AI Model EVE-Vax Predicts Viral Mutations to Enhance Future Vaccine Development

Building vaccines for future versions of a virus

Effective vaccines dramatically changed the course of the COVID-19 pandemic, preventing illness, reducing disease severity, and saving millions of lives. However, five years later, SARS-CoV-2 is still circulating, and in the process, evolving into new variants that require updated vaccines to protect against them. But it takes time to design, manufacture, and distribute a new vaccine, which raises an important question: How can scientists create vaccines for versions of the virus that haven't happened yet? One solution comes from a predictive AI model called EVE-Vax built by a team of scientists at Harvard Medical School, the HMS-led Massachusetts Consortium on Pathogen Readiness (MassCPR), and other institutions. The new model, described May 8 in Immunity, uses evolutionary, biological, and structural information about a virus to predict and design surface proteins likely to occur as the pathogen mutates. The researchers successfully applied EVE-Vax to SARS-CoV-2, designing viral proteins that elicited similar immune responses as the actual proteins that evolved during the COVID-19 pandemic. The research, which was supported in part by federal funding, suggests that the model provides valuable information about the future evolution of a virus and can be used to create panels of "designer" proteins to evaluate the future protection of vaccines. The researchers hope that the model will eventually help scientists develop better vaccines to combat viral outbreaks and pandemics caused by rapidly mutating viruses. "We show that if you can see where a virus is evolving ahead of time, you can begin to make future-proof vaccines," said first author Noor Youssef, a scientific lead for the Predictive Modeling for Vaccine Design group in the Marks lab at HMS. The evolution of EVE Over a decade ago, study co-senior author Debora Marks, professor of systems biology in the Blavatnik Institute at HMS, and her lab began exploring whether they could use millions of years of evolutionary genetic information to make predictions about the structure and function of proteins. In a 2021 paper, the researchers described an AI model they created based on this idea. The model, called EVE (evolutionary model of variant effect), uses large-scale evolutionary data across species to predict whether proteins will be functional. When applied to humans, EVE was able to interpret gene variants as benign or disease-causing. As the COVID-19 pandemic unfolded, Marks and her team adapted their model to predict viral behavior. They built EVEscape, which applied EVE's powers of protein prediction to viral proteins. In a 2023paper, the scientists showed that had EVEscape existed at the start of the COVID-19 pandemic, it would have predicted the most frequent SARS-CoV-2 mutations and spotted the variants of greatest concern that were most likely to cause a spike in human infections. The success of EVEscape led the researchers to wonder whether their model could also forecast the future evolution of rapidly evolving viruses such as SARS-CoV-2. Vaccines for such viruses are updated annually, which requires scientists to make an educated guess up to a year in advance about how viral proteins will evolve. This can lead to a mismatch between the predicted version of a virus used to design the vaccine and the actual version that ends up circulating. The gap between prediction and reality can cause vaccines to be less effective than they would have been if the versions matched more closely. To solve this problem, the researchers developed EVE-Vax, a model that predicts and designs viral proteins that can be used to inform vaccine development ahead of time. "We wanted to see if we could use our methods to create brand-new proteins that would be functional and would have the same immune response that we see with real viruses," Youssef said. Predicting SARS-CoV-2's future maneuvers In their latest research, the scientists used EVE-Vax to design 83 brand-new versions of the "spike" protein on SARS-CoV-2, which is the main surface protein the virus uses to infect human cells. Each new version of the spike protein had a different combination of up to ten mutations. To test the effect of the AI-designed proteins, researchers teamed up with experimental colleagues and co-senior authors Jeremy Luban, professor of molecular medicine at UMass Chan Medical School; Jacob Lemieux, HMS assistant professor of medicine at Massachusetts General Hospital; and Michael Seaman, HMS associate professor of medicine at Beth Israel Deaconess Medical Center. The scientists performed experiments in lab dishes using safe, nonreplicating versions of SARS-CoV-2 engineered for the research. The experiments confirmed that viruses harboring the "designer" spike proteins infected human cells and elicited immune responses that largely matched the real-life immune responses to the virus at five different timepoints during the COVID-19 pandemic. "The fundamental insight here is that evolution tells you what's possible for the virus and its proteins to do and what might happen in the future," Marks said. Finally, the researchers showed that they could easily and cheaply engineer hundreds of new spike proteins that could be readily incorporated into vaccine development for SARS-CoV-2. "Traditional vaccine design uses all sorts of different methods, but nobody's used this approach before," Marks said. "EVE-Vax opens a new field of potential application and design for vaccines." For example, the researchers showed that with EVE-Vax, they could have predicted that there would be considerable immune escape from the COVID-19 booster vaccine targeting the Omicron variant -- knowledge that would have clued in scientists to build the booster differently. "With EVE-Vax, we can predict the immune response, instead of just what the mutations on a virus will be, which is more useful in real-world situations," Marks said. The team is now broadening EVE-Vax to other viruses, including avian influenza, a growing problem in the United States and worldwide. One key advantage of the model, the researchers noted, is that it can work with limited information, which opens it up for understudied viruses such as Lassa and Nipah, as well as newly emerging viruses. Ultimately, the team hopes that EVE-Vax will give scientists critical information about the various ways in which a virus is likely to evolve, enabling them to design vaccines that protect against future versions of the pathogen. Authorship, funding, disclosures Additional authors on the paper include Sarah Gurev, Fadi Ghantous, Kelly Brock, Javier Jaimes, Nicole Thadani, Ann Dauphin, Amy Sherman, Leonid Yurkovetskiy, Daria Soto, Ralph Estanboulieh, Ben Kotzen, Pascal Notin, Aaron Kollasch, Alexander Cohen, Sandra Dross, Jesse Erasmus, Deborah Fuller, and Pamela Bjorkman. Funding for the work was provided by the Coalition for Epidemic Preparedness Innovations (CEPI), MassCPR, and the National Institutes of Health (R37 AI147868). Marks is an advisor for Dyno Therapeutics, Octant, Jura Bio, Tectonic Therapeutic, and Genentech and is a cofounder of Seismic Therapeutic.

New AI tool predicts viral mutations to help future-proof COVID vaccines

By Dr. Priyom Bose, Ph.D.May 11 2025 Researchers have developed EVE-Vax, a computational design tool that creates synthetic SARS-CoV-2 spike proteins mimicking future immune-evading variants. These designed proteins enable early assessment of vaccine efficacy before such variants naturally emerge. Study: Computationally designed proteins mimic antibody immune evasion in viral evolution. Image Credit: TimeStopper69 / Shutterstock A recent study in the journal Immunity reports on a novel computational method (EVE-Vax), which utilizes the EVEscape framework, to design antigens that foreshadow immune escape noted in future viral variants. Novel viral variants undermine existing vaccine efficacy Rapid viral evolution constantly challenges the efficacy of medical interventions and vaccines. We currently evaluate interventions by assessing past or circulating variants. This is exemplified by recurrent breakthrough infections observed during the coronavirus disease 2019 (COVID-19) pandemic, caused by the severe acute respiratory coronavirus 2 (SARS-CoV-2) virus. This underlines the need for more proactive strategies to counter viral evolution. Frameworks predicting immune-evasive mutations could facilitate the development of viral proteins and evaluate the potency of vaccine-elicited antibodies. Both experimental and computational methods have been used to achieve this aim, but experimental methods have limitations, including often relying on patient sera that may be unavailable early in an outbreak, being limited to a subdomain of the antigen, and so on. However, these methods aid the generation of antigens comprising novel mutational combinations that evade neutralization. Computational models help overcome some of the above-mentioned limitations, as evidenced by EVEscape, a computational deep learning model, which accurately predicted immune-evading mutations across influenza, Lassa virus, SARS-CoV-2, and HIV. However, whether computational methods generate functional antigens that foreshadow immune escape has yet to be demonstrated. Graphical abstract About the study This study computationally generates and experimentally tests SARS-CoV-2 spike proteins with novel mutational combinations using the EVE-Vax design pipeline. The assessed mutations were representative of future antigenic evolution. Eighty-three novel versions of the "spike" protein on SARS-CoV-2 were designed using EVE-Vax. Each new version comprised a different combination of up to ten novel mutations relative to its background variant of concern (VOC), and some constructs contained up to 46 mutations relative to the ancestral B.1 strain. The receptor binding domain (RBD) had 57% of the mutations, while 40% were in the N-terminal domain (NTD). Five variants of concern (VOC) backgrounds were used to develop the 83 multi-mutant full-length spike constructs, namely, B.1, BA.4/5, BA.2.12.1, BA.2.75, and XBB. These were engineered as single-cycle infection pseudotypes, a method that allows for safe laboratory evaluation as the pseudoviruses are non-replicative. Computational and experimental researchers teamed up to evaluate neutralization susceptibility against polyclonal immune sera. The sera were derived from nine diverse human serum panels representing varied COVID-19 exposure histories. The spikes that EVE-Vax designed mimicked the emerging VOCs' immune-escape profiles. EVE-Vax scored the probability of antibody escape by considering three biologically relevant constraints: impact on fitness, accessibility to antibodies, and disruption potential on antibody binding. Study findings Ninety percent of the designed constructs were infectious. The eight non-infectious constructs were attributed to two main reasons: four contained a triplet of mutations (L452R, F490R, and Q493S) closer in 3-D structure than pandemic triplets, and the other four were designed using a model trained exclusively on pre-pandemic sequences. However, the 90% success rate is commendable and exceeds expected rates for randomly introduced mutations, offering insights for refining the EVE-Vax design algorithm. Spikes designed on early SARS-CoV-2 variants showed neutralization resistance like subsequent variants, with the highest resistance among actual variants relative to their parent being shown by CH.1.1 and XBB. Most variants showed higher antibody escape relative to their parental variant. The exceptions were XBB.1, BQ.1.1, and XBB.1.5 variants. The emerging variants showed an almost fourfold (3.9-fold) reduction in geometric mean ID50 titers compared to their parent variant, on average. Furthermore, variants with higher antibody escape were noted to have reduced infectivity relative to the parent variants. (A) Schematic overview of EVE-Vax for designing antigenic proteins. Single mutants within the top 1% of highest-predicted escape scores were combined to generate all possible double mutants. Double mutants were scored and further combined to create multi-mutant constructs. Designed constructs were subsequently evaluated for infectivity and neutralization sensitivity using pseudotyped virus assays. Parts of the figure were created with BioRender. (B) Cladogram depicting VOCs and computationally designed constructs (red triangles). Branch lengths are proportional to the temporal order of variant emergence. (C) Mutations across the 83 designed spike constructs mapped onto a representative 3D structure (PDB: 7BNN). Coloring indicates the frequency with which a given residue was mutated across all designed constructs. On average, EVE-vax-designed spikes exhibited a 1.9-fold reduction (with a range of 0.5 to 5.31-fold) in geometric mean ID50 titer relative to the parent variant. EVE-vax-designed constructs on particular backgrounds showed similar neutralization resistance or antibody escape when compared to SARS-CoV-2 variants that evolved naturally from the same backgrounds. Therefore, these constructs can serve as useful proxies for future SARS-CoV-2 evolution. EVE-Vax constructs were able to recapitulate antigenic profiles similar to future variants using only data available at the time of the emergence of VOCs. Earlier variant constructs showed antigenic resemblance to later emerging variants during the pandemic. For example, one B.1-background design (B.1-4a) showed a 3.9-fold reduction in neutralization sensitivity relative to B.1, exceeding the resistance of Alpha, Delta, and Gamma variants. Furthermore, the BA.2.12.1-5a designed construct mimicked the neutralizability of BA.2.75, which emerged later, and XBB designs containing the L452R or S494R mutations resembled the neutralization profile of HV.1 containing L452R. The constructs were used to evaluate the B.1-BA.4/5 bivalent booster vaccine and high titers against the BA.2.75, BQ.1, BQ.1.1, and XBB variants were indicative of ample protection. A range of antibody escape was noted concerning constructs designed on BA.2.75 and XBB. When assessing nanoparticle vaccines, they elicited higher neutralizing titers against future SARS-CoV-2 variants compared to bivalent mRNA boosters. Finally, experimental and computational approaches to predicting pandemic mutations and generating immune escape constructs were compared. Both approaches were able to detect positions that were frequently mutated at greater rates during the pandemic than random selection. Computational methods could identify most escape mutations found in experimental designs by adjusting EVE-Vax's detection threshold. On the contrary, experimental data would not likely have identified the unique mutations present in EVE-Vax constructs, proving the potential of EVE-Vax to be an alternative or complementary approach to high-throughput experimental methods. Acknowledged Limitations While these findings are promising, the researchers highlighted in the paper that the EVE-Vax method, in its current form, primarily focuses on antibody neutralization. Crucial aspects like T-cell-mediated immunity, which is important for long-term protection, have not yet been incorporated. Additionally, the success of these computational approaches hinges on sufficient evolutionary sequence data for effective model training. The generalizability of this method to all viral antigens and the potential for misuse of predictive technologies also warrant careful consideration and further exploration. Conclusions In sum, the AI tool EVE-Vax has demonstrated its capability to predict and design viral proteins that could emerge in the future. These designed constructs triggered similar immune responses concerning the SARS-CoV-2 virus, as noted in actual viral proteins that emerged during the pandemic. EVE-Vax could facilitate the development of vaccines and therapeutics to guard against future variants of viruses that are evolving rapidly. Journal reference: Youssef, N. et al. (2025) Computationally designed proteins mimic antibody immune evasion in viral evolution. Immunity. DOI: 10.1016/j.immuni.2025.04.015, https://www.cell.com/immunity/fulltext/S1074-7613(25)00178-5

Building Vaccines for Future Versions of a Virus | Newswise

AI model EVE-Vax provides clues about how a virus may evolve and the immune response it could provoke Newswise -- Effective vaccines dramatically changed the course of the COVID-19 pandemic, preventing illness, reducing disease severity, and saving millions of lives. However, five years later, SARS-CoV-2 is still circulating, and in the process, evolving into new variants that require updated vaccines to protect against them. But it takes time to design, manufacture, and distribute a new vaccine, which raises an important question: How can scientists create vaccines for versions of the virus that haven't happened yet? One solution comes from a predictive AI model called EVE-Vax built by a team of scientists at Harvard Medical School, the HMS-led Massachusetts Consortium on Pathogen Readiness (MassCPR), and other institutions. The new model, described May 8 in Immunity, uses evolutionary, biological, and structural information about a virus to predict and design surface proteins likely to occur as the pathogen mutates. The researchers successfully applied EVE-Vax to SARS-CoV-2, designing viral proteins that elicited similar immune responses as the actual proteins that evolved during the COVID-19 pandemic. The research, which was supported in part by federal funding, suggests that the model provides valuable information about the future evolution of a virus and can be used to create panels of "designer" proteins to evaluate the future protection of vaccines. The researchers hope that the model will eventually help scientists develop better vaccines to combat viral outbreaks and pandemics caused by rapidly mutating viruses. "We show that if you can see where a virus is evolving ahead of time, you can begin to make future-proof vaccines," said first author Noor Youssef, a scientific lead for the Predictive Modeling for Vaccine Design group in the Marks lab at HMS. The evolution of EVE Over a decade ago, study co-senior author Debora Marks, professor of systems biology in the Blavatnik Institute at HMS, and her lab began exploring whether they could use millions of years of evolutionary genetic information to make predictions about the structure and function of proteins. In a 2021 paper, the researchers described an AI model they created based on this idea. The model, called EVE (evolutionary model of variant effect), uses large-scale evolutionary data across species to predict whether proteins will be functional. When applied to humans, EVE was able to interpret gene variants as benign or disease-causing. As the COVID-19 pandemic unfolded, Marks and her team adapted their model to predict viral behavior. They built EVEscape, which applied EVE's powers of protein prediction to viral proteins. In a 2023 paper, the scientists showed that had EVEscape existed at the start of the COVID-19 pandemic, it would have predicted the most frequent SARS-CoV-2 mutations and spotted the variants of greatest concern that were most likely to cause a spike in human infections. The success of EVEscape led the researchers to wonder whether their model could also forecast the future evolution of rapidly evolving viruses such as SARS-CoV-2. Vaccines for such viruses are updated annually, which requires scientists to make an educated guess up to a year in advance about how viral proteins will evolve. This can lead to a mismatch between the predicted version of a virus used to design the vaccine and the actual version that ends up circulating. The gap between prediction and reality can cause vaccines to be less effective than they would have been if the versions matched more closely. To solve this problem, the researchers developed EVE-Vax, a model that predicts and designs viral proteins that can be used to inform vaccine development ahead of time. "We wanted to see if we could use our methods to create brand-new proteins that would be functional and would have the same immune response that we see with real viruses," Youssef said. Predicting SARS-CoV-2's future maneuvers In their latest research, the scientists used EVE-Vax to design 83 brand-new versions of the "spike" protein on SARS-CoV-2, which is the main surface protein the virus uses to infect human cells. Each new version of the spike protein had a different combination of up to ten mutations. To test the effect of the AI-designed proteins, researchers teamed up with experimental colleagues and co-senior authors Jeremy Luban, professor of molecular medicine at UMass Chan Medical School; Jacob Lemieux, HMS assistant professor of medicine at Massachusetts General Hospital; and Michael Seaman, HMS associate professor of medicine at Beth Israel Deaconess Medical Center. The scientists performed experiments in lab dishes using safe, nonreplicating versions of SARS-CoV-2 engineered for the research. The experiments confirmed that viruses harboring the "designer" spike proteins infected human cells and elicited immune responses that largely matched the real-life immune responses to the virus at five different timepoints during the COVID-19 pandemic. "The fundamental insight here is that evolution tells you what's possible for the virus and its proteins to do and what might happen in the future," Marks said. Finally, the researchers showed that they could easily and cheaply engineer hundreds of new spike proteins that could be readily incorporated into vaccine development for SARS-CoV-2. "Traditional vaccine design uses all sorts of different methods, but nobody's used this approach before," Marks said. "EVE-Vax opens a new field of potential application and design for vaccines." For example, the researchers showed that with EVE-Vax, they could have predicted that there would be considerable immune escape from the COVID-19 booster vaccine targeting the Omicron variant -- knowledge that would have clued in scientists to build the booster differently. "With EVE-Vax, we can predict the immune response, instead of just what the mutations on a virus will be, which is more useful in real-world situations," Marks said. The team is now broadening EVE-Vax to other viruses, including avian influenza, a growing problem in the United States and worldwide. One key advantage of the model, the researchers noted, is that it can work with limited information, which opens it up for understudied viruses such as Lassa and Nipah, as well as newly emerging viruses. Ultimately, the team hopes that EVE-Vax will give scientists critical information about the various ways in which a virus is likely to evolve, enabling them to design vaccines that protect against future versions of the pathogen. Authorship, funding, disclosures Additional authors on the paper include Sarah Gurev, Fadi Ghantous, Kelly Brock, Javier Jaimes, Nicole Thadani, Ann Dauphin, Amy Sherman, Leonid Yurkovetskiy, Daria Soto, Ralph Estanboulieh, Ben Kotzen, Pascal Notin, Aaron Kollasch, Alexander Cohen, Sandra Dross, Jesse Erasmus, Deborah Fuller, and Pamela Bjorkman. Funding for the work was provided by the Coalition for Epidemic Preparedness Innovations (CEPI), MassCPR, and the National Institutes of Health (R37 AI147868). Marks is an advisor for Dyno Therapeutics, Octant, Jura Bio, Tectonic Therapeutic, and Genentech and is a cofounder of Seismic Therapeutic.