New AI tool could help predict viral outbreaks

The model, named EVEscape, works by predicting the likelihood that a viral mutation will enable it to escape immune responses, for instance by preventing antibodies from binding. EVEscape’s underlying framework combines a deep-learning model of evolutionary viral sequences with detailed biological and structural information about the virus. In combination, this enables EVEscape to make predictions about the variants most likely to occur as the virus evolves.

Crucially, the model can anticipate new viral variants before they emerge, solely using information available at the start of an outbreak. This approach could facilitate more effective preventative action, and the design of vaccines which target variants of concern before they become prevalent.

Co-lead author for the study Pascal Notin, a DPhil student in the Oxford Applied and Theoretical Machine Learning group (OATML), part of the University of Oxford’s Department of Computer Science, said: ‘Our study shows that had EVEscape been deployed at the start of the COVID-19 pandemic, it would have accurately predicted the most frequent mutations and the most concerning variants for SARS-CoV-2.’

Predicting the future – before it arrives

This work is of tremendous value, both for pandemic surveillance efforts, but also to inform vaccine design in a way that is robust to the emergence of certain at-risk mutations. The most exciting next step for this line of work is demonstrating how it can be used in practice to inform vaccine design.

  

Pascal Notin, DPhil student, Department of Computer Sciences, University of Oxford.

In the study, the team tested the model’s capability to make early predictions by inputting only information available at the beginning of the COVID-19 pandemic, in February 2020. Based on genetic sequences for spike proteins across the Coronaviridae family of viruses, they asked EVEscape to predict what would happen with SARS-CoV-2.

EVEscape successfully predicted which SARS-CoV-2 mutations would occur during the pandemic and which would become most prevalent. Its accuracy was similar to experimental approaches that test the virus’s ability to bind to antibodies made by the immune system. Furthermore, the model predicted which antibody-based therapies would lose their efficacy as the pandemic progressed and the virus developed mutations to escape these treatments.

The team also demonstrated that EVEscape was similarly effective at predicting immune escape mutations for influenza, HIV, and understudied viruses with pandemic potential such as Lassa and Nipah. In principle, the model could be applied to any virus.

Contributing author Associate Professor Yarin Gal (OATML, Department of Computer Science, University of Oxford) said: ‘The critical aspect that makes our approach very powerful compared to traditional methods is that all the information we use in the EVEscape model is available at the very beginning of a pandemic. We developed new AI methods that do not have to wait for relevant antibodies to arise in the population to predict which variants are the most concerning.’

From EVE to EVEscape

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 – and this is what EVEscape would enable us to do.

  

Associate Professor Yarin Gal, Department of Computer Science, University of Oxford.

The core component of EVEscape is EVE (short for ‘evolutionary model of variant effect’), a deep generative model of protein sequences that helps researchers to understand which mutations preserve the fitness of a given virus. The research team originally developed EVE to predict the effects of genetic mutations on human disease risk, e.g. for cancer or heart diseases. In a previous study, EVE proved extremely accurate in predicting disease-causing mutations in humans, with comparable performance to experimental methods.

Pascal Notin said: ‘Unlike previous machine learning methods that were trained by learning from datasets that had been manually annotated by clinicians, EVE instead learns in an unsupervised manner from a large collection of evolutionarily-related protein sequences. This helps avoid the biases and overfitting risk that supervised methods are subject to, but it also means that the model can be applied effectively to rare diseases for which expert-curated labels are typically scarcer.’

But after the COVID-19 pandemic struck, the team realised that EVE could be readily adapted to predict new viral variants. The team are now applying EVEscape to SARS-CoV-2 in real time, and publish a biweekly ranking of new SARS-CoV-2 variants which is shared with entities such as the World Health Organization. Ultimately, this information could help scientists to develop more effective COVID-19 vaccines and therapies. The complete code for EVEscape is also freely available online.

The study ‘Learning from prepandemic data to forecast viral escape’ has been published in Nature.