AI and the hunt for a COVID-19 vaccine

The use of AI and machine learning will be instrumental in the race to find a vaccine for the COVID-19 pandemic, writes a team of CSIRO scientists.

Artificial intelligence (AI) is a collection of powerful interrelated technologies that can be used to perform complex tasks autonomously and without explicit human guidance. Machine learning (ML) lies at the core of AI, and when combined with other technologies such as computer vision, natural language processing and robotics represents a potential step change in what’s possible for humanity.

Since the launch of Australia’s Artificial Intelligence Roadmap in November 2019, we’ve seen an ever-growing repository of practical AI applications. AI represents a new way of applying science and will help humanity tackle some of our most stubborn and significant challenges. AI will assist us to solve the current COVID-19 crisis and help to protect us from future disease risks.

Perhaps one of the most immediate and urgent applications of AI is the development of COVID-19 vaccines and therapeutics. In January, Google’s DeepMind published an article about its ML system, AlphaFold, in research journal Nature. AlphaFold can predict the 3D structure of a protein based on its genetic sequence far more quickly and accurately than pre-existing approaches. Knowing the structure of a protein associated with a virus allows scientists to develop vaccines and therapeutics. In March, DeepMind released structure predictions of several under-studied proteins associated with SARS-CoV-2, the virus that causes COVID-19.

CSIRO has been working on SARS CoV-2 since January, and on 2 April, announced it was testing potential COVID-19 vaccines at the Australian Centre for Disease Preparedness (CDP) in Geelong, Victoria. The process is expected to take three months. CDP Scientists and technicians are working to produce 50 litres of a candidate vaccine, a specially designed protein developed by the University of Queensland. This will provide about 20,000 vaccine doses for use in the first phase of clinical testing. Additional testing is being done to determine the best way of administering the vaccine.

Underpinning the CPD vaccine trials is research by CSIRO’s Australian e-Health Research Centre (AEHRC) examining the genome of COVID-19. The genome can be thought of as the virus’ blueprint. With several hundred virus strains decoded and deposited into a global repository daily, the dataset for this genomic analysis has rapidly expanded to tens of millions of data points. Each data point potentially holds information about the virus’ future trajectory and virulence of individual strains.

To manually study each data point would be impossible, like searching for a needle in a haystack. Advanced AI-related technologies are needed to attack this dataset. The AEHRC is using ML to map the evolutionary landscape the virus occupies and estimate its trajectory to help choose the most representative strain on which to test the vaccines.

Better forecasting

Within the field of epidemiology — the study of how diseases spread — AI is being used to improve the accuracy and timeliness of epidemiological forecasts. For example, researchers at the University of Texas used an AI ML approach — a “modified stacked auto-encoder” — to forecast COVID-19 cases across China from 11 January–27 February. When actual COVID-19 cases were compared to predicted ones, average accuracy ranged from 97.7 to 99.3 per cent. The researchers conclude that, if the data are correct and there are no “second transmissions” they can accurately forecast transmission dynamics of COVID-19 within Chinese provinces.

Another forecasting system developed by Data61 uses AI tools such as ML, natural language processing, data science and time series modelling to forecast disease spread from social media. The system was used to predict the 2014 Ebola epidemic in Africa three months before it happened. Tracking keywords such as “fever”, “cough”, “headache” and “head cold” has been successful in early outbreak detection of epidemics. In the current COVID-19 crisis, Data61 and AEHRC are working with state health organisations for early outbreak detection from social media, patient testing and hospitalisation data. Better forecasts enable better containment policies and fewer deaths.

There are limitations with the use of social media and internet search terms in predicting influenza. Researchers found Google flu trends alone to be an unreliable forecasting model. However, more recent AI-enabled research, which combines search term data with traditional epidemiological data, is promising. These approaches make use of cutting-edge ML techniques. For example, a recently developed ML forecasting model — ARGO2 — combines Google search terms with disease surveillance data from the US Centers for Disease Control and Prevention. Researchers observed a 30 per cent error reduction in real-time regional tracking of influenza across the US.

Key to reducing the number of deaths from COVID-19 will be ensuring our hospital system is prepared. AI can help achieve this. The AEHRC data analytics team has built models to understand the flow of patients through the healthcare system and the impact on different parts of the hospital. This has shown early discharge planning can improve patient movement from the emergency department to specialist wards, reducing the “access block” in the hospital. The team has also developed forecasting models to predict and track epidemics. In the case of COVID-19, researchers from AEHRC are working with various state health organisations, using various AI techniques to forecast the number of patients likely to need hospitalisation, hospital beds and ventilators, and helping ambulance services predict demand.

Addressing the future

To mitigate the risks of future pandemics, ML is being used for the detection and prediction of reservoirs of zoonotic (of animal origin) disease. As with SARS and H1N1, COVID-19 is a zoonotic disease. According to WHO, 61 per cent of all human infectious diseases are zoonotic — as are 75 per cent of emerging infectious diseases. In a recent study, researchers at the Cary Institute of Ecosystem Studies and the University of Georgia in the US used ML to identify species of rodents, and geographic locations, harbouring zoonotic pathogens. This can be used to help in surveillance, vector control and support research into effective vaccines and therapeutics. A recent review paper published in the Clinical Infectious Diseases journal describes several documented ways ML is being used to combat infectious diseases. These include predicting clinical outcomes for Ebola patients, predicting patients at risk of developing septic shock, and predicting the risk of infectious diarrhoea in patients who are hospitalised.

One of the biggest challenges facing healthcare authorities is ensuring citizens are receiving and using accurate information so they can take precautions. A large amount of misinformation is circulated on social media about COVID-19 — such as ineffective/dangerous treatments and preventative measures and over- or understated risks. Data61 has developed an algorithm that uses ML and natural language processing to identify and filter misinformation on Twitter. It was successfully used to identify bot-generated misleading posts on Twitter during the 2019–20 bushfires. The tool can also alert authorities to misleading information.

AI represents a novel, rapidly growing and powerful tool within the scientist’s toolkit — a new way of thinking and problem-solving. Corporate boards must ramp up science and technology investment to solve the COVID-19 crisis, stop future outbreaks and tackle a wide range of pressing societal and business problems.

This paper was written by a team of CSIRO scientists – Stefan Hajkowicz, David Hansen, Denis Bauer and Sankalp Khanna – working in the fields of AI, strategic foresight, bioinformatics, ML, data science and health informatics. The Australian Centre for Disease Preparedness is tasked with protecting Australians from infectious diseases. Visit the CSIRO website for more information.