The Different Types of Digital Twins

One of the most discussed use cases in pharma and biotech for AI is the creation of digital twins. By using massive amounts and data, the technology can create a digital recreation of a real person, which can then be used for various research purposes.

Pharmaceutical Executive spoke with Hamish Brocklebank, founder of Brox.AI, about some of the ways that digital twins are being used within the industry, and where the technology could potentially head in the coming years.

According to him, these twins have a variety of uses. One of the more commonly discussed use cases is for pure research. In this instance, a researcher creates a digital twin and then uses AI to simulate the effect of various molecules and conditions. However, Brocklebank explains that there are other use cases which may provide just as much, if not more, value.

Pharmaceutical Executive: What are the different kinds of digital twins?
Hamish Brocklebank: There are a few different definitions in this space, so bear with me while I explain how we use the term.

One approach is to take a large corpus of data — imagine you are studying dermatologists. You would gather an enormous amount of data on that specialty: survey responses, behavioral data, prescription data, and so on. That data would not necessarily be tied to specific individuals. From that, you build a persona representing a particular type of doctor or patient, and the resulting digital twin can predict the median outcome for whatever you are modeling reasonably well. The limitation is that it will only give you the median. That is one approach.

The second approach, which I would call truly synthetic, involves almost no grounded dataset at all. You essentially instruct a large language model to behave like a particular type of person based on a set of attributes you define. You can do this with one simulated individual or scale it to ten thousand — the approach is the same either way. You are generating behavior based on assumed characteristics rather than observed ones.

Our approach resembles the second, but with a critical difference: every clone we build is based on a real person who has actively consented to share their data and to participate in the construction of their digital clone. So when we query a thousand digital clones, those represent a thousand real doctors — in our case, actual practicing physicians in the United States whom we have recruited, built a deep and comprehensive dataset around, and trained our models against individually, one to one.

That one-to-one grounding is what separates what we do from the purely synthetic approaches.