CodonBert large language model for mRNA vaccines.

mRNA-based vaccines and therapeutics are gaining popularity and usage across a wide range of conditions. One of the critical issues when designing such mRNAs is sequence optimization. Even small proteins or peptides can be encoded by an enormously large number of mRNAs. The actual mRNA sequence can have a large impact on several properties including expression, stability, immunogenicity, and more. To enable the selection of an optimal sequence, we developed CodonBERT, a large language model (LLM) for mRNAs. Unlike prior models, CodonBERT uses codons as inputs which enables it to learn better representations. CodonBERT was trained using more than 10 million mRNA sequences from a diverse set of organisms. The resulting model captures important biological concepts. CodonBERT can also be extended to perform prediction tasks for various mRNA properties. CodonBERT outperforms previous mRNA prediction methods including on a new flu vaccine dataset.

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

Human H3N2 influenza viruses are subject to rapid antigenic evolution which translates into frequent updates of the composition of seasonal influenza vaccines. Despite these updates, the effectiveness of influenza vaccines against H3N2-associated disease is suboptimal. Seasonal influenza vaccines primarily induce hemagglutinin-specific antibody responses. However, antibodies directed against influenza neuraminidase (NA) also contribute to protection. Here, we analysed the antigenic diversity of a panel of N2 NAs derived from human H3N2 viruses that circulated between 2009 and 2017. The antigenic breadth of these NAs was determined based on the NA inhibition (NAI) of a broad panel of ferret and mouse immune sera that were raised by infection and recombinant N2 NA immunisation. This assessment allowed us to distinguish at least four antigenic groups in the N2 NAs derived from human H3N2 viruses that circulated between 2009 and 2017. Computational analysis further revealed that the amino acid residues in N2 NA that have a major impact on susceptibility to NAI by immune sera are in proximity of the catalytic site. Finally, a machine learning method was developed that allowed to accurately predict the impact of mutations that are present in our N2 NA panel on NAI. These findings have important implications for the renewed interest to develop improved influenza vaccines based on the inclusion of a protective NA antigen formulation.

Two proteins, the hemagglutinin and the neuraminidase, protrude from the surface of the influenza virus. Their detection by the immune system allows the host organism to mount defences against the viral threat. The virus evolves in response to this pressure, which manifests as changes in the appearance of its hemagglutinin and neuraminidase. This process, known as antigenic drift, leads to the proteins evading detection. It is also why flu vaccines require frequent updates, as they rely on 'training' the immune system to recognise the most important strains in circulation ­ primarily by exposing it to appropriate versions of hemagglutinin. While the antigenic drift of hemagglutinin has been extensively studied, much less is known about how the neuraminidase accumulates mutations, and how these affect the immune response. To investigate this question, Catani et al. selected 43 genetically distant neuraminidases from human viral samples isolated between 2009 and 2017. Statistical analyses were applied to define their relatedness, revealing that a group of closely related neuraminidases predominated from 2009 to 2015, before they were being taken over by a second group. A third group, which was identified in viruses isolated in 2013, was remarkably close to the neuraminidase of strains that circulated in the late 1990s. The fourth and final group of neuraminidases was derived from influenza viruses that normally circulate in pigs but can also occasionally infect humans. Next, Catani et al. examined the immune response that these 43 neuraminidases could elicit in mice, as well as in ferrets ­ the animal most traditionally used in influenza research. This allowed them to pinpoint which changes in the neuraminidase sequences were important to escape recognition by the host. Data obtained from the two model species were comparable, suggesting that these experiments could be conducted on mice going forward, which are easier to work with than ferrets. Finally, Catani et al. used machine learning to build a computational model that could predict how strongly the immune system would respond to a specific neuraminidase variant. These findings could help guide the development of new vaccines that include neuraminidases tailored to best prime and train the immune system against a larger variety of strains. This may aid the development of 'supra-seasonal' vaccines that protect against a broad range of influenza viruses, reducing the need for yearly updates.

AF03 adjuvant improves anti-hemagglutinin and anti-neuraminidase immune responses induced by licensed seasonal quadrivalent influenza vaccines in mice.

Seasonal influenza remains a serious public health concern as the viral infection spreads easily from person to person and due to antigenic drift of neutralizing epitopes. Vaccination is the best method for disease prevention, however current seasonal influenza vaccines stimulate antibodies which are often effective against only antigenically similar strains. To boost the immune responses and increase vaccine effectiveness, adjuvants have been used for the past 20 years. The current study explores the use of oil-in-water adjuvant, AF03 to improve an immunogenicity of 2 licensed vaccines. A standard-dose inactivated quadrivalent influenza vaccine (IIV4-SD), containing both hemagglutinin (HA) and neuraminidase (NA) antigens, and recombinant quadrivalent influenza vaccine (RIV4), containing only HA-antigen were adjuvanted with AF03 in naïve BALB/c mouse model. Functional HA-specific antibody titers against all four homologous vaccine strains were enhanced by AF03, indicating potential increase in protective immunity. An increase in HA-specific total immunoglobulin G (IgG) binding titers were detected against homologous HAs, heterologous panel of 30 H3 HAs and seven Influenza B HAs. The neuraminidase inhibition (NAI) activity was significantly higher in IIV4-SD-AF03 group. Use of AF03 adjuvant improved the immune response to two influenza vaccines in a mouse model via an increase in functional and total antibodies against NA and a broad panel of HA-antigens.

A computationally designed H5 antigen shows immunological breadth of coverage and protects against drifting avian strains.

Since the first identification of the H5N1 Goose/Guangdong lineage in 1996, this highly pathogenic avian influenza virus has spread worldwide, becoming endemic in domestic poultry. Sporadic transmission to humans has raised concerns of a potential pandemic and underscores the need for a broad cross-protective influenza vaccine. Here, we tested our previously described methodology, termed Computationally Optimized Broadly Reactive Antigen (COBRA), to generate a novel hemagglutinin (HA) gene, termed COBRA-2, that was based on H5 HA sequences from 2005 to 2006. The COBRA-2 HA virus-like particle (VLP) vaccines were used to vaccinate chickens and the immune responses were compared to responses elicited by VLP's expressing HA from A/whooper swan/Mongolia/244/2005 (WS/05), a representative 2005 vaccine virus from clade 2.2. To support this evaluation a hemagglutination inhibition (HAI) breadth panel was developed consisting of phylogenetically and antigenically diverse H5 strains in circulation from 2005 to 2006, as well as recent drift variants (2008 - 2014). We found that the COBRA-2 VLP vaccines elicited robust HAI titers against this entire breadth panel, whereas the VLP vaccine based upon the recommended WS/05 HA only elicited HAI responses against a subset of strains. Furthermore, while all vaccines protected chickens against challenge with the WS/05 virus, only the human COBRA-2 VLP vaccinated birds were protected (80%) against a recent drifted clade 2.3.2.1B, A/duck/Vietnam/NCVD-672/2011 (VN/11) virus. This is the first report to demonstrate seroprotective antibody responses against genetically diverse clades and sub-clades of H5 viruses and protective efficacy against a recent drifted variant using a globular head based design strategy.

Design and Characterization of a Computationally Optimized Broadly Reactive Hemagglutinin Vaccine for H1N1 Influenza Viruses.

UNLABELLED: One of the challenges of developing influenza A vaccines is the diversity of antigenically distinct isolates. Previously, a novel hemagglutinin (HA) for H5N1 influenza was derived from a methodology termed computationally optimized broadly reactive antigen (COBRA). This COBRA HA elicited a broad antibody response against H5N1 isolates from different clades. We now report the development and characterization of a COBRA-based vaccine for both seasonal and pandemic H1N1 influenza virus isolates. Nine prototype H1N1 COBRA HA proteins were developed and tested in mice using a virus-like particle (VLP) format for the elicitation of broadly reactive, functional antibody responses and protection against viral challenge. These candidates were designed to recognize H1N1 viruses isolated within the last 30 years. In addition, several COBRA candidates were designed based on sequences of H1N1 viruses spanning the past 100 years, including modern pandemic H1N1 isolates. Four of the 9 H1N1 COBRA HA proteins (X1, X3, X6, and P1) had the broadest hemagglutination inhibition (HAI) activity against a panel of 17 H1N1 viruses. These vaccines were used in cocktails or prime-boost combinations. The most effective regimens that both elicited the broadest HAI response and protected mice against a pandemic H1N1 challenge were vaccines that contained the P1 COBRA VLP and either the X3 or X6 COBRA VLP vaccine. These mice had little or no detectable viral replication, comparable to that observed with a matched licensed vaccine. This is the first report describing a COBRA-based HA vaccine strategy that elicits a universal, broadly reactive, protective response against seasonal and pandemic H1N1 isolates. IMPORTANCE: Universal influenza vaccine approaches have the potential to be paradigm shifting for the influenza vaccine field, with the goal of replacing the current standard of care with broadly cross-protective vaccines. We have used COBRA technology to develop an HA head-based strategy that elicits antibodies against many H1 strains that have undergone genetic drift and has potential as a "subtype universal" vaccine. Nine HA COBRA candidates were developed, and these vaccines were used alone, in cocktails or in prime-boost combinations. The most effective regimens elicited the broadest hemagglutination inhibition (HAI) response against a panel of H1N1 viruses isolated over the past 100 years. This is the first report describing a COBRA-based HA vaccine strategy that elicits a broadly reactive response against seasonal and pandemic H1N1 isolates.