Influenza viruses are negative-stranded, enveloped orthomyxoviruses with eight gene segments, each encoding one or two proteins3. The signature antigenicity of the A and B types of influenza viruses is determined by the viral glycoproteins hemagglutinin (HA) and neuraminidase (NA). The annual genetic drift in antigenicity, which is driven by point mutations, is responsible for seasonal influenza epidemics1,3,4. Swapping of gene segments by reassortment between viruses of aquatic birds, swine and humans (genetic shift) produces new type A influenza viruses with novel antigenicity that may cause devastating pandemics1,3,4.

The capacity of influenza viruses for immune escape requires that vaccine strains be updated annually to reflect changes in the HA and NA genes within the impending seasonal strains. Two types of vaccines are currently used: a chemically inactivated virus delivered by injection, and a live attenuated influenza virus vaccine of cold-adapted virus5, delivered as a nasal-spray (FluMist) ( Both vaccines have limitations. Whereas cell-mediated responses are increasingly recognized as a major determinant of influenza immunity6,7,8,9, traditional, killed influenza virus vaccines act mainly by inducing neutralizing antibodies. Unfortunately the killed vaccine appears to have suboptimal efficacy in the elderly population (>65 years old)10, which is the same population most prone to morbidity and mortality from seasonal influenza epidemics. In contrast, live attenuated influenza virus vaccine induces both humoral and cellular immunity but its administration remains restricted to healthy children, adolescents and adults (nonpregnant females), ages 2–49. It works better in immunologically naive young children than in adults11,12.

Here we illustrate the use of SAVE2 to rationally design live attenuated influenza virus vaccines.

The central idea of SAVE is to recode and synthesize a viral genome13 in a way that perfectly preserves the WT amino acid sequence, while rearranging existing synonymous codons to create a suboptimal arrangement of pairs of codons2. For reasons that are not understood, some pairs of codons occur more frequently, and others less frequently, than expected14. This codon-pair bias, which is found in every species examined15, evolves slowly. Yeast and humans have a radically different codon-pair bias, but all mammals share essentially the same codon-pair bias (unpublished results).

Codon-pair bias is independent of codon bias. For example, consider the amino acid pair Arg-Glu. As there are six codons for arginine and two for glutamic acid, there are twelve possible codon combinations that encode this pair of amino acids. Taking into account the frequency of the two contributing codons (codon bias), the pair CGC-GAA is expected 2,397 times in the annotated human ORFeome. However, it is instead observed only 268 times (observed/expected = 0.11). This is an infrequently used codon pair. In contrast, the Arg-Glu pair encoded by AGA-GAA is expected 2,644 times, but is observed 4,195 times (observed/expected = 1.59); this is a frequently used codon pair. By whole genome synthesis13,16 we previously recoded poliovirus to contain 'poor' (that is, infrequently used) codon pairs, and found that this dramatically attenuated the virus2. Although the mechanism of attenuation is unclear, preliminary evidence suggests that translation is affected2. Attenuation can be 'titrated' by adjusting the extent of codon-pair deoptimization2. Because codon-pair deoptimization results from miniscule effects at each of hundreds or thousands of nucleotide mutations (without changing amino acid sequences), reversion to virulence is extremely unlikely2. Aided by computer algorithms2, codon pair–deoptimized viral genomes can be rapidly designed and synthesized, and live virus can be generated by reverse genetics.

To attenuate influenza virus, we redesigned large parts of the coding regions of the polymerase subunit B1 (PB1), nucleoprotein (NP) and HA genes of influenza virus A/PR/8/34 (PR8), using our deoptimization computer program2. Along with other viral genes, these genes play important roles in replication and assembly of influenza virus. Without altering either amino acid sequence or codon bias, the program rearranged existing synonymous codons to deoptimize codon pairs. This resulted in hundreds of silent mutations per genome segment without any amino acid changes. The characteristics of the synthetic genome segments and their changes in codon-pair bias are summarized in Table 1 and Supplementary Figure 1.

Table 1 Characteristics of deoptimized influenza genome segments

The deoptimized segments were synthesized de novo and cloned into a standard ambisense, eight-plasmid system17,18. To generate influenza viruses carrying one or more deoptimized segments, the plasmids carrying the recoded, synthetic segments, together with the complement of the remaining PR8 WT plasmids, were transfected into susceptible cells. Each deoptimized segment PB1Min, NPMin and HAMin in the background of the complementing seven WT segments yielded a viable virus, as did any combination thereof, including the virus simultaneously containing all three deoptimized segments (PR8-PB1Min/NPMin/HAMin, abbreviated PR83F) (Fig. 1 and data not shown).

Figure 1: Characterization in tissue culture of synthetic codon pair–deoptimized influenza viruses.
figure 1

(a) Plaque phenotypes on MDCK cells of PR8 WT virus and synthetic PR8 derivatives, carrying one (NPMin, HAMin or PB1Min), two (NPMin/HAMin or HAMin/PB1Min) or three (PR83F) deoptimized gene segments. (b) Growth kinetics of PR8 WT virus and three synthetic PR8 derivatives in MDCK cells after infection with 0.001 MOI of the indicated viruses. All combinations grew within log 8–9 although only PR8-HAMin, PR8-NP/HAMin and PR83F are shown as compared to WT. (c) Analysis of influenza virus protein expression in infected cells. MDCK cells were infected with PR8 WT virus or synthetic PR8 derivatives, carrying one deoptimized gene segment each (NPMin, HAMin or PB1Min), as indicated. Western blot (WB) analysis of proteins extracted from whole cell lysates was carried out with PB1, NP, HA or actin antibodies. Actin was used to indicate equal protein loading.

Several of these new synthetic viruses were analyzed for their in vitro growth characteristics in Madin-Darby canine kidney epithelial (MDCK) cells. All mutant viruses formed plaques that were either indistinguishable from, or only slightly smaller than that of the WT virus (Fig. 1a). The mutant viruses did not grow as well as WT, but typically only to about tenfold lower titers (Fig. 1b). The plaque and growth properties of viruses carrying combinations of synthetic segments other than those depicted in Figure 1 fall in between those of PR8 and PR83F (data not shown). The deoptimized virus PR83F had about the same response to changes in temperature (to either 33 °C or 39.5 °C) as the WT virus (Supplementary Fig. 2).

Previously, we found that codon- or codon pair–deoptimized polioviruses had a reduced specific infectivity2,19. This was not the case for deoptimized influenza viruses, as their ratio of plaque-forming units (PFU) to HA units was nearly identical to that of WT virus (data not shown).

Our working hypothesis postulates that deoptimization of viral open reading frames reduces protein synthesis, which in turn generates an attenuated virus2. Therefore we used western blot analysis of infected whole-cell lysates to test protein synthesis driven from the deoptimized gene segments in PR8-NPMin, PR8-HAMin and PR8-PB1Min (Fig. 1c). In all three viruses, the deoptimized viral gene product was specifically reduced compared to other proteins from the same virus, or to the same protein in WT-infected cells, or to the actin control in the same cells (Fig. 1c). Protein synthesis from the WT segments in a deoptimized virus was apparently not affected. The exact molecular mechanisms responsible for the reduced protein production remain to be determined.

Despite their reasonably robust growth, codon pair–deoptimized influenza viruses proved to be remarkably attenuated in mice (Table 2). Each individual deoptimized segment had a demonstrable attenuating effect, reducing the median lethal dose (the dose killing half the animals; LD50) 10-, 30- and 500-fold, for PR8-NPMin, PR8-HAMin, and PR8-PB1Min, respectively. Combinations of two deoptimized genes have not been tested, but combining all three attenuating genes into one virus (PR83F) led to a cumulative attenuation of about 13,000-fold (Table 2).

Table 2 LD50 and PD50 of deoptimized influenza viruses

To test the codon pair–deoptimized viruses in animals, BALB/c mice were infected intranasally with 104 PFU of PR83F or PR8, and monitored for disease symptoms (ruffled fur, lethargy, weight loss, death). At this dose, mice infected with PR8 WT virus developed severe symptoms with rapid weight loss and did not survive more than 5 d after infection (Fig. 2a). Mice infected with PR83F, on the other hand, experienced no observable symptoms or weight loss, except for a small, transient delay in weight gain as compared to mock-infected animals (Fig. 2a).

Figure 2: Attenuation of codon pair–deoptimized influenza virus PR83F in BALB/c mice.
figure 2

(a) Body weight curve after intranasal infection with 104 PFU of PR8 WT virus (triangles), 104 PFU of deoptimized PR83F virus (diamonds) or mock-infected (saline; squares). Each time point indicates the average of five mice, with error bars indicating s.d. WT-infected mice did not survive beyond day 5 (indicated by a cross). (b) Virus titer in whole lung homogenate after infection with either 103 PFU of PR8 WT virus (squares) or deoptimized PR83F (circles). Average of three mice per time point. *On day 9 post infection, PR83F was no longer detectable (below 40 PFU/lung).

Live attenuated vaccines in general depend on a limited, yet safe, degree of replication within the host to stimulate the immune system. To assess the replicative potential of our influenza vaccine candidate in an immune competent host, we monitored viral load in the lungs of BALB/c mice infected intranasally with either 103 PFU of PR83F or PR8 WT virus. Within 24 h, WT-infected mice had 3,000-fold higher viral load in their lungs compared to PR83F, leading to death in less than 6 d (Fig. 2b). Conversely, in PR83F-infected animals, amplification of the vaccine virus progressed slowly and peaked at a lower viral load than the WT virus, resulting in a controlled infection with no overt disease symptoms, and virus clearance after 9 d (Fig. 2b).

Infection by a sub-lethal dose of WT virus can in principle provoke protective immunity, which is the usual course of natural human infections. The Chinese scholar Li Shizhen described the art of inoculating humans with live smallpox20. This method of smallpox vaccination, practiced in China for centuries, was very dangerous because the difference between the lethal dose and the immunizing protective dose of WT smallpox is very small. To address the issue of safety margin quantitatively with our influenza viruses, we determined the LD50 and the protective dose 50 (the dose providing protective immunity to half the animals; PD50) for PR8 WT virus and also for the attenuated strain, PR83F (Fig. 3). PR8 had a very low PD50 of 1 PFU, which is equivalent to 40 virus particles when titered on MDCK cells21. The LD50 of PR8 was 61 PFU, resulting in an LD50/PD50 ratio of about 60. This ratio between the LD50 dose and the PD50 dose is the 'safety margin' of a given virus if it were to be used as a vaccine. The narrow safety margin of WT (LD50/PD50 = 60) compromises its suitability for use as a vaccine. In contrast, the attenuated virus PR83F had a PD50 of 13 PFU, which although higher than the PD50 of WT, is still very low. The attenuated PR83F had an LD50 of 790,000 PFU and, thus, an LD50/PD50 ratio (safety margin) of 60,000, which is 1,000-fold better than the WT virus (Fig. 3a versus Fig. 3b, shaded areas under the curve). Thus, it is easy to determine a dose of the attenuated virus PR83F that is both safe to administer and effective in provoking immunity (Supplementary Fig. 3).

Figure 3: Immune responses and protection.
figure 3

(a,b) Vaccine margin of safety for PR8 WT and deoptimized PR83F viruses. The left ordinate indicates the percentage of animals surviving the primary inoculation (black squares) with (a) PR83F or (b) WT PR8, at doses ranging between 100 to 106 PFU. After 28 d, the surviving, vaccinated animals were challenged with a single 1,000× LD50 of PR8 WT virus. Disease and survival were monitored (right ordinate; open circles) for (a) PR83F- and (b) PR8-vaccinated mice. (c) Virus load in mouse lungs after PR8 WT virus challenge of PR83F-vaccinated animals. Twenty-eight days after a single intranasal vaccination with 104 PFU PR83F mice were challenged with 1,000 × LD50 of PR8 WT virus. Three days thereafter the level of challenge virus in lung homogenates was determined. (d) ELISA determination of influenza-specific serum antibodies. Twenty-eight days after a primary infection, serum was collected, and anti-influenza IgG serum titers were determined from animals that had received a primary inoculation of 0.01× LD50 (black diamonds) or 0.001× LD50 of PR83F (black circles), 0.01× LD50 of PR8 (white squares) or saline (black triangles). ELISA antibody titer against PR8 virus antigen is expressed as the lowest reciprocal serum dilution that resulted in a positive ELISA signal (5 s.d. above background). In c and d each symbol represents the data from one animal.

To characterize the protective immunity in more detail, we immunized mice with a single intranasal vaccination of 104 PFU of PR83F. After 28 d, these mice were challenged with 1,000× LD50 of WT PR8. Three days after challenge, PR8 titers in lung homogenates were determined. In 80% of the mice, challenge virus was below the level of detection (suppressed at least 106-fold) (Fig. 3c; open circles), whereas titers of 107 PFU were found in lungs of mock-vaccinated animals (Fig. 3c; open squares).

The mean anti-influenza serum IgG titer in mice immunized with 0.01× LD50 of the respective viruses was 312,500 for PR83F and 27,540 for PR8 (Fig. 3d). At an even lower, and thus safer, vaccine dose of 0.001× LD50, the immune response toward PR83F was nearly unchanged with an antibody titer of 237,500 (Fig. 3d). Thus, at identical doses relative to their respective LD50, PR83F is a more potent inducer of influenza-specific antibodies than WT. These findings attest to the strong immunizing potential of a low-grade influenza virus infection in general, and to the safety profile of codon pair–deoptimized influenza viruses in particular.

Vaccines created in this way encode all viral proteins with the amino acid sequences found in WT viruses. Thus the viruses express the entire WT repertoire of antigenic sites and would have the maximum chance of inducing both cellular and humoral immunity against all epitopes. As attenuation results from hundreds or thousands of nucleotide changes, the probability of reversion to virulence is extremely low, an advantage over other approaches to making live vaccines. Co-infection of the vaccine recipient with a naturally circulating WT virus could lead to gene reassortment, but this is unlikely to produce variants more virulent than the co-infecting virus. The deoptimized segments of the vaccine strain would 'poison' any such reassortant.

In spite of striking genetic dissimilarities, codon-pair deoptimization has given similar results with both influenza virus and poliovirus2. Poliovirus is a plus-stranded RNA virus whose genome encodes a single polyprotein. Codon-pair deoptimization in the 5′-region of the poliovirus genome disturbs the synthesis of all viral proteins with drastic effects on replication2. In contrast, influenza virus, a minus-stranded RNA virus, has a genome divided into eight segments and its proteins are expressed independently from individual mRNAs. The fact that two such extremely different viruses each are highly sensitive to codon pair deoptimization suggests that the strategy may operate quite generally.

As with any new technology, there are issues to resolve before applying this approach for human use. These include our incomplete knowledge of the molecular mechanism(s) involved in attenuation, an in-depth assessment of the genetic stability of the attenuation phenotype and a need to find the right balance between attenuation (for provoking an immune response) and robust replication (for purposes of production). For seasonal epidemics, SAVE may allow a different strategy than the existing live vaccines, such as FluMist. For FluMist, the attenuating mutations (obtained by lengthy selection procedures) map to six viral 'backbone' genes, which therefore must be kept constant every year, with only HA and NA being updated annually (6:2 recombinants). Individuals immunized yearly could develop immunity against 'backbone' proteins, limiting the replication and thus the efficacy of the vaccine (possibly helping to explain the relatively poor efficacy of FluMist in previously immunized army personnel22). In contrast, use of SAVE could attenuate the entire genome of an impending seasonal or pandemic strain, providing a perfect antigenic match between the vaccine and target virus. However, this approach would require an appropriate regulatory environment that might, for instance, approve a method of attenuating each new strain to a certain standardized degree with a standardized change in codon-pair bias (just as the current inactivated virus vaccine uses a standardized method of inactivation). To attenuate in a standardized and predictable way will require many more experiments to exhaustively explore the variables of SAVE.

The recoded influenza viruses described here present a useful paradigm for vaccines, as attenuation is the result of hundreds or thousands of nucleotide changes without the change of a single amino acid. The attenuated phenotype results from large-scale rearrangements of existing synonymous codons, producing under-represented pairs of codons2. Considering (i) the expected high genetic stability of the attenuating genetic changes (“attenuation by a thousand cuts”23), (ii) the possibility of systematically designing such viruses, (iii) the favorable growth kinetics in tissue culture (108 PFU/ml), (iv) the small protective dose of 'deoptimized' influenza viruses, (v) the efficacy of PR83F as revealed in challenge experiments and, particularly, (vi) the wide safety margin, the SAVE technology sets the stage for making efficient live attenuated influenza vaccines. Although it is difficult to extrapolate to human use, in our system 10 ml of culture supernatant apparently contains enough virus to vaccinate and protect 1 million mice with a single vaccination of 100 PD50 doses of PR83F (Fig. 3a and Supplementary Fig. 3).

Vaccines based on changes in codon-pair bias could be generated within weeks for any emerging influenza virus once its genome sequence is known, although of course a further period of testing would be required before the vaccine could be used. The margin of safety appears to be unusually high. This new strategy will be applicable to the rapid development of human vaccines against seasonal flu epidemics and pandemics.


Cell lines.

293T and MDCK cells were obtained from the American Type Culture Collection (ATCC). Cells were grown in Dulbecco's modified Eagle's medium (Invitrogen), supplemented with 10% FBS (HyClone) and penicillin-streptomycin (Invitrogen).

Reverse genetics and generation of synthetic influenza viruses.

All synthetic influenza viruses used here are based on the strain A/PR/8/34 (Mt. Sinai variant, short PR8). The reference sequences of the eight gene segments for this strain are available under GenBank accession numbers AF389115, AF389116, AF389117, AF389118, AF389119, AF389120, AF389121, AF389122). An eight-plasmid ambisense system for this strain cloned in the vector pDZ25 was kindly made available by Peter Palese and Adolfo García-Sastre (Mt. Sinai School of Medicine).

Codon pair–deoptimized genome segments were designed using a computer algorithm previously described2. Coding regions of the segments PB1, HA and NP were targeted to be recoded. A minimum of 120 nucleotides at either segment terminus were left unaltered because these sequences contain packaging signals3. All of the codon changes were synonymous; that is, none of them introduced any amino acid changes. The precise extent of recoded sequence for each target segment and the resulting changes in codon-pair bias are summarized in Table 1.

The recoded gene segments were synthesized de novo (Mr. Gene) and introduced into WT plasmids to replace the respective WT counterpart sequence. Codon pair–deoptimized viruses were derived by DNA transfection of either one, or combinations of two, or three deoptimized segments together with the remaining WT plasmids. For this purpose a total of 2 μg plasmid DNA (250 ng of each of 8 plasmids) was transfected into co-cultures of 293T and MDCK cells in 35-mm dishes using Lipofectamine 2000 (Invitrogen) according to manufacturers recommendations. After 6 h of incubation at 37 °C, the serum free Opti-MEM containing the transfection mix was replaced with DMEM containing 0.2% Bovine Serum Albumin (BSA). After a further 24 h of incubation, 1 μg/ml TPCK-Trypsin was added to the dishes. Two days thereafter virus containing cell supernatants were collected and amplified on MDCK cells.

In vitro growth characteristics and titration of synthetic influenza viruses.

The growth characteristics of codon pair–deoptimized synthetic viruses were analyzed by infecting confluent monolayers of MDCK cells in 100-mm dishes with 0.001 multiplicities of infection (MOI). Infected cells were incubated at 37 °C in DMEM, containing 0.2% BSA and 2 μg/ml TPCK-Trypsin (Pierce). At the given time points 200 μl of supernatant was removed and stored at −80 °C until titration. Viral titers and plaque phenotypes were determined by plaque assay on confluent monolayers of MDCK cells in 35-mm six-well plates using a semisolid overlay of 0.6% tragacanth gum (Sigma-Aldrich) in minimal Eagle medium (MEM) containing 0.2% BSA and 4 μg/ml TPCK-trypsin. After 72 h of incubation at 37 °C plaques were visualized by staining the wells with crystal violet.

Mouse pathogenicity, in vivo virus replication and vaccination.

A minimum of five BALB/c mice (5–6-weeks-old) per group were infected once by intranasal inoculation with doses ranging from 100 to 106 PFU of PR83F or of PR8 WT virus. Inoculum virus was diluted in 25 μl PBS and administered evenly into both nostrils. A control group of 5 mice was inoculated with PBS only (mock). Venous blood from the tail vein was collected from all animals before initial infection for subsequent determination of pre-vaccination antibody titers.

Morbidity and mortality (weight loss, reduced activity, death) was monitored. LD50 of the WT virus and the vaccine candidates was calculated as described24. Mice experiencing severe disease symptoms (rapid, excessive weight loss over 25%) were euthanized and scored as a lethal outcome.

For vaccination experiments mice were infected as above. Twenty-eight days after the initial infection (vaccination), venous blood from the tail vein was drawn for subsequent determination of post-vaccination antibody titers.

The mice were then challenged with 105 PFU of the PR8 WT virus corresponding to >1,000 times the LD50. Mortality and morbidity (weight loss, reduced activity, death) was monitored. The PD50 of codon pair–deoptimized PR83F versus that of the PR8 was determined as the dose required to protect 50% of mice from a challenge with 1,000× LD50 of the WT virus, 28 d after a single inoculation with the vaccine virus.

To assess virus replication in the lungs of infected animals, BALB/c mice were infected intranasally with 103 PFU of either PR8 or PR83F. At 1, 3, 5, 7 and 9 d after infection the lungs of three mice each were collected (WT infected mice did not survive beyond day 6). Lungs were homogenized in 1 ml of PBS and the virus load per organ was determined by plaque assay on MDCK cells, as described above. Similarly, replication of PR8 challenge virus in lungs of PR83F-vaccinated mice were determined. Twenty-eight days after a single intranasal inoculation with 104 PFU of PR83F or saline, five animals each were challenged with 1,000× LD50 of PR8. Three days after challenge (the usual peak of viral replication), lungs were processed and viral load per organ determined as described above.

Western blot analysis of influenza virus proteins in infected cells.

MDCK cells were infected with virus at an MOI of 5 and incubated for 4 h at 37 °C. Subsequently, the cells were harvested in Laemmli buffer. The proteins were resolved by SDS-PAGE, analyzed by western blots with α-HA, α-PB1 and α-actin (Santa Cruz Biotechnology), and α-NP (generous gift from Peter Palese) and HRP-conjugated secondary antibodies, and detected by autoradiography.

Enzyme-linked immunosorbent assay (ELISA) determination of influenza-specific serum IgG-antibodies after vaccination.

Nunc Maxisorp ELISA 96-well plates were coated over night with 100 ng purified influenza PR8 virus in 100 μl PBS followed by blocking with 100 μl 1% BSA in PBS. Serial fivefold dilutions in PBS/1% BSA of mouse sera obtained before and 28 d after a single intranasal vaccination were incubated for 2 h at 22 °C. Mice were previously vaccinated with 0.01 or 0.001× LD50 of PR83F (103 PFU or 104 PFU, respectively), 0.01× LD50 of PR8 WT (100 PFU) or mock vaccinated. After four washes with PBS the wells were incubated with 1:500 of anti mouse-alkaline phosphatase conjugated secondary anti-mouse IgG antibody (Santa Cruz) for another 2 h at 22 °C. After four washes with PBS and brief rinsing with distilled water 100 μl of a chromatogenic substrate solution containing 9 mg/ml p-nitrophenylphosphate in 200 mM diethanolamine, 1 mM MgCl2, pH 9.8 was added. The color reaction was stopped by addition of an equal volume of 500 mM NaOH. Absorbance at 405 nm was read using a Molecular Devices ELISA reader. The endpoint antibody titer was defined as the highest dilution of serum that gave a signal >5 s.d. above background. Background level was determined from wells processed identically to experimental samples, in the absence of any mouse serum.