Job-Killing Computerization Sets Its Sights on the University Researcher

Swikar Patel for The Chronicle Review

After methodically comparing hundreds of drug combinations, an analyst at the University of Chicago has identified a possible cure for an especially challenging form of breast cancer.

The analyst responsible for the breakthrough, however, isn’t a person, but a computer algorithm.

Its discovery of a drug combination that, in lab tests, appears effective against so-called triple-negative breast cancer is one of the most significant breakthroughs to date in a federally sponsored effort to make sense of the vast amount of data now overwhelming human scientists.

That effort is known as the Big Mechanism program, and at a time when scores of traditional industries are being overtaken by computers and automation, it’s a warning to universities and their faculty members that the job of research scientist — at least as it’s generally recognized now — might eventually join the casualty list.

“I don’t really see any limiting factor, to be perfectly honest,” to the evolution of computers as scientists, said Paul R. Cohen, a program manager leading the Big Mechanism project at the Defense Advanced Research Projects Agency, or Darpa.

Computers, of course, have been making steady progress toward replacing human beings in a variety of arenas. Some of the economic upheavals have already been felt, and corporate leaders, including Bill Gates and Elon Musk, have been warning of far greater job losses.

A chief obstacle to computerized systems’ assuming a greater role as scientists concerns the comprehension of vast troves of existing published research. And for that, breakthroughs are just beginning to occur in a key task: reading and understanding the pictures, tables, and charts that accompany journal articles.

Those working on the problem include Jevin D. West, an assistant professor of information at the University of Washington. He’s collaborating on a project called Viziometrics, and he’s begun using it for tasks such as helping a University of Illinois cancer researcher make preliminary predictions about potential drugs that could take months of combing through published literature. Other researchers, he said, have used similar methods to identify new chemical compounds.

Mr. West said he sees no major technical roadblocks to the technique. Using computers to comprehend and interpret all past research, he said, “could transform the scientific method in many ways.”

Nontechnical Barriers

Darpa’s Big Mechanism is also focused, at least initially, on cancer. Successful treatments for breast cancer often work by blocking one of three types of receptors, which are protein molecules that receive chemical signals from outside a cell. Triple-negative breast cancer is challenging because it lacks those three types — estrogen receptors, progesterone receptors, and human epidermal growth factor receptor 2.

The Digital Campus: Tech Innovators 2016

Meet this year’s tech innovators — nine men, women, and projects using education technology boldly and broadly.

• Activist Group Unites via Social Media
• Sanjay Sarma Rethinks the Professor’s Role
• ‘American Panorama’ Uses Digital Tools to Remap History

Some cancer experts are trying to prod breast-cancer cells to produce the estrogen protein. Darpa-funded research at the University of Chicago sped that up by creating a computer algorithm that hunted for any proteins that might affect estrogen receptors and then suggested drug combinations that might help. Initial lab tests show the computer-suggested combinations do work, Mr. Cohen said.

Yet it’s early days for such computerized researchers, and there remain some significant nontechnical barriers to a future in which universities and their faculty members watch and guide while computers do most of the work now performed by human scientists.

One barrier concerns usable access to the millions of pages of already-published scientific data. As publishers are pushed to allow open access to their articles, they’re holding tight to the data and article metrics that computerized systems would need to make full use of them.

The publishers are being wary because they anticipate big profits from creative reuse of their databases, Mr. West said. The windfalls they envision might be even bigger, however, if publishers embraced partners who are now pushing the boundaries of data-mining technology, Mr. West said. “But they’re not convinced of that yet,” he said.

A more fundamental obstacle to an expansion of computer-driven research, Mr. Cohen said, involves how universities train their scientists. Far too few scientists, regardless of discipline, fail to even accept the need learn deeply about computers and networks, he said.

That’s a problem now, Mr. Cohen said, because so many scientists are unable to communicate with colleagues in other disciplines and recognize the abstractions and common relationships they share. And it’s becoming a bigger problem, he said, as huge interacting systems — water, climate, food, energy, and more — increasingly grow too complicated for individual human researchers to fully comprehend.

Whether studying proteins or social relations or predator-prey interactions, the same basic algorithms and mathematics apply, Mr. Cohen said. “The only thing that’s changed is the nouns,” he said. And yet sciences are not taught that way, he said.

Human Brain Skills

For the immediate future, however, the job of university researcher appears to be relatively safe, said Erik Brynjolfsson, a professor of management at the Massachusetts Institute of Technology.

Mr. Brynjolfsson, who studies the growing capabilities of machines, notes their rising ability to make discoveries that seem to have eluded people. A chief example came last year, when a Google-owned computer system defeated Lee Sedol, a champion in Go, one of the world’s most complex board games.

“That said,” Mr. Brynjolfsson added, “one of the comparative strengths of humans is to make these unstructured comparisons and connections that machines have trouble making.”

That should be true for at least the next few decades, he said, meaning the ranks of scientists should grow, aided by computers that can perform certain elements of the research process faster and better than their human counterparts can.

Mr. Brynjolfsson offered his assessment after making a presentation this past weekend in Boston at the annual meeting of the American Association for the Advancement of Science. At the meeting, he warned fellow scientists about the threats to many people in the American work force posed by computerized automation, and he urged them to think more about the implications and their responsibilities to help.

“If you had to look at all the jobs in the economy, research — asking the right questions, coming up with new insights — those are among the hardest things to automate and probably among the most secure jobs in our economy,” Mr. Brynjolfsson said.

Despite its many capabilities, advanced computer technology has “simply no comparison” with a human brain at this point, Randal E. Bryant, a professor of computer science at Carnegie Mellon University, told the AAAS gathering.

One early exception concerns graduate students and postdoctoral researchers, who often handle the sometimes-tedious job of searching past journal articles in return for their educational expenses and small stipends. Automation is eating away at those jobs, both inside the labs and at companies that now perform many lab-related services such engineering test mice, said Paula Stephan, a professor of economics at Georgia State University.

“It’s already been happening in a way,” Ms. Stephan said of automation-related job losses in university labs. “That will really affect the demand for postdocs and graduate students.”

Mr. Cohen nevertheless warns more-senior researchers not to be complacent. Beyond his Darpa position, Mr. Cohen is also a professor and founding director at the University of Arizona’s School of Information. To help build a more interdisciplinary faculty, he had a policy of not hiring anyone unless at least one other department also wanted that person.

He said he also emphasized the hiring of “tweeners,” such as biologists who are also terrific computer scientists. Such a person, he said, has trouble getting hired at most institutions — “the biologists won’t recognize what they’re looking at, and the computer scientists will say, Well, he’s not a computer scientist.”

That’s led to the problem his Darpa-funded computers are now trying to solve: far too many published research articles that add little or no scientific value. Universities encourage that through hiring practices that emphasize volume rather than value, Mr. Cohen said. “My machines have to read 300,000 articles,” he said, “because the universities reward people for writing lots of articles about very little.”

Paul Basken covers university research and its intersection with government policy. He can be found on Twitter @pbasken, or reached by email at paul.basken@chronicle.com.