Coronavirus Research Tracking - 11 July

Highlights of emerging COVID-19 research from the Science Media Centre team.

In this week’s Research Tracker the focus is on research on how artificial intelligence is being used to help in the pandemic. But first a mention of a useful description of how PCR works from the University of Otago.

The Research Tracker is prepared by Dr Robert Hickson for the Science Media Centre. As this is a new service, please don’t hesitate to provide feedback.

University of Otago explainer on PCR testing

Otago’s Biochemistry Department has recently created a site that nicely describes how PCR testing for SARS-CoV-2 works.

What is Artificial Intelligence?

Artificial Intelligence (or “AI”) covers a range of approaches that enable a computer system to “learn” rather than just execute simple programmes. For example, labelling new images based on images it has previously analysed or been trained with. Or answering customer questions. There are different types of AI methods. These include machine learning, neural networks, and deep learning.

Machine learning (where algorithms learn by analysing data sets labelled by humans) is a subfield of artificial intelligence. Deep learning (which can use unlabelled and unstructured data) is a subfield of machine learning, and neural networks (which attempt to mimic human brain processes using several layers of calculations) usually underpin deep learning algorithms.

Google also has an AI explainer.

High Performance Computing

While not AI, a consortium of US and other countries’ research and federal organisations and information technology companies has combined computing resources. Researchers can apply to use the computing power for Covid-19 related projects. Current projects include simulating viral spread, modeling viral proteins to identify possible drugs, and antibody design 

The pandemic is accelerating development of AI applications in healthcare

Last year an article in Clinical Infectious Diseases noted the potential that machine learning has for helping make sense of the increasing availability and complexity of health data. The authors suggested that healthcare epidemiology was on the verge of a major shift toward use of AI. The pandemic has accelerated that shift in some respects. 

AI to “tame” the literature

As mentioned in the first weekly highlights, AI is being used to analyse and categorise research papers. There are several more now available, as described in this Nature article.

Helping improve clinical management

A study published in Cell used 1182 Covid-19 patient records to compare several statistical methods and machine learning models to predict the time till patients were discharged from hospital. It found that one machine learning model outperformed the other methods in accurately predicting when patients would be discharged. It is helpful to note too that one of the algorithms tested gave random results when the model’s assumptions did not hold.

Detecting biomarkers

A well-known use of AI is in the detection of patterns that may be hard for people to spot.  Chinese researchers used machine learning to find potential biomarkers for disease progression in the blood of 485 Covid-19 patients. The research, published in Nature Machine Learning, found three useful markers – the levels of lactic dehydrogenase, lymphocytes, and high-sensitivity C-reactive protein. These could help clinicians anticipate which patients have a greater risk of dying 10 or more days ahead, and so focus more care on them. However, further research is needed.

Classifying images

Before the pandemic machine learning was already being used to help classify X-ray images. A study recently published in PLOS One tested a model on two data sets containing images from Covid-19 patients and uninfected people were tested. The model had accuracies (correctly detecting infected patients) of 96% and 98% for the two data sets, and precision (how precise the categorisation was) of nearly 99%. 

Another study (not yet peer reviewed) compared a deep learning method with two other approaches to predicting the severity of pneumonia in Covid-19 patients. The comparisons involved analysis of computer tomography (CT) images from patients with moderate or severe Covid-19. The deep learning method proved to be more efficient in identifying severity (that is, it required less information to make an accurate assessment). 

Using AI to find useful drugs

Several studies are reporting using AI methods to predict compounds that could be useful in eliminating the virus or retarding cell infections. 

Chinese researchers used a deep learning model to predict the binding affinity of 3,400 FDA approved drugs to several SARS-CoV-2 proteins. Their model compared well to other AI and traditional methods, without the need to know structural information. The model predicted that Atazanavir, remdesivir, and Kaletra can inhibit SARS-CoV-2. But further laboratory and clinical research is required to validate the model’s findings. The study is published in the Computational and Structural Biotechnology Journal.

In this pre-print paper machine learning was used to identify previously approved compounds that could inhibit SARS-CoV-2. Sixty three compounds (not a large number) were assessed. Two of seven compounds tested using laboratory assays demonstrated antiviral activity against SARS-CoV-2. 

An approach called IDentif.AI analysed data from twelve drug candidates to see which combinations could be helpful for treating Covid-19. There were over 530,000 possible drug combinations. The algorithm suggested that the optimal combination therapy against SARS-CoV-2 was remdesivir, ritonavir, and lopinavir. 

The combination was tested on cell lines (although not ones that seem to be infected with the virus) and the authors suggest that cytotoxicity is 6.5-fold better than remdesivir alone. Clinical testing is required to demonstrate effectiveness and safety. The authors are vague on how IDentif.AI works. This paper hasn’t yet been peer-reviewed. 

Benevolent AI is also using AI to predict which already approved drugs could be used to treat Covid-19. One compound, Baricitinib (a kinase inhibitor) used for rheumatoid arthritis, showed promise. Subsequent testing on four patients with Covid-19 pneumonia demonstrated that it had anti‐cytokine effects and reduced viral replication. Further testing of the compound is necessary. Published in EMBO Molecular Medicine

An initiative called AI cures, based at the Massachusetts Institute of Technology, has been established to help foster collaboration to find antiviral treatments using machine learning. They too are looking at SARS-CoV-2. While some projects search for useful drug combinations, others are using AI to design new and safe compounds. They have not yet reported promising new compounds. 

Devices and AI diagnostics

A neural network is being used to detect unusual heart contractions from electrocardiogram data. This approach, developed by Eko Health, was first described in Nature last year. In May the algorithm was given emergency use approval from the FDA for Covid-19 patients.

Researchers at Northwestern University in the US are developing a small wireless device that is worn on the throat. It continuously monitors coughing intensity and patterns, chest wall movements, respiratory sounds, heart rate and body temperature. The data is uploaded to a secure cloud and algorithms analyse the information. The patient's doctor or a hospital can be notified if symptoms seem to be getting worse. 

Analysing phylogenies

An interesting approach for machine learning was to compare virus sequences without aligning them. This study, published in PLOS One, reported accurate classifications of over 5000 viral sequences (including 29 SARS-CoV-2 genomes). Personally, I’d be cautious of this approach until the models can account for different patterns and rates of change across genomes and between strains and species. 

Predicting new epidemics

Ecologists and epidemiologists are also using AI methods to analyse data on environmental conditions, animal ecology and zoonotic diseases to forecast future outbreaks

But not all AI methods are better than existing methods

An article in Science in May highlighted that care needs to be taken when assessing the latest results of advances in artificial intelligence (generally, not just in relation to Covid-19). The “advances” may in fact only be minor, and comparisons between models may not be legitimate or accurate.  

AI methods also can’t distinguish cause from effect, yet. 

More collaboration required for applications of AI

A paper in Nature Machine Learning notes that the urgency in applying data analytics and AI to help address the Covid-19 pandemic creates ethical issues and risks through how patient data and privacy are used and protected.

The authors propose a new approach that they call “ethics with urgency” to ensure AI can be safely and beneficially used in COVID-19 responses. They do not describe how to do it, but note that a diversity of perspectives needs to be involved. 

A commentary in Nature also notes that while AI approaches to Covid-19 are increasing there is the need for greater cooperation. This involves not only the sharing of data sets, but also models & algorithms, and more multidisciplinary research so that the methods address important clinical challenges and have clinical utility.

A similar perspective is presented in The Lancet. This cautions that there is a risk of creating machine learning models that don’t properly consider practical application and potential biases. For example, algorithms may be trained to detect a specific abnormality, while radiographers and clinicians look for a range of symptoms or signs to inform diagnoses. Data sets that are not representative of the population they are intended to be used for also creates biases. 

Questions to ask when reading applications of AI 

Technology Review suggests five questions to ask when reading about developments in artificial intelligence: 

  1. What is the problem it’s trying to solve?

  2. How is the company or lab approaching that problem with AI methods?

  3. How do they source the training data?

  4. Do they have processes for auditing the products and results?

  5. Should they be using AI methods to solve this problem?

Additional questions should include:

  1. Do they explain their method(s) clearly for a more general audience?

  2. Do they discuss limitations and potential biases?

Other SMC resources