New approaches for unravelling reassortment pathways

Abstract

Background: Every year the human population encounters epidemic outbreaks of influenza, and history reveals recurring pandemics that have had devastating consequences. The current work focuses on the development of a robust algorithm for detecting influenza strains that have a composite genomic architecture. These influenza subtypes can be generated through a reassortment process, whereby a virus can inherit gene segments from two different types of influenza particles during replication. Reassortant strains are often not immediately recognised by the adaptive immune system of the hosts and hence may be the source of pandemic outbreaks. Owing to their importance in public health and their infectious ability, it is essential to identify reassortant influenza strains in order to understand the evolution of this virus and describe reassortment pathways that may be biased towards particular viral segments. Phylogenetic methods have been used traditionally to identify reassortant viruses. In many studies up to now, the assumption has been that if two phylogenetic trees differ, it is because reassortment has caused them to be different. While phylogenetic incongruence may be caused by real differences in evolutionary history, it can also be the result of phylogenetic error. Therefore, we wish to develop a method for distinguishing between topological inconsistency that is due to confounding effects and topological inconsistency that is due to reassortment.

Results: The current work describes the implementation of two approaches for robustly identifying reassortment events. The algorithms rest on the idea of significance of difference between phylogenetic trees or phylogenetic tree sets, and subtree pruning and regrafting operations, which mimic the effect of reassortment on tree topologies. The first method is based on a maximum likelihood (ML) framework (MLreassort) and the second implements a Bayesian approach (Breassort) for reassortment detection. We focus on reassortment events that are found by both methods. We test both methods on a simulated dataset and on a small collection of real viral data isolated in Hong Kong in 1999.

Conclusions: The nature of segmented viral genomes present many challenges with respect to disease. The algorithms developed here can effectively identify reassortment events in small viral datasets and can be applied not only to influenza but also to other segmented viruses. Owing to computational demands of comparing tree topologies, further development in this area is necessary to allow their application to larger datasets.

Figure 1

Simulated data with manually introduced SPR modifications. The HA data was simulated on a tree modified by moving taxon ‘G’ to group with taxon ‘B’. (A) Output from the ML analysis for seven segments: MP, NA, NP, NS, PA, PB1 and PB2. A significant SPR was detected that would require moving taxon ‘G’ to group with taxon ‘B’, as suggested by the HA segment (direction of arc from empty to filled circle). Colours of arcs correspond to specific SPR operations. (B) HA tree: seven segments propose a significant SPR modification on the HA tree that would require moving taxon ‘G’ to group with taxon ‘F’. (C) Frequency network from Bayesian results. Edges point from segment proposing an SPR, to the segment whose tree needs to be modified (filled circle). Legend shows SPRs corresponding to the coloured edges. HA proposes moving taxon ‘G’ to group with taxon ‘B’ for the other seven segments. Conversely, the rest of the segments suggest that ‘G’ should move to cluster with ‘F’ on the HA tree. (D), (E) Overlap between MLreassort and Breassort. The x-axis represents the segments that propose the SPR move, whereas the y-axis represents the segments whose trees need to be modified according to that SPR. The name of a tree segment is greyed out in the case where the SPR move is irrelevant, i.e. when the taxa involved in the move are sister-taxa.

Figure 2

Robinson-Foulds distances between trees of the Hong Kong dataset. The intensity of the squares corresponds to the degree of distance. Distances range from 0 to 6, representing the number of bipartitions present in one tree but not in the other. Some trees have the same topology (NP and PB1, NS and PB2) whereas the NA tree seems to be most distant to the other trees (distance of 6).

Figure 3

Trees NS and NA, confidence intervals and SPR modifications. Subtree pruning and regrafting (SPR) modifications that the NA tree proposes on the NS tree, and the confidence interval around each tree (coloured shapes). Three paths are possible. The labels on the arrows refer to nodes involved in a move: m1 - move outgroup to cluster with hk1774, m2 – move hk1073 to cluster with hk1774, m2r – reverse of m2, m3 – move hk1073 to cluster with outgroup, m4 – move env99 to cluster with quail99/sh39/hk1073 group. t1 - t6 are trees resulting from applying these SPR modifications to the NS tree. Arrows between two trees in the same confidence interval (CI) reflect trivial differences (e.g. m1, black arrow), whereas ones between trees from different CIs are considered significant (e.g. m2, red arrow). We consider m2 as significant as we’re interested in the minimum amount of significant branch moves between NS and NA.

Figure 4

Shortest edit path between NS and NA. Trees depicting the m1, m2 SPR edits from Figure 3. Arc direction is from empty to filled circle. NS (t0) is the start tree (−ln L = 2761.72), on which we apply modification m1 (moving A/redknot/NJ/325/1989 H7N7 to A/HongKong/1774/99 H3N2). The log likelihood of the intermediate tree (t1, -ln L = 2761.87) is close to that of NS (this tree is inside NS’ CI). The second modification, m2, is applied to t1 and results in a topology identical to that of the NA tree (t2, -ln L = 2859.68). The rise in the negative log likelihood of the tree after m2 is 97.81 (the resulting tree is outside NS’ CI).

Figure 5

Network of the most frequent SPRs from Breassort. Each node represents the set of trees for the corresponding segment. Edge colours correspond to different types of SPR operations, as shown. Edges point from a segment that proposes the branch swap, to the one that needs to be modified (ending in filled circle). For example, the orange edge going from NA to NS depicts the following operation: cutting the branch leading to A/HongKong/1073/99 H9N2 and reconnecting it to the branch ending in A/HongKong/1774/99 H3N2. The NS tree is the one being modified, and the NA tree proposes this modification.

Figure 6

Combined SPRs from MLreassort and Breassort. Each plot depicts an SPR move, with symbols indicating the cases when this move is significant. Circles represent results from the maximum likelihood-based approach, while crosses represent results from applying the algorithm based on a Bayesian framework. The x-axis shows the segments that propose the specified SPR, while the y-axis shows the segments whose trees need to be modified. The name of a tree segment is greyed out in the case where the SPR move is irrelevant, i.e. when the taxa involved in the move are sister-taxa. For example, moving hk1073 to group with hk1774 in the NA tree is irrelevant, as the NA tree already has these grouping together.