Scientists discuss use of artificial intelligence to combat COVID-19 and future pandemics

Models capable of predicting the spread of pathogens, algorithms that monitor health complaints on social media, and the use of big data and machine learning to speed up drug development were some of the issues discussed by experts who took part in a webinar organized by FAPESP and the Global Research Council.

By José Tadeu Arantes  |  Agência FAPESP – How were the models that predicted the worldwide propagation and evolution of COVID-19 developed? Is it possible to create artificial intelligence (AI) algorithms that can forecast and prevent future pandemics? How can AI for health be protected against misinformation and “fake news”? What contributions can AI make toward optimizing drug discovery? These were some of the questions answered by the experts who took part in a webinar entitled “Artificial Intelligence and COVID-19” and held on October 6, 2021. 

Produced by FAPESP in partnership with the Global Research Council (GRC), the event was moderated by Roberto Marcondes, a member of FAPESP’s Steering Committee for the Research, Innovation and Dissemination Centers (RIDC) Program and a professor of computer science at the University of São Paulo’s Institute of Mathematics and Statistics (IME-USP) in Brazil. The speakers in the webinar included Solange Rezende, a professor at the same university’s São Carlos Institute of Mathematics and Computer Science (ICMC-USP), and Alexandre Chiavegatto Filho, a professor at the university’s School of Public Health (FSP-USP). The rest of the lineup consisted of researchers Wagner Meira Junior and Carolina Horta Andrade, affiliated respectively with the Federal University of Minas Gerais’s Institute of Exact Sciences (ICEx-UFMG) and the Federal University of Goiás’s School of Pharmacy (FF-UFG) in Brazil.

Rezende recalled how in 2020, when little was known about the dynamics of the pandemic, her group worked on the enrichment of forecasting models by expanding databases to include data collected from news and social media, reports, bulletins, and other sources.

To glean the information of interest from the selected texts, the group used an AI tool known as a “named activity extractor”, which uses rules similar to the five Ws of good journalism (who, what, when, where, why). “News reports collected from highly selective platforms protected against fake news were tracked on an ongoing basis. Data and text mining were used for disambiguation of the terms found, and to identify, georeference and correlate events so as to detect contagion curve patterns. We adjusted the models we built on a daily basis and produced seven-day forecasts,” Rezende explained (more at: agencia.fapesp.br/33174). 

Chiavegatto Filho’s presentation focused on machine learning, and how algorithms can contribute to healthcare management by making predictions and orienting decisions that should prevent a new pandemic.

This can be done in two complementary ways, he suggested. One would be to look for relevant information on social media. “People frequently use these platforms to complain, and algorithms can detect anomalous complaints in certain regions and at certain times,” he said. “In 2019, for example, a large number of complaints about respiratory problems were detected in the region of Wuhan in China. Unfortunately, they were not taken seriously at the time.”

The other way entails monitoring health systems to find out what kinds of patient are checking in to clinics, and how frequently. “The discovery of unexpected symptoms would enable us to identify the possible emergence of a novel disease,” he explained.

Meira Junior noted the complex issue of data sources in the case of COVID-19. The issue was rapidly politicized and polarized on social media, with abundant dissemination of fake news and allegations that cases were being under- or over-reported. 

The quality of models depends on three criteria. “The first is algorithm fairness or bias, meaning that models shouldn’t discriminate,” he said. “The second is algorithm responsibility. Who or what is responsible for errors? The third is algorithm transparency, which lets health workers understand what comes out of the model and empowers them to make better decisions.” Considering these criteria, he stressed the need for existing systems to be revamped.

In the last presentation of the webinar, Andrade discussed AI-driven drug discovery with the aim of abbreviating a process that can take 15 years or more, and cutting its traditionally high cost. “For every 10,000 compounds considered in the first stage, only one drug is approved, and the investment needed to bring a drug to market is estimated at between USD 2 billion and USD 13 billion,” she said.

AI can be used in several stages, such as planning and molecule design, for example. “A new drug was recently designed, synthesized and validated in only 46 days thanks to the use of AI,” Andrade noted.

As a starting point, public repositories such as PubChem and ChEMBL already have huge databases containing chemical and biological data on more than 100 million compounds, each with at least one recognized biological activity. “With supervised machine learning, it’s possible to use this data to build models that can be statistically validated and potentially used to predict novel molecules,” she said. 

A recording of the webinar on “Artificial Intelligence and COVID-19” can be watched at: www.youtube.com/watch?v=vKJMuWz2IpA.