The Stanford HAI associate director discusses how algorithms can track infections and identify vaccine or drug candidates.
NIAID-RML / Flickr
Since the COVID-19 health crisis began, researchers have explored how artificial intelligence can track and treat the virus. Although models haven’t yet become widespread, data scientists have experimented with tools to estimate needed hospital beds, predict which patients could become critically ill, and search for possible vaccines. Stanford HAI associate director and medical doctor Russ Altman, a professor of bioengineering at Stanford University, answered readers’ questions on Quora about the role AI could play in this pandemic. Here are some of his main takeaways (and read all his responses or follow him here).
What’s the possible algorithm for AI to find potential infections?
There is fascinating work going on trying to use AI methods for detecting early infections. The basic idea is that sensors in wearables (phones, watches, etc.) may be able to pick up subtle changes in physiology (change in heart rate, change in blood pressure, fever) that are early warning signs for COVID infection. This would be great because it would allow the use of wearables to find patients early, bring them to treatment, and isolate them from others.
The technical challenges for doing this research right now are substantial. First, AI programs rely on a “training” set of known positive cases and negative cases so they can compare the sensor data to discover the differences. Assuming the differences are subtle, it may require many cases and controls to learn the reliable signals that differentiate them. Second, any algorithm that is created needs to be carefully checked in different settings and with different populations to be sure that it doesn’t just work for one kind of person/setting but not for others. Third, we need to make sure that the performance of the algorithms has a tolerable rate of “false positives” (where the algorithm says “infected” but the person isn’t) and “false negatives” (where the algorithm says “patient is okay” but the person is infected).
The possibilities for these kinds of algorithms are very powerful and not to be dismissed, but they must be built and evaluated very cautiously because premature deployment could cause more harm than good.
What is AI’s role in finding a vaccine for COVID-19?
The application of AI to vaccine development is still at an early stage, but a variety of AI-based approaches have the potential to aid COVID-19 vaccine development.
First, a 3-D structure of viral proteins can be very useful for vaccine design. Sometimes the 3-D structures are not available. Machine learning models can predict viral protein structure before molecular biologists have experimentally solved it. Understanding such protein structure provides an important foundation for reasoning about the development of vaccines and other drugs. For example, SWISS-MODEL, a structural bioinformatics web-server focused on making protein modeling accessible to researchers, computationally predicted the structure of all SARS-CoV-2 proteins (SARS-CoV-2 is the virus that causes COVID-19). These predictions were later shown to have high consistency with experimentally solved structures.
Second, other machine learning models can predict which components of a virus are more likely to be immunogenic — or create a good immune response that is critical for long-term vaccine effectiveness. As one example, our research team at Stanford predicted several antibody binding sites on the receptor binding domain of the viral spike protein and recommended that this region be included in COVID-19 vaccine attempts (these are preliminary findings and still must be peer-reviewed). Currently, there are more than five vaccine trials focusing on SARS-CoV-2 spike protein.
Third, in the longer term, AI tools might help us understand how humans develop immunity against COVID-19. Neural network models like NetMHCpan and MARIA can predict which parts of a viral protein seem to attract the attention of the human immune system, which decides what cells are infected and may be destroyed. This provides a starting point for a much deeper understanding of why some people are not protected by a vaccine and how its protective capacity might be improved.
How is AI used in drug discovery?
Drug discovery is typically characterized by the following phases: In target discovery, researchers look for the molecule that the drug will bind and whose function the drug will modulate. During preclinical studies, experimental tests in cells and animals are performed to get a sense of whether the drug will work and whether it will be safe. Phases I, II, and III test the drug’s effectiveness and safety in humans, and Phase IV tracks the drug after it is on the market to be sure that the drug does not cause side effects.
AI is potentially useful in all these phases. For example:
- AI can help target discovery by looking at large amounts of physiological and cellular data, the interactions between all the molecules, and examples of previous successful drugs to seek targets that seem to share characteristics with successful targets for other diseases.
- AI can help preclinical studies by sometimes replacing actual experiments with computational experiments that may approximate the same quality of results, but be much cheaper and faster.
- AI can help in Phase I, II, and III studies by helping the drug developers make sure that they are not missing subtle patterns of toxicity in the patients or signs that the drug may not work.
- AI can help in Phase IV by looking for patterns in patient response that may indicate harmful side effects that are either not obvious to clinicians or occur so rarely that physicians miss them.
There is great variability in the degree to which drug companies have adopted AI methods, and the jury is still out about how beneficial AI technology will be in making drug discovery faster and more successful. But there is no doubt that improvements are needed because current drug discovery methods are expensive and have a pretty low success rate.
What is AI’s role in disseminating information about COVID-19?
On the positive side, AI technologies can be used to sift through information and identify articles, websites, videos, and other media that may be of interest. There has been an explosion of information about COVID-19, and the science community, the public, and governmental leaders have a real challenge keeping on top of it. AI systems that “read” or “watch” the content, filter it, summarize it, and bring it to the attention of the appropriate people are very useful.
On the negative side, AI can be used to spread bad information and target it to people in order to change their behavior (to create social discontent, to sell things of dubious value — all the usual ways that humans sometimes try to take advantage of other humans). This would be a class of AI applications that are purposefully nefarious. There are also negative impacts of well-intentioned AI if it doesn’t work well. If the AI systems are not validated or show biased behavior based on imperfect design or implementation, they may confuse or mislead people.
Like almost anything else, AI needs to be used carefully, monitored, and validated to ensure that it is behaving as designed and is making a positive contribution to the flow of information.