Analysis of virulence determinants in full length H5N1 influenza genomes using computational modelling

H5N1 bird 'flu has caused massive mortality in poultry since its emergence in the far east in 1997. Particularly since 2002, increasing numbers of wildfowl have been found dead and dying, which is very unusual for bird 'flu as it does not normally cause severe disease in wild birds. It has also given rise to an estimated minimum of 250 human deaths. As has been seen in the news, wild birds have recently spread the virus into Europe along East-West migration routes from Asia. We clearly urgently need to identify what has changed in the virus which has led to this increase in mortality. This is a difficult task for several reasons. Firstly, the risks associated with the virus make it difficult to work with experimentally in the laboratory, as we must be certain the virus cannot escape. Secondly, influenza evolves naturally very rapidly. This applies whether or not the mortality associated with it is changing and makes it difficult to identify the exact genetic factors responsible. Thirdly, some genetic changes which have been important in the past have been identified, but recent laboratory studies suggest many more could also be important in the current epidemic. We will study this by taking a computer-based approach. A very large amount of information on the genetic make-up of several thousand different strains of H5N1 'flu. On average each one of these may differ at about 700 different amino acid sites in its proteins from any other. To find out which of these sites are important, we use a new statistical approach which is designed to cope with very large complex bodies of data and identify critical components that are responsible for the increase in mortality. These new models require very fast computers and we will be using one of the fastest in the UK: the IBM BlueGene supercomputer at the University of Edinburgh. Our aim is to identify the important components of the H5N1 'flu genome that can be most clearly associated with the mortality change. In this way we can assist with the design of very specific and controlled laboratory experiments which can be constructed to test our conclusions, and which may lead to identification of the key genetic elements for intervention through drug development.

High pathogenicity avian influenza type H5N1 has caused massive mortality in poultry since its recent re-emergence in the far east in 1997. Increasingly, in recent years, extensive mortality among wildfowl has also been recorded in association with H5N1 infection, which is unusual for avian influenza and a serious cause for concern. It has also given rise to an estimated minimum of 250 human deaths. There is a clear need to understand the mechanisms underlying the increase in pathogenicity. It has been known for some years that an alteration in the cleavage motif of the haemagglutinin protein, the incorporation of several basic amino acid residues, underlies the distinction between low pathogenicity and high pathogenicity avian influenzas in general. However, recent studies of H5N1 have identified several possible contributing factors to the further rise in pathogenicity. The segmented nature of the influenza genome allows replacement of genome elements from the reservoir of strains in wild birds, and several gene elements including the polymerase, M and NS proteins as well as HA and NA have been indicated as distinguishing the more pathogenic from the less pathogenic strains in laboratory studies. We will take an informatics-based approach to the question, using the novel methodology of Bayesian Graphical Models, which we will implement on the University of Edinburgh BlueGene supercomputer, to identify and rigorously test the associations of specific protein variants across the H5N1 genome with the increase in pathogenicity. The analysis will be based on the nearly 4000 H5N1 sequences, and particularly the 239 complete viral genomes, already deposited in GenBank. With this approach we will identify a quantitative globally optimal model which will include all genomic elements significantly associated with outcome, and their links with each other, which can be tested in laboratory systems.