Evolving pandemic, evolving questions
As pandemic stretched on, newbie modelers dropped out, leaving experienced modelers applying their knowledge in an environment of evolving information and evolving questions.
As a member of the World Health Organization’s epidemic response R&D Blueprint, Halloran was briefed on the emerging coronavirus on January 9, 2020. Officials wanted to know where the virus could spread. Halloran immediately reached out to Dr. Alessandro Vespignani, a collaborator at Northeastern University, to use their modeling tool, the Global Epidemic and Modeling project, or GLEAM, to project where in the world the virus could be headed as it spread from China. The team began the work before the virus received its official name, SARS-CoV-2, and submitted their first paper, which appeared in the journal Science, in late January 2020.
Schiffer knew that the same question he’d addressed with HHV-6 — how much virus it takes to infect a new host — would be an even more important issue with SARS-CoV-2. So he, Dr. Ashish Goyal, a postdoctoral research associate in his lab, and Dr. Bryan Mayer, a BBE staff scientist with whom Schiffer had worked on HHV-6, drew on what data they could to create a similar model for the novel coronavirus.
Matrajt models how to optimize vaccine distribution — not a top priority when none exist. She initially thought her scientific involvement with SARS-CoV-2 would end with the paper that resulted from her tweet. But as vaccine developers did their best to outrace the pandemic, and candidates moved into trials at record speed, Matrajt realized she had more expertise to offer.
She’s published work in the journal Science Advances that addresses under what conditions coronavirus vaccines should be targeted at those most at risk of disease, or those most likely to spread the infection. Matrajt is currently investigating the conditions under which strategically deployed single-dose vaccines could help policy makers, faced with vaccine shortages, optimize vaccine allocation.
One of the biggest challenges to modeling the coronavirus at the beginning was the dearth of information about the unknown virus. A lot of the information that scientists studying the flu, for example, have been able to glean from past outbreaks was entirely missing.
“We knew it was a coronavirus, but we didn't know how infectious is it? What’s the natural history you have to put in your models? … We had to work with a lot of uncertainty because we didn't know anything about the disease in January 2020, really,” Halloran said.
In the earliest days of the pandemic, it wasn’t even clear whether the novel coronavirus spread by jumping from human to human, as opposed to jumping to humans from its animal hosts — let alone whether contaminated objects or coughed-up mucus droplets were primary virus carriers. It would be nearly impossible for a virus that couldn’t spread between humans to spread around the world, so scientists trying to model a potentially global pandemic added the presumption of human transmission to their initial models. They also presumed that all transmission originated from symptomatic infections. Now, said Matrajt, it appears that roughly 40% of infections may be asymptomatic.
Over time, as SARS-CoV-2 became one of the world’s most-studied viruses, scientists had to contend with a different problem: reams of data, some of it of dubious quality.
“The availability of good data has been very hard to predict,” said Schiffer.
Prior to the pandemic, his models drew on data produced by scientists with whom he had a longstanding relationship, whose scientific standards he trusted.
“Now we are emailing people we don't know across the globe and asking for data — and getting every possible type of response you could imagine,” Schiffer said.
Even as the data that modelers could draw on changed daily, so did the world they were trying to model. Halloran and Vespignani’s first model of SARS-CoV-2 spread used transportation data to project where the virus would travel.
“But by the time it got reviewed, then came all the travel restrictions. And so then we had to change the focus of the original paper — it was already outdated,” she said. They published the work in the journal Science in April 2020.
The scientific questions Halloran explored evolved with the pandemic. She and Vespignani published several studies addressing new COVID questions. They used mobility data to examine the impact of testing, contact tracing, and quarantine on the second wave of COVID-19, work that appeared in the journal Nature Human Behavior. With collaborators at the University of Florida, the team published a study in the journal Vaccine that examined the use of ensemble forecasting, a modeling strategy that uses a set of forecasts to explore a range of possible outcomes, in the design of vaccine efficacy trials.
She, Vespignani and several other colleagues also used the GLEAM model to estimate the establishment of local transmission and cryptic phase of the COVID-19 epidemic in the U.S. and Europe, and are currently looking at different targeted vaccination strategies on a global scale. Halloran also lent her expertise in the design of vaccine trials to development of the protocol for the WHO Solidarity Vaccines Trial.
Schiffer also had to grapple with the pandemic’s slippery nature. He was working to figure out which variables would contribute the most to COVID-19 cases and deaths and his team’s models pointed to speed of vaccination and lockdown as the main ways society could reduce these.
“Then, of course, the news about the new variant popped up. And so we've had to re-equilibrate that whole paper,” he said.
The questions his team tackled also changed over the course of the pandemic, Schiffer said, ranging from why superspreader events were occurring to the effect masks and vaccines might have on case counts.
“It's been a very strange process to react to the news cycle,” he said.
Models aren’t a crystal ball
Many people anxiously watching the pandemic unfold expected more forecasting accuracy than the models could provide, said Etzioni. Schiffer agreed.
“Philosophically I am not in favor of using models to project what's going to happen next,” he said.
There are too many unknown variables to project a specific number of deaths within a specific time period.
“Our approach has been more to say, ‘What are the variables that will determine how many cases and deaths there will be?’ so we can go to the Department of Health and say, ‘Listen, these are the two things that matter,’” said Schiffer, who prefers to describe his models as “hypothesis generating” rather than “hypothesis validating.”
But, said Etzioni, that doesn’t mean models can’t be helpful. Early in the pandemic, many in the public struggled to understand the implications for society if the number of infected continued growing exponentially. But models could help people understand, she said.
“In a way, it didn't really matter what number the model gave to predicting how big the problem was going to be, as long as it was a big number,” Etzioni said. “The models were saying different things, but they were all saying it was going to be bad.”
In this way, models can help support policy decisions, or confirm what we know but don’t want to admit, such as that if numbers of novel coronavirus cases are climbing, we need to implement stricter social-distancing measures, she said.
At one point, Etzioni asked modelers at the IDM to do a retrospective model, comparing the numbers of infected people in Washington state to how many there might have been if lockdown measures had never been implemented.
“It really was striking, that we were was about at half of where we would have been in terms of the prevalence,” she said. The IDM model helped confirm that the difficult, unpopular restrictions had worked as intended.
Models are also often better at addressing qualitative questions, such as whether to social distance now or later, than quantitative questions, such as how many will become infected, Etzioni added.
Understanding the assumptions included in each model is key to interpreting it, the researchers said.
Modelers should be upfront about what they’ve assumed as they created their model, and how sensitive it is to those assumptions, Etzioni argued.
“We all have to tell the public the one thing that they should know that we are assuming, but that we can't really verify,” she said.
Halloran noted how much different assumptions can change the outcome of models. Take vaccines, for example. A model that assumes that vaccines prevent infection, and therefore transmission, will give dramatically different results from one that assumes that a vaccine allows breakthrough infections in vaccinated people who may then transmit the virus.
“The further you get out from the day the simulation starts, the greater the variability,” Schiffer said.
It’s a little like weather forecasting: predictions for tomorrow’s weather are much more accurate than predictions for next week. This is in part because, Schiffer noted, pandemic modelers don’t have answers to many of the parameters controlling that variability, such as how new viral variants will affect vaccine efficacy, or whether people will adhere to yet another lockdown.
“People expect big mathematical models to predict exactly what's going to happen, as if we had a crystal ball,” Matrajt said. “That’s just not possible. It would be much better if everyone, including the modelers, were a little bit more open about the limitations.”