Throat swabs from patients with influenzalike syndromes were collected from the Clinical Virology Laboratory, Chang Gung Memorial Hospital. The specimens were inoculated in MDCK cells. Typing for influenza A virus was then performed with immunofluorescent assay by type-specific monoclonal antibody (Dako, Cambridgeshire, UK). Subtyping was conducted by reverse transcription (RT)–PCR with subtype-specific primers.

The RT-PCR product was purified by using the QIAquick Gel Extraction Kit (Qiagen, Valencia, CA, USA). The nucleotide sequence was determined with an automated DNA sequencer. Sequence editing and processing were performed with Lasergene, version 3.18 (DNASTAR, Madison, WI, USA). Multiple sequence alignment was performed with ClustalW version 1.83 (ftp://ftp.ebi.ac.uk/pub/software/unix/clustalw). Global sequence comparison that yielded pairwise sequence identities used in histogram analysis was done with the program Needle in the EMBOSS package (5). Amino acid sequences were translated from coding sequences and aligned by BioEdit (6). An entropy value was defined at an aligned amino acid position according to the formula ΣP_i*log(P_i), in which i is the observed probability for each of the 20 amino acids (aa) (7). A graphic tool was developed in Java for displaying the entropy plot used in this work. All amino acid numberings are based on influenza virus A/Puerto Rico/8/1934 (PR8).

To show the host-associated amino acid signatures, we retrieved full genome sequences (as of August 22, 2005) from the genome browser at Influenza Sequence Database (ISD) (8). To differentiate between avian and human influenza viruses, we excluded human-isolated avian influenza viruses from the human dataset and examined those sequences separately. Altogether, we had 95 avian and 306 human influenza viral genomes, henceforth termed "primary dataset." All 11 viral proteins encoded by the 8 genomic RNA segments were compared: PB2, PB1, PB1-F2, PA, HA, NP, NA, M1, M2, NS1, and NS2.

Avian influenza viruses from human influenza patients were separately retrieved from NCBI as well as from ISD. Altogether, we had 417 protein sequences from 60 avian influenza strains, in which 21 strains contain sequences (full or nearly full length) from all 8 genomic RNA segments.

For validating the signatures obtained from analyzing the primary dataset, we further retrieved 15,785 human or avian influenza A viral protein sequences from NCBI's Influenza Virus Resources. Details for the sequences used can be found in Appendix, Supporting Materials and Methods, as well as in Table A1 and Table A2. Eleven Taiwanese genomes produced in this work have been deposited in GenBank with accession numbers DQ415283 through DQ415370.

Using previously described methods (7), we separately calculated an entropy value for every aligned amino acid position for 95 avian influenza viruses and 306 human influenza viruses. Those amino acid residues with an entropy value between 0 and –0.4 for both the human and avian strains were identified as most highly conserved. We chose this entropy threshold on the basis of the entropy value –0.379, calculated at position 627 of PB2 for the 95 avian viruses. This widely reported, species-associated residue is highly conserved; it has E (Glu) in 83 and K (Lys) in 12 avian isolates and Lys in all 306 human isolates. We then selected those conserved positions with distinct amino acid residues between human and avian influenza viruses as potential host-associated signatures. An entropy plot for identifying such signature residues for avian versus human influenza virus NP segments is shown in Figure panel A. In each aligned position, we placed an avian consensus residue on top and a human consensus at the bottom. For example, the entropy value is zero at amino acid position 283 for both avian and human strains, in which all 95 avian influenza viruses contain L (Leu), whereas all 306 human influenza viruses contain P (Pro). The other 2 residues with zero entropy value in avian and human viruses are located at position 55 of PA, in which we have D (Asp) in avian viruses and N (Asn) in human viruses, and position 121 of M1, in which we have T (Thr) in avian and A (Ala) in human viruses. Entropy plots for all 11 influenza viral proteins can be found in Figure A1.

Figure panel B shows a genomewide view of the entropy plots for 11 influenza A viral proteins. The amino acid sequences of hemagglutinin (HA), with an average entropy value of –0.524 within avian viruses and –0.158 within human viruses, exhibit much more diversity than other open reading frames (ORFs). PB2, PB1, PA, NP, and M1, on the other hand, are more conserved (i.e., they have less negative entropy values).

In addition to the previously mentioned 3 positions with distinct amino acid residues between avian and human strains, we found 225 additional positions with nearly distinct amino acid residues, with their computed entropy values less negative than –0.4 in both the 306 human and 95 avian strains that we analyzed. To assess the robustness of those 228 residues used in differentiating human from avian influenza viruses, we further examined 15,785 influenza A protein sequences from NCBI. After validation, 52 positions still showed an entropy value less negative than –0.4 and conserved to distinct amino acid residues between human and avian viruses (Table 1). From this entropy analysis, we identified an additional 51 aa positions that may be as important as the well-known position 627 of PB2. We designated these 52 positions as "species-associated" signatures. Among 11 ORFs, NP contains the highest number of such signatures (15 positions), followed by PA (10 positions), PB2 (8 positions), PB1-F2 (5 positions), M2 (4 positions), M1 (3 positions), PB1 (2 positions), HA (2 positions), NS2 (2 positions), and NS1 (1 position). No signature was found in the NA gene. We also summarized the related functions of those species-associated signatures in Table 1. The complete results of genome scanning and validation can be found in Table A3 and Table A4.

We examined how the amino acid sequences varied at those proposed signature positions for avian influenza viruses isolated from humans. At 9 of these 52 positions, residue changes were characteristic of human rather than avian viruses (Table 2). For example, 34 sequences (27 H5N1, 3 H9N2, and 4 H7N7) were available for inspection at position 199 of PB2 (data not shown). Aside from 10 sequences with gaps (sequences did not cover this position), 19 of the remaining 24 still have Ala, which is typical for avian viruses. Five of them (all H5N1), on the other hand, have this residue changed to Ser, which is mostly seen in human viruses. At the well-known position 627 of PB2, 5 sequences had gaps, 22 retained Glu (typical for avian virus), while the other 7 changed to Lys, which is typical for human virus. Among those 7 mutated sequences, 6 were from H5N1 human isolates (A/Hong Kong/483/1997, A/Hong Kong/485/1997, A/Vietnam/1194/2004, A/Vietnam/1203/2004, A/Vietnam/3062/2004, and A/Thailand/16/2004), and the other 1 was A/Netherlands/219/2003(H7N7), which was isolated from a fatal human case of pneumonia in the Netherlands (32).

To understand how mutations had accumulated within a specific virus, we summarized the amino acid changes for 21 of these avian viruses that contained full or nearly full-length sequences for each segment (Table 3). We found that 19 of 21 strains contained >1 species-associated amino acid change, and 7 of them contained >2 substitutions; A/Netherlands/219/2003(H7N7) had the highest count for mutation accumulation (3 positions). Among these 52 species-associated signatures, the mutation combinations at positions PB2 199 and PA 409 were most commonly seen in H5N1 human isolates from Hong Kong in 1997.

Our observation that NP contained the highest number (15 of 52) for species-associated amino acids suggested that NP might serve as a molecular target for differentiation between human and avian influenza A viruses. To indicate such host specificity, or the "genetic boundary" between these 2 viruses at the nucleotide level, we performed a pairwise sequence comparison for all 11 ORFs on our 401-genome primary dataset and produced histograms on their computed pairwise identities. In Figure A2, pairs with 2 sequences of the same host species (human to human, or avian to avian; termed homopairs) and pairs for sequences that cross host species (human to avian, or avian to human; termed heteropairs) are shown. HA and NA genes exhibited considerable sequence differences between strains, with identities as low as 47%. Also noted was a wide spectrum of percent identities (e.g., 55%–95% in the horizontal axis) containing few sequence pairs for these 2 genes. For both of these proteins, some strains from the same species can have identities as low as 50%. However, the ORF of another surface protein, M2 ion channel protein, is relatively conserved (>74% identity for viruses across species). The histograms for the polymerase genes (PB2, PB1, and PA), NP, and M1, on the other hand, are much less varied (mostly <20% variation). In particular, the NP gene was found to exhibit a fairly clear boundary between homopairs and heteropairs, at ≈86%.

In the main text we have mentioned an entropy value was defined at an aligned amino acid position according to the formula ΣPi*log(Pi), where i is the observed probability for each of the 20 amino acids. An entropy value defined like this is at most zero when all amino acids at this position conserve to the same residue, while a more negative value indicates that the residues are more divergent for containing more residue types. Although BioEdit also includes a module with similar formula in computing entropy values for aligned sequences, we chose to develop our own software for more streamlined data manipulation and subsequent analysis and interpretation.

To reveal the host-associated amino acid signatures, we have retrieved full genome sequences (as of August 22, 2005) from the genome browser at Influenza Sequence Database. Strains containing all eight RNA segments and for each segment a minimum 90% long of the coding sequence based on PR8 were included, which serve as the primary dataset for full genome scanning. Altogether, we have 95 avian influenza genomes (including 60 H5N1, 8 H6N1, 6 H6N2, 1 H7N1, 1 H7N3, 2 H7N7, 17 H9N2) and 306 human influenza genomes (8 H1N1, 2 H1N2, 3 H2N2 and 293 H3N2), the latter include 11 complete genomes of Taiwanese strains from 1996 to 2004 (newly sequenced data from this study). See Supporting Table 1 for a complete listing of accessions for these 401 genomes. Coding sequence alignments for each genomic segment were compiled: PB2, 759 aa; PB1, 757 aa; PB1-F2, 90 aa; PA, 716 aa; HA, 591 aa; NP, 498 aa; NA, 480 aa; M1, 252 aa; M2, 97 aa; NS1, 230 aa; and NS2, 121 aa.

Human-isolated avian influenza viruses from human flu were separately retrieved from NCBI as well as from ISD. Altogether we have 417 accessions from 60 avian flu strains (48 H5N1, 6 H9N2, 5 H7N7 and 1 H7N2), in which 21 strains (17 H5N1, 3 H9N2 and 1 H7N7) contain sequences (full or nearly full-length) from all 8 genomic RNAs. See Table A2 for a complete listing of these accessions.

For validating the obtained signatures from analyzing the mentioned 401-genome primary dataset, we have firstly retrieved 14,057 human or avian influenza A protein sequences from NCBI's Influenza Virus Resources (as of January 17, 2006), including 5,468 avian and 8,589 human sequences (786 H1N1 sequences and 7,097 H3N2 sequences among the others). At the stage of revising this manuscript, we have included more H1N1sequences (2,514 in total, as of April 20, 2006) for validation to relieve the limitation that may be caused by the unbalanced sequence counts between H1N1 (786 sequences) and H3N2 (7,097 sequences) previously used, thus making the results more robust. Altogether we have used 15,785 influenza protein sequences for confirmatory analysis.