Learning from prepandemic data to forecast viral escape

1Viral diseases involve a complex interplay between immune detection in the host and viral evasion, often leading to the evolution of viral antigenic proteins. Antibody escape mutations affect viral reinfection rates and the duration of vaccine efficacy. Therefore, anticipating viral variants that avoid immune detection with sufficient lead time is key to developing optimal vaccines and therapeutics.

2Ideally, we would be able to anticipate viral immune evasion using experimental methods such as pseudovirus assays and higher-throughput deep mutational scans– (DMSs) that measure the ability of viral variants to bind to relevant antibodies. However, these experimental methods require antibodies or sera representative of the aggregate immune selection imposed on the virus, which become available only as large swaths of the population are infected or vaccinated, limiting the impact for early prediction of immune escape. In addition, as pandemic viruses can evolve rapidly (tens of thousands of new SARS-CoV-2 variants are sequenced each month), systematically testing all variants as they emerge is intractable, even without considering the effects of potential mutations on circulating strains.

3It is therefore of interest to develop computational methods for predicting viral escape that can be used to identify mutations that may emerge. An ideal model would be able to assess escape likelihood for as-yet-unseen variation throughout the full antigenic protein, would inform the design of targeted experiments, would be revised with pandemic information and would make predictions with sufficient lead time for vaccine development (that is, before immune responses to the virus are observed). However, previous computational methods for forecasting viral fitness or immune escape depend critically on real-time sequencing or pandemic antibody structures, limiting their ability to predict unseen variants and making them impractical for vaccine development during the onset of a pandemic–.

4In this work, we introduce EVEscape, a flexible framework that addresses the weaknesses of previous methods by combining a deep generative model trained on historical viral sequences with structural and biophysical constraints. Unlike previous methods, EVEscape does not rely on recent pandemic sequencing or antibodies, making it applicable both in the early stages of a viral outbreak and for continuing evaluation of emerging SARS-CoV-2 strains. By leveraging functional constraints learned from past evolution, as successfully demonstrated for predicting clinical variant effects–, EVEscape can capture relevant epistasis– and thus predict mutant fitness in the context of any strain background. Moreover, EVEscape is adaptable to new viruses, as we demonstrate in both our validation on SARS-CoV-2, HIV and influenza and in predictions for the understudied Nipah and Lassa viruses. This approach enables advance warning of concerning mutations, facilitating the development of more effective vaccines and therapeutics. Such an early warning system could guide public health decision-making and preparedness efforts, ultimately minimizing the human and economic impact of a pandemic.

5Viral proteins that escape humoral immunity disrupt polyclonal antibody binding while retaining protein expression, protein folding, host receptor binding and other properties necessary for viral infection and transmission. We built a modelling framework, EVEscape, that incorporates constraints from these different aspects of viral protein function learned from different data sources. We express the probability that a mutation will induce immune escape as the product of three probabilities: the likelihoods that a mutation will maintain viral fitness (‘fitness’ term), occur in an antibody-accessible region (‘accessibility’ term) and disrupt antibody binding (‘dissimilarity’ term) (Fig. 1a and Extended Data Fig. 1). These components are amenable to prepandemic data sources, allowing for early warning (Fig. 1b).

6First, we estimated the fitness effect of substitution mutations (subsequently referred to as mutations) using EVE, a deep variational autoencoder trained on evolutionarily related protein sequences (Supplementary Tables 1 and 2) that learns constraints underpinning structure and function for a given protein family. Consequently, EVE considers dependencies across positions (epistasis), capturing the changing effects of mutations as the dominant strain backgrounds diversify from the initial sequence–. We demonstrate the efficacy of EVE by comparing model predictions and data from mutational scanning experiments that measure several facets of fitness for thousands of mutations to viral proteins–. Model performance approaches the Spearman correlation (ρ) between experimental replicates, including viral replication for influenza (ρ = 0.53) and HIV (ρ = 0.48) (Extended Data Fig. 2 and Supplementary Tables 3 and 4). For SARS-CoV-2, we trained EVE across broad prepandemic coronavirus sequences, from sarbecoviruses including SARS-CoV-1 to ‘common cold’ seasonal coronaviruses including the alphacoronavirus NL63 (Supplementary Tables 1 and 2), and compared predictions with measures of expression (ρ = 0.45) and receptor binding (ρ = 0.26) (Extended Data Fig. 2 and Supplementary Table 4). We note that sites that expressed in the DMS experiments but were predicted to be deleterious by EVE were frequently in contact with non-assayed domains of the Spike protein or with the trimer interface, interactions not captured in the receptor-binding domain (RBD) yeast-display experiment (Extended Data Fig. 2f).

7The second model component, antibody accessibility, is motivated by the need to identify potential antibody binding sites without previous knowledge of B cell epitopes. The accessibility of each residue is computed from its negative weighted residue-contact number across available three-dimensional conformations (without antibodies), which captures both protrusion from the core structure and conformational flexibility (Supplementary Table 1). Finally, dissimilarity is computed using differences in hydrophobicity and charge, properties known to affect protein–protein interactions. This simple metric correlates with experimentally measured within-site escape more than individual chemical properties, substitution-matrix derived distance or distance in the latent space of the EVE model (Extended Data Fig. 3f). To support modularity and interpretability of the impact of each component, each term is separately standardized and then fed into a temperature-scaled logistic function (Supplementary Methods and Supplementary Tables 5 and 6).

8Extensive surveillance sequencing and experimentation prompted by the COVID-19 pandemic have presented a unique opportunity to assess the ability of EVEscape to predict immune evasion before escape mutations are observed. To test the model’s capacity to make early predictions, we carried out a retrospective study using only information available before the pandemic (training on Spike sequences across Coronaviridae available before January 2020; Supplementary Tables 1 and 2). We then evaluated the method by comparing predictions against what was subsequently learned about SARS-CoV-2 Spike immune interactions and immune escape.

9The top predicted escape mutations for the whole of Spike were strongly biased towards the RBD and N-terminal domain (NTD), coincident with the bias for antigenic regions seen in the pandemic (Fig. 2a,b and Extended Data Fig. 4). Within these domains, EVEscape scores were biased towards neutralizing regions—the receptor-binding motif of the RBD and the neutralizing supersite in the NTD (Fig. 2c and Extended Data Fig. 4d). The ability of EVEscape to identify the most immunogenic domains of viral proteins without knowledge of specific antibodies or their epitopes could provide crucial information for early development of subunit vaccines in an emerging pandemic.

10We next compared model predictions with mutations that were subsequently observed in the pandemic as deposited in GISAID (Global Initiative on Sharing All Influenza Data), which contains more than 750,000 unique sequences. For this analysis, we focused on the RBD of Spike, as this domain has been the most extensively studied owing to its immunodominance.

11Fifty percent of our top RBD predictions were seen in the pandemic by May 2023 (Fig. 3a; this proportion is robust to the threshold defining top escape mutations). The more often a mutation occurred in the pandemic, the more likely it was to be predicted by our method—66% of high-frequency observed substitutions were in the top EVEscape predictions (Fig. 3b). We expect that the highest-frequency mutations, seen in historical variants of concern (VOCs), will be enriched for escape variants that provide a fitness advantage in an immune population (while not expecting that all single substitutions in the VOCs will contribute to escape) (Fig. 3c and Extended Data Fig. 5).

12Not surprisingly, the fitness model component alone (here EVE) was better that the full EVEscape model at predicting mutations seen at low frequency in the pandemic (that is, identifying 357 versus 298 of mutations seen 100–1,000 times in the pandemic in the top quartile), probably because these mutations retain viral function but do not necessarily affect antibody binding or have a strong fitness advantage over other strains. This indicates that the immune-specific components of EVEscape may reflect important pandemic constraints not represented in models of fitness alone, and allow for mutation interpretability. For instance, VOC mutations R190S and R408S, with high EVEscape but low EVE scores, are in hydrophobic pockets that may facilitate significant immune escape (Extended Data Fig. 3c). Meanwhile, the few VOC mutations (A222V and T547K) with significant EVE—but not EVEscape—scores have known functional improvements such as monomer packing and RBD opening but do not affect escape, (Extended Data Fig. 3c). Furthermore, the proportion of EVEscape predictions seen during the pandemic increased over time—from 3% in December 2020 to 50% in May 2023 (Fig. 3a)—and should continue to increase, an expected trend both as more variants are observed and as adaptive immune pressure increases with the growing vaccinated or previously infected population. Similarly, the fraction of mutations in VOC strains with high EVEscape scores has also increased over time (Fig. 3b).

13Our model also predicted escape mutations that were subsequently observed in the pandemic in the epitopes of well-known therapeutic monoclonal antibodies under current or former emergency use authorization (Supplementary Table 7), for example, N440, E484A/K/Q and Q493R. These predictions demonstrate the interplay of our three model components; for instance, the high accessibility as well as mutability of E484 results in 50% of all possible mutations at this site in the top 2% of EVEscape predictions and includes E484A/K mutations in the top 1%—notable for escape from bamlanivimab (Fig. 3d)—because of their high dissimilarity scores. We also identify candidate escape mutations in these therapeutic epitopes that have not yet been observed at frequencies higher than 10,000—for instance variants to K444 and K417 (Supplementary Table 7), a subset of which are beginning to appear. This result indicates that escape sites could be well predicted before a pandemic and may have concrete applications for escape-resistant therapeutic design and early warning of waning effectiveness.

14EVEscape represents a significant improvement over past computational methods. EVEscape is more than twice as predictive as previous unsupervised models, both at predicting pandemic mutations (50% versus 24% of top predictions observed in the pandemic and 66% versus 17% of highest-frequency mutations predicted) and at anticipating experimental measures of antibody escape (0.53 versus 0.24 area under the precision-recall curve (AUPRC)) (Fig. 3a–c,e, Extended Data Fig. 5 and Supplementary Tables 4 and 8). All EVEscape components play a part in these predictions, with fitness predictions and accessibility metrics identifying sites of escape mutations, whereas dissimilarity identifies amino acids that facilitate escape within sites (Extended Data Fig. 3). Moreover, other computational methods, focus on near-term prediction of strain dominance rather than longer-term anticipation of immune evasion, as they rely on pandemic sequences, antibody-bound Spike structures or both, limiting their early predictive capacity. It is therefore notable that EVEscape outperforms even supervised approaches at predicting mutations seen in the pandemic (Extended Data Fig. 5 and Supplementary Table 8).

15We contextualized the performance of EVEscape in comparison with DMSs, which have been invaluable in identifying and predicting viral variants that may confer immune escape–. However, these experiments require polyclonal or monoclonal antibodies from infected or vaccinated people, limiting their early predictive capacity. For example, the DMS experiments conducted by 17 months into the pandemic (using 36 antibodies and 55 sera samples) were a third more predictive (46% versus 32% predicted mutations observed in the pandemic) than the experiments conducted 7 months previously (using just ten antibodies) (Fig. 3a, Extended Data Fig. 5 and 6 and Supplementary Table 6).

16Despite being computed on sequences available more than 17 months earlier, EVEscape was as good as or better than the latest DMS scans at anticipating pandemic variation (50% versus 46% predicted mutations observed, respectively, when considering the top decile of prediction) (Fig. 3a). As we considered higher-frequency mutations, EVEscape increasingly predicted a greater portion of pandemic variations than experiments (Fig. 3b) and predicted a larger fraction of mutations in VOC strains (Fig. 3c).

17Discrepancies between EVEscape and experiments shed light on the complementary strengths of these approaches. EVEscape and experiments missed 43 and 48 pandemic mutations, respectively, that were predicted by the other method (Fig. 4a,d). These differences could indicate model inaccuracies, or they could reflect sparse sampling of host sera response in DMS experiments, as well as artefacts from experiments testing only the RBD domain and missing the full set of in vivo constraints. Indeed, as more antibodies were incorporated in experiments, the agreement between EVEscape and experimental predictions increased (Extended Data Fig. 6d). Most of the high EVEscape predictions that were not observed in experimental predictions were in known antibody epitopes (Fig. 4b and Extended Data Fig. 3e). By contrast, those mutations identified by the experiments that were below the threshold for EVEscape predictions were often predicted to have low fitness owing to high conservation in the alignment at those positions (Supplementary Table 6).

18The consensus between EVEscape and experiments is also of interest. Agreement was especially strong for polyclonal patient sera (Supplementary Table 8); in fact, half of the top 10% of EVEscape RBD sites were sera escape sites from experiments–,, (Fig. 4c). Whereas antibody mutational scans are biased towards antibodies with potential therapeutic relevance, the escape mutations from polyclonal sera are of particular interest as they depict real pandemic selection pressures in convalescent patients and are thus crucial to considerations of reinfection and vaccine design. For instance, E484, mutated in several VOCs, had the highest experimental sera binding and was the top EVEscape predicted site.

19The modular design of our framework facilitates its adaptability to the specific characteristics of a pandemic and to new data as they become available. To consider the effects of insertions and deletions (indels) on SARS-CoV-2 Spike immune escape, we replaced the EVE fitness component with TranceptEVE, a recently developed protein large language model that has previously shown state-of-the-art performance for prediction of the effects of mutations, including indels, which both previous computational models and high-throughput experiments have been unable to capture for SARS-CoV-2. When applied to the pandemic, this model captured the most frequent single insertion and deletion, both at site 144, and each in the top decile of pandemic and random indel predictions (Extended Data Fig. 7). We also found that including glycosylation in the dissimilarity component for HIV Env, for which glycans play an important part in immune escape, improved model predictions of high-throughput experimental escape (the AUPRC increased by 10% when glycosylation was included for HIV; Extended Data Fig. 7). We also retrained EVE models with the addition of 11 million new sequences collected during the pandemic, which improved agreement with fitness DMS experiments by 20% (Extended Data Figs. 2 and 8 and Supplementary Tables 1 and 2). This model captured epistatic shifts between Wuhan and BA.2 strains, identifying changes in mutation fitness in the RBD and near BA.2 mutations, and predicting positive epistatic shifts for known convergent omicron mutations and probable epistatic wastewater mutations (Extended Data Fig. 8).

20A key application of an escape prediction framework is to identify circulating strains with high immune escape potential soon after their emergence, enabling the deployment of targeted vaccines and therapeutics before their spread. Although the World Health Organization seeks to identify new high-risk variants as they arise, new strains are occurring at an increasing rate, with tens of thousands of new SARS-CoV-2 strains each month now, a scale unfeasible for experimental risk assessment. To create strain-level escape predictions, we aggregated EVEscape predictions across all individual Spike mutations in a strain. We evaluated EVEscape strain predictions for their alignment with experimental measures of strain immune evasion, as well as their identification of known escape strains from pools of random sequences and from other strains observed at the same pandemic timepoint.

21First, we found that prepandemic EVEscape strain scores correlated well with the results of experiments quantifying vaccinated sera neutralization of 21 strains (ρ = 0.81; Fig. 5a and Supplementary Table 9) and were better than those obtained with an existing computational strain-scoring method (ρ = 0.77), even though that method used 332 pandemic antibody-Spike structures for the prediction. Second, we found that EVEscape strain scores for VOCs were consistently higher than random sequences at the same mutational depth; in particular, the Beta, Gamma, Delta, Omicron BA.4, BA.2.12.1, BA.2.75, XBB.1.5 and CH.1.1 strain scores were in the top 1% of these generated sequence scores (Extended Data Fig. 9). EVEscape strain scores for Delta and the later Omicron VOCs were also in the top 1% against sequences composed only of mutations already known to be favourable—mutations sampled from other VOCs (Extended Data Fig. 9).

22Last, we examined the ability of EVEscape to identify immune-evading strains as they emerged in the pandemic. EVEscape scores increased throughout the pandemic and were higher for more recent VOCs, reflecting their increased propensity for immune escape (Fig. 5b). Moreover, EVEscape scores for newly emerging VOCs were higher than those for almost all strains in previous time periods (Fig. 5b). Taken together, these results indicate the promise of EVEscape as an early-detection tool for picking out the most concerning variants from the large pool of available pandemic sequencing data. We therefore examine the utility of EVEscape as a tool to identify strains with high escape potential as they emerge. We classify ‘high-escape strains’ as the top decile of sequences with the highest EVEscape scores of all new and distinct strains present during a two-week surveillance window. These high-escape strains were consistently the predominant variants throughout the pandemic, constituting on average more than 40% of circulating sequences (Fig. 5c). Moreover, in the two-week windows in which the VOC strains Alpha, Beta, Gamma and Omicron BA.1 emerged, each VOC ranked first of hundreds or thousands of new strains (Fig. 5d and Extended Data Fig. 9). This demonstrates the ability of EVEscape to forecast which strains will dominate as soon as they appear after only a single observation, even as experimental testing of all emerging strains has become intractable.

23To enable real-time variant escape tracking, we make monthly predictions (Supplementary Table 9) available on our website (evescape.org), with EVEscape rankings of newly occurring variants from GISAID and interactive visualizations of probable future mutations to our top predicted strains. In sum, the EVEscape model captures relative immune evasion of successful strains and can identify concerning strains from pools of random combinations of mutations as well as from their temporal peers.

24Most viruses with pandemic potential are subjected to far less surveillance and research than SARS-CoV-2. One of the main features of EVEscape is the ability to predict viral antibody escape before a pandemic—without the consequent increase in data during a pandemic—to select vaccine sequences and therapeutics most likely to provide lasting protection, to assess strains as they arise and to provide a watch list for mutations that might compromise any existing therapies. As one of the first comprehensive analyses of escape in these viruses, we applied the EVEscape methodology to predict escape mutations to the Lassa virus and Nipah virus surface proteins; these viruses cause sporadic outbreaks of Lassa haemorrhagic fever in West Africa and highly lethal Nipah virus infection outbreaks in Bangladesh, Malaysia and India. Crucially, the three mutants present in Lassa that are known to escape neutralizing antibodies were all in the top 10% of EVEscape predictions, indicating that EVEscape captures features relevant to Lassa glycoprotein antibody escape (Fig. 5e and Supplementary Table 6). EVEscape predictions also identified ten of 11 known escape mutants to Nipah antibodies– (Extended Data Fig. 10).

25Moreover, we demonstrate generalizability to influenza hemagglutinin and HIV Env using DMS evaluation (Extended Data Fig. 6). On the basis of these findings, we provide all single mutant escape predictions for these proteins (Supplementary Table 6) to inform active and continuing vaccine development efforts with the goal of mitigating future epidemic spread and morbidity.

26One of the greatest obstacles to the development of vaccines and therapeutics to contain a viral epidemic is the high genetic diversity derived from viral mutation and recombination, especially under pressure from the host immune system. An early sense of potential escape mutations could inform vaccine and therapeutic design to better curb viral spread. Computational models can learn from the viral evolutionary record available at pandemic onset and are widely extensible to mutations and their combinations. However, new pandemic constraints (such as immunity) are unlikely to be captured. To achieve early escape prediction, EVEscape combines a model trained on historical viral evolution with a biologically informed strategy using only protein structure and biophysical constraints to anticipate the effects of immune selection. Through a retrospective analysis of the SARS-CoV-2 pandemic, we demonstrate that EVEscape forecasts pandemic escape mutations and can predict which emerging strains have high escape potential. This computational approach can preempt predictions from experiments that rely on pandemic antibodies and sera by many months while providing similar accuracy.

27EVEscape provides surprisingly accurate early predictions of prevalent escape mutations but cannot anticipate all constraints unique to a new pandemic to determine the precise trajectory of viral evolution. This method will be best leveraged in synergy with experiments developed to measure immune evasion and enhanced with pandemic data as they become available. Early in a pandemic, EVEscape can predict probable escape mutations for prioritized experimental screening with the first available sera samples—validated escape mutations could be strong candidates for multivalent vaccines. EVEscape can also identify structural regions with high escape potential, so therapeutic antibody candidates with few potential escape mutants in their binding footprint may be accelerated. Later in a pandemic, EVEscape can rank emerging strains, as well as mutants on top of prevalent strains, for their escape potential, flagging concerning variants early for rapid experimental characterization and incorporation into vaccine boosters. The model could also be augmented to leverage current knowledge on virus-specific immune targeting and mutation tolerance from experimental and pandemic surveillance data. In return, our computational framework can inform this collective understanding by proposing escape variant libraries for focused experimental investigations.

28EVEscape is a modular, scalable and interpretable probabilistic framework designed to predict escape mutations early in a pandemic and to identify observed strains and their mutants that are most likely to thrive in a populace with widespread preexisting immunity as the pandemic progresses. To this end, we provide EVEscape scores for all single mutation variants of SARS-CoV-2 Spike to the Wuhan strain, as well as scores for all observed strains and predictions of single mutation effects on the most concerning emerging strain backgrounds, with plans to continuously update with new strains. As the framework is generalizable across viruses, EVEscape can be used from the start for future pandemics, as well as to better understand and prepare for emerging pathogens. To further accelerate broad and effective vaccine development, we provide EVEscape mutation predictions for all single mutations to influenza, HIV, Lassa virus and Nipah virus surface proteins. Methods are provided in the Supplementary Information.

29Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

30Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41586-023-06617-0.