NH:     Hi, and welcome to this ecancer activity on molecular testing in breast cancer. We’re going to review some abstracts presented at SABCS 2025. My name is Nadia Harbeck, I’m the Director of the Breast Centre at LMU University Hospital in Munich, Germany and I’m very happy to be joined by my colleague, Giuseppe Curigliano, he’s the Director of the phase I programme at the European Institute of Oncology and the incoming ESMO President Elect. Giuseppe, welcome, great to have you on board.

GC:     Thank you very much for the invitation, Nadia.

NH:     So what we’re going to do is look a little bit at tests that have been presented with results at the San Antonio Breast Cancer meeting. I think the first test that we could talk about is the Oncotype DX and the Oncodetect test. Oncotype DX is the 21-gene assay and Oncodetect is a whole exome sequencing tumour-informed circulating tumour DNA MRD assay. Giuseppe, can you summarise some of the abstracts that were presented for the Oncotype DX – I’m thinking about its utility in the US context as well as in other countries?

GC:     Absolutely, thank you very much. So I will try to summarise the data that have been presented on Oncotype DX. The first one was the multimodal artificial intelligence model integrating imaging, clinical and molecular data from the TAILORx that was presented by Joseph Sparano. What they did, finally, is to integrate the data of digitalised haematoxylin and eosin slides and integrating this data finally with a multiple gene analysis of the primary tumour using Caris MI Tumor Seek Hybrid. Finally, of course, integrating this data with the data of Oncotype DX. If you remember, in that presentation five commercial gene signatures have been included, the most important ones are ODX, MammaPrint, Prosigna, EndoPredict and BCI, with 57 high-variance genes. At the end of the study, finally, with this multimodal analysis, molecular features primarily were driven by prognostic accuracy of Oncotype DX where histological features strengthened the prognostic accuracy for late risk of relapse. So, according to this data, artificial intelligence can be integrated to the data of Oncotype DX and finally this can really support the opportunity to use AI, starting from haematoxylin and eosin digitalised tumour slides, to predict the prognosis of patients.

The second study that used Oncotype DX is a multimodal, multitask, deep-learning model from NSABP B-42 and validated in the TAILORx from late distant recurrence risk. This is a completely different approach because in this study investigators included 6,500 TAILORx patients with digitalised haematoxylin and eosin slides and the relevant clinical data. So here the study is focussed only on Oncotype DX. In the overall cohort 1,134 patients were classified as Clarity BCR high risk and 5,000 as Clarity BCR low risk, according to this algorithm. The prognostication discrimination C-index was 0.5 for Clarity BCR and 0.5 for Oncotype DX. So also in this case an AI-generated algorithm demonstrated a robust, independent prognostic performance.

Finally, I would go on the evaluation of whole-exome sequencing tumour-informed circulating tumour DNA MRD. This was evaluated in the GeparDouze trial. Finally, this is a really large evaluation as a prospective sub-study that included patients included in the NSABP B59 GeparDouze trial. Finally, the idea was to have a prognostic impact of these ctDNA results on the long term. So the median age of the patients was 50; 78.9% had high tumour grade, 61% not positive at presentation. This type of test, finally, was available on all patients with both pathological complete response and residual disease. Distant recurrence was predicted in 9.5% of the patients. So this MRD assay was highly prognostic for distant recurrence in triple-negative breast cancer patients following neoadjuvant therapy and surgery. The assay demonstrated a high prevalence of ctDNA in those patients that had been stratified in the context of this study.

Finally, I would like to mention, of course, the use of the liquid biopsy molecular profiling using Guardant CDx at progression in patients on CDK4/6 inhibitors plus endocrine therapy in the CAPTOR study. I believe this is an important study because this is a real-world study in which patients have been evaluated in advanced breast cancer after first-line treatment with ribociclib plus endocrine therapy. ctDNA was evaluated in these patients following progression to CDK4/6 inhibitors and the tests really evaluated the presence or not of important genes that can really inform on the biomarker driven approach following progression to CDK4/6 inhibitors. So PIK3CA was detected in 48% of the patients, ESM1 in 32% of the patients, some patients also assessed the presence of somatic BRCA mutation and also germline BRCA mutation. An important study that can be integrated, of course, in the real world setting.

NH:     Thank you so much, Giuseppe. Those were great abstracts and the AI-driven biomarkers were also presented – I think one was an oral presentation by Dr Sparano. Let’s split these topics a little bit and talk maybe first about the Oncotype DX which is a prognostic assay mostly. There were a couple of abstracts also looking at the cost efficacy in different settings in the US as well as outside of the US, and a small Swiss study that showed that in their context, in Switzerland, Oncotype DX influenced adjuvant decisions in about 20% of the patients, mostly leading to chemotherapy de-escalation. How do you view these multigene tests? Do you use them in your clinical practice and what is your experience?

GC:     No, they are used, of course, in my country in clinical practice. They are completely reimbursed from the national health system. We did a study of cost effectiveness in our country and it’s very clear that if you use Oncotype DX you can de-escalate chemotherapy and you can reduce the cost of treatment for patients with HR+/HER2- disease. So in the experience of my country and my clinical practice, the use of Oncotype DX is cost effective and, in the majority of cases, this can help you in the adjuvant treatment decision making, sparing a lot of chemotherapy for patients that don’t need chemotherapy. So at the end of the story we suggest the use in clinical practice guidelines because this approach is cost effective.

NH:     Yes, I was very much impressed by the numbers from the abstract from Ireland where they said they had €22 million savings over a period of ten years in Ireland. But they not only looked at the treatment costs but also the society costs, like the loss in productivity in the workplace as well as caregivers and that makes sense in this setting. These costs are sometimes disregarded because they are not paid by the health insurance companies, so different payers, but I thought that was a very interesting abstract as well.

Let’s look a little bit at the abstracts that looked at the AI component where they showed that Oncotype DX, for example, in Joe Sparano’s abstract was mostly predictive for the early recurrences because naturally that’s how the test was developed whereas the histology seemed to be more important for the later recurrence. The AI model picked that up quite nicely. Do you think that these AI models are the future in breast cancer biomarkers?

GC:     Personally yes, because once you will digitise haematoxylin and eosin slides and you have the opportunity, of course, to correlate digitalisation of the old pathology with the new data related to whole-genome sequencing but also Oncotype DX, I am really sure that in the future this will give the opportunity in many centres to have a prognostic indication on our patients based maybe only on haematoxylin and eosin. So personally the data of Sparano integrated genomic data with haematoxylin and eosin data, it may be this type of model can refine better the prognosis. But what if we use only haematoxylin and eosin data and we conjugate with the data of Oncotype DX and we try to address morphological features that can have the same prognostic impact as Oncotype DX? So personally I believe AI will really change the way we are going to define prognosis of our patients with early breast cancer. I really believe in your large study, the ADAPT ER, you need also to validate the use of artificial intelligence because the more I look at the data, the more I believe we should integrate in clinical practice also outside Germany. Because in many cases you can avoid chemotherapy to many patients also in the pre-menopausal setting.

NH:     Yes, I completely agree. We have three disclosures at San Antonio with a combination of the WSG data, the ABCSG data and NSABP data, just looking at a multimodal AI model that was developed in the PlanB and ADAPT study then further refined in the NSABP dataset and then validated in the ABCSG datasets. We saw very good prognostication just from an H&E section in hormone receptor positive early breast cancer. I think the small start-up that we did the research with is now going to take this to the FDA and hopefully there will be a product that will be available for patients globally. I think that the global outreach of these H&E-based AI prognosticators is great because it doesn’t require much technology, just scanning a slide and sending it for analysis.

Just maybe, as you’re the ESMO President Elect, I just want to highlight an ESMO paper that just came out before the Christmas holidays looking at the EBAI system, which is a framework addressing AI-based biomarkers for clinical use, where the expert group classified these biomarkers in class A, B and C, A being AI quantification of established biomarkers, class B indirect measures of non-biomarkers using AI technology, and class C these novel AI-derived biomarkers that we’ve been talking here from the San Antonio disclosures. There is some guidance on the rigour of validation before we can use these in clinical practice. Do you think that we should apply these recommendations for future AI-based biomarkers?

GC:     Yes, absolutely. Yes, because the major problem is that the majority of the AI biomarkers are generated in retrospective analyses. So what ESMO tried to do with this important position paper is to generate a level of evidence for AI biomarkers, as you say they’re rated class A, B and C, in order to provide indication on how we should validate prospectively biomarkers generated by AI in a retrospective setting. The EBAI framework addressed AI biomarkers for clinical use, so performance, generalisability, concordance study, with class A that require concordance studies, class B with analytical validation, and class C1 with high-quality retrospective real-world clinical data. So personally I believe this is going to change the way we approach in terms of prognosis and prediction of response the treatment of our patients with breast cancer.

NH:     I completely agree. So let’s use the last part of our conversation to talk about the liquid biomarker data that we saw at San Antonio. We’ve already summarised the data, the validation of this assay in the GeparDouze data which Marija Balic presented. There was also another dataset from another group, from the MONITOR trial, looking at a different technology, also patient- or tumour-informed whole-exome sequencing. What I thought was really interesting was they both looked at the neoadjuvant setting, both found over 90% ctDNA at presentation and then rapid clearance and only those patients that did not clear or were ctDNA positive after surgery, they had poorer outcome, be it that they had residual disease or they had distant recurrences. So I thought that that was very strong data confirming that this MRD assay may be helpful in the neoadjuvant setting. How do you think we’re going to use this in the near future?

GC:     Personally I believe we are going to generate a new patient population that is a patient in the curative setting with tumour removed. After surgery, without evidence of radiological distant metastasis, the patient can be ctDNA positive or ctDNA negative. So ctDNA can be used for adaptive trials in order to understand how to treat those patients without ctDNA clearance but can be used also to personalise adjuvant treatment in those that despite treatment there is no ctDNA clearance. So I really believe it’s time to think about changing also the TNM. As we incorporate Oncotype DX in the TNM, it’s time maybe to consider to incorporate ctDNA positive versus negative or patients with ctDNA clearance versus no clearance in order to generate a new patient population and to design new clinical trials for the future.

NH:     I think that may be the best way forward, to look in the neoadjuvant population and the post-neoadjuvant setting first and then also take this to the adjuvant setting. The data from Sherene Loi from the monarchE study, also, I thought was very intriguing where they showed that basically everybody who was positive before starting with abemaciclib and then remained positive had relapsed after the first two years. So I think the data is now strong enough and the sensitivity of the assays seems to be good enough. The last point is the predictive gene testing which you summarised from the CAPTOR study with the Guardant assay. I think they had a lot of targetable genes detected; I think in the majority of patients they had an alteration, about 72% had a targetable alteration. When do you think is the best time to use these assays in the metastatic setting?

GC:     This is a good point. If you add inavolisib in the first-line setting I would test also in the first line because we need to detect those patients with PI3 kinase inhibitor eligible for inavolisib plus palbociclib plus endocrine therapy. In the majority of the older patients with endocrine-sensitive disease of course you can do after progression to first line CDK4/6 inhibitors. So you have a group of patients that are resistant to adjuvant endocrine therapy in which you can test for PI3 kinase, and then you have another class of patients that can be tested after progression to first-line CDK4/6 inhibitors. But now that we have patients progressing to adjuvant CDK4/6 inhibitors, I believe also testing with the multigene liquid biopsy test that you proposed in your study is a good opportunity to try to understand how many patients may have ESR1 mutation or PI3 kinase mutation or also AKT, PTEN, because we know perfectly that the biomarker-driven approach will be the future.

NH:     So I think maybe even we’ll not just test once but we’ll see if the SERENA-6 principle is approved by the FDA we’ll then test sequentially and act upon molecular progression if we see more solid follow-up data that this will actually impact survival of patients. So I think it’s an interesting future with these liquid tests, both for the MRD in the early breast cancer setting as well as for the targetable alterations in the metastatic setting. Any last comments from your side, Giuseppe?

GC:     No, I believe the future really is an integration of artificial intelligence and super-sensitive models and technologies to detect minimal residual disease or molecular residual disease.

NH:     A very nice concluding statement. Thank you so much also for summarising the disclosures at San Antonio. It’s always a pleasure to work with you. Thank you to our audience and thank you for watching this ecancer short webinar.