How AI can speed scientific discovery, from predicting virus variants to vital protein research

The rapid ascent of generative artificial intelligence (AI) over the past year has raised many red flags, from the risk of disinformation to replacing jobs. But could AI be a force for good, too? This was a topic debated at the World Economic Forum’s recent Annual Meeting of the Global Future Councils. Here are some of the practical examples speakers highlighted during the session.

Viruses are masters of disguise. They are evolving to evade the human immune system by spawning new variants to get around our defences, as the COVID-19 pandemic has demonstrated. What, though, if you could predict a virus’s next move before it makes it? A new AI tool promises to do just that.

Developed at Harvard Medical School, EVEscape combines evolutionary and biological information to predict how a virus might adapt to escape our immune defences. In a study published in Nature, researchers report that if the tool had been deployed at the outset of the pandemic, it would have predicted the most frequent mutations and picked out the most serious variants of Sars-CoV-2. It was similarly on the mark about the evolution of other viruses such as HIV and influenza.

EVEscape is now being used to anticipate future variants of Sars-CoV-2, and the researchers are also applying the tool to other viruses.

Anyone who has left their PC on at night to provide computing power to Folding@Home as a citizen scientist will be aware of the challenge of protein “folding” – finding the structure of proteins.

With hundreds of millions of known proteins – and new ones added every year – determining their specific shape and function is a lengthy and costly exercise using multi-million dollar equipment such as nuclear magnetic resonance. This has impeded research and, consequently, the ability to treat diseases, according to Alphabet-Google.

The company applied its DeepMind AI technology to the protein-folding challenge with AlphaFold, teaching it the structures of about 100,000 known proteins. This enabled the tool to predict the shape of a protein almost immediately, with atomic accuracy. In July 2022, Google announced that the tool had predicted the structures of virtually all known proteins. Today, it’s being used to discover new diseases, develop a more effective malaria vaccine and new cancer drugs as well as tackle antibiotic resistance.

Before any original research happens, scientists start with a literature review. Exploring the results of existing studies, to form their research premise and find their starting point, involves trawling through reams of material for days or even weeks.

With the Elicit AI research assistant, scientists can significantly speed up this gruelling process. Elicit is a large language model that helps automate the literature review process. All they have to do is enter a research question and the AI will identify the top papers on this topic. It will then summarize the key findings and extract information into a research matrix that scientists can work from.

This is particularly useful for meta-analyses. Here, researchers explore an existing body of research on a particular condition, disease or drug, for example, and aggregate the results. These analyses not only reflect the current understanding of the topic in question but often also deliver new insights based on the analysis.

Elicit claims that more than 800,000 researchers have used Elicit to date.

While participants in the session flagged many of the questions that still need to be addressed when it comes to AI governance, there was agreement that AI has great potential to be a force for good.

Increasing the pace of scientific discovery by an order of magnitude is not limited to the life sciences. The Guardian reported earlier this month that researchers are using AI to decipher ancient scrolls from Herculaneum destroyed after the eruption of Mount Vesuvius in AD79. The AI system was able to identify the first word inside their blackened remains, 2,000 years on.