Charles Fisher, Ph.D., is the CEO and Founder of Unlearn, a platform harnessing AI to tackle some of the biggest bottlenecks in clinical development: long trial timelines, high costs, and uncertain outcomes. Their novel AI models analyze vast quantities of patient-level data to forecast patients’ health outcomes. By integrating digital twins into clinical trials, Unlearn is able to accelerate clinical research and help bring life-saving new treatments to patients in need.
Charles is a scientist with interests at the intersection of physics, machine learning, and computational biology. Previously, Charles worked as a machine learning engineer at Leap Motion and a computational biologist at Pfizer. He was a Philippe Meyer Fellow in theoretical physics at École Normale Supérieure in Paris, France, and a postdoctoral scientist in biophysics at Boston University. Charles holds a Ph.D. in biophysics from Harvard University and a B.S. in biophysics from the University of Michigan.
You are currently in the minority in your fundamental belief that mathematics and computation should be the foundation of biology. How did you originally reach these conclusions?
That’s probably just because mathematics and computational methods haven’t been emphasized enough in biology education in recent years, but from where I sit, people are starting to change their minds and agree with me. Deep neural networks have given us a new set of tools for complex systems, and automation is helping create the large-scale biological datasets required. I think it’s inevitable that biology transitions to being more of a computational science in the next decade.
How did this belief then transition to launching Unlearn?
In the past, lots of computational methods in biology have been seen as solving toy problems or problems far removed from applications in medicine, which has made it difficult to demonstrate real value. Our goal is to invent new methods in AI to solve problems in medicine, but we’re also focused on finding areas, like in clinical trials, where we can demonstrate real value.
Can you explain Unlearn’s mission to eliminate trial and error in medicine through AI?
It’s common in engineering to design and test a device using a computer model before building the real thing. We’d like to enable something similar in medicine. Can we simulate the effect a treatment will have on a patient before we give it to them? Although I think the field is pretty far from that today, our goal is to invent the technology to make it possible.
How does Unlearn’s use of digital twins in clinical trials accelerate the research process and improve outcomes?
Unlearn invents AI models called digital twin generators (DTGs) that generate digital twins of clinical trial participants. Each participant’s digital twin forecasts what their outcome would be if they received the placebo in a clinical trial. If our DTGs were perfectly accurate, then, in principle, clinical trials could be run without placebo groups. But in practice, all models make mistakes, so we aim to design randomized trials that use smaller placebo groups than traditional trials. This makes it easier to enroll in the study, speeding up trial timelines.
Could you elaborate precisely on what is Unlearn’s regulatory-qualified Prognostic Covariate Adjustment (PROCOVA™) methodology?
PROCOVA™ is the first method we developed that allows participants’ digital twins to be used in clinical trials so that the trial results are robust to mistakes the model may make in its forecasts. Essentially, PROCOVA uses the fact that some of the participants in a study are randomly assigned to the placebo group to correct the digital twins’ forecasts using a statistical method called covariate adjustment. This allows us to design studies that use smaller control groups than normal or that have higher statistical power while ensuring that those studies still provide rigorous assessments of treatment efficacy. We’re also continuing R&D to expand this line of solutions and provide even more powerful studies going forward.
How does Unlearn balance innovation with regulatory compliance in the development of its AI solutions?
Solutions aimed at clinical trials are generally regulated based on their context of use, which means we can develop multiple solutions with different risk profiles that are aimed at different use cases. For example, we developed PROCOVA because it is extremely low risk, which allowed us to pursue a qualification opinion from the European Medicines Agency (EMA) for use as the primary analysis in phase 2 and 3 clinical trials with continuous outcomes. But PROCOVA doesn’t leverage all of the information provided by the digital twins we create for the trial participants—it leaves some performance on the table to align with regulatory guidance. Of course, Unlearn exists to push the boundaries so we can launch more innovative solutions aimed at applications in earlier stage studies or post-hoc analyses where we can use other types of methods (e.g., Bayesian analyses) that provide much more efficiency than we can with PROCOVA.
What have been some of the most significant challenges and breakthroughs for Unlearn in utilizing AI in medicine?
The biggest challenge for us and anyone else involved in applying AI to problems in medicine is cultural. Currently, the vast majority of researchers in medicine specifically are not extremely familiar with AI, and they are usually misinformed about how the underlying technologies actually work. As a result, most people are highly skeptical that AI will be useful in the near term. I think that will inevitably change in the coming years, but biology and medicine generally lag behind most other fields when it comes to the adoption of new computer technologies. We’ve had many technological breakthroughs, but the most important things for gaining adoption are probably proof points from regulators or customers.
What is your overarching vision for using mathematics and computation in biology?
In my opinion, we can only call something “a science” if its goal is to make accurate, quantitative predictions about the results of future experiments. Right now, roughly 90% of the drugs that enter human clinical trials fail, usually because they don’t actually work. So, we’re really far from making accurate, quantitative predictions right now when it comes to most areas of biology and medicine. I don’t think that changes until the core of those disciplines change–until mathematics and computational methods become the core reasoning tools of biology. My hope is that the work we’re doing at Unlearn highlights the value of taking an “AI-first” approach to solving an important practical problem in medical research, and future researchers can take that culture and apply it to a broader set of problems.
Thank you for the great interview, readers who wish to learn more should visit Unlearn.
