COVID-19 had a scientific head start; the next pandemic virus may be a stranger. NVIDIA and Google DeepMind are using AI to map 2,800 viruses before the next outbreak

The outbreak of COVID-19 demonstrated a potential issue that might be even more difficult to handle in future pandemics - scientists do not necessarily have many years of research ready for them once a new virus comes along.

With SARS-CoV-2, researchers were given a significant advantage because decades of previous work on coronaviruses provided them with many biological insights about the viruses, such as the structure of essential proteins.

Also Read | Scientists found something T. rex bones could never show: First known trackway reveals how the dinosaur actually walked 66.5 million years ago


The next pandemic pathogen may not offer the same advantage.

Scientists are now attempting to get ready for such situations by creating a big structural database of viruses before they turn into a health crisis. An international study group has used their technology to predict the 3D structure of protein complexes of more than 2,800 viruses which are known to infect humans.

Also Read | This atomic clock is so precise it can detect how gravity changes time. Scientists now want to take it outside the lab and measure tiny shifts across Earth

The predictions are being stored in the open-access AlphaFold database where researchers can study the structures without needing to make the predictions from scratch.

This project involves combining AlphaFold2, which is Google DeepMind’s AI-based technology to predict the structures of proteins, with NVIDIA’s BioNeMo Inference Runtime for studying viral proteomes on a large scale.

The concept is simple; if scientists already have information about the structures of proteins of the virus, they can start studying the virus even faster after the outbreak.

What scientists may learn before the next outbreak

Proteins seldom work alone. Proteins within viruses and virus-infected cells cooperate with other proteins to perform tasks that are required by viruses for replication, entry into host cells, and evasion of the immune response.

These interactions may serve as targets for drugs or vaccines.

Yet for thousands of viruses, scientists do not have experimentally determined structures for many of these molecular interactions. The new dataset attempts to close part of that gap by predicting the structures of protein complexes encoded within viral genomes.

This is because the researchers state that about 30% of the interactions that have been captured in the latest release do not have structures in the Protein Data Bank.

It does not mean that all predictions will turn out to be correct. AI-generated structures carry confidence estimates, and researchers can use laboratory experiments to test predictions that matter most.

But having a computational prediction available can give scientists a starting point rather than leaving them with a protein sequence and little information about its three-dimensional behaviour.

Joe Grove, a professor of molecular virology at the University of Glasgow’s Centre for Virus Research and a collaborator on the project, described the approach as a way of building knowledge before it is urgently needed.

“When the next pandemic happens, there may be something that comes out of the blue,” Grove said. “What we’re trying to do is stockpile some of that knowledge ahead of time.”

Why 2,800 viruses matter for pandemic preparedness

The importance of the database is partly in what it does prior to an epidemic.

Once an unknown pathogen becomes an epidemic problem, scientists must quickly determine basic information about the pathogen, including such questions as: What are the proteins that it produces? What do these proteins look like? With what molecules does it interact? What parts of the virus can be used as drug or vaccine targets?

Structural biology can supply the answers, but finding out the protein structure can take much time.

AI prediction changes the starting point.

AlphaFold2 can deduce the probable three-dimensional structure of the protein from the sequence of its amino acids. With the use of GPU computing technology, scientists can apply this method to thousands of viral proteomes instead of working on proteins individually.

The newest version of AlphaFold2 includes viral families that have the ability to infect humans, including common viruses responsible for respiratory infections and the mpox virus.

These structures are not meant to replace lab experiments; they are just experimental suggestions.

That distinction matters.

The prediction of the structural relationship between two viral proteins may provide a suggestion for how the proteins can interact; however, further study is required to verify that such an interaction exists in nature and holds significance.

However, for scientists who are facing a novel virus, a hypothesis supported by facts serves as a good starting point.

From an AI prediction to a global research resource

The project brings together Google DeepMind, NVIDIA, EMBL-EBI, the Coalition for Epidemic Preparedness Innovations, Seoul National University, Sungkyunkwan University, the Swiss Institute of Bioinformatics and the University of Glasgow.

Rather than keeping the resulting predictions within the project, the collaborators are making them available through the AlphaFold Database.

The database already contains more than 260 million predicted protein and protein-complex structures, covering a vast share of catalogued proteins known to science.

The new viral dataset adds another layer: instead of looking only at individual proteins, researchers can investigate how multiple viral proteins may fit together.

That could prove useful beyond pandemic preparedness. Structural predictions can help researchers formulate questions about viral replication, host interactions, diagnostics and possible therapeutic targets.

The collaboration is also releasing the BioNeMo Structure Prediction Pipeline used to generate the dataset. The researcher can thus utilise the GPU-aided workflow for making predictions for their protein targets.

Jo McEntyre, acting director at EMBL-EBI, said that providing the information publicly would assist researchers in virus studies in areas where fewer resources were available.

For researchers who face a new pathogen, existing structural information becomes another obstacle to be overcome.

Instead of beginning every investigation from a blank page, researchers may be able to compare an unfamiliar virus with structures already predicted for related proteins or complexes.

The bigger question: can science prepare for an unknown virus?

There is no database capable of predicting exactly which virus will cause the next pandemic.

The value of this work is more practical. It is a reservoir of biological data that may prove to be useful in case an unknown disease agent appears.

A study conducted by the Center for Global Development put forward an estimate that by the year 2050, there is a fifty per cent probability that a pandemic will occur that will be as damaging as COVID-19. The above estimation pertains to a threat in the future in regard to pandemics rather than a prediction of a disease agent.

The structural predictions should hence be regarded as a form of foresight and not as a crystal ball.

Some of the predicted complexes will eventually be confirmed experimentally. Others may be revised as new evidence appears. Some may lead researchers towards biological questions that had not previously been considered.

Chris Dallago, applied research science team lead in digital biology at NVIDIA, described the database as an engine for generating biological hypotheses.

That may be the most useful way to view the project. The next pandemic pathogen could belong to a virus family that scientists know well, or it could come from a biological corner that has received comparatively little attention.

If it is the latter, researchers will still have to do the difficult work of understanding it.

But starting with a map of possible protein structures is different from starting with a blank sheet.

And that is the bet behind the new dataset: prepare the knowledge before the emergency, because the next virus may not give scientists the same head start that COVID-19 did.

FAQ

1. What discoveries have been made with the new research on viruses?

Scientists have developed AI-predicted three-dimensional models of protein complexes in over 2,800 human-infecting viruses. The structures are being made openly available through the AlphaFold Database.

2. How does AlphaFold help study viruses?

AlphaFold2 predicts how proteins are likely to fold into three-dimensional shapes from their amino acid sequences. In this project, the approach was applied at large scale to viral proteins and their potential complexes.

3. Are the AI-generated viral structures proven to be accurate?

They are predictions, not experimental measurements. Each structure is accompanied by confidence information, and scientists can use laboratory experiments to confirm predictions, particularly those relevant to medicines, vaccines or viral biology.

4. Why map viruses before a pandemic happens?

An emerging virus may have proteins and interactions that scientists have never studied. Having predicted structures available in advance could give researchers useful starting points for understanding the pathogen and identifying questions or potential targets for further investigation.