A goal of genomic epidemiology is to infer epidemiological and emergence dynamics from virus genome sequences obtained over short epidemic timescales1. Rapid in situ sequence generation and phylogenetic inference is based on the detection of genetic changes in pathogen sequences. However, during outbreaks there are many unknowns. The outbreak of coronavirus disease 2019 (COVID-19) which originated in Wuhan, China, was reported in December 2019 (ref. 2). By January 2020, the genome of the causative novel coronavirus, named severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), had been sequenced and made publicly available2. Virus sequences have underpinned the development of diagnostics and vaccines and have been used to assess patterns of transmission and spread. Although sequence data were used to answer crucial epidemiological questions during the Ebola and Zika outbreaks3,4, the pace of SARS CoV-2 genome data generation is unprecedented and is informing public health policy in real time.

Importantly, it is not only sequences that inform phylogenies; multiple factors contribute to the outputs including model assumptions, sampling density, the timing of sample collection, the portion of the viral genome sequenced, quality of sequencing data and the mutation rate of the virus itself. Although it is important to extract as much information as possible from sequence data as outbreaks unfold, it is imperative to bear in mind that the historical relationships of strains (phylogenies) are hypotheses that can be challenged as more data become available. Here, we highlight some of the challenges of genomic epidemiology during outbreaks such as SARS-CoV-2 and advise that the interpretation of findings from phylogenies needs to assess all epidemiological and supporting information, and consider sources of bias.

During outbreaks, we want to know if cases are linked and if this implies transmission. Most viruses can be separated into strains and, if two infections are caused by dissimilar strains, one can rule out transmission. The often-forgotten point is that phylogenies can rule out transmission, but if infections are caused by the same strains or identical viruses it does not definitively prove transmission. During an emerging outbreak when pathogens have not yet diverged into different strains, phylogenetic information is too weak to hypothesize transmission linkage — which, in turn, can be used for geographic inference, as even if the phylogenetic information is stronger, the same phylogeny is consistent with multiple transmission histories and there may be missing links due to incomplete sampling5. Consequently, we need to combine phylogenetic findings with epidemiological and supporting information such as environmental factors and human air-travel data before we draw any immediate conclusions regarding transmission. This was the case with Zika virus in Africa, where epidemiological, human mobility and climatic data supported the phylogenetic hypothesis that the outbreak was likely imported from Brazil6.

In the first stage of an outbreak, we can use phylogenetics to discern possible zoonotic sources, as in the case of the 2018 Lassa fever virus outbreak where phylogenetic patterns indicated independent spillover events from rodent hosts7. The crucial observation was that the correct identification of the source of zoonotic transmission relies on the availability of viral genome sequences from potential animal reservoirs. If the source of any virus has not been sampled it cannot be inferred, because phylogenetic linkage alone does not prove it. This limited knowledge of viral abundance from potential animal reservoirs is the reason for the uncertainty surrounding the zoonotic source of SARS-CoV-2 (ref. 8). The generation of additional viral genome sequences from an outbreak, coupled with virus-specific and epidemiological knowledge, provides insight into whether or not multiple ‘jumps’ occurred from a reservoir that might warrant appropriate control measures. Identical or near-identical virus genomes are expected from early transmission chains if a single spillover occurred recently, unless multiple zoonoses originated from the same low-genetic-diversity virus pool. In contrast, higher diversity in the early stage of human-to-human transmission is expected if multiple zoonoses have occurred or if there is significant within-host evolution9.

Geographical inferences (where and when) are feasible as more representative viral genome data — in temporal and spatial scales — become available. In order to date epidemiological events, we can hypothesize the location of common ancestors using ancestral reconstruction methods and infer phylogenies scaled to time. Such analyses require a molecular clock, which models the rate at which mutations accumulate with time and how this varies across the branches of a phylogeny. However, during the early stages of an outbreak there may not be sufficient signal to accurately estimate the clock rate. If this is the case, then it may be appropriate to apply an estimate from another closely related virus10. If temporal signal is present and a clock rate can be estimated, results need to be reported as credible intervals (instead of point estimates) to account for uncertainty in both the data (as incomplete, biased or improper sampling can lead to misleading phylogenies) and many aspects of the methods.

When investigating the dissemination of an emerging virus, the number of sequenced viral genomes may not be representative. Even as the outbreak unfolds and more genomes are obtained, they only represent a snapshot of the underlying genetic diversity. If phylogenies alone are considered, we cannot conclusively assert the geographical origins of the virus — or the extent of community transmission — as we are unable to distinguish between local transmission events and multiple introductions of genetically similar viruses from geographically distinct sources if one aspect has not been sampled. In this way, uneven sampling can also lead to misleading conclusions on the geographical source, the number of introductions and the size and duration of local transmission chains11. The significance of these associations is harder to ascertain when the phylogeny is reported without any assessment on the reliability of internal branches. Therefore, phylogenetic interpretation from ongoing outbreaks (for example, SARS-CoV-2) needs to be performed in the context of all available information such as temporal and spatial distribution of cases and travel patterns, and any evidence of epidemiological linkage, sampling uncertainty and other sources of bias need to be carefully considered and reported.

The methods for valid phylogenetic inference require multiple assumptions which are unlikely to be met during emerging outbreaks. Examples include, but are not limited to: adequate phylogenetic signal, which is low when strains have not yet diverged; geographical representation and effective sampling time points with sufficient molecular clock signal, which only become feasible as the epidemic unfolds; and random mixing, which may be violated under certain circumstances, for instance when mitigation strategies are set in place. Estimates from phylogenies may be sensitive to one or more of these assumptions and conclusions need to be made and shared with caution. Another essential consideration during an epidemic is accurate rooting of the phylogeny as it determines the direction of transmission over time12.

There are also genome features that are intrinsic to the biology of the virus that may impact the extent and applicability of phylogenetics during outbreaks. For instance, the presence of recombination or reassortment and low diversity (due to the virus’ rate of evolution, selective constraints and transmission bottlenecks) complicate the resolution of phylogenetic relationships, but the incorporation of within-host viral diversity may provide greater resolution in understanding transmission dynamics13. Moreover, some mutations in the viral genome sequence can be due to the error rate of the sequencing technology, recurrent sequencing issues, hypermutability or contamination — issues which warrant caution with interpretations and especially with those concerning selection and recombination.

Genomic epidemiology has supported public health outbreak responses. Indeed, the ability to exploit viral genome sequences has allowed us to characterise early patterns of SARS-CoV-2 transmission in China, New Zealand and Australia14,15. In the midst of an outbreak, sharing data is both necessary and important for an effective response, but sharing the associated metadata is also necessary to aid interpretations (such as how representative the data are of the country-wide situation) and to avoid creating sampling bias by researchers that are not doing the sequencing themselves.

The emergence of SARS-CoV-2 has presented a series of challenges about how we reliably extract information from phylogenies to gain insights into virus transmission and spread, and how we responsibly present our findings. Owing to low genetic diversity and uneven sampling, several controversial hypotheses have already been put forward. One cautionary tale involves how an outbreak in Bavaria seeded the epidemic in northern Italy and the subsequent wider outbreak in Europe. This notion was based on a small sample of very similar sequences. However, it overlooked a more likely scenario in which this virus was already circulating in China and that European regions had multiple introductions from China. At this early stage, conclusions about the impacts of mutations on transmission and disease (for example, a D614G mutation in the spike protein16) should not be made on the basis of phylogenies alone, but with separate evidence supporting not only a phenotypic difference but the resulting consequences for epidemiology.

The SARS-CoV-2 pandemic has highlighted the importance of providing a comprehensive rationale for any conclusions about the spatiotemporal dispersal of the virus. Phylogenies represent hypotheses that encompass different sources of error and this uncertainty needs to be visualized and communicated far more transparently. Another challenge is how we facilitate the dissemination of metadata and integrate this with phylogenetic trees. Incorporating host characteristics (such as age, onset date and exposure history) to aid phylogenetic interpretation would undoubtedly result in more reliable inferences.

Now more than ever, careful reporting of phylogenetic interpretations while safeguarding the privacy of infected individuals would ensure that both policy makers and the public have the best possible information during an outbreak. Failure to balance these issues could jeopardise both scientific integrity and public confidence in the field of genomic epidemiology.