240: Computational Pathology Is Changing Companion Diagnostics

Can a treatment decision depend on whether one pathologist sees 45% biomarker positivity and another sees 55%?

Visual immunohistochemistry scoring helped establish precision oncology. But as targeted therapies become more sensitive to subtle biological differences, categorical scores such as 0, 1+, 2+, and 3+ may no longer capture the information needed to identify the right patients.

In this episode, I speak with three Roche experts:

• Gordana Juric-Sekhar, MD, anatomic pathologist
• Saleh Miri, PhD, Director of Digital Pathology AI Algorithms
• Purvi Gaglani, Regulatory Affairs Lead for Digital Pathology

We discuss how computational pathology is changing companion diagnostics by moving biomarker assessment from visual estimates to continuous, cell-level measurements.

The conversation examines the limitations of manual IHC scoring, including interobserver variability, intraobserver variability, visual fatigue, borderline cases, tumor heterogeneity, and the inability of the human eye to measure complex spatial relationships.

Using TROP2 scoring in advanced non-small cell lung cancer as an example, Saleh explains the normalized membrane ratio. This computational metric measures protein expression at the cell membrane relative to total expression within the cell—something that can’t be reproduced through conventional visual scoring.

We also clarify the difference between computer-assisted scoring and a fully computational companion diagnostic. An assisted tool supports a pathologist’s visual interpretation. A computational CDx generates the biomarker measurement through algorithmic, cell-level analysis.

That doesn’t remove the pathologist.

Pathologists remain responsible for evaluating tissue quality, staining quality, scan quality, tumor selection, image analysis results, and the final clinical context. They can reject a stain, request a rescan, exclude inappropriate regions, question the result, or seek a second opinion.

The episode also examines the regulatory implications of computational companion diagnostics. Instead of evaluating a single IHC assay, regulators may need to assess the complete system - from tissue preparation and staining to scanning, image management, algorithmic analysis, display, and the final biomarker report.

Finally, we discuss what laboratories will need to implement these workflows, including validated scanning infrastructure, cybersecurity, tighter preanalytical process control, and training that helps pathologists interpret continuous computational measurements.

Episode Highlights

• 00:00 — Pathologists remain central to computational CDx
  Why computational tools provide more precise measurements without replacing pathology expertise.
• 01:09 — Why companion diagnostics are changing
  Visual IHC scoring helped launch precision oncology, but the model is approaching its limits.
• 04:53 — The current companion diagnostic landscape
  How IHC, next-generation sequencing, liquid biopsy, and visual biomarker scoring are used today.
• 07:01 — The mathematical burden placed on the human eye
  Why manually assessing tens of thousands of tumor cells requires pathologists to estimate rather than calculate.
• 08:39 — The borderline patient dilemma
  A digital tool can distinguish measurements such as 74% and 76%, while that difference is difficult to reproduce visually.
• 09:27 — Why spatial context matters
  Computational pathology can measure biomarker heterogeneity, clustering, and relationships between tumor and immune cells.
• 12:17 — Where manual scoring reaches its limits
  Interobserver variability, intraobserver variability, fatigue, staining interpretation, and heterogeneous tumors.
• 18:29 — Moving from judgment calls to quantified measurements
  Why the next stage of precision oncology requires information beyond human visual perception.
• 19:14 — Computer-assisted scoring versus computational CDx
  The important distinction between helping a pathologist calculate an existing score and generating a new algorithmic measurement.
• 23:22 — Computational pathology and decentralized workflows
  How digital images can support remote review, access to expertise, and second opinions.
• 27:48 — Why therapies require higher-resolution biomarkers
  Modern targeted treatments may respond to biological differences that categorical scoring can’t capture.
• 32:09 — TROP2 in advanced non-small cell lung cancer
  The episode’s example of a biomarker requiring computational measurement.
• 33:26 — Understanding the normalized membrane ratio
  How the algorithm measures membrane expression relative to total protein expression at the individual-cell level.
• 35:46 — Working with regulators on a new diagnostic model
  Purvi discusses global health authority engagement and the FDA Breakthrough Device Designation.
• 38:33 — The computational CDx as a system of systems
  Why staining, scanning, image management, algorithms, displays, and reporting must be evaluated together.
• 40:11 — Changes to validated workflow components
  How using a different scanner, monitor, or other component could fall outside the defined device configuration.
• 43:30 — Why computational pathology is becoming necessary
  Continuous measurements can reveal biomarker-treatment relationships that may remain hidden within categorical scores.
• 49:12 — The pathologist’s role in the workflow
  Reviewing sample, staining, scan, image, algorithmic analysis, and the final biomarker result.
• 53:39 — Digital second opinions
  How image management systems can simplify collaboration without physically transporting glass slides.
• 56:43 — What laboratories need to prepare
  Validated infrastructure, cybersecurity, preanalytical control, training, and digital pathology literacy.
• 58:49 — Learning to interpret computational results
  The shift from visually estimated categories to continuous, quantitative biomarker measurements.

Resources Mentioned

• Full discussion on YouTube: https://youtu.be/oKW1xC6TTZg

Listen to the full discussion to understand how computational pathology could change companion diagnostics—and what pathologists, laboratories, and regulators must prepare for next.

Transcript

00:00:00

The role of the pathologist in the CDX kind of a device. We are not replacing pathologist whatsoever. >> When a project starts as a manual score, it can fail because the human eye imposes artificial categories and biological [music] gradients. >> One of the key capabilities of computational pathology is also the ability to precisely assess the entire region of a protein expression intensity within a tissue section. we are just empowering them with more quantitative, more precise and reproducible tools. So

00:00:33

 that will require a major cognitive shift. >> What we see is that the more we refine our algorithms, the more we understand how to train these algorithms to quantify staining heterogenity, providing detailed and objective measures of uniformity or non-uniformity of a biomarker distributed across a tissue. The manual score it's too noisy to find a correlation. So the biomarker looks nonpredictive. But measuring protein density with the algorithm on a continuous scale eliminates the subjectivity, this variability and noise

00:01:09

 present in human assessment. Welcome everyone and thank you for joining me for this webinar. I'm Dr. Alex Zura, founder of Digital Pathology Place. And today we're talking about something that is quietly changing how companion diagnostics work and what that means for pathologists. For decades, companion diagnostics have rested on visual IHC scoring. And that model launched precision oncology, but it's now reaching its limits. uh to help us understand what's happening and where it's going. I'm joined by three

00:01:50

 experts from Raj. First, Dr. Gordana Uritar, anatomic pathologist at Raj with deep involvement in digital pathology projects across the company. Gordana, welcome. >> Thank you so much for the invitation to be here, Alex. And uh I'm as you mentioned I'm Gordon Aurori Shaker and I'm a physician board certified in anatomic pathology and neuropathology. I began my career in academia before transitioning to in industry five years ago when I joined Raj. I'm involved in various aspects of companion diagnostic

00:02:30

 development through multiple avenues primarily the evolution of computational pathology and the development of new biomarker scoring algorithms. >> Thank you so much. I'm also joined by Salai Miri, PhD, the director of digital pathology AI algorithms at Rush H, where he spent over eight years leading algorithm development for hystopathology. Sally, welcome. >> Thank you, Alex, for the invitation and having me on this webinar. As you mentioned, I'm leading the clinical digital pathology algorithm development

00:03:07

 at RO for uh past eight years and very excited to be here talking about computition pathology as the next generation of comput companion diagnostics. I am super excited. And we also have Purvy Gaglani, regulatory affairs lead for digital pathology at Rush. And uh in the past three years, she was working closely with the FDA on submissions across computational pathology, companion diagnostics and software as medical device. Uh Pervy, welcome to the webinar. >> Thank you, Alex. Excited to be here. uh

00:03:44

 very very interesting uh topics today and uh being part of the team that has been able to interact with health agencies across the world when it comes to computational pathology uh CDX and uh software as a medical device I think we're at the cradle of bringing something very new for our patients and everyone across the globe. So very excited. Thank you. I am I am so excited about this topic because I've been following the development of this new way of uh doing companion diagnostics for I don't know two years already. I

00:04:18

 don't remember when was the first time that it was like publicly mentioned that this is happening. H but uh the format today is a little bit different. We don't have presentation but we've got eight questions to work through over the next hour. H and then we'll open up live Q&A. So, let's get into it. I'm going to uh start with the question to you, Gordana. My first question is um what does the companion diagnostic landscape look like today and uh what data are we using and how it's being scored because

00:04:53

 that's going to set the stage for what's new. >> Thank you. So despite the emergence of new technologies like next generation sequencing or let's say liquid biopsy, iminoistochemistry has been the foundation of companion diagnostics for decades largely due to its practical advantages. While next generation sequencing processes the sample to identify mutations, you knowchemistry allows a pathologist to review the sample composition and confirm protein expression either on the cellular

00:05:31

 compartment or on infiltrating immune cells. Also, iminohistochemistry is a standard procedure in nearly all pathology laboratories globally and it's a generally faster and cheaper than new technologies. I have to say that a clinical landscape has shifted from binary, positive, negative to a spectrum of expression. Biomarkers are all visually scored. Pathologists currently use semiquantitative scoring method methods for companion diagnostic uh tests which involve estimating ratios such as tumor

00:06:11

 proportion score and combined positive score or assigning categories one uh zero 1 plus 2 plus 3 plus based on staining intensity. So the era of onesizefits all pathology has ended. The expertise of a pathologist is also pertinent and training on specific companion diagnostic biomarkers is now a regulatory and a clinical imperative. For example, of FDA labels for almost all new CDX essays explicitly state that the test must be interpreted by a qualified pathologist. >> Semiquantitative scoring is the current

00:07:01

 gold standard for precision medicine, but it highlights a significant bottleneck. When a drug approval is tied directly to a specific essay, the pathologist's device become the ultimate gatekeeper for therapy. However, the challenge is that a single tissue section can contain let's say five 50,000 or even 100,000 tumor cells. For a human is truly calculate, it takes hours per slide or whole slide image. Instead, We pathologists use us use visual bins or percentage ranges. Let's say we'll

00:07:50

 say the tumor shows positivity between 40% and 50%. >> So, but also we are generally better at identifying positive cells than actu accurately uh accurately counting the total number of viable tumor cells. the denominator if but however if the denominator is wrong the entire [clears throat] score is wrong >> in the context of current CDX landscape I would also emphasize the role of computational pathology as its advancements have uh taken it beyond the experimental phase. So computational pathology can address

00:08:39

 some significant challenges such as I call first point the borderline dilemma. In many clinical trials the cutoff for a drug is set as a specific percentage for instance 75%. Humans cannot acu accurately differentiate it between 74 and 76% tumor cell expression in tens of thousands of tumor cells. However, digital tools can perform an exact cell count and the borderline patient receives a more accurate uh classification reducing the risk of denying potentially beneficial treatment. The other point which I want to also

00:09:27

 stress is beyond protein expression. Most uh CDX essays only look at two things. How many cells are stained and how dark is the stain. This ignores the spatial context of the tumor. So computational pathology can measure spatial heterogenity. It can answer the protein expressing tumor cells touching immune cells. Are they clustered together or scattered? Trials are already showing that these spatial metrics are often better predictors of drug response than simple positive negative bits. So to summarize this overview, we are

00:10:13

 still taking about man we are still talking about manual scoring but the era of eyeball bin is coming to an end. The CDX landscape is evolving toward an integrated multimodel system shifting from I would say qualitative to quantitative science. >> Thank you so much. So I want to highlight a couple of things that you said uh because you started with estimating rat ratios and well like I have research experience with PDL1 scoring >> and um I remember like being trained on how to do the borderline cases and I

00:10:57

 went to a specific training and they would show us these borderline cases. doesn't matter how long we trained and how they like how many hours they trained us at the end of the day half of the uh room was showing that it was below the cutoff and the other half it was above the cutoff. So like it's impossible. Um and the the the scores based on calculation you said but we never calculate right if we were calculating every every single cell it would take us um pathologists hours per slide. There is no hour per slide. So

00:11:36

 you do these bins um and and the b and most of the time we're okay but the borderline cases who are borderline samples from patients um and the decision on the borderline case is going to change okay what kind of treatment they're going to get. So um that brings me to the second question that I have for you as well. Uh where does the current companion diagnostic model begin to reach its limits? The current companion diagnostics model which relies heavily on the manual interpretation of imuninohistochemistry

00:12:17

 it's approaching its limit because it it essentially demands human vision to perform fundamentally mathematical operations. Yeah. So although this model initiated precision oncology its current sensitivity it's insufficient for next generation treatments such as we are talking about let's say antibbody drug conjugates. >> So uh the current model has critical limitations that are evident at certain points. Foremost is interobserver variability. Variability can be attributed to differences in path a pathologist

00:13:02

 assessment including tumor proportion score or combined positive score interpretation, staining intensity interpretation and membranous versus cytoplasmic staining interpretation for the tumor percentage positivity. uh for example the cut off is let's say 50% and one pathologist would assess 45% and another would call it 55 positive tumor cells for the staining intensity interpretation the cutoff it's let's say two plus and three plus uh at 50% and the distinction between one plus and two plus staining

00:13:49

 can lead to inter observer variability Sometimes it's challenging. And the third point for example for membranous versus uh cytoplasmic staining interpretation for example uh the cutff is 50%. Membranos staining at any intensity. The essay is challenging to distinguish between membranos and cytoplasmic staining. This can lead to wide differences in interpretations. But I also would like to emphasize other points which may reveal limitations. Intraobserver variability. The same pathologist may score the same tumor

00:14:32

 differently. For example, today and tomorrow morning. Search fatigue on large tumor resection. A small cluster of positive tumor cells could be easily overlooked altering a score. Another point which I also want to stress it's a low sensitivity assessment uh a tra the traditional staining intensity. We always talk about zero one plus two plus 3 plus system has evolved. Now pathologist need to distinguish in her too and also evaluate ultra low zero plus >> which includes any faint barely

00:15:20

 perceptible membrano staining in tumor cells. So next it's a average positivity and heterogenity tumor may reveal heterogenous biomarker expression with areas of intense so-called hot and minimal or no co so-called so-called cold in uh immuno reactivity. So a pathologist often averages these areas for a single score. For example, 15% average could mean 15% of cells are staying evenly throughout or it could mean one small corner in 100% positive area while the is zero. So these patients respond differently to

00:16:09

 treatment but the bin treats them equally at some point. >> And one point which also want to mention it's failure to capture spatial context. A manual score provides a single number. For example, combined positive score it's 30. But it doesn't specify where the protein is expressed. Two patients with the same 30% score can differ. Positive tumor cells might be clusters clustered near immune cells in one patient but isolated in the tumor core in the other patient. We now know that spatial proximity

00:16:55

 distance between T-C cell and tumor cell predicts successful immunotherapy better than protein density alone. Manual scoring cannot measure these micro level distances. I also want to add the decades old staining intensity scale system which is the gold standard for breast and gastric carcinoma is an inconsistent predictive biomarker in advanced non small cell lung carcinoma. In breast carcinoma, her two aberrations are typically associated with an overexpression of receptors on the cell surface. On the other hand, in non small

00:17:41

 cell carcinoma, her two is generally characterized by a structural abnormality and a genetic mutation causes the receptor to be permanently active regardless of receptor quantity. >> So, High three plus score doesn't always mean that for example her two gene is amplified and many patients in advanced non small cell lung carcinoma with a IHC score of zero or one plus carry her to exm 20 mutations and respond to targeted drugs. Interesting. Yeah. And in so I put many points together but in summary I can say

00:18:29

 that the current model reaches its limit at the threshold of human perception. To unlock the next level of precision, the diagnostic must shift from being our judgment call to a quantified measurement. So that's like the clear uh visual limitation and I also like see the evolution of as we um like as a research community I would say as we discovered more more biology we tried to apply the uh pretty decent robust method of visual assessment to something that's too granular for this assessment right and

00:19:14

 the first level is okay let's have us not guesstimate let's have us uh have something some computer aided tools right to show us um like how much score is there and then we can confirm it's still like the same of visual principle but now computer is calculating this and this has been there for her two for PDL1 um these computational um companion But that's that's different from computational companion diagnostic. I want to differentiate two things because this is computer aid for pathologists.

00:19:56

 It basically like does what this score was designed. It calculates instead of estimates and this is the computer aided AI tool. And this question is going to be for you pervy. Um I want you I would love you to explain to us um how does a computational companion diagnostic so like truly computational differ from this assistive or computer aided AI tool that we are pathologists are mostly familiar with because a lot of these scores have something a computer algorithm that can calculate right but we're going to be talking about

00:20:33

 something that is new and that differs and I want to highlight this difference. >> Thank you, Alex. Yes. Uh what we've seen with the advent of computational pathology is the changing role that the pathologist has in the overall clinical workflow. And as you rightly put it in your question, the difference between what we're doing with computational pathology versus creating tools that have aided pathologists in in making uh a diagnosis or an assessment is starkly different when it comes to the way the

00:21:06

 pathologist is involved and the amount of work they get to do when they're interacting with such a uh CP based device. uh this particular device that has been uh the top of the town. We see that the role of the pathologist is more focused on ensuring that the sample quality, the scan quality, the image quality is of a certain nature and is allowing the pathologist to gain that confidence that whatever analysis comes out from a computational pathology workflow or an algorithm is something that they can trust. Whereas if this

00:21:42

 were a computer aided scoring algorithm or or something that was done to aid the pathologist in enhancing what they're doing would probably be restricted to giving them an idea about I think this is where the tumor cells lie. Now it's up to the pathologist to figure out whether they agree or disagree with that uh scoring and also make amends to that. that is a very distinct um identifier between what a computational pathology based algorithm for a CDX workflow does versus what it would do for a computer

00:22:18

 aided uh program or a tool that would just allow them to make their work a little bit easier. But what we also see is that with the advent of computational pathology uh that is facilitated by the dual tools that we have in digital image analysis uh has has marked a transformative shift in the way IC images have been uh viewed. There are better scanners now. There are better image management systems and a lot more sophisticated algorithms that involve artificial intelligence. Deep learning models that allow us to say that

00:22:52

 whatever scoring we're going to receive at the end of that workflow is is a thoroughly vetted model that is giving you the confidence to accept the score that the algorithm is giving you. The challenges that Gordana described are exactly the challenges the computational pathology algorithms are trying to solve. This makes our solution more scalable. It allows us to get diagnosis to people around the world where pathology diagnosis have been challenging. They don't probably have the right kind of tools. They probably

00:23:22

 don't have enough amount of doctors who can help them. Having a decentralized workflow where you're sitting anywhere in the world and as a doctor, as a pathologist, you can view an image from any patient anywhere is where this marks the the advent of what we of the future looks like for digital pathology. Uh, one of the key capabilities of computational pathology is also the ability to precisely assess uh the entire region of a protein expression intensity within a tissue section. What we see is that the more we refine our

00:23:57

 algorithms, the more we understand how to train these algorithms to quantify staining heterogenity. Uh providing um detailed and objective measures of of uniformity or non-uniformity uh of a biomarker distributed across a tissue. And it also allows to gain a level of resolution which has in the past been either challenging to the human eye in terms of assessment or in terms of quantifying it to a level that would become universal. And hence the challenges that Gordana uh mentioned where you see a lot of path variability

00:24:32

 is is the problem that this potentially takes away. When you couple softwares like this with uh AI and deep learning models, we see that it creates a very very powerful diagnostic tool and it permits the efficient systematic investigation of any disease any biioarker that has so far either eluded us or has been difficult to diagnose and hence quantify. Uh we have also seen that with a new generation of drugs coming in, the need for precision medicine, the need for having very precise outcomes of your diagnostic

00:25:08

 test, of your companion diagnostic test, CP allows us to define those and gain insights into these new biomarkers and allow us to have companion diagnostic devices that can work on new drugs. So this indeed is is that full solution where you not just get to see we are creating a tool that allows pathologists to do their work a little more efficiently but actually taking away the burden of having to go through those 10 15,000 cells allowing to see them the membrane intensities the stain intensities that is quite challenging

00:25:44

 and burdensome in terms of a workflow. uh at the same time allowing them to then come up with a diagnosis that is very direct, precise and has high sensitivity to a drug that we know would work for the patient. So all in all when we compare a CPbased diagnostic algorithm versus a computer uh aided tool for the pathologist there is a clear uh strength in in developing tools like this and developing algorithms like this that truly allow the pathologist to uh scale their work and make it more sensitive and precise as we move towards

00:26:19

 new drugs and biomarkers. >> Thank you so much. Um so the role of the pathologist in this um in this landscape changes h because they need to kind of collaborate with the new u way of diagnosing right it's going to be okay we have a precise tool but is this tool measuring in the right place was this material even diagnostic did we do the correct sample and basically it becomes collaborative. One thing that um you said uh basically both of you Gordana and you Purvy said Gordana you said IHC is everywhere right and and we

00:27:05

 just mentioned and computational pathology doesn't have to be everywhere to help uh those patients that would have IHC done on their uh samples. So I just want to um have that in the back of our head um because and we'll have questions and answers for everybody but basically um we can help uh patients somewhere else than where their IHC was done. Um and there is this theme okay biology is getting more complex and we're discovering more biology and we're going to be getting to what particular test

00:27:48

 we're talking about. But before we get there, h Sally, I have a question for you. Why does personalized on oncology now require higher resolution biomarker measurement? What like what is there? What are we discovering? Right? We just said at the beginning that actually this visual scoring did launch precision oncology but uh I'm super happy that we're advancing and and why why these higher resolution biomarker measurement requirement right now. >> Yeah thanks for the question Alex. I

00:28:23

 think uh with what Gordana and Pervy mentioned they set the stage nicely for us to relate to this question. Why do we need um higher resolution biomarker measurement? And I want to just highlight a few points. Um from my perspective um I guess the modern targeted therapies uh nowadays they are um highly sensitive to um subtle shift and respond to much um finer biological um gradient of the expression. And because of that um We we do need precise tools uh precise tools like advanced technologies that can um essentially uh

00:29:12

 you know capture these minor fluctuations that ultimately uh they predict the um therapeutic benefits and the patient response. Um so besides that Gordon talked about this quite nicely this um continuous u versus categorical change that we have. So that tumor biology itself is inherently continuous and heterogeneous, right? But the current companion diagnostics um is this visual assessment and estimations that we have, it reduces uh this crucial data into simplistic categorical bins like 0, 1 plus, 2 plus,

00:29:58

 3 plus. And this categorical scoring um frequently fails to consistently predict therapeutic um benefit especially when we deal with uh more sophisticated and u complex cases um such as advanced NCLC. Um one more additional point that I wanted to highlight also here is the and Gordana touched on it the limitations of this visual story. I guess achieving a truly reproducible and u cell level quantitative signal um requires moving beyond this visual estimation because this visual estimation and categorical scoring

00:30:45

 inherently introduces um high observer variability as you also mentioned from your experience in PDL1 scoring. Uh so no matter how much you train, no matter how much you um you know read cases, there's still some inherent variability between and within pathologies. So that reader variability is always there and computational methods are necessary to standardize uh the measurements and deliver the precision that is required nowadays for these modern therapies. Thank you. Thank you so much. So um the

00:31:30

 biology that we discover then translates into the therapies that are super sensitive, right? That that's why we have a companion diagnostics. That's why we want to check okay who is this uh drug going to be good for and who's who is not going to react in the desired way to the drug, right? and and the cutoffs are like very um granular how to say and you have you granularity like we said is beyond perceptions. Um and um that brings us to the the thing that you said continuous versus categorical. Categorical doesn't cut it

00:32:09

 anymore. Uh we need continuous um and um you mentioned also okay at the cellular level. So we are now and that takes us to the marker that is kind of the hero of this webinar. That's the marker where we have this new companion diagnostic TROP 2 in advanced non small cell lung cancer. This is going to be our example. Um and what is a computational pathology based companion diagnostics for this particular marker? We kind of hinted to it, but I uh want you to explain what it does. And when you explain that, it's

00:32:50

 going to be clear why it's visually impossible to to replicate it because that is the the thing that's happening now. Before we were kind of decent in estimating and here we reached a point where the granularity is so that we cannot estimate anymore and we have to have this collaboration with the computer with the computational tools. So let's talk about the trop uh scoring how does it look right now? Sure I can start uh from my side and and we'll pass it on to Pervy to also talk about the

00:33:26

 regulatory and uh on this >> um so a computational companion diagnostics um um is a new class of regulated diagnostics where the final score um is derived from as I mentioned algorithmic cell level quantitative analysis and This uh analysis quantitative analysis completely replaces subjective visual estimation rather than dividing the results um into discrete categories like 0 1 plus 2 plus 3+ this will be a truly continuous um measurements of the protein expression for trop 2 in ncl specifically

00:34:18

 uh this uh involves calculating the normalized membrane ratio or NMR in short. This is a metric that um precisely quantifies the amount of trap 2 protein at the cell membrane relative to the total expression within the entire cell. And as you can see this is completely different from uh the current standard the visual estimation that Cordonna explained quite nicely. So this goes beyond uh the current standard and goes to the individual cell level quantitative analysis and measure uh this ratio which quantifies the membrane

00:35:08

 staining intensity versus the entire um cell which includes membrane and cytoplasm together. Mhm. >> Um so I would like to also invite uh Pervy to uh talk about from a regulatory standpoint how is the entire system defined? >> Yes. Thank you. >> This is super interesting because that's like a new frontier of um yeah how do how how does it look from the regulators perspective? What do regulators do with such new discoveries and how did you interact with them as well? interactions

00:35:46

 I can say. Yeah, the interactions have been uh >> very very uh interesting. Let's start with that. We've learned a lot in in our uh discussions in our uh communications with the agency not just here in the US but across the world. So all health authorities one common theme that we see when we interact with health authorities across the world is everybody wants better solutions. Everybody wants precise solutions and health authorities are willing to engage with the manufacturer to ensure whatever help

00:36:19

 they can provide, whatever guidance they can provide, if a guidance is non-existent or if the device is as unique as ours, uh they're there to support you, they're there to help you. So it's been an extremely enriching experience uh especially working very closely with the FDA on on this particular device uh gaining the trust and understanding how they review the device knowing it's a very novel uh concept it's a very novel technology uh having the experience to work with them

00:36:49

 on uh getting the BDD which was the breakthrough device designation that is granted to medical devices that are meeting meeting an unmet medical need with a unique techn technology is is uh what gave us the uh validation and the impetus to move forward in a direction where we know we're helping patients uh receive diagnosis that probably has been missing so far. One of the most interesting things that we learned is if I compare the device that we're talking about right now to a traditional IHC

00:37:20

 based device that is visually reviewed under a microscope, we would probably be talking to FDA about an IHC kit for the assay. >> FDA would review the reagents. FDA would review the uh objective evidence associated with how the slides are stained. uh how the WNB was performed and all of this would be by a human being looking at a microscope. That idea that approach is completely changing here. sustaining uh workflow or or getting your slide ready has become a part or a component or an input to this

00:37:55

 new diagnostic device where up until now whatever they had been reviewing uh is is now a system within a system of systems. So the perspective has largely changed and what otherwise used to be an assay kit uh is is now just the starting point for what the diagnostic workflow would look like. Uh what we have seen is that one of the strongest recurring theme across our discussions with the health agency was ensuring that we do not work outside defi uh device definition and for especially this was a

00:38:33

 device definition that started at us staining a slide all the way through the digital pathology workflow generating a PDF with your biomarker uh results. And so you had a staining workflow and then you had a digital pathology workflow where you're taking your glass slide sca scanning it for uh for creating a whole slide image for your man image management platform and then applying the algorithm to that image to come up with a biomarker score. Of course there is the pathologist in the loop for the

00:39:06

 digital pathology workflow as well. And that is where the pathologist is determining if the algorithm is doing what it's expected to do. They do have a lot of power. They do have a lot of uh opportunities to agree or disagree with how the algorithm proceeds. But at the same time, the algorithm has been designed to ensure that the variability that we have otherwise seen or the sensitiv sensitivity that we would like to see for a device like this is enhanced and that takes away this human variability that we've seen. It enhances

00:39:37

 the sensitivity and the precision that we want these unique biomarkers to have. And when we submit this or the expectation that uh FD especially has had now is that when you submit the data that objective evidence should support the end to end workflow. Everything that you did from the start of the time you got your tissue the way you created your slides all the way to the point you got your biomarker report is what we would like to see as an end toend system verification and validation uh data set and you need to submit the

00:40:11

 whole of it to us. Any changes in the way the device is currently described would be considered an off label used if you're using a component in that entire workflow that is not defined in the device description. And that is a very very important thing to understand that thing things like a display monitor. >> We know that pathologists would use different use different types of display monitors. uh but we have to ensure that if they are using the stroke to device for NCLC with the particular algorithm

00:40:47

 that ro is developing the expectation is that you use it with the monitor with which it was verified and validated. Any changes to any component of the device you scanning the images or scanning the slides to get that image on a different scanner that is not the one that we we end meet with would be considered an off label use. And so the realm of how FDA has always seen seen device configuration, how they look at objective evidence that supports your intended use, how they look at off label usage has not changed. But the content

00:41:22

 of what that device description looks like has moved from us creating one assay kit to probably having seven to eight components that are systems of their own. And hence this becomes a system of systems now which is a very very high complexity high-risk CDX uh diagnostic device. So that's >> that is a to that is a very big shift in perspective from the test side but it kind of is in parallel how digital pathology was being viewed by regulators from like the first scanner clearance right it was a workflow and now even the

00:42:06

 tests are part of the workflow which I see as something pushing dig digital in general, digital pathology forward because now we have something highly complex and highly granular cannot be uh reproduced like in a DIY fashion and this is something that helps patients. So if you want to help patients, you need to expand to be able to provide this kind of workflow and and it's you know this is this is my perception my like joy why we discovered this biology in terms of how it's going to go backwards to promote digital

00:42:52

 pathology because this is a level of care that we can now provide who like everybody wants to provide that level of care and now let's find a way to actually do that um so Gordana back to you um in in this kind of landscape regulatory biological like the limitation of our visual interpretation when I hear like okay it's a ratio of membrane uh relative to the total stain like even if I want it and if I count like if I drew every single cell I would be tired after 15 minutes of counting these pixels so that's basically not not

00:43:30

 feasible So uh I want to hear your perspective on why is computational pathology no longer optional for certain companion diagnostics and here we're talking about the specific companion diagnostics but I am hoping for more discoveries like that and also I am hoping for the other estimates that pathologists have to do to no longer be estimates but have um computational support for that. Thank you, Alex. [clears throat] So, as we move toward precision medicine, the shift to quantitative data in precision medicine makes manual

00:44:12

 biomarker scoring with necessary reproducibility nearly impossible. So, the human eye as as remarkable as it is has physical and cognitive limits for the precise quantification of modern precision demands. So this challenge is leading the field to transition toward computational pathology and uh as even Salah mentioned instead of tearing scoring system algorithms can generate a continuous score by counting every single pixel and cell. Basically a computational pathology produces a continuous scale.

00:44:54

 It was mentioned earlier and the key metric is a normalized membrane ratio which precisely quantifies the cell surface to internal cytoplasmic uh or internal cytoplasmic protein proportion. So when we talk about pathology in traditional pathology the instrument is the microscope uh but the measurement happens in the pathologist visual cortex in the algorithm integrated system. Every link in the multi-step computational pipeline impacts the final measurement. When the project starts as a manual score, it can fail because the human eye

00:45:37

 imposes artificial categories and biological gradients. So for example, pathologist might score a tumor as one plus or two plus if the true biological threshold for drug benefit is actually 1.4. The manual score it's too noisy to find the correlation. Half the plus patients respond and half the two plus patients don't. So the biomarker looks nonpredictive. But measuring protein density with the algorithm on a continuous score scale uh eliminates the subjectivity, this variability and noise present in human

00:46:22

 assessment. In this case, patients with a score equal or above 1.4 four can show a survival benefit. Further example, the industry saw this play out in real time with drop two. We already mentioned in small cell lung carcinoma and the clinical utility of tropion initially appeared to be inconsistent when assessed with manual IHC scoring. The results didn't clearly show that more drop to better response. So the manual score couldn't find the right patients. However, the algorithm could identify

00:47:05

 spatial patterns and membrane specific intensities that pathologist couldn't quantify. The drug bcame unlocked and to specific subop that manual scoring failed to identify. So this is leading to let's algorithm only biomarkers in clinical trials which have no manual equivalent. So but I as a pathologist could say that uh and also want to emphasize that the pathologist is not replaced and instead of digital tool as a biomark calculator providing the numerical aspects while the pathologist focuses on

00:47:53

 the complex diagnostic integration. Mhm. And I love that because burdening the pathologist with these like calculations that we are like human vision is not made for >> is basically taking cognitive power from all the other things that a pathologist is trained for and is very good at. Um and that kind of supports the theme of um collaboration human with machine or however we want to call it. Uh but basically now okay um person the pathologist is very good at hystopath morphology uh recognizing which cells

00:48:34

 recognizing is it from the correct compartment is the calculation being done on the correct things that the device is supposed to do calculations and that's easy that's like no cognitive burden no like depletion from your diagnostic superpowers as a pathology ologist and then you have this calculator like you mentioned that was designed as a calculator is a good calculator for whatever we designed it uh to be a calculator for um and um following on that one h I want to ask you pervy about what is the role of

00:49:12

 pathologist in a computational companion diagnostic workflow because um you were mentioning okay before it would be a pathologist under the microscope and you know uh they would that would be the ground truth, the estimation of whatever the positivity was and now we have a full workflow and a pathologist has a role in this workflow. How what is this role and also how is it viewed differently by regulators and I would also add to this question by the team that is working on this workflow. It's

00:49:48

 no longer like a slide stained with IHC and the pathologist is the test. there's a lot more to it. So um what would you say the role of the pathologist is in this comp computational companion diagnostic workflow? >> A very important one. So what we know is that there is nothing that we are creating with these tools that takes away the the work of a pathologist or the expertise of a pathologist. Like you and Gordana mentioned earlier, this is more like outsourcing that calculator job to a

00:50:23

 tool that we know is trained and is precise enough to give us a good output that we can trust. These were the questions also highly discussed in our interactions with FDA and the emphasis on clarifying the role of the pathologist in this clinical workflow was was paramount to us explaining what the device was all about. The first the first step that we would see as what how is the pathologist involved and how does the pathologist interact with a device like this is what we call the assessment portion of how

00:50:58

 the algorithm is performing. So the pathologist is in charge here. The pathologist is controlling what is going to be the outcome here. The pathologist is able to review if or not the scanner that was used to scan the slide and give the image to them on the computer screen was equivalent if not better to what they would have otherwise seen on a very very sophisticated microscope. They would also look at the kind of staining that was done for that image. If they do not agree with the kind of scan or the stain, they are in complete

00:51:32

 power to reject it and ask for either a restain or a rescan if they don't agree with the kind of quality of image or the kind of quality of staining that they're seeing on the computer screen. More importantly, they also have the ability to judge the sample quality and what the tissue looks like because what the human eye sees under a microscope, even if this was a 40x zoom versus what you can see on a computer screen at a pixel level, is is a whole uh difference. It's it's a whole new world of looking at

00:52:03

 your uh sample and your tissue. the slightest of the fold, the the smallest of the air bubbles, the tiniest of a pen mark, anything that could have smudged your slide from anywhere, all this gets captured and amplified because you're looking at this on a high resolution screen. The pixel pathway from your medical grade scanner to your image management platform has been evaluated, vetted, and tested so many times to ensure that the image quality that you get on your screen is doing justice to

00:52:36

 what a human eye would want to see and explore further. So the pathologist is in a lot of control when they get to see the image for the first time, the scan, the tissue for the first time, the staining intensity for the first time. Once the algorithm starts doing its job, the pathologist also has the power to to reject areas that they do not agree with which the algorithm might have selected. So they have a lot of control there and at the end of the workflow after they see the biomarker status because they

00:53:07

 are the doctors they are the people who know exactly what the patient history is, what the clinical history is, have the ability to still go ahead and reject the outcome if they don't agree. The the plus point here is they can reject but they would be asked to tell why they would be rejecting a case and and why they don't agree with the biomarker analysis to ensure that even even the workflow is is you know able to understand and justify why this was not the appropriate diagnosis. Additionally,

00:53:39

 it is tools like these and having a platform for an image management system that allows you to make your education or getting a second opinion so much easier than what it used to be where we would have physically transported a slide from lab one to lab two which was probably miles away. U pathologist looking at it under a different kind of a microscope to get that second opinion is now happening at a push of a button. And so the pathologist is in a lot of control in trying to also get a second opinion and find out if if what they're

00:54:11

 looking at or if they are not agreeing to the algorithm analysis where are those differences lying. And when these things were explained to the agency in the in terms of our end to end clinical workflow uh they they realized that this is not a black box where pathologist does not have the power to control how the diagnosis is coming out. They have a lot of control. They have a lot of power and at the end of the entire workflow they still have the option to say no and get a second opinion if they like.

00:54:43

 >> This is um and Gordana tell me what you think about it briefly as well. But I think uh what you just said um pervy is very empowering to pathologists who sometimes come into this space with a little bit of fear. uh when a new technology that is a technology that um provides something that they cannot provide um there is fear okay will we even be relevant um but this again highlights it's a tandem work and the um like the good thing the the thing that is empowering is that the pathologists

00:55:28

 already have the knowledge necessary to be a partner in this workflow. It's not that oh now everybody has to get a fellowship in computational pathology. No the the the expertise the training what pathologists bring into the equation they already have that as physicians as pathologists. So the missing piece of the puzzle is okay what does this test do and how do we like provide the expertise and that we already have. It's gonna require Gordana, what do you think about that? Do you feel that way when you talk to

00:56:06

 fellow pathologists who who don't know about it? Do you feel like they think it's empowering or do you still uh see some hesitation? And we're going to have questions in a Q&A. I'm super curious what the propos have some hesitations, >> you know, moving forward. >> That's why you're here, right? That's why we're trying to explain it and it will require doing things differently by the pathologists but also in the labs. Um so Sally a question to you what must

00:56:43

 laboratories consider if computational companion diagnostics become routine? What do they need to now think about that they didn't have to think about before? >> Yes. No, thank you for the question. very relevant and important question actually. So as um computational CDX becomes more routine there are a few um categories or items that labs need to keep in mind when transition or adapt to more digit digital you know workflows. Uh first and foremost is the infrastructure and security. Of course

00:57:21

 validated um digital scanning and uh computational infrastructure uh will become an essential component of these u diagnostic standards. Um so they need to implement rigorous measures to address any cyber security for example throughout the entire um end to end system to make sure the data integrity and in general you know system integrity all the patient information is all uh kept intact. Uh the other um important aspect is process control and by that I mean uh the pre-analytical staining um consistency and all these uh

00:58:08

 steps that may not have been as important as before and we had more wiggle room and more buffer to allow for variability. Now they are significantly uh more critical than before and the reason for that is uh these quantitative algorithms they are very precise and highly sensitive and they can measure a slight variability in sample preparation or staining protocols. uh and these variabilities um they directly and perhaps negatively um impact the algorithm performance, its accuracy and also its reproducibility in

00:58:49

 generating the same results. Um the other aspect that I wanted to also touch on um is the training and interpretation by pathologist. I think in the previous question uh Pervy highlighted the role of the pathologist in the the CDX uh kind of a device and uh as Gordon mentioned we are not replacing pathologist whatsoever not at all we are just empowering them with more quantitative more precise and reproducible tools. So that will require uh in my opinion a major cognitive shift uh both in pathology teams and perhaps

00:59:32

 even oncologist. So they they must uh move away from relying on visually estimated categories like TPS or 2 plus 3 plus bins to interpreting now precise quantitative data. And this might demands as you mentioned it doesn't uh demand to go back and get a degree in computational pathology now in the universities but it dem it it might demand new skills such as >> you know like in precore digital refinement you know pervy mentioned a few examples like they need to now ensure the scanning was done properly

01:00:16

 the quality of the scan is on par with their expectation the staining the sample you know um it's adequate that we have enough tumor um you know tumor cells present and also they might be asked to review the image analysis results to just like you know do QC and say okay like yes the algorithm is paying attention to the right part of the tissue has identified the proper enough um tumor coverage is there the purity of the tumor region segmentation for example is there and they approve And this is something that

01:00:52

 they didn't have to do it before. Uh but of course it doesn't need a you know separate degree just that shift in their training and their mindset also you know uh kind of a postcore interpretation. You know before they used to look at this categorical uh you know scoring now they need to understand how the um continuous metrics such as the normalized membrane ratio that we talked about you know looking at all cells the expression of the protein in membrane versus uh the entire um tumor cell and

01:01:27

 then look at the 75th quantile of all the tumor population that exists on the slide. That may not be something very intuitive that you know they can quickly confirm um just like before that they could say yes I agree this is a one plus or yes I agree this is a two plus. This is a new way of scoring um the IHC slides and they just need to um I guess get comfortable with and understand how these new computational CDX um compute the final binary therapeutic results. And I' >> I'm actually Go ahead. Go ahead.

01:02:09

 >> I'd like to add this is this is where the challenge lies for us as manufacturers, right? The entire adoption of creating and having a digital workflow in your lab is not going to come easily unless you make that workflow easy to adapt. If you put too many challenges, if you make it very hard, the hesitance will only increase, the adaption will only get delayed. So it it it is striking that fine balance where it's it's not it's not an app on your phone. you can't pick it up and

01:02:42

 take a picture and analyze tumor cells but at the same time it's not even that hard uh that you cannot you know get used to it and and going back to your point Alex where we don't want them to go and have a whole new skill that they need to learn and like sie mentioned have that paradigm shift in the way they have been analyzing uh or diagnosing patients earlier to to moving to a new way of diagnosing patients but having the same power that they had back then. >> I think so this um kind of parallels the

01:03:20

 push on adoption of digital. So here uh it's on the training level on the uh literacy level. There is going to be a requirement uh of digital pathology literacy at a specific level um artificial intelligence literacy, image analysis literacy. But this is no different than it was like IHC literacy in the 70s. Molecular pathology literacy uh which like that's far more complex for me at least as a veterary pathologist because this is not as developed as in um human pathology. Um but basically that's another level of

01:04:01

 literacy that the pathologist need to acquire. Um and you know there are different ways there are different uh like like this webinar, right? This is one way of acquiring this level of literacy. But thank you so much. Thank you so much for answering all the questions.