Phylodynamics: The Discovery Phase
The advent of extremely high throughput DNA sequencing ensures that genomic data from
microbial organisms can be acquired in unprecedented quantities and with remarkable
rapidity. Although this genomic revolution will affect all microbes alike, our focus
here is on RNA viruses, as the rapidity of their evolution, which is observable over
the time scale of human observation, allows phylodynamic inferences to be made with
great precision. In the foreseeable future it is likely that complete genome sequencing
will become the standard method of viral characterization, providing the highest possible
resolution for phylogenetic studies. The rapidity with which genome sequence data
were generated from the ongoing epidemic of swine-origin H1N1 influenza A virus [1]
is testament to the power of this technology.
Understandably, pathogen discovery is a major focus of this new-scale genome sequencing
[2]. It is now possible to sequence the entire assemblage of viruses in a particular
tissue type or host species [3]–[5], as well as all those viruses that are associated
with specific disease syndromes [6],[7]. In essence, this new era of metagenomics
constitutes a crucial taxonomic discovery phase in virology and epidemiology that
allows the genetic characterization of new viruses within hours of their isolation.
Assembling an inventory of viruses that may emerge in human populations is of major
importance to public health and to students of biodiversity. However, it is only the
first step in developing a full quantitative understanding of the processes that shape
the epidemiology and evolution—the phylodynamics—of RNA virus infections [8]. To achieve
this goal, we argue here that the field of viral phylodynamics requires its own discovery
phase; that is, a comprehensive and quantitative analysis of the interaction between
the ecological and evolutionary dynamics of all circulating RNA viruses from the molecular
to the global scale. Such a marriage of phylogenetic and epidemiological dynamics
is currently only potentially possible for the select few human viruses for which
large genome sequence datasets have been acquired, such as HIV and influenza A virus,
and even here fundamental gaps in our knowledge remain (see below). Indeed, it is
striking that so few complete genome sequences are currently available for viruses
whose epidemiological dynamics are known in exquisite detail, such as measles [9],[10];
these sequences have been so sparsely sampled in both time and space that a full phylodynamic
perspective has not yet been achieved. We contend that a better understanding of RNA
virus phylodynamics will allow more directed attempts at pathogen surveillance, facilitate
more accurate predictions of the epidemiological impact of newly emerged viruses,
and assist in the control of those viruses that exhibit complex patterns of antigenic
variation such as dengue and influenza. Just as PCR and first-generation DNA sequencing
ushered in the science of molecular epidemiology, so next-generation sequencing may
herald the age of phylodynamics. Box 1 lists a number of key questions that can be
addressed within this phylodynamics research program.
Box 1. Key Research Questions in RNA Virus Phylodynamics
What is the range of phylodynamic patterns observed in RNA viruses? Can they be categorized
into specific groups? How do these patterns relate to other “life history” variables
exhibited by RNA viruses?
What epidemiological and evolutionary processes give rise to these phylodynamic patterns?
What generalities can be drawn?
How commonly does natural selection (compared to neutral evolutionary processes) determine
the population dynamics of pathogens? On what scale does natural selection act? How
does viral immune escape reduce herd immunity at the population level and allow the
persistence of viral lineages in epidemic troughs?
What is the range of spatial patterns exhibited by RNA viruses? What epidemiological
factors are responsible for these patterns?
How do different viral species (various respiratory viruses, for example) interact
in host immunity?
A number of important advances are needed to meet our goal of a comprehensive catalog
of the diversity of phylodynamic patterns in RNA viruses. Because answers to many
of the most interesting research questions depend on sufficiently large sample sizes,
we require large numbers of sequences that have been rigorously sampled according
to strict temporal, spatial, and clinical criteria, and that as much of these data
are publicly accessible as possible. A phylodynamic analysis has little value unless
viral genomes are sampled on the same scale as the epidemiological processes under
investigation.
The only acute virus for which a suitably expansive genome dataset currently exists
is influenza. In this case, the >4,000 complete genomes generated under the Influenza
Genome Sequencing Project [11] have provided important new insights into the evolution
and epidemiology of this major human pathogen [12]. To highlight one key insight here,
these genome sequence data have revealed that multiple lineages of influenza virus
are imported and circulate within specific geographic localities (even within relatively
isolated populations), generating both frequent mixed infections [13] and reassortment
events [14]. Even so, the sampling of these genome sequences (and associated epidemiological
covariates) may not be dense enough to fully capture spatial dynamics [15]. There
is also a marked absence of samples from asymptomatically infected patients (or those
with mild disease), so it is impossible to link genetic variation to clinical syndrome.
Such a bias against viruses sampled from individuals with asymptomatic infections
is a common problem in molecular epidemiology.
Epidemiological Factors
It is also clear that for many RNA viruses we need to better understand a number of
key epidemiological factors, such as the interaction between local persistence, epidemic
dynamics in both time and space, the impact of measures to control the spread of infection,
and the consequences of adaptive evolution in those viral genes that interact most
intimately with the host immune response. It is instructive to imagine the ideal database
for addressing these issues. In the case of acute infections, the goal would be to
collect four parallel datasets on the appropriate scale of interest during outbreaks
(Figure 1). This database would comprise, first, epidemic dynamics in time and space,
ideally at a comparable or higher frequency than the generation time of individual
infections. Second, and in parallel, our ideal study would collect viral genome sequence
data at these time points, sampling both within and among infected hosts. Both disease
incidence data (bolstered by contact tracing) and viral sequence data furnish information
on the transmission network traced by an outbreak. Third, we would need to know the
underlying contact network of susceptible individuals, which serves as fuel for the
epidemic. This is a difficult structure to measure directly, although novel measurements
of human interactions are increasingly shedding light on the problem [16]. Finally,
measurements of the immunity structure of our contact network [17]—reflecting the
past history of the virus in the population—are key for understanding both the dynamics
of epidemic spread and the evolutionary pressures that shape virus diversity.
10.1371/journal.pcbi.1000505.g001
Figure 1
Sampling scales for acute RNA viruses and the associated phylodynamic processes that
viral genome sequence data and host sampling can elucidate.
The outbreak of foot-and-mouth disease (FMD, an RNA virus infection of cattle) in
the UK in 2001 resulted in a database that is arguably closest to our ideal on the
epidemiological scale [18],[19]. Notwithstanding a variety of gaps in data from the
epidemic [20], it is one of the most well-documented large outbreaks in terms of the
availability of spatiotemporal incidence data in parallel with contact tracing and
the underlying spatial pattern of the susceptible farms as a measure of the contact
network. In addition, analyses of viral sequences from relatively small samples of
farms have drawn important conclusions about epidemic spread and allowed the testing
of new methods to recover the spatiotemporal patterns written into sequence data [18],[20].
Importantly, samples exist from over half the ∼2,000 confirmed infected premises in
2001: sequencing whole FMD virus genomes from these samples would provide a vast resource
for basic and applied developments in integrating epidemiological and phylogenetic
information to dissect spatiotemporal spread. We suggest that achieving this task
would be a huge contribution to understanding the phylodynamics of acute viruses.
Another virtue of animal infections like FMD is that the relationship between the
determinants of viral variability within and between hosts can also be dissected by
experimental infections (see [21] for another example).
A parallel limitation of many phylogenetic approaches to viral epidemiology is that
they have often proceeded in the absence of the necessary metadata, such as the precise
time and place of sampling or those that relate to clinical syndrome [22]. A perhaps
more challenging goal for phylodynamics is therefore to integrate phylogenetic patterns
with other biological variables, such as the nature of antigenic variation, the capacity
for drug resistance, or the clinical syndrome of the host, as well as the spatial
host network data outlined above. Cohort studies may be the most productive way to
link genomics with epidemiological variables.
The lack of a synthesis of phylogenetic and phenotypic/epidemiological data is reflected
in the current debate over the mode of antigenic evolution in human influenza A virus.
Although it has long been known that the hemagglutinin (HA) and neuraminidase (NA)
proteins of human influenza A virus evolve by strong natural selection to evade the
host immune response—a process commonly called antigenic drift [23],[24]—the precise
mechanisms by which such drift occurs are uncertain. From a phylodynamics perspective,
the key observation is that over long time periods a single lineage of HA sequences
from subtype A/H3N2 influenza viruses links epidemic to epidemic [23], although intensive
sampling has revealed that single populations may harbor far higher levels of genetic
diversity [25]. Rather different phylodynamic patterns are seen in other influenza
viruses, including those sampled from birds (Figure 2). Three models have been proposed
to explain the distinctive phylodynamic pattern observed in human A/H3N2 viruses:
(i) that there is short-lived cross-immunity among viral strains [26], (ii) that the
HA evolves in a punctuated manner among antigenic types that are linked by a network
of neutrally evolving sites [27], and (iii) that the virus continually reuses a limited
number of antigenic combinations [28].
10.1371/journal.pcbi.1000505.g002
Figure 2
Phylodynamic patterns of human and avian influenza viruses.
The left diagram shows the phylogeny of the hemagglutinin (HA) gene of human H3N2
influenza A viruses sampled between 1985 and 2005, revealing the “ladder-like” branching
structure indicative of antigenic drift. By comparison, the phylogeny of the HA gene
of human influenza B virus sampled over the same interval (center diagram) shows the
co-circulation of the antigenically distinct “Victoria 1987” and “Yamagata 1988” lineages,
as well a shorter length from root to tip, reflecting a lower rate of evolutionary
change. Finally, the phylogeny for the HA gene of H4 avian influenza virus (right
diagram) reveals the deep geographic division between the Eurasian and Australian
versus North American lineages of this virus.
To determine which combination of these models best explains influenza phylodynamics
will require more expansive genome sequence data, as well as focused sampling and
epidemiological surveillance in Southeast Asia, which is likely the global source
population for the virus [29]. More importantly, it is also crucial that these phylogenetic
data are combined with detailed, spatiotemporally disaggregated antigenic information.
Indeed, it is remarkable that despite the abundance of information on the antigenic
characteristics of individual influenza viruses, most notably through the use of the
hemagglutinin inhibition (HI) assay [17], these data have not been routinely linked
to phylogenetic information. It is clear that both antigenic and phylogenetic analyses
would greatly benefit from each other.
New-Generation Computational Tools
Another important challenge for phylodynamics is to match the remarkable ongoing developments
in genome sequencing technology to the increase in the power of the computational
tools available to analyze these sequence data. Crucially, in phylogenetics, the size
of the space of possible trees increases faster than exponentially with the number
of sequences, such that the availability of datasets comprising thousands of complete
genomes [30] presents a major combinatorial problem. This problem creates a growing
discrepancy between our ability to generate genome sequence data and our capacity
to analyze them using the most sophisticated methods. Redressing this balance should
be the major goal of bioinformatics in the future; and in fact some progress has been
made recently [31].
It is also clear that improvements need to be made to the methods that are available
to analyze genome sequence data. A powerful set of research tools in this area comprises
those based on coalescent theory, as this provides a natural link between the analysis
of epidemiological and phylogenetic patterns [8],[32]. In particular, the coalescent
allows the demographic characteristics of viral populations (particularly population
size and growth rate) to be inferred directly from gene sequence data. Coalescent
analyses are especially powerful in the case of RNA viruses, because their rapid evolution
means that temporal and spatial dynamics are discernable over the period of human
observation [33] and can in theory be combined with time series epidemiological data.
However, currently available coalescent methods are restricted by the limited scope
of demographic models and their inability to fully incorporate spatial information.
In particular, most acute RNA viruses have complex population dynamics that combine
distinct periods of growth and decline. The most commonly used phylodynamic tool available
in such cases is the Bayesian skyline plot (and the related Bayesian “skyride” [34]),
which represents a piecewise graphical depiction of changes in genetic diversity through
time [32]. In the case of neutral evolution, such changes in genetic diversity also
reflect underlying changes in the number of infected hosts. Although the Bayesian
skyline plot can reveal unique features of epidemic dynamics (Figure 3) [30], precise
estimates of parameters such as population growth rate are not yet possible.
10.1371/journal.pcbi.1000505.g003
Figure 3
Fluctuating genetic diversity of influenza A virus.
The figure shows a Bayesian skyline plot of changing levels of genetic diversity through
time for the HA gene (165 sequences) of A/H3N2 virus sampled from the state of New
York, US, during the period 2001–2003. The y-axes depict relative genetic diversity
(N
e
t, where N
e is the effective population size, and t the generation time from infected host to
infected host), which can be considered a measure of effective population size under
strictly neutral evolution. Peaks of genetic diversity, reflecting the seasonal occurrence
of influenza, are clearly visible. See [30] for a more detailed analysis.
The coalescent methods commonly used to study RNA virus evolution focus largely on
temporal dynamics (a natural function of the rapidity of viral evolution), with little
consideration of patterns of spatial diffusion. Although these phylogeographic patterns
are becoming increasingly well described for RNA viruses [35], few methods effectively
recover the spatial component in genome sequence data. For example, commonly used
parsimony-based approaches consider a single phylogenetic tree without an explicit
spatial model (see, for example, [36]). In addition, these methods usually describe
the place of origin and direction of spread of viral lineages without formal tests
of competing spatial hypotheses. As a specific case in point, although gravity models
(in which patterns of viral transmission reflect the size of and distance between
population centers) have been applied successfully to morbidity and mortality data
from human influenza A virus to describe its spread across the United States [37],
they have yet to be interpreted within a phylogenetic setting. A clear push for the
future should therefore be the development of coalescent tools that integrate the
analysis of spatial and temporal dynamics within a single framework, with a focus
on those that combine phylogenetic data and information on the dynamics of the host
contact network of susceptible, infected, and immune individuals.
Looking beyond the Consensus Sequence
The vast majority of studies of RNA virus evolution undertaken to date, particularly
of those viruses that cause acute infections, rely on the analysis of consensus sequences
in which the nucleotide shown for any given site is the most common among all the
genomes within a patient. Although the use of consensus sequences is adequate for
many aspects of molecular epidemiology, in which complete genomes may suffice to determine
even tight transmission chains [20], there is growing evidence that key evolutionary
processes occur beyond the consensus. In particular, extensive intra-host gene sequencing
has revealed the existence of minor viral subpopulations within individual hosts that
are not detected by consensus sequencing and that are sometimes of great phenotypic
importance [38],[39]. Given the intrinsically high mutation rates of RNA viruses,
as well as the immense size of intra-host populations, such extensive genetic and
phenotypic diversity is only to be expected.
A full description of the extent and structure of intra-host viral genetic variation
is critical for understanding evolutionary dynamics, informing on such issues as the
frequency of mixed infection, and hence the degree and extent of cross-immunity; the
frequency with which antigenic variants are produced and whether antigenic evolution
can occur on the time scale of individual infections; and the size of the population
bottleneck that might accompany inter-host transmission. As a case in point, it is
commonly assumed that viruses experience a severe population bottleneck as they are
transmitted to new hosts, a phenomenon that greatly restricts the power of natural
selection to fix advantageous mutations. Although this assumption appears to be true
in some cases [40], whether this is a general property of RNA viruses is unclear;
the evidence that multiple viral lineages can be transmitted among hosts argues against
a narrow bottleneck in all cases [41]. To more accurately determine the size of the
transmission bottleneck, analyses of intra-host genetic diversity along known transmission
chains will be essential. On a larger scale, it is unclear whether phylodynamic patterns
differ within and among hosts, and whether any differences among these scales of analysis
are qualitative or quantitative.
Intra-host sequence data are also essential for understanding the process of cross-species
virus transmission and emergence. Key parameters in determining whether a virus will
adapt successfully to a new host species include the extent of intra-host genetic
diversity, the fitness distribution of the mutations produced, and how many of these
mutations will assist adaptation to new host species [41]–[43]. No such data are available
for any acute RNA virus, so testing models for viral emergence is difficult. We believe,
however, that understanding the mechanics of this adaptive process is at least as
important as surveying for new emerging viruses.
Challenges for the Future
Our discussion has highlighted a number of key challenges for a successful phylodynamic
research agenda. These challenges comprise data, theory, and methodological issues,
and are briefly summarized as follows. First, with respect to data, it is clear that
more genome sequences must be acquired and with increased temporal and spatial precision.
For example, wherever possible, GenBank records should contain the exact day and precise
latitude and longitude of sampling. In addition, it is essential that these sequence
data be linked with the relevant metadata, such as the associated clinical syndrome
and (if applicable) measure of antigenicity. Similarly, it is essential that equivalent
genome sequence data be acquired from multiple time points within individual hosts.
Second, in terms of theory, it is crucial that we fully integrate patterns of viral
evolution across multiple epidemiological scales, from within hosts, to local outbreaks,
and on to global pandemics. Although the coalescent is hugely useful in this respect,
it is essential that its theoretical framework be extended to incorporate models of
population growth and decline that most accurately reflect the population dynamics
of acute RNA viruses, in particular the dynamics of the susceptible “denominator”
that fuels epidemics. Sequencing of all available samples from the UK 2001 FMD epidemic
would yield great scientific dividends here. Third and finally, with respect to methodology,
new computational tools are needed to rapidly make phylodynamic inferences from genomic
datasets that may contain thousands of sequences and that efficiently integrate genomic
with other forms of biological data. We hope this review will stimulate research in
all these areas.