An AI-derived digital pathology-based biomarker to predict the benefit of ADT in PC with validation in NRG/RTOG 9408

Bookmark and Share
Published: 3 Mar 2022
Views: 255
Prof Daniel Spratt - University Hospitals Seidman Cancer Center, Cleveland, USA

Prof Daniel Spratt speaks to ecancer about an AI-derived digital pathology-based biomarker to predict the benefit of androgen deprivation therapy in localised prostate cancer with validation in NRG/RTOG 9408. Initially he talks about the background of the study.

Prof Spratt says that in this study they trained and validated the first predictive biomarker for ADT use in prostate cancer using multiple phase III NRG Oncology randomised trials. He further talks about the methodology and the key results from the study.

The study successfully validated in a phase III randomised trial the first predictive biomarker of ADT benefit with RT in localised intermediate risk prostate cancer using a novel AI-derived digital pathology-based platform.

Prof Spratt concludes that this AI-derived predictive biomarker demonstrates that a majority of patients treated with RT on NRG/RTOG 9408 did not require ADT and could have avoided the associated costs and side effects of this treatment.

Right now, one of the standards of care in the treatment of localised prostate cancer is the combination of radiation and androgen deprivation therapy, people sometimes call it hormone therapy. Hormone therapy has many side effects and although it has been shown, we have shown in a MARCAP meta-analysis, that it significantly improves survival in an unselected population of patients, it’s very clear that not all patients derive benefit from it. The side effects, because it lowers testosterone, are bothersome to many men – decreased libido, hot flashes, weight gain, loss of muscle mass. So there has been a desire for many years to identify predictive biomarkers to be able to identify which men should or should not have hormone therapy. However, there have been no validated predictive biomarkers to do so, so instead we rely on prognostic biomarkers to identify patients that have a very low risk of recurrence to have a shared decision making with patients to the pros and cons of receiving hormone therapy.

So we’re in an era now that due to computational power and the era of artificial intelligence and deep learning that we can now go far beyond our typical variables in prostate cancer – PSA, Gleason score – by capturing a wealth of unused data. In this study I’ll be focussed on the actual histopathology slides that right now are only used for Gleason grading by a pathologist but there’s a wealth of information. So we had the hypothesis that using five phase III randomised trials that we could go back and get all of their histopathology slides, digitise them, and using AI technology to be able to create and train a model and validate it to potentially be the first predictive biomarker to be able to offer to patients who should or should not receive hormone therapy.

What was the methodology used in this study?

What was done is we worked with the NRG Oncology, they’re a NCI operated clinical trials group in primarily North America. We got access to five randomised phase III trials that treated patients with radiation or with radiation and various durations of hormone therapy. We accessed the NRG biobank to actually access all the slides and we had them digitised so that they could now be in a quantitative state that can then be fed through a series of novel pipelines that a team of artificial intelligence scientists that we worked with at a company called Artera. Using both the clinical data like their age, their Gleason score, T stage, PSA, as well as processing all these slides through and having feature extractions and these were non-annotated slides, meaning that it wasn’t that we were telling it to only look at certain features, it was an unbiased, I guess you could say, screening of the slides to pull out relevant features across the thousands and thousands of samples. In total there were about 16 terabytes of imagery data so four of the trials were used to train the model to predict differential benefit or differential outcomes in patients receiving radiation versus radiation plus hormone therapy.

The primary endpoint was distant metastasis, a very clinically meaningful endpoint. A one-step model was optimised and locked. We then independently with an entirely separate randomised trial, a separate biostatistician we applied that locked model to RTOG 9408, it’s the largest randomised trial and it’s very important because it’s randomised men to radiation with or without short-term hormone therapy. The vast majority of the patients on that trial had available slides to be imaged and we then applied the model to that trial to look at what we call a biomarker-treatment interaction to see would there be differential benefit based on patients who are biomarker positive or biomarker negative for the receipt of hormone therapy.

What were your findings?

The key findings are, number one, we were able to successfully train a model that was able to identify differential benefit of hormone therapy. Probably even more important than that, because you can train things to do a lot of things, is we independently validated for the first time ever in a randomised trial that this biomarker, this predictive AI-derived biomarker, predicted differential benefit from hormone therapy. So about one-third of patients on that validation trial were biomarker positive and they had a substantial benefit of hormone therapy, about 10% with long-term rates of improvement in distant metastasis. Whereas biomarker negative patients, which was two-thirds of the patients on this trial had basically no… the hazard ratio was 1.00, there’s effectively no benefit at all from the receipt of hormone therapy on metastasis.

So this is very powerful and we looked at other endpoints as well – death from prostate cancer, metastasis free survival – and all of those endpoints also showed that there was significantly greater benefit of hormone therapy in the biomarker positive patients and effectively no benefit in the biomarker negative patients.

How can these results impact the future treatment of prostate cancer?

This is a very rare project in the sense that usually when you create a biomarker it sometimes is many, many years until you have sufficient evidence to where you would say it’s clinically ready to be deployed. Our group, combined with NRG and Artera and a very large team learned from prior biomarkers we’ve worked with to go straight to high level, high quality, randomised data. So what we have now is a predictive biomarker with independent validation in a randomised trial. So this really is the bar that most of the time is used to say that this is clinically ready for deployment. So in 2022, hopefully by the end, we’ll be able to have this available for people to order so they can send their slides to be digitised and this algorithm run on their slides. This is going to be profound for patients because they can now rather than just telling all patients with predominantly what we call intermediate risk prostate cancer to receive hormone therapy, the majority of them can probably successfully omit hormone therapy with confidence that they will derive no intrinsic biological benefit.

So this is going to be a major game changer and this is really the entrance, the beginning, of a whole new era of AI-derived biomarkers that are going to really help personalise therapy.

Is there anything else important that you would like to add?

There were almost 6,000 patients that enrolled on these trials so really we have to thank the patients’ families for enrolling and for participating in research, NRG Oncology, all of the investigators, collaborators, Artera. Really, it’s a huge, huge team effort and very exciting work.