Mapping of H3N2 Influenza Antigenic Evolution in China Reveals a Strategy for Vaccine

Transcription

1 Supplementary Information Mapping of H3N2 Influenza Antigenic Evolution in China Reveals a Strategy for Vaccine Strain Recommendation Xiangjun Du, Libo Dong, Yu Lan, Yousong Peng, Aiping Wu, Ye Zhang, Weijuan Huang, Dayan Wang, Min Wang, Prof Yuanji Guo, Yuelong Shu, Taijiao Jiang 1

2 Supplementary Figure S1. Comparison of antigenic patterns of influenza A (H3N2) viruses between China, the USA and Europe during Laboratory-confirmed H3N2 infections for each month, normalized by total infections within one season time period (6 months before and after), are shown month by month (background lattices). Winter seasons for the Northern Hemisphere are indicated by grey backgrounds. 2

3 Supplementary Figure S2. The influenza transmission patterns within mainland. China can be divided into Northern China (Temperate regions) and Southern China (Tropic regions) according to Qin Mountain and Huai River. Antigenic clusters are extracted from sequence dataset with month information in China. X label is arranged as calendar time, which is only for the periods with sufficient data with detailed isolation month information. Y label is the laboratory confirmed H3N2 infections for each month normalized by total infections within one season time period (6 months each before and after). Winter seasons for the Northern Hemisphere are indicated by grey backgrounds. 3

4 Supplementary Figure S3. Comparison of antigenic patterns of influenza A (H3N2) viruses between Coastal China and Inland China during Laboratory-confirmed H3N2 infections for each month, normalized by total infections within one season time period (6 months before and after), are shown month by month (background lattices). Winter seasons for the Northern Hemisphere are indicated by grey backgrounds. 4

5 Supplementary Figure S4. Cluster size curve for ACnet clustering of viruses after 1990 from Smith data 25. Mean cluster size was calculated as the probability of two randomly chosen influenza viruses that lie in the same predicted antigenic cluster 19,61. The threshold in the beginning of the first plateau of the cluster size curve (indicated by an arrow) was used for identifying predicted antigenic clusters. 5

6 Supplementary Table S1. HI assays of influenza A (H3N2) viruses collected in mainland China from March to July Vaccine strain A/Brisbane/10/2007 A/Brisbane/10/ Recent Viruses Cluster Isolation Time HI Titer Fold differences compared to the vaccine strain* A/Tianjinnankai/155/2009 BR / A/Sichuanwenjiang/55/2009 BR / A/Shanxichengqu/1180/2009 BR / A/Jiangsugulou/134/2009 BR / A/Tianjinnankai/167/2009 BR / A/Jiangsuchongan/38/2009 BR / A/Shaanxiqindu/159/2009 BR / A/Sichuanyucheng/135/2009 BR / A/Fujiantongan/142/2009 BR / A/Shandongsifang/183/2009 BR / A/Shanghainanhui/171/2009 BR / A/Jiangsuchongan/314/2009 BR / A/Sichuananyue/121/2009 BR / A/Shanghainanhui/177/2009 BR / A/Shanghainanhui/184/2009 BR / A/Jiangsuquanshan/116/2009 BR / A/Fujianjinjiang/51/2009 BR / A/Jiangsuquanshan/114/2009 BR / A/Shanghainanhui/190/2009 BR / A/Jiangxidonghu/1100/2009 BR / A/Shanghainanhui/1106/2009 BR / A/Zhejianghaishu/1252/2009 BR / A/Yunnanchuxiong/1101/2009 BR / A/Tianjinjinnan/177/2009 BR / A/Zhejianghaishu/1234/2009 BR / A/Zhejianghaishu/1259/2009 BR / A/Yunnanchuxiong/190/2009 BR / A/Hebeiqiaoxi/53/2009 BR / A/Tianjinnankai/181/2009 BR / A/Shanghainanhui/1132/2009 BR / A/Fujianshishi/53/2009 BR / A/Hubeiwuchang/1140/2009 BR / A/Hunanwuling/128/2009 BR / A/Hunanshigu/129/2009 BR / A/Fujiantongan/196/2009** PE / A/Fujiansiming/1121/2009 PE / A/Fujianfengze/311/2009 PE / A/Fujianhuian/31/2009 PE / A/Fujianhuian/39/2009 PE / A/Fujianjinjiang/35/2009 PE / A/Guangdongmaonan/37/2009 PE / A/Guangdongmaonan/33/2009 PE /

7 A/Guangdongluohu/1256/2009 PE / A/Guangdongluohu/1257/2009 PE / A/Fujianxiangcheng/52/2009 PE / A/Guangdongmaonan/38/2009 PE / A/Jiangxizhanggong/31/2009 PE / A/Hunanfurong/1189/2009 PE / A/Fujiansiming/1254/2009 PE / A/Fujianfengze/316/2009 PE / A/Shanghainanhui/1208/2009 PE / A/Guangdongfutian/1253/2009 PE / *Fold differences compared to the vaccine strain were calculated as follows: Titer Titer vaccine vaccine virus vaccine where Titervaccine vaccine was calculated as the homologous HI titer of the vaccine strain (A/Brisbane/10/2007) and was 320 here, and Titervirus vaccine was calculated as the heterologous titer of the circulating isolate against the reference strain. **A/Fujiantonggan/196/2009 was selected as a candidate vaccine strain by the China CDC 7

8 Supplementary Table S2. Summary of 20 predicted antigenic clusters for influenza A (H3N2) viruses isolated from mainland China during ID Cluster* Circulation time Strain number World Health Organization (WHO) recommended vaccine strains Other candidate vaccine strains used by the China CDC Time for the First isolation** Time for the Last isolation** 0 PE A/Perth/16/2009 A/Fujiantongan/196/ /4 2009/8 1 BR A/Brisbane/10/ / /7 2 JX A/Jiangxidonghu/312/ /6 2007/4 3 CA A/California/7/2004 A/NewYork/55/2004 A/Jiangxi/424/2004 A/Wisconsin/67/2005 A/Yunnan/1145/ /3 4 FU SY After WU95 A/Hiroshima/52/2005 A/Fujian/411/2002 A/Kumamoto/102/2002 A/Wyoming/03/2003 A/Wellington/01/2004 A/Sydney/5/1997 A/Moscow/10/1999 A/Panama/2007/1999 A/Guangxi/194/2003 A/Fujian/445/ /4 A/Fujian/151/ /6 2002/ / WU A/Nanchang/933/1995 A/Wuhan/359/ /1 8 BE A/Beijing/32/1992 A/Shangdong/9/ A/Johannesbury/33/ BE A/Beijing/353/ SI87 A/Sichuan/2/ A/Shanghai/11/ A/Guizhou/54/ A/Philipines/2/ PH A/Christchurch/4/ A/Leningrad/360/ BK A/Bankok/1/ TX A/Texas/1/ VI A/Victoria/3/ EN A/England/42/1972 A/PortChalmers/1/ HK A/HongKong/1/ Minor*** Minor Minor * Clusters were named according to the first vaccine strain or candidate vaccine strain that was included in the cluster. ** Month information is given if have *** Minor clusters contained less than 50% of the viruses in any year. 8

9 Supplementary Table S3. Contribution of individual features or group of features in our antigenic relationship prediction model. Accuracy** Features* Use individual feature Epitope A 79.00% 83.65% Epitope B 70.67% 89.39% Epitope C 68.15% 89.48% Epitope D 68.15% 89.11% Epitope E 68.15% 89.54% Leave one feature out Hydrophobicity 68.15% 89.45% Volume 71.84% 89.08% Charge 75.37% 88.89% Accessible Surface Area 73.50% 88.97% Polarity 67.32% 89.55% Use feature group Leave one group out 85.15% 81.08% 79.73% 86.28% Receptor Binding 71.88% 88.50% 71.88% 88.48% Glycosylation 68.15% 89.34% 68.15% 89.34% * The model include all the features can achieve an accuracy of 89.54% ** The contribution was quantified as the 10-fold cross validation accuracy of the model constructed using each single feature for the training dataset. 9

10 Supplementary Table S4. Cutoff values obtained during discretisation for the 12 features. Features Discretisation cutoff * Epitope A 0 Epitope B 2 Epitope C 0 Epitope D 1 Epitope E 0 Hydrophobicity 1.82 Volume Charge 2.49 Accessible Surface Area Polarity 0.10 Receptor Binding 1.79 Glycosylation 1 * For receptor binding, changes less than the cutoff value are related to antigenic change; for the other features, changes more than the cutoff value are related to antigenic change. 10

11 Supplementary Table S5. Predicted antigenic clusters for the training data using either the odds ratio or the logarithm of the odds ratio as the edge weight. The seasons covered by the predicted antigenic cluster are also given. log(odds ratio) Odds ratio Season HK68* HK EN72 EN VI75 VI TX77 TX BK79 BK79(A/Philippines/2/1982**) BK79 BK79(A/Leningrad/360/1986) SI87 SI87(A/Sichuan/2/1987) SI87(A/Guizhou/54/1989) BE89 BE BE92(A/Beijing/32/1992) BE92 BE92(A/Shangdong/9/1993) BE92(A/Johannesbury/33/1994) WU95 WU SY97 SY97(A/Sydney/5/1997) SY97(A/Panama/2007/1999) FU02 FU * The predicted antigenic clusters were named according to the work of Smith et al 1. When the logarithm of the odds ratio (log (odds ratio) was used as the edge weight, the predicted antigenic clusters matched those antigenic clusters defined by Smith et al, and were named accordingly. ** WHO-recommended vaccine strains contained in Smith s antigenic clusters. When the odds ratio was used as the edge weight, some antigenic clusters identified in work of Smith et al were divided into 2 or more sub-clusters. However, we found these predicted sub-clusters actually corresponded to the vaccine strains used by the WHO (indicated in parentheses). 11

12 Supplementary Methods HA Sequence Data and Sequence Analysis 1071 influenza A (H3N2) viruses sampled by the China CDC from diverse and representative regions of mainland China between 1968 and 2009 were analyzed by sequencing HA1 (Supplementary Data 1). Detailed (monthly) sampling time information was available for 682 strains (Supplementary Data 1). All viruses were propagated in embryonated eggs or by kidney tissue culture. Viral RNA was extracted from clarified supernatants using Qiagen Rneasy Mini Kits. HA1 nucleotide fragments were amplified by reverse transcription-pcr using the oligonucleotide 5'-AGCAAAAGCAGG-3. The PCR products were purified and subjected to sequencing with oligonucleotides 5'- AGCAAAAGCAGGGGATAATT-3 and 5'-CCTGCGATTGCGCCGAAT-3. The HA1 sequences of influenza A (H3N2) viruses other than those determined in this study were downloaded from the National Center for Biotechnology Information Influenza Virus Resource 16 ( HA1 sequences shorter than 100 amino acids, and anomalies were excluded. In total, HA1 sequences for 8315 influenza A (H3N2) viruses were obtained from NCBI, including 5594 strains for which detailed information on the month of isolation was available. All HA1 sequences were aligned with ClustalW without insertions or deletions 42. Undetermined positions in the coding region of HA1 were then corrected based on the consensus sequence 20. Phylogenetic trees were constructed based on protein sequences using the Maximum Likelihood (ML) method of PhyML 43 version The model of amino acid substitution and frequency was estimated from the protein sequences using ModelGenerator 62. Additionally, the proportion of the invariable sites was also estimated and four substitution rate categories were used with the gamma distribution parameter estimated to account for variable substitution rates among sites. All other parameters were set to default. HI Assay The antigenicity of the viruses was tested with ferret post-infection sera by HI assay using 0.5% (v/v) turkey red blood cells (TRBC). The ferret sera were treated with receptor-destroying enzymes overnight. 4HAU of viruses 12

13 were added to the serial 2 fold dilution ferret sera from 1:10 to 2560 and reacted with each other for 20 minutes in the 96-well micro-plates. Then 0.5% TRBC were added to each well and the results were recorded when all the control TRBC settled after 30mins. The HI titer was recorded as the reciprocal of the last dilution of antiserum that completely inhibits hemagglutination. Performance Assessment of Feature-based Model The model performance was assessed using cross validation and retrospective testing on the training data. For cross validations, the training data were randomly divided into 10 parts: 9 parts were used to build the model described as above and the remaining one part served as the testing data. The retrospective testing is a test that is based on historical data to predict the future outcome. It could also be done on the training data. To perform the retrospective testing on the Smith data 25, the antigenic pairs of viruses isolated before a given season were used to train the models by following the procedure described above, and the remaining antigenic pairs for viruses isolated after the given season were used to test the performance of the trained models. The accuracy was calculated as the ratio of the number of correctly predicted antigenic pairs (including both antigenically similar and antigenically distinct pairs) to the total number of antigenic pairs. Identification of Antigenic Clusters by Using MCL program There are two parameters in the MCL program that can be tuned: one is the scheme parameter, and the other is the main inflation value. A high scheme value results in more expensive computations that may possibly be more accurate and we always used the highest value. The inflation value is the main handle for affecting cluster granularity. It is usually chosen somewhere in a range: a bigger value will tend to result in fine-grained clusterings, and a smaller value will tend to result in very coarse grained clusterings. We tested different values of the main inflation parameter and try to find the proper partition following the method proposed by Plotkin et al 19,61. In brief, changes in mean cluster sizes (the probability that two randomly chosen influenza viruses lie in the same cluster) were plotted as changes in the main inflation values. In the cluster size curve (see Supplementary Fig. S4 for Smith data after 1990), the first plateau represents coarse-grained predicted antigenic clusters, while the last plateau represents the fine-grained predicted antigenic clusters. For convenience, the main inflation value at the beginning of the first plateau was called the coarse-grained clustering value. In addition to the MCL parameter, the main inflation 13

14 value, the edge weighting of the ACnet also affects cluster size. Clusters identified from the ACnet using the same MCL parameter, and the odds ratio as the edge weight, were finer than those generated from the ACnet using the logarithm of the odds ratio as the edge weight (Supplementary Table S5). For example, when the odds ratio was used to characterize the overall antigenic patterns for the viruses used in Smith et al. s work 25, we found that four of the 11 antigenic clusters were further split into two or more sub-antigenic clusters, indicating that the odds ratio can provide a finer separation of antigenic clusters than the logarithm of the odds ratio (Supplementary Table S5). As can be seen from the above, the edge weight of the ACnet, and the clustering parameter MCL, both control the cluster size. In characterizing overall patterns of influenza antigenic evolution, we sought to predict coarse-grained antigenic clusters by using the logarithm of the odds ratio as the ACnet edge weight and the coarse-grained clustering value in the MCL program. To assess the performance of PREDAC in predicting antigenic clusters, it was applied to the five antigenic clusters defined by Smith et al 25. Five antigenic clusters were predicted (Fig. 1c) using a main inflation value of 1 24 (Supplementary Fig. S4). To characterize the antigenic evolution of human A (H3N2) viruses in mainland China, 1438 HA sequences of influenza A (H3N2) viruses isolated during were used and 20 predicted antigenic clusters (Supplementary Table S2, additional vaccine strains were included) were obtained using a main inflation value of To compare antigenic patterns between mainland China and other countries, 9386 HA sequences from all over the world during were used in our analysis and 15 predicted antigenic clusters were identified using a main inflation value of Network Visualization All ACnets in this work were visualized using the yfiles Organic layout in Cytoscape 63,64. To better display predicted antigenic clusters, the positions of some nodes were manually adjusted so that each predicted antigenic cluster had a clear boundary from the others. Analysis of the key parameters on the efficiency of vaccine strain prediction To propose an effective vaccine strain prediction strategy, we tested several parameters. The first parameter was the edge weight of the ACnet and affected the sensitivity for detecting novel antigenic clusters. When the logarithm of 14

15 the odds ratio was used as the ACnet edge weight and a coarse-grained clustering value was used, the predicted antigenic clusters captured the actual antigenic evolutionary patterns observed in history with high accuracy (Fig. 1d and Fig. 2c). However, in the coarse-grained predicted antigenic clusters, closely related antigenic clusters, like A/Sydney/5/1997 and A/Panama/2007/1999 in the Smith data (Supplementary Table S5), and A/Wisconsin/67/2005 and A/California/7/2004 (antigenic distance < 4) 57 in our analysis could not be separated (Fig. 2b and Fig. 3), suggesting that there is a lack of sensitivity in coarse-grained prediction. For vaccine strain recommendation, novel antigenic clusters must be detected with high sensitivity. This can be achieved by either using the odds ratio as the ACnet edge weight, or by using higher main inflation values that generate fine-grained clusters (Supplementary Table S5). Since the choice of the main inflation value requires the inspection of the cluster size curve, for convenience, we chose to keep the main inflation value untouched. Therefore, to enhance sensitivity for the detection of novel antigenic clusters, we used the odds ratio instead of its log derivative as the ACnet edge weight. Indeed, when the odds ratio was used for vaccine strain recommendation, WI05 (A/Wisconsin/67/2005-like strains) could be separated from CA04 (A/California/7/2004-like strains) (Fig. 4). The other two key parameters are threshold percentage and the date to make prediction. The effect of the two parameters on efficiency of vaccine strain prediction were analyzed for the seven seasons from to based on viruses sequenced by China CDC during regular influenza surveillance. The flu seasons in China usually start in October and peak in the winter (from December to the following March). Due to the lengthy process of current vaccine production which usually takes ~6 months to produce a sufficient stockpile for mass vaccination, the vaccine strains need to be recommended no later than April. We tested different dates starting from the beginning of January (Jan 1) to the end of April (April 30) with an interval of half a month (see Table 1). We further explored different threshold percentages required for a novel antigenic cluster to be recommended (see Table 1). The numbers of correctly predicted vaccine strains were recorded in Table 1 by comparing the predicted vaccine strains to the dominant strains monitored by the China CDC. 15

16 Supplementary References 61 Plotkin, J. B., Chave, J. & Ashton, P. S. Cluster analysis of spatial patterns in Malaysian tree species. Am Nat 160, ,(2002). 62 Keane, T., Creevey, C., Pentony, M., Naughton, T. & Mclnerney, J. Assessment of methods for amino acid matrix selection and their use on empirical data shows that ad hoc assumptions for choice of matrix are not justified. BMC Evolutionary Biology 6, 29,(2006). 63 Wiese, R., Eiglsperger, M. & Kaufmann, M. yfiles - Visualization and automatic layout of graphs. Math Visual 378, ,(2004). 64 Shannon, P. et al. Cytoscape: a software environment for integrated models of biomolecular interaction networks. Genome Res 13, ,(2003). 16