By Gerard Oakley, MD
What is IHC 0 in HER2 breast cancer patients, and how can it help? So the current study is really an expansion on previous press releases that we've made about the development of an algorithm that we call HER2 complete which is designed to evaluate HER2-positive and negative breast cancer cases. And we've really aimed it for the IHC 0 category within HER2 testing. And so what we've done is we were able to access an additional 900 plus cases within IHC 0 from our collaboration with Memorial Sloan Kettering, as well as take a look at those cases within TCGA and just ex expand on the performance and validation of our algorithm. And essentially what we find when we do this is that the algorithm that we have developed is very adept at picking up what we are calling true HER2-negative cases or HER2-nullHER2-null, and what we call HER2-null or HER2-negative, we've used the two interchangeably.
Is those cases that are IHC 0 and have essentially no expression of ErbB2 by Messenger RNA (mRNA). And so that forms the ground truth of what the algorithm is taught to recognize as a true HER2-negative. And the reason that we focus there is that we believe that this is a level of expression that will ultimately be too low for even some of the newer HER2 targeted therapies like the antibody drug conjugates to find, bind, and deliver payload to tumor.
And so we wanna be able to really, pick those cases out, especially within IHC 0 disease, and the algorithm is able to do that. It's able to identify a phenotype within those, 900 additional cases and certainly the several thousand cases that we had previously studied across IHC staining spectrums.
And when it does, so it identifies two patient populations within that IHC 0 category. It finds those cases that we identified as HER2-negative, so IHC 0, very low level of mRNA expression, and it finds another phenotype within IHC 0 staining breast cancers. And those look to the algorithm much closer to cases that should have stained at least IHC 1+ or better.
And when you look at their mRNA expression as you might expect, they tend to have higher levels of mRNA expression. Not only is it detectable, it tends to be, again, in the range that overlaps what with cases that had stained, bonafide on IHC within that, HER2-low category or above. And so ultimately what we find is that we arrive at a really interesting hypothesis with this algorithm in this work.
Within IHC 0 disease, we find some cases that the algorithm can clearly pick out as a phenotype that is associated with essentially negative HER2 expression. And our hypothesis is that, Cases are probably enriched for non-responders to some of the antibody drug conjugate treatments that are out there that have got recently got approval like in HER2 simply by virtue of the fact that they probably don't express enough of the target for these agents to be effective against them.
But it also raises the possibility that within IHC zero disease, there's this separate patient population that looks to the algorithm closer to cases that should have stained HER2-low and it's a tantalizing hypothesis that some of those cases that the algorithm identifies actually have sufficient residual HER2 express, to be able to respond to some of these therapies. Now of course, the proof of that is, is in the eating, just like the proof of the pudding is in the eating. And so our next step with this particular research that we've published is not only to continue to look at its performance in external dataset such as TCGA but also to attempt to get some of the clinical outcome data particularly with an IHC 0.
To test this hypothesis and see if it is also associating with response in a way that we think it could potentially do. And in terms of the outcome with the TCGA the cases that we looked at the long story short is that it turns out there's just not a whole lot of IHC 0 cases, particularly ones with low mRNA expression.
We were able to come up with a few different methods for determining what essentially zero expression of Erb B-2 was from the RNA-seq data that TCGA has available to it. The method that we presented that we felt worked the best was actually in comparison to three reference genes in breast cancer (cells) that we know have essentially no expression.
And unfortunately when we did that, there's just not a whole lot of what we would consider true, negative cases with both IHC 0 as well as very low levels of mRNA expression. However, in the, 30 odd cases that we were able to identify, the algorithm was able to pull out the majority of those that were, our definition of HER2-negative, so IHC 0. Very low levels of expression of mRNA. So we have at least some early indication that it will generalize to other data sets as well. But that's something that we hope to follow up on again with a larger set of cases. Hopefully ones with more true negative examples, as a future next step for this currnet study.
The most common question that your colleagues ask you anytime you have a poster presentation is tell me about this poster. So I, I think we addressed that in the first question. The second most common though really relates to the potential clinical applications for this assay.
Because with the approval, in HER2-low disease. I think everybody recognizes, and certainly my colleagues who were stopping by the poster recognized that there really is a new diagnostic dilemma that is facing the breast cancer community around HER2. The classic definition of HER2-positive and negative was, of course, therapeutically driven. The IHC and INSIGHT 2 hybridization that we use now to define classic HER2-positive disease was those levels of expression that identified a population of breast cancers that had overexpressed an overactive HER2. Because these are the, tumors that would respond well to Herceptin and Projeta.
And so because of that, the immunohistochemistry in particular was optimized for that very high level of expression. And now that we have a new class of drug that doesn't require blocking, overactive and overexpressed HER2 but instead just uses any amount of HER2 expression as a way to guide the antibody drug conjugate to the tumor and deliver payload directly to the tumor.
And is shown efficacy at lower levels of HER2 expression down into what they're now calling HER2-low disease, which would be IHC 1+ or 2+ without evidence of amplification. We have a redefinition of what it means to be HER2-positive, because after all that IHC low category, at HER2-low category IHC 1+ and 2+ without amplification was something we used to call up until really just this year, and never thought anything more about it. And we know from previous clinical trials that there is, certainly significant inter and even intra observer variability for pathologists who are attempting to evaluate immunohistochemistry on the 0 to 1+ cutoff, even for the FDA approved IHC kits.
It's not something that any of us have ever really had to train, been focused too hard on in the past and now that is a clinically critical cutoff, to be able to identify and to adjudicate correctly on either side of that. And we also know from some early reports that are starting to break into the literature and pre-public prints and other abstracts, that there may be some variability in the, test results on the same case for different IHC kits that are available as well. And so it speaks to the need of an improved diagnostic approach in how we are identifying and classifying HER2-positive disease. And so again, that's, it's why we tried to focus our algorithm on where the immunohistochemistry by definition cannot go at the moment because we're focused in IHC 0 disease where, there's already essentially no or very little expression of HER2 according to the current diagnostic criteria.
And so the way that I was discussing it with my colleagues at the poster at the San Antonio Breast Cancer Symposium (SABCS). Was that we really have two possible roads into clinical (practice) use for this kind of algorithm. The earlier possibility I would say would be something like what we call an injunctive use case because the algorithm is very, highly trained to identify truly HER2-negative cases, so IHC 0 and essentially no mRNA expression. We believe there's a possibility that it could help pathologists who are dealing with a particularly tricky case from the new diagnostic cut point between IHC 0 and 1+. By being able to be run on the scanned H&E image, because that's where our algorithm works.
So you do your IHC stain as usual, and if it was a really difficult borderline case or if the oncologist was calling and saying, Hey, I'd really like to try an antibody drug conjugate for this patient. If you can tell me that there is, a level of standing that is appropriate, that would show that this patient should be treated with this.
So are you really sure that this case is a true IHC 0? So in those settings, the pathologist can now go back to a scanned version of the H&E. They can run our algorithm on it and it can tell them if the algorithm thinks that this is a case that looks. , it has the phenotype of what we defined as HER2-negative.
And in that case, we think they can be fairly confident that case really and truly is an IHC 0 and is unlikely to respond even to some of the newer agents. Now, on the flip side of that, we don't really have a good sense yet what the best course of action is if it looks at one of those borderline cases and thinks that it is, HER2 expressed, which is the other potential outcome in that setting from the algorithm.
Is that a case that you would automatically go directly to calling it IHC 1+? I'm not sure, without, additional clinical trials. Would the next best course of action be a consensus conference with your other pathologist so that everybody can look at it and decide along with the algorithm if they agree that it's now, over that borderline and is now true HER2-low. Again, because of the therapeutic implications, is it best to try and go back and stain another block of the same tumor if there's another tissue block available? Should you try a different IHC kit? Should you try another method entirely? We don't really know, so we're trying to launch, hopefully early next year a clinical study with some partners that we're trying to work with to look at some of these cases that are on the borderline, use the algorithm on them and see if we can get some clear clinical direction for how to best handle, cases that are called HER2 expressed in that sort of borderline setting, in addition to you verify its performance as a useful adjunctive tool to give pathologists confidence that an IHC 0 is really and truly IHC 0. The other, use that we're exploring with this algorithm is what I like to think of as the more complete version of HER2 complete. Like I said, we have this really tantalizing hypothesis in front of us where we think the cases that it's calling HER2-negative with an IHC 0 disease are likely enriched for non-responders to drugs like in HER2.
Some of the ones that it's calling HER2 express within IHC 0 may be potential responders, and could be a population of patients that are not currently covered by the existing label for in HER2 but could potentially receive benefit. But that needs a clinical trial to be able to examine that possibility. And so we're hoping to get access to clinical outcome data for some of these drugs alongside some of, potential pharmaceutical partners who have, drugs targeted to antibody drug conjugates targeted to HER2. And of course that study is going to take a little bit longer to be able to do.
Right now there's just simply not a whole lot of cases that we're aware of where that were really and truly IHC 0 and received in HER2 and have, mature enough outcome data to really validate that hypothesis. We need a bigger pool of those patients before we can be confident in a more complete version of HER2 complete.
Where now you, we would know if a, HER2 expressed call by the algorithm for instance, means that's a patient who's likely to respond like a patient who is truly HER2. It's also an opportunity to train directly to outcome, which is something that we'd like to add to this particular version of the algorithm.
It does a good job in detecting HER2 expression at the moment, particularly HER2-negative expression. But of course when you were thinking about the antibody drug conjugates, HER2 is only part of the story. And while it's the major focus at the moment for the breast cancer community, it's important to remember that the payload matters too.
And there may be phenotypic signals that we can identify if we can train directly to outcome with these drugs that also predict off of the same scanned H&E slide, not only is it expressing enough HER2 to be targeted. But if it has indications that the payload is likely to be effective, and in this regard, it really speaks to the power of artificial intelligence for problems like this, because if we can do it, and if those phenotypes exist for payload response prediction, then this kind of approach becomes a one-stop shop for the best way to identify patients who are likely to respond to these drugs.
And that, to me is, the most complete version of our, HER2 complete algorithm. And that's where we hope to steer this on into the future. And so it, a lot of the discussion was, this is, Interesting tantalizing data and what are the next steps and how do we get there for them?
Read and Share the Article Here: https://oncologytube.com/v/41598
Listen and Share the Audio Podcast Here: https://oncologytube.com/v/41599
Antibody drug conjugates (ADCs) against HER2 have demonstrated clinically relevant activity in HER2 low breast tumors, defined as 1+ or 2+ staining on immunohistochemistry (IHC) without gene amplification using in situ hybridization (ISH) techniques.
Based on the detection of invasive breast cancer in whole slide images of H&E staining, a model was constructed and then trained using a computational neural network with multiple instance learning for binary categorization of cases as HER2 "negative" or HER2 "expressed" (low).
Concomitant transcriptomics data (RNASeq) were retrieved for the TCGA cohort as a reference for HER2 mRNA expression, and "HER2 expressed" was defined as RNASeq expression of HER2 greater than the 90th percentile of the geometric mean of expression of three reference genes not expressed in breast tissues (TTN, MUC13, OR10A6). In the TCGA cohort, values less than this reference cut-off were classified "HER2 not expressed."
The model identified 82 IHC-0 test cases from the MSK cohort as 'negative,' while 819 were determined to contain HER2 expression traits (HER2-Low). Except for 13 instances, all of the 82 negative cases in the MSK cohort had mRNA levels greater than 9, and 786/819 of the HER2-low cases had mRNA levels greater than 8. Our reference-based RNASeq expression cut-off resulted in 33 "HER2 not expressed" cases among the 52 IHC 0+ cases in the TCGA cohort. Our approach classified 15 of these 33 instances as 'negative,' while 15 of the 19 TCGA cases with IHC 0+ and HER2 'expressed' by our cut-off were classified as HER2-Low.
AI algorithms that analyze WSIs from routinely generated H&E sections may be able to predict HER2 status in breast cancer. Further research employing treatment response data is needed to indicate that instances with morphologic features of low level HER2 expression will react to ADCs.
AI more broadly has a couple of different inroads into oncology at the moment. I would say the field of medicine, and it of course touches oncology. That far and away is using the most artificial intelligence out of the gate at the moment is actually our radiology colleagues.
There's quite a number of artificial intelligence algorithms that are active and they're being used in the clinic at the moment to help guide their interpretation, focus their interpretation to suspicious lesions on the various CT and MRI scans. The challenge with bringing this into pathology and of course, pathology provides a lot of the laboratory support diagnostic support, prognostic support and treatment predictive support increasingly to oncology.
Is that the, there's a huge difference in the, just the file size between your average CT scan and MRI scan and just what is the scanning size of a pathology slide. It's orders of magnitude, more data. It requires more to compute and it's taken a little bit longer to get there. But I think that we certainly have the computing power, we have the algorithms, we have the models to be able to start to do this, and already you can see it start to move in on the diagnostic side. Paige, for instance, the company I work for as a pathologist, we have FDA approved diagnostic products already out and in the market, and they're actively being used on patients around the world with academic and reference laboratory partners here in the us several across Europe, several through Central and South America as.
And so the lead product at the moment is what we call Paige Prostate Detect. And so that's the one that has FDA approval and there it's using artificial intelligence to help make sure that pathologists do not miss very small focuses, Foci of prostate cancer on core needle biopsies. So the algorithm will scan the biopsy and it will flag areas that it thinks are suspicious for, focus of cancers, so the pathologists can go right there and see if they agree. And we actually have some pretty strong and robust utilization data on that as well. Showing for instance, that it will take, an average community pathologist performance. In that setting with the Paige Prostate Detect Assistance up to the level of someone who's a specialist pathologist in terms of being able to find those small Foci and a specialist pathologist who does nothing but prostate, all day long.
And even some of the specialists it's been able to prefer, improve their performance as well. They don't miss many, it has helped them pick up a few here and there. And in terms of cost benefit to the entire system. It's also been shown to reduce the amount of immunohistochemistry that is needed to investigate those kind of small Foci to prove that they're cancer, which represents the cost savings as well.
There's already ways that artificial intelligence has moved into some of the ancillary field supporting oncology. I would say those inroads are likely to continue. There's certainly more, diagnostic applications that we'll be able to identify. I think the work that we presented here through this poster represents the next level step up where artificial intelligence will be able to identify biomarkers or make direct outcome predict.
In and of itself and will only help to increase the adoption of these kind of tools, particularly within pathology where it can be applied relatively easily to, to the images now. And so it can help standardize calls between pathologists, it can certainly make pathologists life it easier.
And it offers the ability, such as what we present here in this poster to go where some of the existing testing methods currently cannot or might not be able to. It can be more comprehensive than some of what's available, or it can even compliment some of them. Other applications that we've looked at, for example and we have press releases and disclosures about these are where you can use artificial intelligence biomarker assays to essentially triage for the presence of genomic alterations.
And in doing so, you can flag those cases that are highly likely to harbor a targetable mutation and, when you do that you only really need to send those cases that the algorithm has identified as highly likely to have that mutation. So it saves time, it saves tissue, it saves money, and there's a, I would argue a potential as well to improve the response prediction. Again, that's gonna need further clinical trials, but the tantalizing possibility is there, because these H&E algorithms that we develop here at Paige are really recognizing an activated phenotype so you can marry the activated phenotype that the artificial intelligence is recognizing to the genomic alteration and not only tells you that this is a case that you need to look at to see if it has the genomic alteration that you want to target. But if it does, it means it's highly likely that genomic alteration is, truly active in driving that tumor because the expected phenotype is now being seen by the H&E analyzing algorithm.
And so I think, it's early days yet we're still climbing the adoption curve overall. But, there's a lot of momentum and I think a lot of good reasons to expect the artificial intelligence tools to continue to move into the clinic.
For pathologists I think, the takeaway is quite simple. It. Additional evidence that some of these tools to make what I like to think of as making everything that was old and pathology new again are on the way. When I trained back in residency and fellowship, I was fortunate enough to overlap with that generation of pathologists who did not have immunohistochemistry and sequencing and a lot of the ancillary techniques that we have now available to us. And I'm sure my fellow pathologist watching this will agree. It was amazing to see what those guys and gals who did not have the advantages that we have in terms of ancillary testing techniques, could pick up off just the H&E slides.
And so with artificial intelligence looking at H&E, we have the ability to now. Let the computers look at tens of thousands of cases, for us, in a week, a 2 week period and pick up patterns that, it would've taken us years to stare at those same tens of thousands of cases and identify potentially clinically relevant phenotypes.
They'd either predict other biomarkers or our directly predicting response outcomes or other clinical significance in outcomes that would address real clinical problems, and help patients and our, clinical colleagues in oncology with clean and clear therapeutic decisions to treat that patient's individual cancer.
And so if some of the follow on work that we have planned, continues to deliver on the promise that we think that this poster represents. We may also have some tools that may help them gain some confidence in some of the borderline cases that they're already getting calls about for in HER2 treatment and HER2 IHC 0 1+, as well as some new tools that, that may ultimately we would hope Be a little bit more comprehensive in response prediction coming down the future.
And help, with the redefinition of what true HER2-negative breast cancer (cells) will be on into the future. For the oncologists same kind of story. In clinical oncology, they're struggling just as much with, this exploded population of women with HER2-positive breast cancer who could potentially benefit from this new therapy, but they're all well aware that diagnostic cutoff is now a little bit more subjective. And so there's definitely concern among the oncologists as well as to how best to address that into the future. Hopefully, the work that we've begun in this poster, like I said, continues to deliver on its promise and we have some tools to be able to help provide confidence and better direction on into the future.
The best message for the pathologist, oncologist watching this video is, to just stay tuned to this space. I think we all know that the excitement around this new class of drugs. That's available to treat an entirely new category of patients in breast cancer is going to force us to completely rethink what we call HER2-positive and HER2-negative disease.
Now, the classic definition was therapeutically driven. I would argue that very soon in the near future, we're going to have a new therapeutically driven definition of what HER2-positive is and HER2-negative is. And that HER2-positive is again, going to be redefined. that level of HER2 expression that, Is likely to respond to at least an antibody drug conjugate. And we may even have to differentiate between, HER2-positive, those cases that have overexpression and overactivation and are likely to respond to Herceptin and Projeta. And whether we mean HER2-positive in terms of, just enough expression to respond to an antibody drug conjugate.
And then of course, the redefinition of HER2-negative. Is probably the next frontier for breast cancer. If we think just about the triple negative breast cancers there's a number of those that are now arguably not really triple negative anymore. They had IHC scores that were 1+ or 2+ without amplification that were called negative under the old testing regime.
But now are potential responders to some of these new drugs, and are they really triple negative? And so how do we best identify a true triple. Case into the future, and what are the therapeutic implications for that? So like I said I think, the work that we're starting on here, the work that many others are starting on here to better understand and establish these important thresholds and cutoffs is going to be certainly the work that we're gonna be all focused on within breast cancer for the next several years.
Dr. Gerard Oakley is a board-certified pathologist who specializes in anatomic, clinical, and molecular genetic pathology. He also holds a medical degree. His professional background includes academic clinical practice, the pharmaceutical industry, and the information technology sector, among other fields. In the context of personalized medicine, he has supervised and coordinated the development of liquid biopsy, molecular, and immunohistochemistry assays for both diagnostic and therapeutic predictive purposes. He has a history of success in a variety of therapeutic fields that may be demonstrated. Having said that, the processes of metastasis and the clonal evolution of cancers are two of his key research interests. His objective is to tailor patient treatment and bring superior regimens to clinical practice by linking a mechanistic understanding of metastasis and clonal evolution in tumor pathology. During the present time period, he is employed by Paige in the capacity of Medical Director of Biomarker Development.