How AI Evolved Drug Redevelopment: Q&A with Dr. Friso Postma

Dr. Friso Postma
VP of AI drug development
BioXcel

A unique way that that the Pharma industry has embraced AI is by using it for drug re-innovation. This is the process where existing drugs are re-examined, using AI-assisted software, in order to determine novel indications. Dr. Friso Postma, VP of AI drug discovery at BioXcel Therapeutics, spoke with Pharmaceutical Executive about how use of the technology in this area has developed in recent months.

Pharmaceutical Executive: How did you get involved in AI?
Dr. Friso Postma: I got into AI while working in digital health, looking at how to deploy wearable devices to inform health status. That was the first time that I had to interface with data scientists and AI engineers, and that was a very interesting experience. I learned very quickly that if you have a large amount of data, you can ask an AI engineer to extract features out of that. However, if you don’t contextualize it, it’s not necessarily biologically or clinically relevant.

What I did there was focused on trying to create that level clarity and contextualize the AI data sets. In the last three years, I’ve been building an integrated AI platform. I often refer to AI as augmented intelligence because it augments our decision-making process in the drug reintegration process.

My current position is a sweet spot for me because I can combine my interests in technology and biology to address unmet medical needs.

I’m not an AI engineer and I’m not a data scientist. I do lead a multi-disciplinary team with people in Prague, Europe, and India. The team includes MDs and PhDs. We’re trying to re-innovate existing compounds that have safety profiles associated with them to capitalize on the intrinsic value of these assets. We’re identifying these compounds and looking for signs of novel indications. We leverage AI throughout the entire process.

PE: Can you discuss the current state of AI usage in drug redevelopment?
Postma: Things are moving very quickly. In a way, the AI is not one thing. AI is very often considered a black box that sits in the corner of a room that you ask questions to. That’s not really what it is. Our AI is a platform that we refer to as composite AI. It’s a number of different utilities that very specifically address the facets and aspects of the drug re-innovation process.

We often refer to it as the holistic use of AI. We don’t just do one little thing, like try to identify a concept in biology, and then leave it at that. There’s a whole bunch of different things to happen subsequently that allows you to rank order the concepts you have generated.

For us, a concept has been defined as a hypothesis for use of an existing drug for novel indication. These different elements need to be integrated and updated, both in terms of capabilities and the kind of data that we accumulate into the platform.

The AI platform is not a static thing. It’s multi-modal, composite, and it grows. It’s ability to effectively identify these opportunities expands. Things that we’re working on right now are different regarding the automation of that entire process. We have reduced the steps where there would be human interface, and instead have people looking at whether the data makes sense.

PE: Can you discuss explainable AI and multi-modal AI?
Postma: There are some things that are referable as explainable AI. What you have is a system that makes a recommendation. It’s similar to what Netflix is using to recommend movies and shows based on what you have seen in the past. We have set up a system that does something similar. Based on the interactions between drugs and the targets, along with the links to indications, the platform helps to determine if there anything that we can predict might be effective in this particular indication.

Sometimes, AI can do that, but you don’t always know why. Explainable AI is something that we’re looking into right now. Another thing that we feel that we’ve successfully implemented is multi-modal AI. Different types are being interpreted by deep-learning networks.

What we are dealing with is a AI-assisted pipeline where we identify high-value assets and then put that information into a knowledge graph that connects entities through relationships. For example, the data would look for a drug that combines with a receptor that is involved in a particular neuro-circuitry that then mediates a behavior or symptom.

From these knowledge graphs, we come up with ideas on how to reposition a particular drug for a new indication. That’s not where it stops, though. That concept then needs to be validated clinically. If it can’t be, then we can’t test it and it’s not a good concept. If I have 25 concepts, we need to rank them in terms likelihood of validation, which is a difficult proposition.

Once we have validation, we also want to understand what the IP landscape is. If we want to develop this, we need to determine if we can create a patent. You have to understand what the regulatory pathway is. What is the landscape with similar companies and compounds? You also must determine what a clinical trial would look like and how many patients you would need.

All these different types of decisions that go into drug redevelopment must be considered.