In this episode of Eureka!, a McKinsey podcast on innovation in life sciences R&D, hosts Navraj Nagra and Alex Devereson speak to Kim Branson, senior vice president and global head of AI and machine learning at GSK. They discuss how these emerging technologies are being used to potentially reshape how drugs are developed and about the complexities AI and machine learning introduce in data management, cross-functional collaboration, and responsible innovation.
Driving new approaches to pharmaceutical research and development
Alex Devereson: What is the primary focus of the work you’ve been driving at GSK?
Kim Branson: It’s across the discovery and development continuum. We use machine learning for genetic analysis, cellular and clinical imaging, and active learning. For clinical applications, we’re thinking a lot about building models to predict the effect of individual treatment and how to discover disease heterogeneity. We do a lot of work in computational pathology, where we’re increasingly doing a lot of multimodal models because there’s lots of information present. The dream is for our medicine to be the well-known blockbuster for all possible people. But the diseases we go after are quite complex, with sometimes narrow populations. So we’re looking for ways of identifying those people and building support.
Extracting value from AI and machine learning
Navraj Nagra: Where have you seen the most value extracted from AI and machine learning?
Kim Branson: I’ll give you a good example, which is bepirovirsen. It’s a hepatitis drug that’s in development, and we ran a heavily instrumented phase two trial for it, collecting a lot of different blood-based biomarkers and things like that to work out what fraction of people will respond to it. You could have the best molecule ever, but you won’t see its clinical effect if it’s the wrong target population. And it’s machine learning that helps us bring all the data sources together and link them back to specific populations.
Alex Devereson: Are there different levels of understanding, willingness, and excitement to use these methods across the value chain?
Kim Branson: GSK decided machine learning was a core part of its strategy back in 2019. The first place where that really started to come together was oncology because there’s a large amount of data there, it has a lot of heavily sequence-driven methods, and it’s an area where people are quite familiar with the idea of precision medicine. But people are now realizing more broadly that performance matters. Other areas haven’t had companion diagnostics and things like that because there hasn’t been any investment in or data for them. But I think that commercial pressure is coming, particularly because we are affecting a lot of lifelong chronic conditions. That’s a macro trend everybody’s seeing: the idea of precision medicine moving beyond oncology to everyone else.
Understanding the organizational implications
Navraj Nagra: How do you think about managing the scope of AI and machine learning capabilities and expertise needed to achieve your aim of treating the underlying disease pathway while prioritizing safety and efficacy?
Kim Branson: Multimodal data sets require deep expertise. Domain knowledge absolutely matters for assessing how the data is generated. So, for example, in clinical imaging, I have people who are PhDs, postdocs in clinical imaging, and machine learning people. You really need to know what’s going on as well as what you will collect. So we have people that are both machine learning people and deep experts in some of these domains.
The big thing for us was generating data with the explicit purpose of building models, because we believe that’s a source of advantage. We’re thinking about the data we need to collect and generate as well as the measurement technology we want to bring to bear. People forget that a machine learning organization is not a data engineering organization—these guys are not data engineers. That’s why we created an organization called Onyx to do the data engineering at scale. Some companies have partnerships, but we needed to build this muscle internally.
What separates tech bios from big pharma is scale. Big pharma can deploy capital rapidly, put a lot of capital across a lot of different disease areas at once, and generate data in humans. That’s key, because data with real-world outcomes is gold. You can do all of it theoretically. But if you haven’t got the outcome data from patients and can’t build that linkage model, it’s very difficult to calibrate and use it for discovery.
Driving competitive advantage through data
Navraj Nagra: How do you ensure the organization is generating data that’s useful for you and for the models you’re building?
Kim Branson: There’s always competition for internal resources. And it’s a strange concept for some people when I tell them I need all possible data, and don’t stop generating it. It’s a huge mindset shift for sure.
Alex Devereson: Let’s imagine merging the big pharma and tech bio cultures. What would excite you most about that?
Kim Branson: It still comes down to data. What I’m excited by is the ability to add a lot more sensors through wearables and the like to collect a lot more diverse patient information. We collect a lot of data from medicine EHRs [electronic health records] that are for management and changing things, but it’s not for understanding a disease after you’ve been diagnosed. The explosion of collecting data at scale is going to be the most interesting thing, and I’m really excited by being able to integrate information from different disciplines, such as clinical imaging plus other things we have never combined before.
Balancing innovation with ethics and regulation
Navraj Nagra: How do you ensure regulatory compliance with some of the cutting-edge AI technologies? And how do you think about the ethical implications of applying some of these approaches more broadly to patient well-being and health?
Kim Branson: We have a company-wide set of trainings, and we’ve been interfacing and integrating with regulatory bodies from the FDA [US Food and Drug Administration] and others to embed regulation from the start, such as standards that inform how we build machine learning models. For us, it’s about making sure we balance the return from machine learning with the risk while realizing there will be differential trade-offs.
Where we’re pushing on regulators is around how we can think a little bit differently and why the machine learning models are a little bit different than the pharmaceutical process. Traditionally, we take a medicine, run various stepped-up trials, and then do large-scale trials with a representative population to conclude that a drug seems to be reliable, safe, and effective. Now we’re asking how we can monitor predictions and outcomes in real time. We don’t do that right now for a lot of diagnostics once they’re certified and calibrated, but we easily could.
