STAT is reporting that IBM has stopped trying to sell their “Watson for Drug Discovery” machine learning/AI tool, according to sources within the company. I have no reason to doubt that – in fact, I’ve sort of been expecting it. But no one seems to have told IBM’s website programming team, because the pages touting the product are still up (at least they are as I write this). They’re worth taking a look at in the light of reality. Gosh, it’s something:
Watson for Drug Discovery reveals connections and relationships among genes, drugs, diseases and other entities by analyzing multiple sets of life sciences knowledge. Researchers can generate new hypotheses using the resulting dynamic visualizations and evidence-backed predictions. . .
. . .Pharmaceutical companies, biotech and academic institutions use Watson for Drug Discovery to assist with new drug target identification and drug repurposing. Connect your in-house data with public data for a rich set of life sciences knowledge. Shorten the drug discovery process and increase the likelihood of your scientific breakthroughs.
Well, no, apparently they don’t use it much, because no one seems to have felt that they were increasing the likelihood of any scientific breakthroughs. The IBM pages are rather long on quotes from the Barrow Neurological Institute, about how they can make such breakthroughs “in a fraction of the time and cost”, but it looks like they’re going to have to get along without the product unless IBM is providing support to legacy customers. And since the STAT piece says that they’re halting both development and sale, that seems unlikely. Barrow and IBM press-released some results in late 2016, and there’s a promotional video from a month or two later, but that was both the first and last announcement from that collaboration.
What happened? Reality. As this IEEE Spectrum article from earlier this month shows in detail, IBM’s entire foray into health care has been marked by the familiar combination of overpromising and underdelivery. To their credit, the company made a very early push into the area (2011 !) with a lot of people and a lot of money. Unfortunately, they also made sure that everyone knew that they were doing it, and what a big, big deal it all was.
The day after Watson thoroughly defeated two human champions in the game of Jeopardy!, IBM announced a new career path for its AI quiz-show winner: It would become an AI doctor. IBM would take the breakthrough technology it showed off on television—mainly, the ability to understand natural language—and apply it to medicine. Watson’s first commercial offerings for health care would be available in 18 to 24 months, the company promised.
In fact, the projects that IBM announced that first day did not yield commercial products. In the eight years since, IBM has trumpeted many more high-profile efforts to develop AI-powered medical technology—many of which have fizzled, and a few of which have failed spectacularly.
Watson for Drug Discovery is just one of that suite of tools (well, potential tools). The idea was that it would go ripping through the medical literature, genomics databases, and your in-house data collection, finding correlations and clues that humans had missed. There’s nothing wrong with that as an aspirational goal. In fact, that’s what people eventually expect out of machine learning approaches, but a key word in that sentence is “eventually”. IBM, though, specifically sold the system as being ready to use for target identification, pathway elucidation, prediction of gene and protein function and regulation, drug repurposing, and so on. And it just wasn’t ready for those challenges, especially as early as they were announcing that they were. I first wrote about the company’s foray into drug discovery in 2013, and you’ll note that nothing really came out of the GSK/IBM work mentioned in that post. To the best of my knowledge, the two companies never really collaborated on drug discovery at all, but hey: they did team up on more targeted ways to advertise flu medicine.
Meanwhile, attempts at diagnostic and drug-therapy recommendations in oncology have been not only unproductive, but (according again to earlier reporting at STAT) even worse than that. The Spectrum article linked above goes into more details on those and other efforts all over the health care area that have come to naught, along with a few limited successes. And oddly enough, I’m going to finish off thinking about those. I still believe that machine learning is a perfectly good idea, with potential applications all over the field. But it ain’t magic. The areas where it’s worked the best so far are the ones with well-defined outcome sets based on large and very well-curated data collections, and where people have not been expecting the software to start spitting out golden insights and breakthrough proposals. It’ll get better – with a lot of work.
Just because people tried to sell the world on the idea that we’d moved past that stage years ago (A) does not make that so but (B) does not mean that we’ll never move past that stage at all. Next week I’ll have a post about machine learning and AI that goes into the real state of the field, from practitioners who have been spending their time whacking away at the code rather than generating promotional videos. IBM, though, has so far been doing the entire field a disservice with the way that they’ve spent too much time on the latter and not enough on the former.