Precision Pathology Using Explainable AI

Precision Pathology Using Explainable AI

[00:00:00]

[00:00:28] Don Davis PhD, MBA: Welcome to life science success podcast. My name is Don and today we're going to be talking about precision medicine and especially talking about using explainable AI with precision medicine. If for those of you don't know me, I'm a consultant in life sciences. I help companies scale and manage complexity and increase performance today.

I'm joined by Lans, Taylor Chakra Chennubhotla and Dusty [00:01:00] Majumdar welcome guys.

All right. So if we could just start with you, Lance, just really quickly, could you do an introduction to yourself for the audience?

[00:01:13] Lans Taylor, PhD: Sure. Pleasure to take part in the podcast. I'm Lance Taylor, I'm the executive chairman of SpIntellx, and I'm also the director of drug discovery at the university of Pittsburgh drug discovery Institute.

[00:01:28] Don Davis PhD, MBA: Very good. Welcome Chakra. Could you do a brief introduction to yourself?

[00:01:34] Chakra Chennubhotla, PhD: Thanks, Don. Thanks for the invite. My name is Chakra Chennubhotla, and I'm the CEO of SpIntellx.

[00:01:41] Don Davis PhD, MBA: All right. And dusty.

[00:01:45] Dusty Majumdar PhD: Hey, Don. Good to be back again on your podcast. Really appreciate it. I've been in the area of precision medicine now for over 20 years.

Been in companies like GE exact sciences. IBM. My journey with [00:02:00] the AI as started when I was with IBM and I've been fortunate to see this whole cycle over the last decade, if you will. And currently working with spent Lex heading up the strategy on marketing with Chakra and Lance and really enjoying positioning AI in position pathology.

[00:02:20] Don Davis PhD, MBA: All right. Very good. Yeah. I almost feel like I should give you a frequent visitor card. I think this is podcast number four for you. You did one by yourself and a couple other ones and then a couple other panels. And now this one, so thanks for coming back and especially exciting to get to talk to Chakra and Lance's.

[00:02:40] Dusty Majumdar PhD: Yeah, there's going to be exciting.

[00:02:41] Don Davis PhD, MBA: All right. So maybe we just start with kind of a, a brief overview of this area. So what trends do you see in terms of precision medicine and AI and dusty? Why don't I toss it to you first?

[00:02:56] Dusty Majumdar PhD: Okay, sure. I'm sure that lands in Chakra will have a lot to say about that too [00:03:00] specifically around a pathology.

So I'll start a little bit of a higher level as in between 2014 and 2018. There's a lot of hope. There's was a lot of hype as well. And what we see now are real applications with AI emerging. And the reason for that is we are now living. The era of tremendous convergence, if you will, we have more data than ever coming at us.

We have more computing power than ever before, and we have AI machine learning, deep learning algorithms, which are getting more and more robust. So things are changing over the last three or four years. This is the time when AI is becoming real and specifically in oncology. As there's so many variables that enter into a clinician's decision for a diagnosis and treatment, and AI can play a huge role in oncology.

And we'll talk more about that. For example, in analyzing complex and heterogeneous data coming from [00:04:00] intermix, if you will data integration to make to see a holistic view of the disease monitor patient's response, for example, in a complex disease like cancer, you got to check many different parameters to monitor.

And there is a consensus that for precision medicine to succeed, you have to look beyond just genomics at this point of time and take into account the different omics, including the pathological observations that we'll talk about quite a bit today in the case of solid tumor cancers and Dani know this, a major issue is that patients don't respond very predictably to treatment.

And I think one of the estimates I've seen lately at spent Alexis that 90% of the drugs are effective for fewer than 50% of the patients. So think about that, right? And immunotherapy many things is is is the ultimate miracle and that's less than 20% effective in most cases. So a [00:05:00] fundamental driver of the lack of effectiveness of immunotherapy is the heterogeneity.

And so we were composition and signaling network upon cells. And we're going to talk about how spent Elix addresses. And into tumor and the intro tumor heterogeneity plays a huge role in drugs, not being efficacious and many different kinds of cancer. So that's what I see right now. And honestly, it's been a great journey with AI starting, with the initial forays into care with Watson health and then working with a range of startups and now, really enjoying working with SpIntellx and getting precision pathology to the next level.

[00:05:48] Don Davis PhD, MBA: Yeah, it is for sure, a very exciting time from both where we stand with the science, as well as where we stand with the technology. It's just a, a great pivotal [00:06:00] point. Chakra, do you have anything to add to that? What trends are you seeing in terms of precision medicine and AI as a leader of spent.

[00:06:10] Chakra Chennubhotla, PhD: Thanks Don. So one of the things dusty alluded to is who do we really, what are the solutions supposed to address? And then if you think about both the clinicians and the oncologists and the pathologists, even we use this there's a big gap in the trust of AI algorithms and how much they trust it.

And we saw a huge gap in the explainability of these algorithms. The clinicians don't like to use black box algorithms, it spits out a decision without actually taking you through the thought process for how it made a decision, and it might show maps, and that's a very popular approach, but heat maps have a trouble that you project European and on this heat maps, as opposed to actually the willingly underlying.[00:07:00]

And so when we looked at the evolution of technology, we found these two being a big gaps of not having the explainability and how you build the trust. And how do you exploit the tumor microenvironment in a more holistic sense to build the technologies. And that's where that brought us to spent the Lex lads.

[00:07:23] Don Davis PhD, MBA: It's really truly exciting and great to know that, from a scientific standpoint, and this is where I would definitely go to you, lands is in terms of the sort of understanding how AI is making the decision. I would imagine that is pretty important overall and having the ability to really understand explainable AI versus, a black box as well.

But I would ask the same question, what are you seeing in terms of precision medicine in your space? And especially with.

[00:07:57] Lans Taylor, PhD: So building on the idea [00:08:00] of explainable AI in our discussions with potential customers going back to the beginning of the formation of SpIntellx we were told over and over again, that explainable AI would be a game changer in the field.

For the reason Chakra mentioned people don't like to take direct decisions from an algorithm without understanding why the algorithm made the decision that it did. We think the future clearly is with explainable AI so that the machine learning can occur. But then the output to the end user, which would be the oncologist or pathologist or other clinician, they will be able to look at what the algorithms decided and agree with it or not.

And they're still in control and that's critical in this field.

[00:08:56] Don Davis PhD, MBA: Yeah, very important. And with the [00:09:00] hype, dying off around, what I would say, AI, could be, and it actually becoming a more of a reality. What are you seeing? Some of the real applications now PI powered by AI that are on the horizon and Chakra will come back to you.

Maybe for the first answer on.

[00:09:24] Chakra Chennubhotla, PhD: What we observed but to clearly I'll take the pathology field as the, as a topic of interest here. 2017 was when FDA cleared given approval for scanning a digital scanner where you can use scan the slides, glass slides primary diagnosis. And this was the same time within a machine learning, particularly in the form of deep learning started picking up steam and in the first generation of tools that came out from these efforts was to [00:10:00] cough these digitized images as data and then run.

And then these are vast, huge amounts of data that you get out of this scanning.

And the deep learning algorithms are directly on these images. And then what we saw was exactly something that we already spoke about is that there's a gap in explainability that even though these algorithms can run these data sets, they don't quite meet up the expectations. So with that in mind, I think where the given the hype and give them the reality of what people want.

There are many applications now where you can use AI, if it is just the basic accounting of the cells and that's happening in some domains, particularly in the pathology. We had an, all you need is a cell count, how many 3d deposits through there, for example. But that again falls [00:11:00] very short of understanding in great depth, the deceased biology.

And then, so it was clear for us that if you want to get to the next stage, which is the precision pathology you need this additional tools. And I, my take is that something that, that does the prototype that this petition pathology will address many of these needs in designing, finding out why some patients respond to therapy and some don't what kind of, what, how we can build really powerful companion, diagnostic tests and and so on.

So there will be a range of applications where you imagine a position pathology making inroads

[00:11:40] Don Davis PhD, MBA: very good and lands. Why don't we come to you next with regards to this? What are some of the real applications that you're seeing now that are powered by AI that are on the hernia.

[00:11:53] Lans Taylor, PhD: In the broadest sense, I think there is an understanding that none of the individual [00:12:00] Omix are going to be the only part of precision medicine up there's power in genomics and metabolomics proteomics, and in pathology, in in precision pathology and more and more data will be integrated from a patient using multiple omics.

And it's going to be then critical to have explainable AI to again tell the clinicians of the reasons for the algorithms making a particular recommendation, and that recommendation could be a particular therapeutic strategy. It could be identifying a potential novel target for a drug. Could be an important and powerful companion diagnostic to drug that a pharmaceutical company is [00:13:00] developing.

And it can also be, is is basic as putting together a full patient

or what kind of comorbidities exist. It's going to play a central role.

[00:13:20] Don Davis PhD, MBA: Absolutely. And dusty, I would ask you the same question. What are the real applications that you're starting to see on the horizon in this space as

[00:13:30] Lans Taylor, PhD: well?

[00:13:32] Dusty Majumdar PhD: Yeah. So we're going to talk a lot about position pathology today, if you think about the real applications that you know, we are seeing today, out there, At the emergence of chatbots that actually interact with patients is on the rise in robotic surgery.

We see application of AI to reduce variation. It's a, there's a lot of data coming in and AI can definitely help and [00:14:00] making sure that it captures that, 3d 4d data and helps to make robotic surgeries more accurate, more precise. The emergence of virtual nursing assistants Lance talked about the use of precision medicine across the continuum of the drug discovery and development process.

So from identifying the right targets to the right biomarkers in discovery, to identifying certain populations of response in the clinical trial phase, all the way to market access, to really understand the efficacy of certain drugs in certain populations and opening up patients with different clinical indications.

I see application of AI throughout this whole spectrum. And the last thing I think that Lance was referring to was this whole concept of digital twins, where now, you can take all this data that you're collecting from various patient populations, or maybe even individual [00:15:00] patients and come up with a virtual patients on whom you can actually try different interventions.

I would just pass it on to you Dom and I'll mute

[00:15:11] Intro Music: my

[00:15:14] Don Davis PhD, MBA: okay. Very good. So as a follow-up, one of the, one of the things I know I saw years ago, Was this idea of having computer-aided diagnosis in radiology and, just this, the whole idea around, having something, help to read the images.

Really radiology and pathology were early adopters, overall of AI. Where are we going to now?

[00:15:46] Dusty Majumdar PhD: Dusty? I'm in radiology. If you think about it when it do into AI almost 10 years ago in, in quite a big way, and they have made some progress in terms of [00:16:00] helping the radiologists get augmented with AI in one of the coats I heard sometime back was in a radio AI is not going to replace the radio.

But the radiologists who don't use AI would replace radiologists or radiologists who use AI would replace radiologists to do. So Don, that's that's been more and more clear with the passage of time that AI is going to be a critical component of radiology and just like radiology and pathology and more and more digitization of images has happened over the last 20 years or so.

And it's no surprise that now, but how would you use emerging as a key area in which AI is being deployed?

[00:16:49] Don Davis PhD, MBA: Yeah, I just wanted to to bring up some of the chat that's going on currently while we're while we're on here. We have Michael's on he's [00:17:00] putting on your your website address so people can follow up with you there.

It also looks like Monisha is on from Boston as well. So just just wanted to mention them just briefly as well before we get to the next question. Welcome. And so with that Lance or Chakra I guess any other thoughts in terms of where we're headed next in terms of early adoption of AI in radiology and pathology?

[00:17:40] Lans Taylor, PhD: Certainly as you pointed out radiology has been using digitization for the last decade. And so they got into the use of machine learning and AI very early on pathology had a lag because [00:18:00] it came to digitization later. And it's only been in the last four years or so where the FDA started approving particular scanners as tools that were approved for doing diagnoses, not through the microscope, but from a screen.

And that of course opened the door to the first generation of computational pathology tools. And those tools have been very valuable. And I think this next generation. We'll really use more sophisticated spatial analytics to look at spatial biology, and then tie that to explainable AI so that the clinicians can understand what the algorithms are doing.

And that gets to the point that these tools aren't designed to replace the pathologist or the oncologist, but to give them added support in [00:19:00] terms of understanding, especially complex patient's results from a complex patient. There's a lot of pathology. That's very easy to spot something that's obviously normal or obviously cancerous.

It can be done by any algorithm on the other hand, some intermediate and difficult to identify things like in breast cancer with ADH, it takes the sophistication of the algorithms in order to pick up. And one of the things that we're seeing happen now is that it can be used as a pre a sign-out tool for pathologists to make sure the sub-specialists get those difficult cases.

Because right now, 90% of the effort of pathologist is spent on 10% of the samples. And that's because it's more or less a random handout of the patient samples to pathologist. This can help [00:20:00] guide the right patient to the right subspecialist.

[00:20:05] Don Davis PhD, MBA: Yeah. I just, I look at the, at this idea of augment.

Physicians who are intelligent and already applying complex, thought, to their field. And yet there's so much more information coming at everybody, with regards to all the technology that we have available and having an assistant that can help us sort through things would be important.

Chakra, anything else that you would add to this point as well in terms of, where we're headed? Yeah. So

[00:20:39] Chakra Chennubhotla, PhD: darn I feel like both radiology and pathology being the key Points for the patient care. With radiology, you're mostly looking for patterns that are either there or not there, but when it comes to pathology you have this extra there's so much information.

First of all, the tissue is at the premium. [00:21:00] And then there's so much more information and knowledge that you can next part properly indicating what might be the reason for the disease and where it might be heading next. And I think that sort of not just the diagnosis, but also the prognostic aspect of the pathology make a very interesting, makes it very challenging for these algorithms.

And I think explainability besides being very critical will entirely change the workflow for the propeller, just in release that they'll appreciate more. It's already, they have to go. Even when you have the slides digitized, they have this workflow solution that they have to in a log in. And then Y where the explainability comes in is that the touch of a button they can have.

Something that can prove the disease, they can prove the tissue and then ask what might be happening here as a landlord alluding do the cases in that tissue could be atypical ductal hyperplasia, particularly if it is a breast biopsy and there could be places where there is a cancer and it turns out [00:22:00] a difficult ductal hyperplasia happens about 15 to 20% of the.

For the breast a biopsy of breast cancer patients and the trouble here is that the discordance between pathologists in diagnosing atypical ductal hyperplasia for some of the studies that have been deported recently is extremely high as 52%. So now the idea is that if it is ADH, you're definitely want to get a second.

And the third opinion, just to confirm that it is eight years, but if that is not diagnosed, the chance of that patient getting cancer is about 50% within 20 to 25 years. So knowing this video early on and knowing that this is actually a difficult case is a key component of how these diagnostic tools will play a role in assisting the pathologists and the technicians in the, in this, in the space of the topic.

[00:22:56] Dusty Majumdar PhD: And then one thing, one thing, one thing I want to hit upon before [00:23:00] you go to the next question is what Lance and Chakra both alluded to. So in radiology, one of the reasons why we don't see even now widespread adoption of AI after all these years, after hundreds of companies getting into radiology and coming up with these AI algorithms is to a great degree, the lack of explainability.

Now there's a balance between explainability and accuracy, right? You can have a result that's 99% accurate, but you can't explain why, because you have a black box now, is that completely useless? Probably not. You can still view, but you need to have explainability, especially when you're trying to get to the root cause of a disease, a complex diseases like cancer.

And one of the critical obstacles in the widespread adoption of AI has been the perception of AI as a black box. And that's what these guys that spent Alexa trying to get away from. When clinicians see the [00:24:00] result, they're able to answer the question. Why are they seeing what they're seeing from an AI algorithm?

And I think that's a that's critical. I would just say that as we move along this discussion, you'll hear this theme over and over again, how explainability is making a difference.

[00:24:19] Don Davis PhD, MBA: And I guess from that standpoint, I'm gonna take a bit of an aside as well. Why is it that explainability is important?

I overall I could envision a few reasons, but would love for one of you to, just maybe, provide a little bit more thought in terms of, what you're thinking in terms of why is explainability is more, is so important, especially to the physician that's looking to care for them.

[00:24:54] Dusty Majumdar PhD: No. I know those Chuck Ryan landscape speak about a pathology. I'll just say from [00:25:00] a related field in drug discovery, when you're trying to understand complex interactions between genomic data lab data, the clinical data of the patient, and, you want to develop a virtual arm or a in silico arm of a clinical trial, you need to understand the underlying circuitry of the disease.

So if you don't really know a whole lot about, why the AI algorithm is saying, that this is where, you should be going and you don't understand the underlying mechanisms, you cannot treat that disease and you cannot develop a next generation drug without understanding the basic underlying circuitry of the complex disease.

That's, that's one of the rationale for really understanding, going beyond just a couple of layers of the CNN network at the top of the barn, but we'll get deeper into it.

[00:25:51] Chakra Chennubhotla, PhD: Yeah. So explainability, I think is on that spot towards causation in the end, I think the need to understand why [00:26:00] something happened is critical across these disciplines across pathologies.

As I said, the tissue holds so much information as to where the disease started, where it is now and where it is going next. And although we are only looking at one image or maybe a subgroup of cross sections of the at the biopsy, but there's a central team that we're driving hard at. And that's where they will appreciate this this assistance.

And that's coming through explainability from the AI algorithms. And we have seen that we have spoken to with the market research that we did there was overwhelming impression that that sort of explanations when we showed them the demo of how explainability. They would come back and say, this is going to be in our adoption of these systems.

And the best part is that they could override that and say, you're wrong to the machine and let the machine continue to learn for that particular practice. [00:27:00] What kind of reconcile being made and why sort of interaction between the clinician in this case, the pathologist and the AI algorithms with the key, for how AI could be adopted in a practice.

[00:27:16] Don Davis PhD, MBA: So important and land site, anything you'd like to add to this as well?

[00:27:21] Lans Taylor, PhD: Yeah, if we think about this, it's true for radiology as well as pathology, but a pathologist in particular trained for many years to look for patterns within a tissue sections their brains where the algorithm. And they could do pretty well.

One of the challenges with looking at blast slides through a microscope is that if people got very tired and the reproducibility was not all that great between pathologists, whether they were looking at the slide early in the morning or at the end of the [00:28:00] day. And in fact, some critical studies have shown that it's significantly not good in terms of having a dozen different pathologists look at exactly the same samples and make the same call.

It can be as bad as 50%. So the ability to have an aid, if you will, and have that aid use the same language that you do. In other words, when their brains look at the slide and make a call based on their vision algorithms then if the explainable AI machine does the same thing, tells them the answer and why it's reproducing, what the pathologist brain was doing.

This becomes then believable, and they can have confidence in it. Now it becomes a true aid to the process rather than just a simple we can [00:29:00] replace the pathologist with the.

[00:29:03] Don Davis PhD, MBA: So important overall and just this idea of, I think turning over some of these very critical, elements of healthcare as well is from a patient perspective.

I would say, I would want my, I would want to know that my physician really knows the background of the decision that was made at the end of the day, instead of fully believing at the end of the day, whatever it was that was given to them. You've gone to school you've done a lot to to look at all of the information that's out there.

Why not just double check it with with a computer with AI. Yeah, I agree. We see quite a few companies out there really trying to implement AI in pathology. Is there anything really that's groundbreaking that's come out so far and why don't we go to you Chakra for the.

[00:29:58] Chakra Chennubhotla, PhD: Thanks, [00:30:00] Don.

As I said, I think given the the nature of how the technology has evolved starting from the FDA approval of the scanners and this immediate obligation of the deep learning algorithms, particularly the black box and other black box algorithms are full on this data. It was clear that you need this ladder to climb to the next version, which is the position, the apology, and the components of that is one, of course, the explainability.

And we talked about that. The second is this unbiased wasteful analytics and here what I meant to say is that you run the deep learning. We can see that it misses out on some key aspects of the biology. For example, it will never be able to tell you that this is a fusion cell within the fusion cell has properties of both the tumor cell and the Macrobid cell.

And then what we realized is that [00:31:00] Having this ability for AI to recognize at this additional communication patterns between cells, because this fusion cell is a consequence of one such communication, and then keeping track of how this communication changes over time tells you how the disease has progressed.

So this is, and to enable that you have technologies, which do the spatial imaging of the tissue with the several biomarkers. Now there are platforms that can go up until a hundred biomarkers. So pretty soon this protein expression, these are mostly protein, but we can also have nuclear assets here.

You'll see that the the protein expression to start matching up with the transcriptomics in not to distance. And there are many studies that show you that just because you have the MRI genetics. Does not necessarily mean that the protein expression follows from the MRR net and RNA, especially. So eventually what the proteins do is going to be a very key [00:32:00] component and measuring it has been hard before I've gotten somewhat easier and hence the technologies are rapidly progressing.

So both in the AACR, we saw that the American association of cancer research meeting, and within the last two years, about half a dozen platforms already in the market. I love when you do visualize these patient properties, but then there are analytics that go with it to actually fully explore it.

The information that you're getting and that's where we come.

[00:32:31] Don Davis PhD, MBA: Very good. Lance, anything that you're seeing that's groundbreaking that's come out so far?

[00:32:38] Lans Taylor, PhD: I think it is important that the first generation computational pathology. I have been adopted rather rapidly during, especially the last two or three years.

And now there's been a great demand and people are beginning to see the importance of [00:33:00] analyzing the spatial relationships between cells in tissues and in particular, in solid tumors, the appearance of micro domains or specialized regions that have different types of cells in close association.

And that now enables us to investigate within those regions, the complex biology that is driving the cancer. And I think now the challenge is how much of the pathology and biology can be extracted. And right now we would say with the first-generation tools, a lot is left on the table, if you will.

And so on a natural progression. The power of the first generation tools as yield or our great deal. Now, I think there's pressure to extract even more knowledge from these samples

[00:33:57] Don Davis PhD, MBA: and lands. It looks like you have a [00:34:00] co-host over there on, on your side. Look like your cat was without rages.

Join us.

[00:34:05] Lans Taylor, PhD: And I had the notice how definitely I moved her away.

[00:34:11] Don Davis PhD, MBA: It won't be the first day. It won't be the first time here that this, that's happened. And I have great Danes that I have to prevent from entering the room because they would completely take the shot. So very well.

[00:34:23] Dusty Majumdar PhD: And also just to build on, what Lance just said, Chakra said this is really a quantum leap in terms of where the first-generation tools of the technology work and.

Chuck Ryan, Lance in Abbott, able to take these tools with spent Lex the last few years, because understanding the spatial network and the cell interactions and discovering those unique spatial collection of cells and cell states that Ms share, something in common, the micro domains, if you will, I think that's huge.

And that is where [00:35:00] explainable AI comes in, because now you're explaining why a particular recommendation was made, instead of just saying that the black box spit it out. And we believe that clinicians ontologists would really trust this algorithm a lot more than what the first generation tools put out there.

[00:35:19] Don Davis PhD, MBA: Yeah. And it, it sounds to me like, I guess correct me if I'm wrong, but it sounds to me like Spatial understanding in this space is what has been missing or more or less where the gap has been. Lance I'd come back to you. Where has the gap been in terms of this, the overall space here for in precision medicine and AI,

[00:35:46] Lans Taylor, PhD: And specifically for a pathology, but it's true for any of the omics cell communications and sub of different kinds of cell populations that are [00:36:00] communicating are crucial in disease progression.

We have some good tools for single cell genomic analysis and you can get a lot of deep information from individual cells. The challenge there is that you can go deep on an individual cells, but how many cells are communicating. And can you get all of those cells with that kind of. Using pathology, which is a spatial Omix approach to begin with because you can look at multiple cells in parallel.

And if you have enough biomarkers, you can define these domains. And in fact, you can in an unbiased way, have the data, tell you what the micro domains are in the first-generation tools, that's more biased. You might identify a couple of key biomarkers and then extract information from around those [00:37:00] biomarkers.

But in this second generation approach, the data itself in an unbiased way tells the user what are the domains that you should be looking for? And then it automatically identifies those domains and then extract great depth of information from those microglia. But once again there's going to be a deeper integration of multiple Omix.

And it, I think it will be even more focused regionally. So I think the most spatial analysis will come from the spatial biology in new precision pathology, but also being able to direct the other OMAC extraction of information based on this unbiased identification of key micro domains, where you focus the metabolomics proteomics or genomic analysis,

[00:37:59] Dusty Majumdar PhD: and [00:38:00] AI would be critical and integrating all these different, a complex data set.

And, one thing to remember in terms of what Lance just said, Don, is that you got to go micro, real small, and you got to look at the organization of cells at the macro level as well. So you got to look at the forest and the trees. Yeah. All right. So that's the key here, Chakra.

[00:38:24] Chakra Chennubhotla, PhD: So I was going to add to that sort of discussion is that Don biology is not static.

And the tissue actually tells you a story. And the emergent behavior is the key here. And you need tools that actually capture that emergent behavior like Lance was pointing out. Some of these micro domains might be unique to this patient. And heterogeneity is not random. Heterogeneity is organized as instructional units that we're calling as micro domains and they emerge and you don't know beforehand how to, what to look for and you should let the data drive for [00:39:00] the emergence and for the capturing of these micro domains.

And I think that's the key for how best to use this technology is going.

[00:39:09] Don Davis PhD, MBA: The best thing that I'm hearing too, is that you're not only looking at what's happening today. It's, you're giving information for what might happen tomorrow which really for, from what I've seen in my personal experience with with cancer with family members is that, so often you see this sort of this turn where they're treated one way and all of a sudden you see that it looks like, their treatment is progressing and so quickly it can turn around and all of a sudden can be back in a worse spot than you were originally.

And from that point standpoint alone, I, I definitely think that we need all the tools possible to look at what might possibly happen in this world of ours.

[00:39:58] Dusty Majumdar PhD: And you need to capture that [00:40:00] insight that could be hidden. It could be so subtle, right. So that's what that's what a spatial biology and really AI.

GaN can yield and sometimes it's not obvious, but the solution maybe right there, you're just not seeing it. There's a famous quote by Einstein that God is subtle, but not malicious. So it's the hint maybe already there.

[00:40:25] Don Davis PhD, MBA: Lance, can you tell us a little bit about the evolution of pathology and why digital digitization is such a big deal?

[00:40:34] Lans Taylor, PhD: Yeah pathology was essentially the last imaging medical platform to be digitized. And part of that was, again, what I mentioned before that pathologists are well-trained and their brains are the computer and the AI, and they didn't believe early on that a machine [00:41:00] could do as well as they.

But it is a fact that pathologists are human. They get tired. You look at N number of patient samples a day. You just can't be perfect on every call coupled with the fact that there are subtleties as Chakra said earlier, in some of these biopsies that it would be very difficult to identify. So I think two things happen.

A new generation of pathologists came along that were used to doing video games. And so dealing with a computer monitor was was simple to them. And the other thing was the fact that in some well published studies, it was shown that the concordance between pathologists looking at the same samples, wasn't nearly as high as you'd expect it to be.

And that has nothing to do with the quality of the pal pathology, [00:42:00] but the complexity of the heterogeneity in these samples. And there's the new generation of pathologists said we should start looking at making these measurements. The first thing that had to happen for them to be involved was the FDA had to approve of the first platform for doing diagnoses.

And that's been done and multiple platforms are now improved, and then it became feasible. Once you add a digital image of the path sample to start doing quantitation. And it started out doing very simple thing, counting the number of cells in cell division counting the number of T cells within the region relative to tumor cells.

These are pretty straightforward measurements and they're all throw tedious for the pathologist to do themselves, which they had [00:43:00] done in the past. Once that was shown to be useful this first generation is computational pathology. We've been talking about then started demonstrating capability.

And in fact that the FDA has approved the first couple of tests based on computational pathology. And so that's made a major move forward. And then I would say this next generation that we've introduced of. So we're calling it precision pathology because we've taken it a step further in making an unbiased and.

[00:43:41] Don Davis PhD, MBA: Yeah, I think I said this to dusty previously that the I was speaking to somebody that just came out of school on the radiology side. And she was essentially telling me, look, you can't take away. What we have in terms of digitization and in AI and radiology, it's it's [00:44:00] something that we feel like we need to have now.

And because the, there's this just this general discussion, which I think you guys hit on earlier, which is, Hey, is AI going to come and take our jobs away? But as we're seeing data just continues to grow at an exponential rate and you just have to have, people that, that are helping to interpret the results.

[00:44:25] Chakra Chennubhotla, PhD: Darn that was going to have that COVID the pandemic has seen some other practices. There's a relaxation for FDA as to how and when you can diagnose you can test it on your laptop without having necessarily an approval from the FDA. So that also increase the acceptance of checking out and, being able to view this images on digital platforms and be able to make diagnosis.

So that part of the impact in the last two years,

[00:44:56] Don Davis PhD, MBA: so important. So dusty [00:45:00] I'll come to you. So in terms of in radiology, the fact that images were ready, digital design, digital digitized, and accelerated the adoption of AI did a similar thing happened in PA pathology struggling to get the words out, but

yeah.

[00:45:19] Dusty Majumdar PhD: As Lance just said pathology followed radiology in terms of digitization, they were a little bit behind, but then I think it was three or four years ago when FDA actually said that you would read from digitized images, that's when it really took off. And that's where you see the emergence of a lot of these AI powered pathology companies which are currently, as Lance said, mostly first generational in terms of capability.

Most of them don't really look at this from an explainable AI perspective and the spatial element seems to be missing in a lot of cases. And that's [00:46:00] where, spent Alexis come in and really taking this to the next level with the explainable AI and understanding deeply spatial biology, geospatial analytics, Lance, you're ready to add, since you have seen the field for a long time,

[00:46:17] Lans Taylor, PhD: Yeah, no, I think I it, rather than adding to that, cause I think you summarized it.

I just want to pick up on something you mentioned earlier, dusty, where is all of this going? I hope I didn't take one of your questions away from you, Dawn. And that w what is the potential outcome from where we are in using machine learning and artificial intelligence. I think the future of precision medicine, and this is just a personal view.

Other people share it, but is that the patient, digital twins are going to be emerging. Of course, digital twins as a [00:47:00] technology has been around for a long time and in various industries, but now there's a growing amount of clinical data on every patient that is, can be managed over time for that patient.

And now we also are in the realm where using induced pluripotent stem cells and organoids, you can actually create what I like to call biomimetic twins of the patient by making a patient on a chip. And that technology is begun to explode. And now we have the potential to have the actual patient data, and you can't do experiments on patients, but you now have the biomimetic twin that you can do experiments on test drugs on and do other manipulation wrapping all around.

That is machine learning tools to integrate all of this data and [00:48:00] explainable AI to bring that patient to a kind of a point. Now, this is a huge, this is a moonshot, I would say, maybe it's a Venus. But it's something that we have the beginnings of the tools to approach. And I think more and more places will begin putting this together.

And because of the spatial element of doing pathology precision pathology will be a central component of this direction.

[00:48:29] Don Davis PhD, MBA: Yeah. I've had trivia Frazier on the podcast before she's the CEO of Obatala sciences. They have Oregon annoy Donna on a chip, sort of technology. And for sir, I could see the future of, what this might be able to contribute in general to the field, especially if you could have digital twins of, patients that, that are, maybe you have multiple routes of care that you could possibly go down.

Wouldn't it be nice to be able to test [00:49:00] some of them in the future?

[00:49:01] Lans Taylor, PhD: Yeah. That's actually coming to life as we speak

[00:49:06] Don Davis PhD, MBA: very exciting Chakra. Anything else that you.

[00:49:10] Chakra Chennubhotla, PhD: I was going to say that clearly there's this multidimensionality of data coming from many different places and it wouldn't be nearly impossible without the computational aid to process all of this information and get the help of the understanding of this data set.

So that's the other place to get in? I don't think so. You can live without those computational aids. And then the question is that how good are these aids and what exactly are they adding at your decision-making? Are they helping you get to the next place out of the stopping you or slowing you down?

I think the explainability there, again makes sense across these all these domains that you mentioned.

[00:49:54] Don Davis PhD, MBA: Yeah. And it seems in terms of literature, spatial biology is pretty [00:50:00] something that, that a lot of people are taking a look at it as well. Why is why now, I guess is the key question that I would have

[00:50:15] Dusty Majumdar PhD: a Lance? Yeah, go ahead.

[00:50:17] Lans Taylor, PhD: Yeah, that's a great question. It is surprising how something, once it, people start doing it, how obvious it is, but what the the mountain that has to be climbed to get people thinking like that. So in pathology, pathologists knew they had to scan a slide and look for patterns.

So it's been clear for a long time that spatial relationships in standard transmitted. H and E staining is important. What's happened and maybe what's really triggered. This is the development of these multiplex than HyperFlex labeling [00:51:00] schemes. And as Chakra mentioned, there are now four or five very good platform companies that can produce multi or hyper Plex labeled samples from just three or four biomarkers all the way up to a hundred biomarkers.

So you can really define in molecular terms the patient sample on a slide. So then the challenge is okay, how do you capture all of that information in an unbiased way to find the spatial relationships? And how do you explain what's actually there and how does it relate to the disease? So probably the development of the imaging platforms with the different reagents was the real trigger that opened the.

[00:51:48] Don Davis PhD, MBA: Very good Chakra. Did you have anything to add? I th I thought I saw, so you start to answer as well.

[00:51:53] Chakra Chennubhotla, PhD: No, I think I was going to second that and again, go back to the idea that [00:52:00] how the, in the case of tumor, in the case of. What is the host response and the host response, that communication, which involves communication between salts and this communication is spatial.

So there's no going around it. And the idea is how well can you capture this communication and how do you the emerging patterns of this communication and enhance this additional interest in spatial biology now and knowing how something that dusty mentioned before the heterogeneity is healing.

And then the fact that the the drugs don't work uniformly well. So knowing the microenvironment and understanding the micro domain has become the critical aspect from immunotherapy to all the standard therapies out there. And hence this addition, I feel that there's an additional focus now on spatial biology to really handle.

The micron man,

[00:52:55] Dusty Majumdar PhD: maybe Chakra, we can get a little bit more granular here. So talk about [00:53:00] immunotherapy and the failures and immunotherapy right now. And why you think understanding the spatial environment, understanding a little bit more from an explainable standpoint, the functionality of cells, if you will, how the development how's that going to help in today?

The response rate for patients with cancer stage four cancer to immunotherapy is less than 20%. How's that going to help and improve?

[00:53:27] Chakra Chennubhotla, PhD: Yeah, I was, I'm thinking of an example. We talked about micro domains so we should lean forward structures, TLS. This is something that people have noticed the renal cell carcinoma and other places wherein they actually count the number of tissue lymphoid structure, and they found a prognostic value for how many of these tissue lymphoid structures out there and what they are a collection of the T cells and the B cells.

But nobody knows what the [00:54:00] role really of these cells are, but not the suppressor. I went to the promote cancer. And so this really requires you to understand the spatial nature of the interactions between these salts and with the cancer cells right outside or out surrounding them. And understanding that network biology definitely means that you are found a way to explain the behavior of these micro domains in the form of digital infrastructures.

Let's say in predicting the response of this patient with particular therapy and this I think has become the key across all therapeutic modes.

[00:54:39] Don Davis PhD, MBA: Yeah. So one follow-up that I would also have is what feedback do you have from customers?

[00:54:47] Chakra Chennubhotla, PhD: Yeah. We two things that we noticed with the first-generation platforms, one is that if you look at this continuum of imaging, you have the transmitted light [00:55:00] platforms that generate clansmen who liked the major states in the IFC.

And then you have these other platforms with gender and the multiplex and HyperFlex datasets and made solutions that can cover this continuum of datasets. So we have two offerings his pro mapper. And that rumor map is the map that is for the transmitted light dataset and tumor map, but also the multiplex in HyperFlex datasets and what the customers have wanted up until now.

And I think that will be the trend is that they want to be able to use both the platforms in in understanding the tissue because each one of these compounds. Different information, morphological ideas with the transmitted light and the biomarker expression. What is the spacial statistical relationship between the biomarkers and on the prosecuted light side B what are the histological structures and what sort of modeling can we do work?

This is the logical structures and the prohibition with respect to the progression [00:56:00] of the disease. So the customer feedback really is that use of both these platforms in analyzing and extracting knowledge of the data.

[00:56:09] Dusty Majumdar PhD: Maybe Lance, you can talk a little bit about the colorectal cancer results that you have, where you actually had a direct comparison between what you guys do and the results from genomics.

[00:56:21] Lans Taylor, PhD: So sure. One of the projects that we did was on colorectal cancer and one of the challenges was in being able to make decisions on recurrence and using the primary tool. We were able to look at in retrospective study, which was the first one that was done. We could identify those characteristics of these micro domains and the biomarkers and their interactions that were consistent with of recurrence within eight years.

And those that were [00:57:00] consistent with no recurrence over that period of time and digging deeper within these micro domains, using what we call systems pathology. We can define the network biology within these micro domains and actually identify key changes that occurred that are consistent with recurrence or not.

This then changes the approach that you would take in therapeutic strategy. And actually one of our collaborators wanted to be able to have a a test to choose what treatment would go to what patient based on what is predicted, because up until that time, it was basically hit or miss. And you know that from, it sounds like your own experience in your family with cancer.

A lot of time, the patients are the experimental objects because they don't know what the [00:58:00] optimal treatment would be if treatment a doesn't work then you go to treatment B. So a great opportunity here is to be able to make predictions about what's going to happen. That can then guide the therapeutic strategy for those individual patients.

And in fact, in terms of. The, how well the precision pathology work compared to genomics tasks the presently reimbursed genomics test had a area under the curve which is one of the metrics use of about 0.72, which is a good enough that it was approved and reimbursed people use the test, but using the spend telex approach of unbiased, spatial analytics, unexplainable AI, we could push it over 0.9.

And I [00:59:00] think the impact is significant and that even brings bigger focus to the fact that spatial relationships are important. And just sampling the different regions of genomics. But what I would call low spatial resolution is not powerful enough, which is why people started to move to single cell genomics, but you'd want to do that relative to a high resolution, spatial biology approach.

And that's what we bring to the table.

[00:59:35] Don Davis PhD, MBA: I think that the follow on question that I would have for you though, to Lance's where do you see this going next? Where, what then happens and how does this evolve going further?

[00:59:45] Lans Taylor, PhD: I think this is going to be one of the important tools in precision medicine, where a patient is going to be analyzed and treatments [01:00:00] designed in advance based on their.

And I think in his cancer, as an example, we'll stop doing the experimental testing and try to identify those patients based on their genetic makeup and environmental characteristics have a higher probability of responding to drug a than another subpopulation. So it's going to improve patient outcomes as well as quality of life in, in, in cancer, going through whatever treatment.

There are, the other thing is using the same technology. One can select optimal cohorts for clinical trials. So we know that a lot of these drugs that don't get approved. Did have an effect on, 20, 30, 40% of the population of patients, but they were random population. If you could [01:01:00] identify cohorts of patients that have characteristics that increase the probability that the drug will have an effect and the patient will respond, you'll collect those patients and make a clinical trial cohort.

Now we can have drugs that are really targeted to the right subpopulation of patients. And as Chakra already said, going along with that, the same technology can be used to develop a companion diagnostic, which especially in cancer. This is something the FDA is very interested in, but it's all from that unbiased, spatial analytics and the explainable.

[01:01:42] Don Davis PhD, MBA: So I want to start wrapping us up, but in terms of just, other foul ons, Chakra, or dusty, do you have anything to add to that? I do have one last question that I want to come to, but anything else to add on this one?

In

[01:01:57] Dusty Majumdar PhD: general what I would [01:02:00] say is that down, we have talked about this this is the century of biology, right?

When Salman Nash. And I were the last podcast, we talked about the fact how the Marsha discovery is continuing through the century. And if you were to call 2003, when the human genome Atlas was published, the headlines in New York times and time magazine was that we have unraveled the mysteries of life, the picture of Francis Collins and Greg Venter.

And we know now that, we barely scratched the surface with genomics, so now.

[01:02:35] Don Davis PhD, MBA: As we are progressing, humans are more complex. Dusty

[01:02:39] Dusty Majumdar PhD: ones are way more and cancer is a very complex disease, right? So as we are looking into genomics from point of view, a mutation methylation RNA expression proteomics looking at metabolomics, looking at the microbiome environment around the tumor micro at, on the funeral, the microenvironment of the tumor [01:03:00] and looking at the tology in terms of development of the functionality of cells in the tumor microenvironment, combining that with robust explainable AI and powered by next generation, computational technology, whether it's quantum computers, you might need, very powerful computers to do all the computation.

Ultimately, when you have all this information together that is going to result in the outcomes that Lance is talking about. And then of course, once you can develop the digital twin based on all of this information, You can actually have interventions on digital twin to see, the efficacy or the safety of drugs.

So that's the promised land that we are working towards, hopefully in the next 10 years, we'll get closer to that.

[01:03:49] Don Davis PhD, MBA: Yeah. It's well, I was talking to somebody earlier today about Amazon's compute because they they constantly, whenever, whenever you [01:04:00] need more compute power, you just turn to Amazon and just say, Hey, I need more compute and they turn it on.

So this idea of having super and other things, becomes much more of a reality for people that are developing fabulous technologies. And so Chakra, I would come back to you maybe for my last question here. As the CEO of spent Lex, how do you see this evolves with regards to the patient and how things get used into the future?

Where do you see this, the clinical workflow. For patient outcomes. Yeah.

[01:04:36] Chakra Chennubhotla, PhD: Happy to respond to that. A add-on so out on that and how it's going to change their workflow. One way to put all this in context is if you have a cohort whose clinical profile is roughly the same, that molecule or phenotypes are roughly the same, some responded to therapy and some did not there's nothing you can do.

And that's when I think by [01:05:00] looking at the T. By doing this unbiased, spatial analytics and explainability. I, we have this power to to dig deeper into why some responded to therapy and some did not. And I think this basic idea that part of it, all the applications, scenarios that we painted, whether it is companion diagnostics or whether it is drug discovery or suggesting novel combination of this idea that you look at space and you have to look at the space with the idea that biology is going to emerge, and you want to capture this emerging biology in an unbiased manner.

And I think that's where this is going to head up

[01:05:37] Dusty Majumdar PhD: And on just in breast cancer, there are 700,000 biopsies done every year in the U S and if you look at all cancers, probably a few more. And I would say that at this point we use less than 1% of the information that is extracted from them in standard pathology and ultimate the gen one AI and pathology [01:06:00] tools that we have

[01:06:02] Don Davis PhD, MBA: until that's where that definitely where this could contribute lands.

How about from you?

[01:06:08] Lans Taylor, PhD: So this is even going to permeate the practice of pathology, whether it's in medical centers or in private practices. We have the collaboration with the large west coast pathology practice with greater than 60 pathologists. And one of the challenges is how do they get, and they have a specialist subspecialist as well as general.

How do they get the right patient sample to the right pathologist and with a workflow in place where the images are managed, the metadata is managed. The digital images are managed. And with the application of the precision pathology initial decisions are made and they can be directed to the right pathologist who can then look at the data, [01:07:00] ask the question, why make sure they agree with what the algorithm is said or not.

The efficiency is going to go up dramatically. Now the pathologist aren't going to be replaced by these tools, but their talents are going to be focused on the more difficult, challenging sample. And so the efficiency and accuracy will go up. We've already done some studies on some early work of that showed both of those occurred and the patients will be happier because they'll get a faster, more accurate diagnosis.

[01:07:39] Don Davis PhD, MBA: Yeah. So thank you so much. Any final comments from anybody before I wrap us up,

[01:07:46] Dusty Majumdar PhD: this was a great conversation. I'm glad that I got to be in this with the folks from spent Lex and really appreciate Don giving us this opportunity.

[01:07:58] Don Davis PhD, MBA: Thanks so much. [01:08:00] And Lance Chakra and dusty, thank you so much for being a part of the why science success podcast.

I greatly appreciate you being here.

[01:08:09] Lans Taylor, PhD: We enjoyed it. Thank you. Thank you.

[01:08:12] Don Davis PhD, MBA: Thanks so much.