Top Secret Plan for AI in Medicine, Part 1

by Charles K. Fisher

Everyone who knows Unlearn is aware that we currently offer products and services to biopharma companies that help them run more informative, more efficient, and more ethical clinical trials. However, that’s not exactly where

we started, and it definitely isn’t where we want to stop. Our offerings aimed at clinical trials are our initial products, not the only products we’re working on, which is a key part of our go-to-market strategy. Not that many people seem to know about this strategy so I suppose it is a secret plan (this is obviously inspired by The Secret Tesla Motors Master Plan, Master Plan, Part Deux).

Unlearn’s mission, in its most recent form, is to advance artificial intelligence to eliminate trial and error in medicine. This is an evolution of the mission statement that Aaron, Graham, Jon, and I wrote down when we first started the company in 2017, to create artificial intelligence to simulate biology. Let’s break it down a bit.

The first three words of our mission are to ‘advance artificial intelligence.’ 

Biology is one of the remaining sciences that hasn’t been placed on a mathematical foundation. There are almost no quantitative, predictive theories of major phenomena. Of course, biological systems are complex. In fact, the status quo belief in much of the biological and medical sciences is that biological systems are so complex that it’s a fool’s errand to try to turn it into a mathematical and computational science. I disagree. In fact, I think that developing quantitative, predictive models of all major biological phenomena is one of the most important outstanding scientific problems of the 21st century.

In its current form, AI has already demonstrated the ability to learn quantitative, predictive models of complex phenomena that humans weren’t capable of doing themselves. This is a type of narrow superhuman intelligence in contrast to artificial general intelligence. For example, hand-crafted models of protein folding didn’t work that well, and methods based on deep learning crushed them. Same thing for language. And all of this progress has come after just a couple decades of deep learning research. I believe that the same thing will eventually be possible for all important phenomena in biology. 

Suppose I’m right. As a business, we still need to ask ourselves three questions: 

• First, how long do we think it will take before we can use AI to predict all biological phenomena that are important for medicine? 

• Second, what types of products and services can we offer along the way while we are still working to build that AI?

• Third, what avenues of research and development should we pursue in order to bring this AI into existence while also offering these products and services?

I don’t know the answer to the first question, but I’m at least 99.99% certain that it will take longer than 5 years (probably a lot longer, but 5 years is roughly the maximum time any business can wait before launching). As in any business, Unlearn needs to sell things in order to make money, and we can’t wait 5 (or 10, or 20) years until we’ve completed our mission to do so. Therefore, we need to turn our attention to the second question—what products and services should we offer along the way?

The sequence in which we bring new products to market will be the key to our success. We will build an amazing business if we get this right, and a terrible business if we get it wrong.

The latter part of Unlearn’s mission is to eliminate trial and error in medicine. Medicine is filled with trial and error today because we don’t know which treatments work best for which patients. We can break that down into three related questions:

• Given an investigational treatment, does it have a positive impact on health outcomes for some patients?

• Given a number of alternative treatments, which one tends to provide the largest benefit to most patients?

• Given a number of alternative treatments, which one will provide the largest benefit to a specific patient at a specific time?

Question 1 is the domain of clinical trials, question 2 is the domain of comparative effectiveness research, and question 3 is the domain of personalized medicine. 

At a high level, our secret plan is pretty simple:

1. Build an AI that can create digital twins of patients in clinical trials and use it to design more efficient, ethical, and reliable clinical studies.

2. Reinvest that money into research and development to improve our AI so that it can also simulate the effects of existing treatments and use it to power an in silico comparative effectiveness platform. 

3. Reinvest that money into research and development to improve our AI so that the predictions it makes for treatment outcomes are accurate enough at the individual patient level that they can be used to guide treatment decisions. 

4. Scale these solutions until we can turn most of medicine into a computational science.

What does this mean for the development of our AI?

Clinical trials compare two potential outcomes: ‘what would happen to a patient if we gave him/her this new investigational treatment?’ vs. ‘what would happen to a patient if we gave him/her the current standard treatment?’.  It’s straightforward (though not easy) to build an AI that can simulate the latter potential outcome. Moreover, clinical trials typically aggregate information across a population rather than looking at individual effects, and we can use various methods to provide reliable population estimates even if our individual predictions are unreliable—this ability to create zero-trust solutions is one of the biggest advantages of choosing clinical trials as our initial market. 

For comparative effectiveness, research will need to expand the capabilities of our AI to simulate multiple potential outcomes: ‘what would happen to a patient if we gave him/her treatment {A, B, C, …}?’. The AI will need to understand cause-effect relationships, not just associations. For some technical reasons, I think we’ll lose the ability to provide zero trust solutions that we have in clinical trials, but comparative effectiveness research doesn’t require the same standards of evidence as clinical trials do (because you’re limiting yourself to comparing treatments that are already in use).

The requirements for our AI for personalized medicine are exactly the same as those for comparative effectiveness; it’s just that the bar is a lot higher. The model needs to be good enough to trust its predictions at the individual level, not just the population level. There’s no zero trust solution here. Also, in order to scale the solution, we would need a model that has a very large scope—covering almost any indication that a patient may show up with.

That is, first, we’ll build an AI to create digital twins of patients in specific indications on specific treatments. We’ll use that to improve clinical trials. Then, we’ll make that AI better by giving it the capability of simulating the effects of various treatments, and we’ll build an in silico platform for comparative effectiveness research. Then, we’ll develop that AI until we trust its individual-level predictions and roll it out to support personalized medicine. Finally, we’ll scale that AI across all indications.

It’s that simple!