DeciBio Q&A

AI’s Impact on the Clinical Pathology Lab : DeciBio's Q&A with Paige’s CEO Andy Moye

November 8, 2022
Print Friendly and PDF

In our latest installment in our digital pathology interview series, we spoke with Andy Moye, the CEO of Paige to discuss how diagnostic algorithms are expected to impact patient care in the clinic.”


Key Takeaways

  • Large hospital systems and reference labs have increasing interest in digital pathology adoption, particularly for AI-guided diagnostic algorithms; the financial incentives are beginning to align to justify adoption in this group
  • The ability to look at biomarkers from H&E-stained slides has the potential to be truly transformative, especially in settings where testing access is more limited; this technology could help bridge the gap between patients that should be getting biomarker testing and those that actually do
  • Ultimately, AI algorithms are expected to become a tool in the toolbelt of pathologists; pathologists will remain an essential aspect of the pathology workflow; however, these tools may allow pathologists to go beyond what can be done by eye

Thank you for joining us today. To start, would you share a bit of your background as it relates to digital pathology? 

I have spent the last 17 years focusing on pathology, oncology, genomics, and precision medicine. Prior to that, I served eight years flying airplanes for the Navy– That's maybe a different blog post.

My involvement in pathology began around 2005, and I have been especially interested in digital pathology ever since. Back in my early days at US Labs, we were offering pathologists the opportunity to send their samples to a central lab. We would do the technical component, scan the image, and potentially return the scanned image back to the pathologist. That pathologist would then sit AOL style where you'd log into a modem from a hospital computer and wait at least 10 minutes, and maybe your image would finally come up on the screen. That was what digital pathology was for many years; it was very slow work.

I was then able to run a lab at Caris Diagnostics where we had some digital pathology capabilities, but I became much more familiar with it when I became head of commercial operations for Philips in 2016. We received FDA approval for an ultrafast scanner at that time. It's been a couple of years now, and digital pathology has really started to mature. We started looking more at how AI applications of digital pathology could really enhance the pathologists’ workflow at Philips. Throughout my experience, I have seen that digital pathology and AI are a part of the cancer diagnostic pathway, not necessarily technologies that are solely part of the pathologist’s workflow.

Thank you for sharing. We've seen Paige become very active on the clinical front with a first-of-its-kind FDA approval for Paige Prostate Detect. We'd love to hear a bit more about the clinical side of digital pathology. How would you characterize the adoption of AI-based algorithms in the clinic? 

It's been a full year since we got the FDA approval on Paige Prostate. We are still the only company with an FDA approval on an AI algorithm in pathology. We take a lot of pride in that; there's definitely a secret sauce when it comes to our training of the algorithm, and the studies that got us there. [Paige Prostate Detect] has got a ton of excitement in the pathology community as technology that, with regulatory approval, has met that bar where it can now be used as a clinical grade application.

Clinically, what we're starting to see is that more and more hospital systems, particularly the large reference labs and academic medical centers, are really excited about moving forward with digital pathology and AI. For example, we recently announced a partnership with Sonora Quest Laboratories. Particularly, these groups are building their own AI algorithms and want to understand how they can incorporate them into the digital pathology workflow.

I think we all understand that we're still in the very early days of digital pathology and AI adoption for clinical use. These are mainly the large reference labs and academic medical centers; your community hospitals, smaller IDN's, and smaller community health systems are just not there yet. A lot of that is due to financial and economic drivers. We are seeing some early adoption; as we start to see digital pathology codes come in for the next year, I hope we will see them convert to more category one code coverage and we’ll see some more adoption
 

Paige has also made strong strides in the diagnosis of both prostate and breast cancer. We saw you recently received the CE-IVD mark for HER2Complete™. How do you see the uptake of AI algorithms across clinical applications, from diagnosis with Paige Prostate or Paige Breast, to prognosis and even therapy selection? 

Our vision is to be a cancer diagnostics company; we want to use the power of tissue-based AI to transform cancer diagnostics. We frame this for most of the clinical users out there by reminding them that we already have techniques like IHC, PCR, NGS, ISH, and now we have AI. It's another tool to enable that full workflow in a laboratory that pathologists have at their disposal to get to the right answer and help that cancer patient.

Tools like Paige Prostate, Paige Breast, and Paige Breast Lymph Node can enable pathologists to not miss a small first sign of cancer, or to be able to grade and quantify that cancer in a more standardized way. Perhaps they can use it in a quality control manner so they can go back and see if they are missing any small loci of cancer. Using Breast Lymph Node, pathologists might catch a metastatic lesion that would have otherwise been missed because it was small, and they happened to be tired at the end of the day. This impacts treatment. So, these are tools for pathologists to do their jobs more efficiently, in a better manner for patient care. 

We also are developing transformative and disruptive algorithms like HER2Complete, which help pathologists with information that they otherwise wouldn't be able to get from the H&E slide, or would have to get ancillary studies on, like NGS, or profiling, or whatever the case may be. The vision is that pathologists can sit down at their screen and, on a breast case for example, they can pull up the case and get information on the subtype, where that cancer is, calcifications, and mitotic counts. They will also get information on some of the potential biomarkers that are in that tumor, such as PIK3 kinase, HER2 expression, ER expression, PR expression, and all of this will be on the H&E at the moment that slide got created, stained, and scanned. This is not the case today, but we’re not far off.

How important do you think those types of tools are in driving adoption by community hospitals? 

I think we're going to find a pull from the clinician or oncologist that really needs this information. This is going to be information that they need to treat patients. In the case of NGS, we saw early adoption in academic medical centers who had access to this new technology. They were doing it for research purposes, but the type of comprehensive genomic profiling with clinical utility was really only being done at a Foundation Medicine, Caris or central labs. Now, however, we see that just about every late-stage cancer patient should get a comprehensive genomic profile. We're starting to see some decentralization of this process. I think you'll see the same happen with digital pathology– There's going to be a demand from oncologists. Like with HER2Complete, clinicians really need to understand if a patient has HER2 low. It’s very important; if the pathologist can't provide that information, there may be a central lab that uses AI and might be able to provide it.

 

How can AI tools complement the traditional role of the pathologist? What will be the biggest driver of adoption?

If you think about something like EGFR, we know that EGFR in non-small cell lung cancer is as standard a marker as you can get. Just about every single, and certainly every late-stage, cancer patient should get an EGFR profile. The McKesson/US Oncology Network came out with a study recently in their MYLUNG Consortium and found that only around ~71% of patients were getting an EGFR test before treatment was initiated, so there are still a lot of patients who don't get the EGFR test (Citation below).

Imagine having an EGFR screen on AI at the moment of diagnosis. At the press of a button, we could have AI information. If there is a high probability of an EGFR mutation found by AI on an H&E slide, this will really start to impact patient care. We've got some papers and initiatives underway in response to that. Those are the kinds of things that I think will start to be transformative, where you can look at biomarkers on an H&E. It won’t be today that these biomarkers are going to dictate treatment, but the information should at least be able to point to a recommendation for downstream testing. It should be able to identify potential patients who may have something missing in their standard of care. There may not be as much need in large academic medical centers, where there is fairly integrated care and oncologists, pathologists, and radiologists get together for tumor boards and there's a lot of collaboration going on. However, in the community centers where care is more fragmented, and where the majority of care in the US takes place, I think there's going to be a lot of demand.

As the algorithms continue to get more advanced and go beyond what a pathologist can do manually, it further promotes digital pathology as a must-have tool. However, it also forces pathologists to work with technologies they're not familiar with, to trust algorithms to make diagnoses and prognoses. That's a tough sell to pathologists.

What is the role of Paige and other digital pathology companies to push for DP adoption as it becomes increasingly complex and goes beyond what pathologists can do manually?

Interestingly, if you talk to a lung pathologist who’s been doing it a long time, they can tell you there's something there [as far as EGFR is concerned]. They don't know exactly what it is, but they can tell there's some pattern there. There has to be a pattern– I think a lot of it is there has to be some manifestation of that genomic mutation in the phenotype, that ultimately the computer can see that potentially humans aren’t trained to see. I think the question we're asking is really if there is trust in AI.

That's why we deliver products like the Paige Prostate Suite, which ultimately shows you what a pathologist can see. The pathologist is able to agree with what the algorithm is telling them. We believe that's the best way to build trust in AI, and so we really developed our business model across this continuum. We have a platform for pathologists to be able to read cases, and we have FDA-approved algorithms so pathologists can use AI every day and see it in their cases. Ultimately, there will be novel technologies, and clinicians will then have some faith because they’ve seen that AI works on prostate, breast, and colon cancers. So, they can trust that it probably works for EGFR and picks up on HER2 and other things. We also try to have some of that explainability to show pathologists that the new patterns they see are the new utilizations of the tool. Ultimately, it should be seen as a tool or test in the pathologist or laboratorians tool belt that they can use to provide the best information for oncologists.

Will this technology shift the role of the pathologist in cancer diagnosis over the next 5 to 10 years or continue to be a tool in their tool belt?

There was a famous quote by machine learning pioneer Geoffrey Hinton that said radiologists were running over the open, empty space, like Wile E. Coyote, but they just didn’t know it yet. This was implying that machine learning was going to take over their jobs. It never happened. What machine learning did was actually enable an explosion in radiologists. There are more radiologists than ever coming into the field, and this is because there's so much information available now. It's like everything else in the information age. We don’t just need access to data, but we need to curate it and determine what’s important and how we can use it.

Our hope is to enable and empower these pathologists, to give them the power of all this information at their fingertips, so they can make the best treatment decisions for patients. I do think pathologists will be empowered by DP and AI. I think what we'll see is a whole new generation of pathologists that will come into the field who are more like informatics pathologists. They understand how to use the technology. They understand how to curate the information they're getting from AI and all the different algorithms that are going to be out there. We're not the only ones that are generating algorithms; there are going to be hundreds of different groups that do this. So, I think there'll be a whole new generation of pathologists.

Taking EGFR as a specific use case, digital pathology becomes competitive with existing molecular technology. It becomes competitive with PCR in the long term. The more you can do with it (EGFR, ER, etc.) the more competitive it becomes with NGS and other technologies. In the longer term, how do you see these technologies complementing or competing with each other? Are patients going to get an increasing number of tests to determine which therapies they should be getting, or will this result in consolidation because you can determine more from the H&E than you ever thought possible?

What has often been seen in the laboratory space over the years is that even with the advent of new technology, very little has gone away. Every technique that's come into the lab or clinic is almost always added, so IHCs did not replace H&E and NGS did not replace PCR. In fact, PCR has gotten crazy; it's probably way bigger than it ever was. Maybe the only thing that's gone away is Sanger Sequencing– that was slow and painful. Generally speaking, almost everything's been added.

If you have this conversation surrounding liquid biopsy and think about something like the Galleri test from Grail, you might think this will reduce the need for biopsies. I completely disagree. You're going to find that as that explodes in the next five years, if you get a test that says you might have prostate cancer, you can absolutely bet your money that you're going to visit a urologist who will take a biopsy. This is only going to increase the number of patients with a need for pathologists. In the short, maybe five-to-10-year range, I think AI will be complimentary. It will be an opportunity to screen for biomarkers or determine where ancillary testing needs to be done for whatever you think you found.

In 15 or 20 years, who knows? I think we will see companion diagnostics come around in AI. What pharma is looking to say is that DP or AI will be the companion diagnostic for new drugs, probably in the next five to seven years. Certainly before 10 years is out, I think we'll see companion diagnostics come in that are just AI driven.
 

How do you see biopharma using digital pathology algorithms today?

We're really excited about working with biopharma. I think they are a lot farther along in this journey. They're looking at digital imaging and AI as another technique to find patients that could potentially benefit from their novel drug. We recently announced a partnership with Janssen on an FGFR bladder algorithm to stratify patients in clinical trials. One of the values to Paige is that we can deploy those algorithms at CROs and different clinical trial sites. We can help find these patients, either as an exploratory algorithm alongside another methodology, whether that's PCR, IHC, or something else, to see if it is just as good as the SOC or if the algorithm is identifying more patients.

Ultimately, pharma sees the value in an H&E algorithm that's global and easily deployable. At the end of the day, in  certain places around the globe, you don't have access to IHC; maybe you do an H&E and that's all you can get. If you can get that image scanned, you can get access to the types of algorithms that will help dictate treatment. I think there's a lot of excitement in the pharma world. We were recently with ESMO in Paris. There was a lot of discussion surrounding different biomarkers, and there was a lot of angst around why people aren’t testing more now that so many biomarkers have been identified. Why can't people test for the KRAS G12C mutation or the exon 14 skipping mutation? Well, there's a lot that goes into that. Frankly, there are a lot of drugs available for patients that work really well, like Larotrectinib for NTRK; the challenge is finding these patients. I think pharma is really excited about seeing AI as another technique that can help find patients who would potentially benefit.
 

What do you see as the key trends moving towards CDx development?

There's a lot of excitement about the field, and there are also a lot of questions about it. For AI to be a good companion diagnostic, like Paige Prostate in the clinic, it has to be generalizable and robust. It has to function across genders and ethnicities and be free from bias as much as possible. It has to be deployable globally. Can you get these algorithms into the hands of labs who will ultimately run the test and provide answers for oncologists and clinicians? These are the kinds of questions pharma is trying to answer right now. We see ourselves as part of that continuum of care. It's a workstream, starting in development and moving to clinical trials with academic medical center sites and CROs. Ultimately, we end up in the clinic and can impact patient care. That's really what we care about, that everything we build and touch here will ultimately impact patient care.

Let’s broaden the scope of the discussion a little bit to consider different analyses and types of digital pathology. We've come across digitizing single plex, multiplex, and deep H&E / finding novel markers within H&E. What do you expect to be of the most interest among Biopharma and also within the clinic?

I think biopharma is exploring all relevant aspects, whether that's spatial omics, multiplexed immunofluorescence, single cell RNA seq, DNA seq– There is a universe of what's possible to explore in this space. AI and machine learning could be incorporated into each of these, but in some cases, such as humble H&E, we just need a stain. We don't need all the extra steps. At Paige, we feel like these will probably be the most accessible in the long term versus other technologies…We don’t get too involved in a lot of early-stage translational research. We really want to develop algorithms that can ultimately stratify patients for clinical trials and predict treatment responses or genomic mutations.

Finally, what are some of the new technologies or general trends that you're keeping an eye on over the next five years?

I'm not an AI scientist, but the fields of computer vision, AI, deep learning, and machine learning are advancing at such a rate that I suspect they’re going to be pretty impressive in five years. We're starting to see some really interesting foundation models from Open AI, like GPT-3, which use natural language processing and are trained using 10 to 15% of the internet. This the kind of thing that we're really getting excited about here. Everything that we do at Paige is about science, but science is about being able to help patients. If we can create a model that is more performant in predicting biomarkers or treatment response, we're going to do that. The technology is currently far ahead of what clinicians want, need, or understand. Ultimately, in five years, hopefully, we'll see the broader adoption of technology like Paige Prostate Detect, where pathologists will trust AI and feel comfortable using it. Then there can be new developments over the next five years.

Thank you so much for your time, Andy.

 

Nicholas J. Robert, Janet L. Espirito, Liwei Chen, et al;

Biomarker testing and tissue journey among patients with metastatic non-small cell lung cancer receiving first-line therapy in The US Oncology Network, Lung Cancer,  Volume 166,  2022,  Pages 197-204, ISSN 0169-5002, https://doi.org/10.1016/j.lungcan.2022.03.004.

Authors
Katie Gillette
Senior Project Leader
Tina Wang
Project Leader
Related Insights

Precision Medicine is evolving at a rapid pace

Discover how we can help

Contact Us
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.