In this series of posts, I will be telling Prism’s story. How we came to be and why we work to solve the problems that we do.
It is often difficult to trace the origins of ideas that end up shaping our lives. However, in the case of Prism, there is a clear line of inquiry going back from where we are today (building software solutions for epistemological problems in biomedicine) to a philosophy of probability class that I took in graduate school.
One of the papers we read in that class was Jon Dorling’s “Bayesian Personalism, the Methodology of Scientific Research Programmes, and Duhem's Problem,” published in 1979 in the journal Studies in the History of Philosophy of Science. As the title suggests, this paper shows how a Bayesian epistemic calculus could be used to solve the “Duhem Problem”.
But unless you are a philosopher of science, you are probably asking yourself: What is the “Duhem Problem”?
Well, I’m glad you asked that question! Because the “Duhem Problem,” and its solutions, ended up being the focus in the first chapter of my PhD thesis—and it remains a topic that is near and dear to my heart.
The “Duhem Problem” is named for the French physicist and historian of science, Pierre Duhem (1861-1916). In his 1906 work, Aim and Structure of Physical Theory, Duhem observed that logic, evidence, and the experimental method were insufficient to truly test scientific theories and guide scientific progress.
“When certain consequences of a theory are struck by experimental contradiction, we learn that this theory should be modified but we are not told by the experiment what must be changed. It leaves to the [scientist] the task of finding out the weak spot that impairs the whole system. No absolute principle directs this inquiry, which different [scientists] may conduct in very different ways without having the right to accuse one another of illogicality.
[But] …Pure logic is not the only rule for our judgments; certain opinions which do not fall under the hammer of the principle of contradiction are in any case perfectly unreasonable. These motives which do not proceed from logic and yet direct our choices, these ‘reasons which reason does not know’ … is appropriately called good sense.”
Duhem P (1906). Aim and Structure of Physical Theory, pp.216-217
Thus, the “Duhem Problem” is essentially the challenge of figuring out what is good sense. What are these “reasons which reason does not know” that distinguish good scientific judgments from poor ones? What rules (if any) should guide a scientist’s decisions about whether to reject a theory in the face of disconfirming evidence?
Or to apply this in the context of drug development: How does one know whether to take a drug from phase 2 into phase 3? If a phase 3 trial comes back negative, should we cancel the whole research program or should we design a new phase 3 trial and try again?
At bottom, these are all versions of Duhem’s Problem, and knowing what to do requires good sense.
My term paper for that philosophy of probability class extended Dorling’s ideas to argue that Bayesian epistemology was one way of making good sense explicit—since Dorling shows how, starting from a reasonable set of priors, one should update the degrees of belief in a theory T, given evidence E, according to Bayes’ Theorem.
My thesis chapter (written a few years later) ultimately rejected Dorling’s model, and instead built on ideas from another philosopher, Bill Wimsatt, to argue that good sense could be better understood in terms of scientific heuristics and meta-heuristics.
In the next entry in this series, I’ll go into my heuristics and meta-heuristics model in more detail. But as I wrap up this first part of our story, I want to emphasize two things:
- As we build Prism, I know we are always benefitting from the help and work of others who have come before us. So in that spirit, I want to use this series to highlight some of our intellectual ancestors, like Duhem, Dorling, and Wimsatt (although much more on Wimsatt next time), who laid the conceptual foundations upon which our business and technology stand.
- The core problems that Prism is now solving are not new problems. They are actually very old problems whose difficulties stem from some of the fundamental uncertainties that affect all experimental sciences.
We believe one of our differentiators in this industry is our deep appreciation (and excitement) for these kinds of philosophical challenges. Indeed, it would be a comparatively easy thing to put together a sexy-sounding-but-ultimately-illusory story about how our “super-secret ML algorithm” is going to magically make sense of everything, revolutionize pharma, and turbocharge drug development. That is not the story I’m going to tell in this series.
The story I’m telling is about coming to see how a very old, very hard problem in science can be overcome—and how it desperately needs to be overcome in biomedical research. This is the intellectual core of why we do what we do at Prism. We are building products that are grounded in the fundamentals of science and epistemology—because we know that to truly solve these very old, very hard problems, and to increase the efficiency of biomedical research, a philosophically-grounded solution is required.