Outsmarting Disease — With Artificial Intelligence
WHY YOU SHOULD CARE
Because finding the cures for diseases needn’t be a sluggish, costly process.
By Melissa Pandika
More and more, 67-year-old Washington resident Lon Coleman feels like he’s wandering through a fog. He walks into the living room and forgets why, or makes a phone call only to blank on whose number he dialed. An author of three books who once wrote up to five poems a day, now the lines that spring to his mind often slip away as soon as he puts pencil to paper. Sometimes the fog clears, and when his memory comes back, “it’s amazing,” he says. “Sometimes it doesn’t, I have to admit.”
Coleman is among the estimated 46.8 million people worldwide with some form of dementia, including Alzheimer’s disease. Twice a day he takes drugs that slow, but don’t stop, the disease’s progression. Meanwhile, a cure continues to elude scientists. But some think that artificial intelligence, or AI, could speed up drug discovery for dementia, as well as a host of other diseases. On a race to outsmart themselves, they’ve designed technology that can plumb the genome for mechanisms that underlie disease, or scan millions of molecules to identify those most likely to work as drugs. AI could yield faster diagnoses too, with software that detects cognitive decline from voice recordings and malaria from microscope slides. The proliferation of genomic and other data, plus a dramatic increase in computers’ ability to process and store it, have created “kind of a perfect storm … to tackle these really tough medical problems,” says Courosh Mehanian, principal research scientist at Intellectual Ventures Laboratory in Seattle.
By removing the grunt work, AI technology can free up scientists and clinicians for other types of work, like spending time with patients.
Given that it can take around a decade or more and billions of dollars to develop just one drug, such technology could have huge implications, especially for rare and tropical diseases, which pharmaceutical companies have little financial incentive to produce drugs for. It could also provide a powerful tool for precision-based medicine, meaning diagnostics or treatments tailor-made to individual patients. And pharma companies can use AI technology to predict how a patient or group of patients will respond to a drug based on their genetic profile, for instance, minimizing side effects. All of that potentially means faster, more precise diagnostics that could lower health care costs and free up clinicians to spend more time with patients.
Today, AI broadly encompasses machine learning and deep learning, which both involve training a computer model to recognize an unknown object by presenting it with lots of examples. Winterlight Labs has trained its machine learning-based software, which detects cognitive impairment from one- to five-minute speech snippets, to recognize differences in features such as pitch and grammatical complexity between people with and without Alzheimer’s. Founder Frank Rudzicz envisions users recalling a picture or a story over the Internet. The software would then send its analysis to a doctor. Analyzing one sample takes around five minutes — less than the three-hour-long battery of assessments typically used today — which could allow for earlier diagnosis, quickening access to medical services and cutting costs.
Other systems interpret the genome to understand disease, suggesting more precise diagnostics and even treatments. Deep Genomics in Toronto has developed a system designed to move beyond correlating genetic variations with certain diseases, unraveling how they lead to disease. The company is testing its technology in a study of a child with an immunological disorder who has a never-before-seen genetic mutation, which its CEO Brendan Frey hopes could “change the way pharmaceutical research is done,” in part, by identifying drug targets and predicting how patients will respond to a drug.
Meanwhile, Intellectual Ventures Laboratory’s Autoscope relies on deep learning — which uses networks modeled after those found in the brain — to detect the malaria parasite from blood films on glass microscope slides. Today, the standard method for highly trained microscopists is to pore over slides for the tiny, hard-to-spot parasite. But many areas hard hit by malaria have few microscopists, not to mention training resources. In a recent field evaluation in Thailand, the Autoscope, which is expected to roll out in 2017, correctly classified 170 slides using characteristics such as shape and texture. But since it uses electricity and has a target cost of about $1,500, it’s intended for clinics with sufficient resources — not remote areas.
Then there are systems that streamline the actual designing of a drug. Enter Atomwise’s AtomNet, a deep learning-based system that teaches itself how to recognize medicinal chemistry building blocks based on data from earlier research. When presented with the protein structure of a target (think of it as a “lock”), it considers a million possible keys per day to predict which ones will “open” it effectively. The San Francisco-based company is collaborating with researchers to predict potential drug molecules for cancer, neurological diseases and more.
But Atomwise COO Alexander Levy notes that the candidates AtomNet proposes still need to undergo testing. It “doesn’t solve every problem in the development of medicine.” Ditto for other AI technologies, where Coleman notes any new discoveries will still face a lumbering approval process. “There might be hype, but in the end for it to be truly useful, there needs to be some task not solved by any other approach,” says Yanjun Qi, an assistant professor of computer science at the University of Virginia. Others worry that AI will replace human analysis, possibly leading to more errors as well as job cuts.
Some patients might feel uneasy about clinical decisions based on conclusions drawn by a machine. But Suzie Siegel, a 57-year-old survivor of leiomyosarcoma — a rare, aggressive cancer — points out that “there are doctors out there who are making decisions now based on insufficient information,” she says, “so that doesn’t scare me at all.”
- Melissa Pandika, Melissa Pandika is a lab rat-turned-journalist with an eye to all things science, medicine and more. Likes distance running, snails, late-night Korean BBQ + R&B slow jams.Contact Melissa Pandika