We know that kidney cancer is a distinct disease, biologically, from melanoma, non-small cell lung cancer, and other tumour types, so the quest for biomarkers in kidney cancer is different. We can’t just extrapolate from other tumour types. So many research groups, including our own, have looked at one thing at a time and tried to see if that impacts response or resistance. So is a certain gene mutated or not? Is a particular T-cell there or not? While we found moderate associates with response or resistance, what we really need to do is integrate things together. These are complex systems, and we really need to look at not just one thing at a time, but many factors at a time.
That’s the background to this study. We really hypothesised that integration of multiple factors, things we learn about the tumour, but also things we learn about circulating immunity, that together in an integrated fashion would help to predict response and resistance.
What was the methodology and what were the findings?
We leveraged biospecimens collected as part of the CheckMate 9ER trial, this was a pivotal phase III trial of nivolumab plus cabozantinib versus sunitinib for the first line treatment of metastatic clear cell RCC, and it really showed superiority, long-term benefit, of nivolumab plus cabozantinib.
What we did was use both conventional biomarkers but also some novel biomarkers, initially one at a time, and then using machine learning, integrating them together to see how that might predict things. Within the tumour we looked at staining for PD-L1 by immunohistochemistry, but we also looked at an AI-powered analysis of H&E images using PathAI’s PathExplore platform. Within the peripheral blood, we looked conventionally at the immune subtypes that are there by flow cytometry, but we also looked at a new feature, circulating extracellular matrix proteins which give a window into the biology of the tumour itself. Finally, we used a machine learning-based approach to integrate these thousands of different measurements all into a model that would help us predict response or resistance to these different agents.
What are the clinical implications of these findings?
What we saw was that, while individual features, so just PD-L1 status or just a certain immune population had a little predictive ability, what was critical was that integrating these things together from all four different domains, taking the most informative information, that really outperformed each individual variable by itself.
I really think it provides a proof of concept that this sort of machine learning-based integration is possible, and my hope is the next wave. We have a lot of work to do; we need to improve the features that we integrate into this model. There’s lot of other things we can include – the genetics of the tumour, more details about the immune cell types – and there’s lots of validation that’s needed. This is our first discovery set but we really need to validate this in larger settings.
But the concept of this, that you can really integrate things together using these machine learning-based approaches, rather than looking just at one thing at time, it sort of embraces the complexity of these tumours in a way that I think is very helpful.
