Machine Learning’s Next Trick Will Transform How Research Is Done
Though research is a slow moving and rigid process, one study shows that the rate of scientific study has exploded in the last 50 years. According to the paper, humanity’s scientific output now doubles every nine years. Considering the rigors of science, that’s pretty fast. And it’s just the average rate.
In specific areas like healthcare, the doubling rate is even faster, as much as every 3 years currently with an expected increase to every 73 days by the early 2020s. For overwhelmed researchers navigating the growing stack of science literature, the value isn’t in having so much new information, but finding relevant insights when they need them.
According to Jacobo Elosua, a co-founder of Iris AI, a Singularity University portfolio company, the research process is very often tedious and unfruitful.
“Researchers are genuinely struggling to find the scientific papers, the clinical data, and other information required to do their job. And when they do find it, it’s most often after a painful and time consuming process,” he told Singularity Hub.
Elosua and the team at Iris hope recent advances in machine learning AI might be one way through the noise. Machine learning is powerful because it allows programmers to assign a task to an algorithm, in this case, combing through scientific literature, and then let the code teach itself to improve its model as it is fed more data over time.
Iris works by reading scientific papers and learning to determine what’s being discussed in the text. The goal is to augment the discovery process by leading researchers to relevant papers and new discoveries as they are published. By identifying emerging trends and concepts within the areas of science that may impact a researcher’s domain of interest, AIs can shoulder some of the burden of constantly scanning new literature.
According to Elosua, “Iris users will be able to drop in any scientific text with over 500 words as an input to the tool, say like an abstract of an interesting paper. Iris will then display a visual map enabling an intuitive navigation of the most relevant papers.” Elosua added that, “In terms of time saved we believe it will be more than ten times faster to use Iris.”
The promise of an accelerated research process is exciting, but hurdles remain. Though global trends in academia have shown a shift to open access, many research papers are locked away in closed databases. Also, Iris's proof of concept scans the science literature contained in TED talks, a fairly broad set of areas. Iris is currently working to develop more specialized ways to use their service.
Another Singularity University start up, Miroculus, is hoping that their more targeted machine learning tool may help with their own research needs.
The team at Miroculus, in partnership with Microsoft, have built Loom, a tool that uses machine learning to search papers for the relationship between specific microRNAs and various diseases and genes. Though Miroculus' core business is developing a low-cost cancer diagnostic tool, the Loom project may prove valuable to research efforts in their space.
MicroRNA is a type of RNA found in the bloodstream that delivers genetic instructions telling the body what proteins to build and when to build them. In a TED Talk, Miroculus CTO Jorge Soto explains that microRNAs help regulate gene expression. And since changes in gene expression are a major component of cancer, understanding how microRNAs vary depending on conditions in the body, and measuring these changes, may help us diagnose cancer far earlier than today’s standards.
In the talk, Soto describes how catching cancer early is the closest thing we have to a silver bullet cure against it. But there’s a problem. Soto says, “There is no compelling way to access much of the microRNA research today, other then to manually retrieve relevant papers and read them thoroughly.” This can take days or even weeks in some cases.
He hopes that by having a way to quickly track microRNA literature, his team will be able to understand the latest findings in the space. In an interview with Singularity Hub, Soto said, “With Loom, our objective is to provide a compelling overview of how microRNAs relate to specific diseases and genes.”
Loom is able to locate relevant papers that mention specific microRNAs, extract the relevant parts of the paper, and then score the relationship between the microRNA and the specific gene or disease being studied. According to Soto, Loom was trained by learning from a manually created dataset that curated over 10,000 mentions of microRNAs, and the tool becomes more accurate every day as more literature is published.
As AIs take on more responsibility in managing the discovery process, the science community may free up significant portions of the time they currently devote to scanning for trends. One Canadian AI company, Meta, can already scan for emerging technology trends and predict those technology's future significance.
In parallel, as AIs learn to better navigate the subtleties of language, they may be better equipped to draw meaning from science literature. Earlier this month, for example, Google announced exciting progress in natural language understanding and open-sourced machine learning code in the area, which may further empower AI-assisted research tools.
Though science is moving fast, maybe too fast for our brains to handle, projects like Iris and Loom are out to show how AI can help today’s researchers keep up with today's accelerating pace.