You may have seen articles last week about a breakthrough for artificial intelligence in medicine that managed to be both arcane and exciting at the same time. Google's DeepMind research arm solved a 50-year-old problem related to predicting how proteins fold themselves -- news only for geeks, right? Think again. Understanding how these chains of amino acids fold themselves into 3D shapes, providing the structural components for the tissues in our bodies, opens up all sorts of possibilities for exploring our inner workings and for rapid development of drugs.
What I haven't yet seen explained -- amid all the speculation about just how many Nobel Prizes in Medicine will spring from the work -- is that the type of AI that DeepMind developed to solve the protein-folding conundrum should also provide breakthroughs in insurance. This type of AI can take dead aim at some core issues in insurance, especially in underwriting and claims.
AI is funny. It tends to be talked about as a single thing, but it's really a whole bunch of things, pushing against limits in a wide range of directions. And some of the progress is flashy without being all that important.
For instance, when IBM's Watson defeated the greatest Jeopardy champions in 2011, IBM talked about sending Watson to medical school. After all, if it could beat Ken Jennings, what couldn't it do? But Watson's breakthrough was in natural language processing, a great advance if you want to be able to talk to a computer but little help if you're trying to cure cancer. Similarly, when DeepMind beat the world champion at Go in 2017, the event made for fun headlines but not much more. The AI is terrific for any setting where there are a small number of rules and where the computer can play games against itself ad infinitum to optimize its approach, but how many real-world situations fit that description?
By contrast, what DeepMind accomplished in solving the protein-folding problem is of deep significance because the approach the scientists used -- known as supervised deep learning -- can be applied to so many business situations, including in insurance.
Without getting too deep into the details (which you can find in this excellent piece in Fortune, if you want to geek out like I did), the scientists faced a problem far more complex than businesses face: trying to figure out how a protein folds itself, in the milliseconds after it is created, based on a host of forces. While we've been able to sequence the human genome for more than 15 years now, you also have to know how the string of amino acids folds, because the shape determines so much of how the protein behaves.
Although a famous conjecture in 1972 said it should be possible to predict a protein's shape just from the sequence of amino acids in it, the computation had proved to be too complex. Instead, the shape of a protein had to be determined through a complex chemical process and, often, through the use of a special type of X-ray produced by a synchrotron the size of a football stadium. The process could take a year and cost $120,000, for a single protein.
(I realize I may be giving you flashbacks to high school biology and chemistry and perhaps some unpleasant memories, but I'm just about done with the science and am getting to the implications for insurance.)
What the scientists had going in their favor were two things: a sort of answer key, because of some 170,000 proteins whose shape had already been determined experimentally, and some coaching tips that could help the AI focus on the key variables.
That starts to sound like a business situation, especially, in terms of insurance, in claims and underwriting. If you want to train an AI to take over tasks, you have underwriters and adjusters who can tell you what the right answer is and who can guide the AI's self-training by steering it toward certain variables. Over time, that AI can become as good or better than a human at, say, looking at photos of the damage in a car accident and estimating the damage.
At least, that's how it worked for DeepMind on a much harder problem. On a scale where 100 is perfect accuracy, the previous best AIs scored about 50, well below empirical methods, which scored about 90. But in a recent competition in which AIs predicted the shape of proteins whose forms had been determined experimentally but had yet to be published, DeepMind's median score was 92 -- a computer prediction outscored that year-long, expensive, physical process. Importantly, DeepMind's AI can tell scientists how confident it is about each prediction, so they know how heavily to rely on it.
The immediate application for the DeepMind AI will, of course, be in medicine. There are some 200 million proteins whose shapes haven't yet been determined, and the AI can quickly go to work on those. (The required computing power is only perhaps 200 of the graphics chips used in a PlayStation.) Understanding the shapes will help researchers see what drugs might interact with which proteins, potentially reducing drug development time by years and lowering costs by hundreds of millions of dollars.
However, how this AI moves into the mainstream remains to be determined. DeepMind functions as a research arm of Google, not as a business, and has promised to ensure that the software will “make the maximal positive societal impact,” but you could hardly blame Google if it tried to recoup the development costs through charges to Big Pharma. Only once this AI filters through medicine will it, I imagine, spread to other business problems, such as those that insurance faces.
For me, it's enough to know at the moment that this sort of AI is possible, because that means that a lot of smart people will accelerate their efforts to bring supervised deep learning to insurance. While the wins at Jeopardy and Go were startling, the AI that solved the protein-folding problem will prove to be far more consequential.
P.S. Here are the six articles I'll highlight from the past week:
Smart contracts will likely be used first for simpler insurance processes like underwriting and payouts, then scale as technology and the law allow.
With businesses cutting back, many are asking that question. But there are huge misconceptions about how to think about the issue.
With home-working widespread because of COVID-19, security around access and authentication points is critical.
Chances are, you have natural gifts. However, many of the skills you need must be developed, nurtured and maintained intentionally.
Legacy systems that have evolved over long periods can be bloated and far less efficient and cost-effective than more modern technologies.
Health plans strive to deliver efficiency and great customer experiences and improve care outcomes. But what data are they missing?