Permalink
Browse files

Change references to new figure names in response.

  • Loading branch information...
evogytis committed Dec 12, 2017
1 parent 5d0e1eb commit b1fe9abbd633222342f7850ec01a494812e2ca9b
Showing with 8 additions and 8 deletions.
  1. +8 −8 response_1/response_1.tex
View
@@ -75,10 +75,10 @@ \section*{Reviewer responses}
\section*{Essential revisions}
1. The population genetic model (the particular form of structured coalescent) is highly idealised and this may influence the quantitative conclusions, although we suspect the conclusions are quite robust qualitatively. This model specifically estimates the rate of a lineage moving between demes going backwards in time; the numbers cited for the camel$\rightarrow$human rate is really the rate that a lineage in humans goes to a camel going down the tree. The relationship between these migration rates and the epidemiologically meaningful transmission rate is complex and depends among other things on the ratio of population sizes in both demes. Per-capita transmission rates could be estimated using an epidemiologically structured coalescent model (see e.g.\ papers by Volz \& Rasmussen), which would ideally be stochastic due to bursty dynamics in humans. But this would be a large undertaking and so we suggest that for now the distinction is clarified. Overall, a little more discussion of the complexity and pitfalls when relating idealised population genetic models (like the island model used here) to a noisy nonlinear epidemic like this one might be merited.
\textbf{Yes, we agree that the structured coalescent approach is idealised and does not reflect a meaningful rate of zoonotic transfer of lineages, which is the reason we restricted any mention of rates to supplementary figures and do not attach numbers whenever rates are mentioned, but still report the number of introductions observed in the sequence data. We have altered Figure S2 in the reviewed manuscript to reflect that the rates shown are backwards in time. We have added the following sentences to the discussion to highlight the fact that the coalescent model employed is not ideal:}
\textbf{Yes, we agree that the structured coalescent approach is idealised and does not reflect a meaningful rate of zoonotic transfer of lineages, which is the reason we restricted any mention of rates to supplementary figures and do not attach numbers whenever rates are mentioned, but still report the number of introductions observed in the sequence data. We have altered Figure S2 (now Figure 1-supplement 2) in the reviewed manuscript to reflect that the rates shown are backwards in time. We have added the following sentences to the discussion to highlight the fact that the coalescent model employed is not ideal:}
\begin{quotation}
Although we recover migration rates from our model (Figure S2), these only pertain to sequences and in no way reflect the epidemiologically relevant \textit{per capita} rates of zoonotic spillover events.
Although we recover migration rates from our model (Figure 1-supplement 2), these only pertain to sequences and in no way reflect the epidemiologically relevant \textit{per capita} rates of zoonotic spillover events.
%\end{quotation}
[...]
@@ -94,20 +94,20 @@ \section*{Essential revisions}
% \end{quotation}
2. p4, `Our analyses recover these results despite sequence data heavily skewed towards non-uniformly sampled human cases and are robust to choice of prior'. This is a quite nice result and raises the question if skewed sampling would bias estimates if using a substitution model approach (`discrete trait analysis', DTA). It would strengthen the paper to include a comparison of the structured coalescent estimates to another method for ancestral states; the most popular approach in beast has been substitution models (DTA). These may give divergent results because of skewed sampling. It would be rather easy for the authors to run a DTA and if biased, this would serve as a good cautionary example when sampling is highly skewed towards one deme.
\textbf{An excellent suggestion, thank you. We have run this analysis and include the results as a new supplementary figure S3. As expected, the skewed sampling results in a reconstruction of ancestral states that puts humans as the source of most MERS-CoV lineages in camels. We have added the appropriate description of methods as well as the following paragraph in results:}
\textbf{An excellent suggestion, thank you. We have run this analysis and include the results as a new figure supplement (Figure 1-supplement 3). As expected, the skewed sampling results in a reconstruction of ancestral states that puts humans as the source of most MERS-CoV lineages in camels. We have added the appropriate description of methods as well as the following paragraph in results:}
\begin{quotation}
Our findings suggest that instances of human infection with MERS-CoV are more common than currently thought, with exceedingly short transmission chains mostly limited to primary cases that might be mild and ultimately not detected by surveillance or sequencing.
Structured coalescent analyses recover the camel-centered picture of MERS-CoV evolution despite sequence data heavily skewed towards non-uniformly sampled human cases and are robust to choice of prior.
Comparing these results with a currently standard discrete trait analysis (Lemey \textit{et al.}, 2009) approach for ancestral state reconstruction shows dramatic differences in host reconstruction at internal nodes (Figure S3).
Comparing these results with a currently standard discrete trait analysis (Lemey \textit{et al.}, 2009) approach for ancestral state reconstruction shows dramatic differences in host reconstruction at internal nodes (Figure 1-supplement 3).
Discrete trait analysis reconstruction identifies both camels and humans as important hosts for MERS-CoV persistence, but with humans as the ultimate source of camel infections.
A similar approach has been attempted previously (Zhang \textit{et al.}, 2016), but this interpretation of MERS-CoV evolution disagrees with lack of continuing human transmission chains outside of Arabian peninsula, low seroprevalence in humans and very high seroprevalence in camels across Saudi Arabia.
We suspect that this particular discrete trait analysis reconstruction is false due to biased data, \textit{i.e.} having nearly twice as many MERS-CoV sequences from humans ($n=174$) than from camels ($n=100$) and the inability of the model to account for and quantify vastly different rates of coalescence in the phylogenetic vicinity of both types of sequences.
\end{quotation}
3. A comparison to ML tree reconstruction could potentially be illuminating. We think you could be clearer about what drives the results in your paper. It is unusual for a phylogenetic ancestral reconstruction, that the results seem to be determined as much by the coalescent assumptions as by the tree topology. The two-patch model had a much higher coalescent rate in the human deme than in the camel deme -- so long branches are only really possible in the camel deme. This may be why for example, staring at the top clade of figure 1, one can see camel ancestry to a whole bunch of human sequences that are not topologically separated by camel sequences. If this is correct, these results may not necessarily be wrong, but it made us slightly uncomfortable that the results are driven by the coalescent model, not the tree topology. Please elaborate, either correcting us, or explaining better. A simple test of this hypothesis would be that an ML ancestral reconstruction on the ML tree would not give the same clusters. I don't think that would make the ML result correct, but it might be an enlightening comparison. Or you may prefer another way to address this.
\textbf{A very good point. We share the suspicion that the results are largely driven by contrasts in effective population sizes between demes. In addition to the requested maximum likelihood phylogeny (supplementary Figure S14) we also ran a structured coalescent analysis where deme sizes are enforced to be the same (supplementary Figure S4). This model fails in a similar way to a DTA reconstruction shown in the new Figure S3. We now explain how the structured coalescent arrives at the tree shown in Figure 1 in the discussion:}
\textbf{A very good point. We share the suspicion that the results are largely driven by contrasts in effective population sizes between demes. In addition to the requested maximum likelihood phylogeny (now Figure 1-supplement 5) we also ran a structured coalescent analysis where deme sizes are enforced to be the same (now Figure 1-supplement 4). This model fails in a similar way to a DTA reconstruction shown in the Figure 1-supplement 3. We now explain how the structured coalescent arrives at the tree shown in Figure 1 in the discussion:}
\begin{quotation}
When allowed different deme-specific population sizes, the structured coalescent model succeeds because a large proportion of human sequences fall into tightly connected clusters, which informs a low estimate for the population size of the human deme.
@@ -198,7 +198,7 @@ \section*{Reviewer \#1}
6.1. Dispersion parameter is set at 0.1 based on a previous branching process study, however there's not great support for re-using the value for a different epidemic. Was an effort made to estimate this parameter?
\textbf{The dispersion parameter was indeed chosen from a previous study. We have run small scale analyses over a grid of three bias levels (same as the main analyses), 23 levels of $R_{0}$ (0.50--1.05), and five levels of dispersion (0.004, 0.02, 0.10, 0.50 and 1.0) with 2000 replicate simulations for each combination of variables. Numbers of simulations matching empirical data are shown in Figure S15. In this analysis, we find strongest support for dispersion $k$ = 0.1.}
\textbf{The dispersion parameter was indeed chosen from a previous study. We have run small scale analyses over a grid of three bias levels (same as the main analyses), 23 levels of $R_{0}$ (0.50--1.05), and five levels of dispersion (0.004, 0.02, 0.10, 0.50 and 1.0) with 2000 replicate simulations for each combination of variables. Numbers of simulations matching empirical data are shown in Figure 3-supplement 6. In this analysis, we find strongest support for dispersion $k$ = 0.1.}
6.2. The sampling model accounts for `sequencing capacity', which is good. But is it the case that sequencing probability varied over the course of the epidemic?
@@ -223,7 +223,7 @@ \section*{Reviewer \#1}
6.4. Can you show a comparison of the empirical distribution of cluster sizes and the fitted distribution? A QQ plot? Are there any aspects of the empirical distribution that are not well characterised by the branching process model, eg at the tails?
\textbf{We have generated QQ plots of sequence cluster size distribution observed across the posterior distribution of structured coalescent trees (picked at random) against matching sequence cluster size distributions from the primary Monte Carlo simulation with bias parameter set to 2.0. Simulated sequence clusters tend to exhibit a left skew near the middle of the distribution, but not so much near the tails. We have added figure S9 to show this.}
\textbf{We have generated QQ plots of sequence cluster size distribution observed across the posterior distribution of structured coalescent trees (picked at random) against matching sequence cluster size distributions from the primary Monte Carlo simulation with bias parameter set to 2.0. Simulated sequence clusters tend to exhibit a left skew near the middle of the distribution, but not so much near the tails. We have added Figure 3-supplement 5 to show this.}
\section*{Reviewer \#2}
@@ -305,7 +305,7 @@ \section*{Reviewer \#3}
8. Discussion. I think the results of this analysis on $R_0$ and cluster size distribution could be compared with Cauchemez et al.\ PNAS 2016, who estimated many of the similar quantities using spatiotemporal data to define clusters. An important result of that paper was that the cluster size distributions were very skew - I think it would be good to look at skewness here.
\textbf{An excellent suggestion, however the authors of the Cauchemez \textit{et al}. study did not share any data beyond the raw input file they used. Since neither the actual numbers in Fig 1C of Cauchemez \textit{et al}., nor an implementation of the MCMC algorithm exist in the public domain we have included an additional figure (supplementary Figure S8) that visualises the distribution of cases recovered by simulation. These exhibit a heavy left skew, much like the distribution in Cauchemez \textit{et al}.}
\textbf{An excellent suggestion, however the authors of the Cauchemez \textit{et al}. study did not share any data beyond the raw input file they used. Since neither the actual numbers in Fig 1C of Cauchemez \textit{et al}., nor an implementation of the MCMC algorithm exist in the public domain we have included an additional figure (Figure 3-supplement 4) that visualises the distribution of cases recovered by simulation. These exhibit a heavy left skew, much like the distribution in Cauchemez \textit{et al}.}
% \bibliography{./../mers-structure}

0 comments on commit b1fe9ab

Please sign in to comment.