hypothesis is meant to provide a mechanistic basis for empirical findings that total virus abundances increase with total microbial abundances even as the number of viruses per microbe decreases as microbial abundances increase. This pattern is found across marine, freshwater and other environmental systems (see Knowles et al.2, Wigington et al.3, and Parikka et al.4), with similar patterns found in predator–prey relationships5. Sublinear (or less than proportional6) increases in virus abundances with microbial abundances may arise from multiple governing mechanisms. These mechanisms include PtW, whose underlying mathematical model predicts that viral abundances increase with increasing microbial abundances, albeit sublinearly (see Fig. 1b of ref. 2). Addditional mechanisms that could explain sublinear increases include variation in life history traits in antagonistic virus–microbe dynamics7 or trade-offs in Kill-theWinner models8. As a consequence, the value of these patterns as exclusive indicators of any particular mechanism is disputed (see exchange of Weitz et al.7 and Knowles and Rohwer9, as well as the follow-up work of Knowles et al.10). Nonetheless, the possibility of using metagenomics-based methods to infer virus–host pairs and their abundance relationships could provide insights into viral strategies and their consequences in marine systems. Here, we focus on the empirical findings of Coutinho et al.1 and ask: do the abundance relationships exhibit robust evidence for sublinear increases in virus abundances with microbial abundances? Coutinho et al.1 used multiple approaches, including virus–virus abundance associations, to link viruses and their putative hosts. We use the term “abundances” as a proxy for the metagenomics-inferred relative densities of viruses and host types reported by Coutinho et al.1, consistent with their implementation. Once they estimated abundances, Coutinho et al.1 quantified the relationship between the ratio of virus-to-host abundances vs. host abundances given putative pairs at both the genus and phylum levels. For example, let y be the log-transformed virus abundance and x be the log-transformed host abundance of an identified pair. If y increased sublinearly with x, then one would expect that y ~ xα where 0 < α < 1. The inequality α > 0 implies that virus abundances increase with microbial abundances and the inequality α < 1 implies that the increase is sublinear. It is also possible to evaluate ratio-based fits, i.e., quantifying the relationship between y/x and x. In that case, we expect y/x ~ xβ where β= α− 1. Hence, sublinear power-law relationships between y and x should lead to power-law relationships between virus-microbe ratios and microbial abundances with negative slopes between −1 < β < 0. Coutinho et al.1 examined relationships between y/x vs. x, rather than directly examining y vs. x. If y is unrelated to x then one would expect best-fit curves between y/x vs. x to be statistically equivalent to fitting 1/x vs. x, thereby Corrected: Author correction