The antigenic landscape of human influenza N2 neuraminidases from 2009 until 2017

Reviewer #1 (Public Review):

Summary
The authors investigated the antigenic diversity of recent (2009- 2017) A/H3N2 influenza neuraminidases (NAs), the second major antigenic protein after haemagglutinin. They used 27 viruses and 43 ferret sera and performed NA inhibition. This work was supported by a subset of mouse sera. Clustering analysis determined 4 antigenic clusters, mostly in concordance with the genetic groupings. Association analysis was used to estimate important amino acid positions, which were shown to be more likely close to the catalytic site. Antigenic distances were calculated and a random forest model was used to determine potential important sites.

This has the potential to be a very interesting piece of work. At present, there are inconsistencies in the methods, results and presentation that limit its impact. In particular, there are weaknesses in some of the computational work.

Strengths
1. The data cover recent NA evolution and a substantial number (43) of ferret (and mouse) sera were generated and titrated against 27 viruses. This is laborious experimental work and is the largest publicly available neuraminidase inhibition dataset that I am aware of. As such, it will prove a useful resource for the influenza community.

2. A variety of computational methods were used to analyse the data, which give a rounded picture of the antigenic and genetic relationships and link between sequence, structure and phenotype.

Weaknesses
1. Inconsistency in experimental methods
Two ferret sera were boosted with H1N2, while recombinant NA protein for the others. This, and the underlying reason, are clearly explained in the manuscript. The authors note that boosting with live virus did not increase titres. Nevertheless, these results are included in the analysis when it would be better to exclude them (Figure 2 shows much lower titres to their own group than other sera).

2. Inconsistency in experimental results
Clustering of the NA inhibition results identifies three viruses which do not cluster with their phylogenetic group. Again this is clearly pointed out in the paper. Further investigation of this inconsistency is required to determine whether this has a genetic basis or is an experimental issue. It is difficult to trust the remaining data while this issue is unresolved.

3. Inconsistency in group labelling
A/Hatay/4990/2016 & A/New Caledonia/23/2016 are in phylogenetic group 1 in Figure 2 and phylogenetic group 1 in Figure 5 - figure supplement 1 panel a.
A/Kansas/14/2017 is selected as a representative of antigenic group 2, when in Figure 2 it is labelled as AC1 (although Figure 2 - supplement 4 which the text is referring to shows data for A/Singapore/Infimh-16-0019/2016 as the representative of AC2). A/Kansas/14/2017 is coloured and labelled as AC2 in Figure 2 - supplement 5.
The colouring is changed for Figure 3a at the bottom. A/Heilongjiang-Xiangyang/1134/2011 is coloured the same as AC4 viruses when it is AC1 in Figure 2.
This lack of consistency makes the figures misleading.

4. Data not presented, without explanation
The paper states that 44 sera and 27 H6N2 viruses were used (line 158). However, the results for the Kansas/14/2017 sera do not appear to be presented in any of the figures (e.g. Figure 2 phylogenetic tree, Figure 5 - figure supplement 1). It is not obvious why these data were not presented. The exclusion of this serum could affect the results as often the homologous titre is the highest and several heatmaps show the fold down from the highest titre.

5. The cMDS plot does not have sufficient quality assurance
A cMDS plot is shown in Figure 5 - figure supplement 1, generated using classical MDS. The following support for the appropriateness of this visualisation is not given.
a. Goodness of fit of the cMDS projection, including per point and per titre.
b. Testing of the appropriate number of dimensions (the two sera from phylogenetic group 3 are clustered with phylogenetic group 2; additional dimensions might separate these groups).
c. A measure of uncertainty in positioning, e.g. bootstrapping.
d. A sensitivity analysis of the assumption about titres below the level of detection (i.e. that <20 = 10).
Without this information, it is difficult to judge if the projection is reliable.

6. Choice of antigenic distance measure
The measure of antigenic distance used here is the average difference between titres for two sera. This is dependent on which viruses have been included in the analysis and will be biased by the unbalanced number of viruses in the different clusters (12, 8, 2, 5).

7. Association analysis does not account for correlations
For each H6N2 virus and position, significance was calculated by comparing the titres between sera that did or did not have a change at that position. This does not take into account the correlations between positions. For haemagglutinin, it can be impossible to determine the true antigenic effects of such correlated substitutions with mutagenesis studies.

8. Random forest method
25 features are used to classify 43 sera, which seems high (p/3 is typical for classification). By only considering mismatches, rather than the specific amino acid changes, some signals may be lost (for example, at a given position, one amino acid change might be neutral while another has a large antigenic effect). Features may be highly, or perfectly correlated, which will give them a lower reported importance and skew the results.

Reviewer #2 (Public Review):

Summary:
The authors characterized the antigenicity of N2 protein of 44 selected A(H3N2) influenza A viruses isolated from 2009-2017 using ferret and mice immune sera. Four antigenic groups were identified, which correlated with their respective phylogenic/ genetic groups. Among 102 amino acids differed by the 44 selected N2 proteins, the authors identified residues that differentiate the antigenicity of the four groups and constructed a machine-learning model that provides antigenic distance estimation. Three recent A(H3N2) vaccine strains were tested in the model but there was no experimental data to confirm the model prediction results.

Strengths:
This study used N2 protein of 44 selected A(H3N2) influenza A viruses isolated from 2009-2017 and generated corresponding panels of ferret and mouse sera to react with the selected strains. The amount of experimental data for N2 antigenicity characterization is large enough for model building.

Weaknesses:
The main weakness is that the strategy of selecting 44 A(H3N2) viruses from 2009-2017 was not explained. It is not clear if they represent the overall genetic diversity of human A(H3N2) viruses circulating during this time. A comprehensive N2 phylogenetic tree of human A(H3N2) viruses from 2009-2017, with the selected 44 strains labeled in the tree, would be helpful to assess the representativeness of the strains included in the study. The second weakness is the use of double-immune ferret sera (post-infection plus immunization with recombinant NA protein) or mouse sera (immunized twice with recombinant NA protein) to characterize the antigenicity of the selected A(H3N2) viruses. Conventionally, NA antigenicity is characterized using ferret sera after a single infection. Repeated influenza exposure in ferrets has been shown to enhance antibody binding affinity and may affect the cross-reactivity to heterologous strains (PMID: 29672713). The increased cross-reactivity is supported by the NAI titers shown in Table S3, as many of the double immune ferret sera showed the highest reactivity not against its own homologous virus but to heterologous strains. Although the authors used the post-infection ferret sera to characterize 5 viruses (Figure 2, Figure Supplement 4), the patterns did not correlate well. If the authors repeat the NA antigenic analysis using the post-infection ferret sera with lower cross-reactivity, will the authors be able to identify more antigenic groups instead of 4 groups? Another weakness is that the authors used the newly constructed model to predict the antigenic distance of three recent A(H3N2) viruses but there is no experimental data to validate their prediction (eg. if these viruses are indeed antigenically deviating from group 2 strains as concluded by the authors).

Reviewer #3 (Public Review):

Summary:
This paper by Portela Catani et al examines the antigenic relationships (measured using monotypic ferret and mouse sera) across a panel of N2 genes from the past 14 years, along with the underlying sequence differences and phylogenetic relationships. This is a highly significant topic given the recent increased appreciation of the importance of NA as a vaccine target, and the relative lack of information about NA antigenic evolution compared with what is known about HA. Thus, these data will be of interest to those studying the antigenic evolution of influenza viruses. The methods used are generally quite sound, though there are a few addressable concerns that limit the confidence with which conclusions can be drawn from the data/analyses.

Strengths:
- The significance of the work, and the (general) soundness of the methods.
- Explicit comparison of results obtained with mouse and ferret sera.

Weaknesses:
- Approach for assessing the influence of individual polymorphisms on antigenicity does not account for the potential effects of epistasis.
- Machine learning analyses were neither experimentally validated nor shown to be better than simple, phylogenetic-based inference.