00:00:00
We see the value of AI in doing slide QC, right? It's really important to make sure that we get rid of out of focus areas and all those typical artifacts we see in slide scanning systems using AI to exclude areas that are out of focus or folds. Understanding the complex tissue micro environment, you have tumor, stroma, muscle, fat, you know, you have all these different tissue architectures and typically you're only interested in at least in oncology of measuring the expression [music] in the

00:00:25
tumor cells. So being able to develop an AI model that differentiates all those different tissue classifiers [music] or different classes I mean is extremely valuable. And then you move on to the cell analysis where we're trying to [music] detect nuclei and measure cytoplasmic expression membrane expression you know I feel like have really been a gamecher you know the last 5 or 10 years of being able [music] to accurately segment cells. >> Welcome my digital pathology trailblazers. Uh today I'm here with

00:00:54
Doug Bowman from Indica Labs and our podcast is about companion diagnostic in the new AI era AI powered companion diagnostics. Welcome to the show Doug. How are you today? >> Yep. Doing great. I'm excited to be here. >> So we always start with the guest. H tell the trailblazers about you. >> Yeah. So um my background is in my degree is in biomedical and electrical engineering. >> Um and you know I like to start with my first job. Uh my first job was at a a medical uh research facility or medical

00:01:25
research group uh university hospital. Um and you know looking back at my first job it really was the foundation of like my career in microscopy and digital image analysis. >> So I both learned what did >> you do there? >> Um we had a lab that was pretty um I would say pretty high techch in that um the the principal investigator was wellunded. We built all the own imaging systems that we use. So digital imaging microscopes. This was kind of in the early days of digital imaging.

00:01:54
>> Um so we built all these high-speed timelapse systems and 3D microscopes. >> Oh wow. >> And so the group included um uh physicists, computer scientists, engineers, um optical people, microscopy people, and then of course the scientists that were in the lab doing their PhD research. >> That's a good like background for digital pathology work because this is also very multidisciplinary. >> Definitely. Definitely. Then you you learn I think this is a skill that is

00:02:22
>> underappreciated in the field because you have a lot of a lot of very skilled scientists that still need to talk learn to talk to each other. So that that setup that you had is very much setting you up for success >> and and honestly I did not realize this coming out of university right you know I'm an engineer doing you know programming and designing electronic circuits and stuff like that. And what what really hit me during that first position was the research the work. So I

00:02:47
was basically building things to respond and help the scientists answer their biological questions. >> Right. So my whole career has been around you know really using the technology to support you know biotech, pharma and medical research. >> Yeah. >> And your role at Indica Labs. >> Yeah. So I I'll after that I did spend some time at a a software company before Indica. Um and then after that I actually went into pharm a large pharmaceutical company and so that was here in the Boston area. I spent a lot

00:03:17
of time supporting kind of the whole drug discovery pipeline. So early discovery research you know time-lapse microscopy stuff like that. Um and then moving more towards the translational and clinical space. So >> this was a little bit in the early days of digital pathology. So Apiria was kind of just starting with their slides you know commercial slide scanners. So a lot of the effort early effort was building slide scanning systems out of microscopes. Um but then as Appirio came along came along we built uh basically

00:03:46
the uh digital imaging laboratory that went all the way from a limb system and sample management through scanning and histologology and and um you know delivering you know pharmacodnamic assays to our scientists that we were working with >> and then transition. So that's where I got to know Indica Labs and they were obviously a software provider Halo. Um we started using that at uh at the pharma company and I really enjoyed working with the team at Indica and so ultimately I convinced them to hire me

00:04:17
and build a pharma services group. >> Oh my goodness. So you're the the OG of the pharma services at Indica. [clears throat] >> Correct. So over the past like nine years you know supporting our biotech and pharma clients um you know we saw a real need um you know especially in the world of companion diagnostics a real world a real need to combine groups at Indica. So we now have a precision medicine group that consists of a traditional pharma services team that's supporting pharma in a services role and

00:04:47
then our AI diagnostics team that is building AI commercial solutions. Mhm. >> Um, so kind of a combined that really fits well into the CDX story. >> So you're the boss of which group? >> Of both. We've combined our pharma services and AI diagnostics into a precision medicine group. >> Yeah. >> Amazing. So speaking of precision medicine and let's talk about uh the companion diagnostics. Let's start with explaining uh what it is. What is a companion diagnostic and why they are

00:05:16
becoming critical in precision medicine. >> Yeah. um and especially uh for complex therapies and by complex therapies I mean antibbody drug conjugates. >> Yeah. So com companion diagnostics, the audience is probably pretty familiar with it, but it's essentially what the word says, right? It's a companion to the therapeutic. Um, so in precision medicine, as you get more and more specific and targeted therapies, um, the need for a companion diagnostic is really important because you're going to

00:05:45
you're going to help the clinician understand what patient population might respond to the treatment better. Um, so that companion diagnostic really gets developed in parallel with the therapeutic. Mhm. So you have a the test and then uh depending on the result of the test the patient either will get the therapy or not and what are we assessing in this test? >> I mean it really depends on the therapeutic. I mean a good example is her two for breast cancer. >> Um her two um you know the the targeted

00:06:15
therapies for her two rely on having her two expression. >> So a test for that would measure her two expression in the tumor cells. Mhm. So we have a biomarker and then we decide okay a certain level of biomarker qualifies people to get the drug and below that level there is no correct benefit of the therapy. >> Correct. >> Correct. >> Correct. >> Correct. >> Correct. And that level is really important of course like which level is going to the patient going to respond if

00:06:39
it's low expression versus medium versus high expression. >> And that's probably involved in the development both of the drug clinical trial and like all these combinations. Right. >> Correct. the whole process of of developing pharmaceutical with the test, >> right? I mean, I think in early pre-clinical work, they can get an idea at least in animal models like what the cut off is for for correlating with response and efficacy, but then and they can model that and maybe estimate what

00:07:08
that is in the human, but really you need to do those tests in the human samples to really understand what that cutoff is. >> And that's certainly one of the challenges I think for companion diagnostics in general. Mhm. >> So, you know, companion diagnostics like the her two example, that is one that's traditionally read by a pathologist. Pathologists are very good at um you know, looking at different tissue morphologies and then measuring expression. Um but there are challenges

00:07:34
with some of these biomarkers and her two is a good example where you know more recently they've seen uh clinical responses with her too low and her two ultra low and when you have really low expression >> you know the importance of the um the hisytologology and the um reagent as part of the assay is extremely important but it is an area where quantitative image analysis can really excel at capturing that low expression >> and I personally have a strong opinion on like visual scorings that until we

00:08:04
had good image analysis h was as good as we could get. But the and the strong opinion is basically the biology the physiology of human vision. You can only estimate and if you need like a precise cutoff and decide whether a biomarker level is below or above this cutoff, there's always going to be 50/50 chance that you're going to be wrong around the cutoff if you're doing [clears throat] it visually. Right? If it like if it's obvious then it's going to be fast and reliable but around the cutoff which is

00:08:37
where it matters because this is where uh the decision is being made is the patient going to get the drug or not. human vision is not the most precise instrument to do that. >> There's that classic example where there's a checker, you know, a grayscale checkerboard, right? And you have two like little cylinders in different parts of that and you know visually, you know, 90% of the people are going to say those are different intensities and then they slide the two cylinders next to each

00:09:03
other and indeed they are the same intensity. So the context really matters and when you think about like the checkerboard relative to tissue you know you get really complex heterogeneous genius um expression in a tissue it's a very challenging problem for a pathologist >> visual visually scoring >> there's an interesting paper as well by a colleague of mine Pam Efner where she explains the paradox of ground truth paradox especially in this area of visual illusions of coraler perception

00:09:35
ion and this is where computers are a lot better in doing this. So that brings me to the next question which is AI capabilities. So we have the companion diagnostic uh it it is a test to check okay what's the level of the biomarker because we have a cutoff h and we are not that great in evaluating around the cutoff. So and and that that was the traditional way of doing this. Um also traditionally these tests were uh relying on a single biomarker right we are we are giving this example for two

00:10:09
that's just one biomarker and now in the era of precision medicine there's going to be more there's going to be maybe imunofllororesence at some point basically uh the visual image to assess is getting more and more complex so how is AI expanding the capability of companion diagnostics >> yeah I I think AI I is really impacting all ex all aspects of quantitative analysis. Um we see the need for or the value of AI in doing slide QC. Right? It's really important to make sure that

00:10:43
we get rid of out of focus areas and all those typical artifacts we see in slide scanning systems. They're not really inherent to the slide scanning systems but you do get some artifacts like that. So using AI to exclude areas that are out of focus or folds, understanding the complex tissue micro environment, you have tumor, stroma, muscle, fat, you know, you have all these different tissue architectures and typically you're only interested in at least in oncology of measuring the expression in

00:11:09
the tumor cells. >> Um, so being able to develop an AI model that differentiates all those different tissue classifiers or different classes, I mean, is extremely valuable. Um and then you move on to the cell analysis where we're trying to detect nuclei and measure cytoplasmic expression, membrane expression. Um the AI models, you know, I feel like have really been a gamecher, you know, in the last five or 10 years of being able to accurately segment cells. >> If you look at a tissue like colarctyl

00:11:38
where you've got these really densely packed nuclei and they're elongated and they're all right next to each other, you know, the ai models are very good at capturing those. Mhm. >> They still have their challenges. I mean, throughout that whole workflow I described, there's challenges at each step. >> I think immuncology is is like a classical example where there's too many variables and we try to compact or or you try to find like the one thing that the pathologist can assess visually or

00:12:07
at least historically. Uh this was the tendency because then you can easily deploy it in the clinic. But for [clears throat] immune we know okay different immune cell populations the locations of them how far they are from the tumor margin and all that stuff by the second variable human vision is not good enough anymore. >> Um so with AI and and by AI here our application is going to be AI for image analysis in the broadest sense of of this >> broadest sense of of the meaning of the

00:12:38
word basically like AI for working with these images. any benefits for the patients. How would you describe why is AI more beneficial for the patients? >> You know, good question. I think ultimately it comes down to the the accuracy of the test. You know, whatever test it is, but certainly in image analysis, the accuracy of that test, you need to make sure that it's detecting, you know, only tumor cells or as much tumor as it can. the quantitation of the biioarker itself whether it's a nuclear

00:13:07
cytoplasmic or membrane making sure that's as accurate as possible so that when it gets to that stage where you're defining a cutoff to enroll or not enroll a patient that that's as accurate as possible >> yeah and that basically is a function of the lack of precision of human vision basically we can overcome that hurdle with computers >> right now that's not to say there are numerous challenges in it so you know I think at this point. It's still a pathology assist tool, right? It's going

00:13:36
to be a tool the pathologist uses to make to make the measurements. The pathologist is going to, you know, see markups and in images that give them an idea of what's been scored and how it's been scored so they can, you know, verify >> that they agree with the the decision of the AI model. I think it's a perfect combination of this tandem work uh because you leverage the advantage of the computer of calculating the exact threshold of as exact as we can get and it's a lot more exact than with human

00:14:07
vision but then confirming is it in the right place is it actually doing visually the job that the pathologist would do but with better calculations. So like you said assist tool >> tandem work with the pathologist with pathologists in the loop >> um relevant for the quality of the test itself and also for for patient trust for patient safety I think >> right and I think it's important to note that the pathologist is involved in much of the earlier stages as well right the

00:14:35
pathologist is involved with training the model assessing the model iteratively improving the model so they're they're involved throughout the process >> so specifically for antibbody drug conjugates. What are the challenges there? Like what is why is it more difficult to develop a companion diagnostic for this type of therapeutic? Yeah, I think uh ADCs so antibbody drug conjugate. This is a system where um you know the you identify a target population of cells. You know, hopefully

00:15:04
there's some protein that's being expressed only in tumor cells. >> Um and then the ADC has a drug toxin that's attached to it >> and so it's very precise in identifying the cell population that it's attacking. So as far as companion diagnostic, it's really an ideal companion diagnostic because you know which the target population is. you can develop an assay all the way from you know the antibbody the reagents scanning image analysis workflows specifically to measure that

00:15:32
>> um in most antibodies are looking at a single target um but there are some new bicep specific uh ADCs so they're looking at two different targets so there's always the question do we develop a single assay that looks at two markers or do we just do two in parallel two two side by side I think we're still very much in the early stages of that >> um But I think time will tell as we as we get more and more data um and maybe more and more information if that leads to a more accurate test then we might

00:16:02
move to a multiplex environment or multiplex if environment >> because I think because we're still optimizing for okay how can it be deployed without the computer even if we have the capability but the moment we actually like have proof that this biology cannot be assessed in a different way that's going to be the moment where It's not going to be an option any more to not do it with a computer, not do it with AI. >> Yep. >> So, >> and I think the I think the Astroenica

00:16:31
trope 2 story, if you're familiar with that, is a really good example where they identified a relatively um unique, I guess, signature >> that captures or correlates with efficacy. So, um it's an ADC, so theoretically it should be membrane bound. So you would think that just membrane expression would would really drive that that scoring criteria, but what Astroenica found was that it was a ratio of the nuclear I'm sorry of the cytoplasmic expression and membrane expression. >> So that's a case where image analysis

00:17:02
fits right in. It can analyze those different subcellular compartments, create that ratio that you can then create the score. I am very excited about this one because this is exactly like and even without going multiplex going multiple markers it's the same marker where there is this visual thing that a pathologist can calculate >> very challenging for a pathologist to calculate a ratio in their head while looking at all the heterogeneity in an image. [laughter] >> Yeah, I know. So, so this is a perfect

00:17:31
example of that >> and it falls really well into image analysis because you know image analysis is generating a lot of different metrics with each image. It's not just expression. It could be things like nuclear shape, nuclear size, cell morphology, cell size. So, there are a lot of features that get extracted out of an image analysis algorithm. And you never know, one of those might be important for, you know, correlating with outcome. >> And it's it's a little bit like there's

00:17:55
never a guarantee that you're going to find this biology in there, but if you don't look, then you're not going to find it. And here here they did look uh enough to find something that was better than just a visual cutoff, >> right? Yep. And I think that actually introduces one of the challenges with development, right? Is they don't know if there's going to be a predictive score, >> but in order to try to drive that part of the CDX development earlier, they want to start that at risk earlier in

00:18:25
that development. They can't wait to the end of the clinical trial then to go back and do the CDX. They need to do those in parallel. That is true. And I always make when I I remember this test and then I think okay how many of other biomarkers were tested and how many drugs maybe have failed the clinical trial because there was not enough granularity or not there wasn't a method to extract enough granularity out of the data to actually find some signature and it's you know it's always a funnel you

00:18:55
have you're you're going to like I I say life is a funnel because there is always so much trial and error that you have to do at the top of the funnel to actually get to something and there is never a guarantee that you're going to get to something >> h but there's a chance so definitely leveraging AI to find something that can later help patients um justifies this investment you know in infrastructure in um and you mentioned that um when you're going to be working on an AI based

00:19:25
companion diagnostic it's going to be a system you said scanner and the lab work for IHC and It becomes an end to end workflow, right? Definitely. Definitely. >> It's not just like IHC and you show it to a pathologist. >> And I agree that biomarkerdriven clinical trials are going to become more and more prevalent whether it's just to confirm you know mechanism of action of the drug itself or to look at outcome or you know what patient population for patient stratification. I think it's

00:19:54
going to continue to evolve and grow. Mhm. I think with this with this I don't know proof of concept you can call it proof of con concept that it actually is there because until this uh very thing that they found out I was like oh maybe it's there maybe it's not there [laughter] everybody's looking for the holy grail of signature and they actually found it. So, it's it's worth trying, but it's a workflow. It's a system. And that brings [clears throat] me to the partnerships that uh the

00:20:23
partnership that Indica Labs have with Leica. >> Yep. >> Can you tell us a little bit about that? >> Yeah, I'd love to. So, back in January of last year, 2025, we did a strategic partnership with Leica Bios Systems, >> and it was really twofold. It was it was around our clinical product, Halo AP. Um, so being able to deploy these algorithms in the clinic or the diagnostic labs. Um so that's Appirio Halo APDX um is the software platform there. So really focused on the clinical

00:20:50
market. Um and then combined with our AI diagnostic and AI uh clinical algorithms being able to deploy that in the clinical market. But secondarily and really interesting to me is the companion diagnostic right Leica is a leader in that endtoend solution where Indica fits in really nicely. They can do all the reagent development. They do all the um uh imunohistochemistry and protocol optimization. They have slide scanners. Uh they have the um GT450DX which is um an FDA cleared slide scanner. They have a global footprint.

00:21:22
So combining that capability with into clabs and AI development really allows us to offer that endto-end solution and I think especially true in companion diagnostics where there's going to be a quantitative component >> very much and you have a history with them right because the founders were previously at Leica and then they span off to work on indicabs and now there is um the collaboration back to work together. Yeah, correctly. Correct. Our founder um was it back in 2011 um left Apirio and and built image built Indica

00:21:57
Labs. >> Yeah, it's quite exciting actually. I think and you'll see that at the company in general. A lot of us were in the area whether they were at biotech or pharma or in academia using the products using image analysis and then have joined the company afterwards. Mhm. >> It's also a good fit because of this heritage. You kind of understand uh their technology and like you knew what they were doing. It's not like trying to fit to each other without knowing what it was before.

00:22:25
>> Right. Right. Yeah, you know there are there are a lot of um you know image analysis is really prevalent but I think the domain knowledge and you know when you get when you get a certain um you know tissue type or challenging images um really understanding what those challenges are and what the best solution or what path forward is to try to optimize and remove like artifacts and stuff is that experience is really powerful. >> So when the decision is made to develop an AI powered diagnostic um how does

00:22:53
that actually work? What does the journey look like from early development to regulatory approval? And I want to cover it for like now for single biomarker, but also how would it look in the future if we had multimodel, multiplex, different um data modalities flowing into this because that's an option as well, right? >> It's definitely an option and I think that's the power that we're going to see with precision medicine in general is being able to bring in different assays

00:23:21
that add to that story. Um so the development journey you know it's interesting because at least for quantitative image analysis companion diagnostics right there is not one on the market yet so we're still very much early in that journey. >> Um but the workflows um and the kind of progression of assays from early development uh early discovery through development and stuff like that has been going on for a while. It's just that last piece that we're kind of waiting on. >> So we typically see these image analysis

00:23:49
algorithms being developed in the preclinical models. Um, so many of our pharma customers have our software. Um, they're developing the algorithms maybe internally to help drive their pre-clinical work, maybe their lead optimization, for example, and when they're looking at different preclinical models. So, they're developing the algorithms or some version of the algorithms earlier. um as they start moving into the clinic, that's an opportunity where they might come to someone like us to further develop that

00:24:12
algorithm, maybe do a feasibility in uh human samples, modify or develop a new algorithm and train that algorithm, you know, for those clinical samples. >> And that early feasibility allows them to get like an early story like is this going to work in the in the human setting? And then as they start seeing promise there then you would want to start further developing the al algorithm more towards a research use only or clinical trial assay and seeing if they start seeing results in the clinic. Okay both therapeutic right this

00:24:39
is going along with the therapeutic. So the biomarker assay might work great and the therapeutic does not. So all of these things are working in parallel. >> Um but as that therapeutic progresses and especially in the idea around companion diagnostics and ADC is a good story for that. you know at some point at risk you need to start developing that uh ADC companion diagnostic in parallel with the therapeutic. Mhm. >> So we would be further generalizing the assay looking at different scanners,

00:25:07
different sites, different analytical validation, clinical validation kind of in parallel with the therapeutic development. >> And then once you have like once you know it works, then it's going to be the [clears throat] regulatory path of >> then. Yeah. Exactly. I mean this this you want to you know you want to start that regulatory process as early as possible for AI development. That's typically around the documentation and quality management system that you have while you're doing the development. So

00:25:32
we would typically start the design history file doing all that documentation a little bit at risk earlier so that if it does lead more towards the regulatory process, we've got a jump start on that and it's going to be a lot quicker to get there. Mhm. And the documentation is a key word I I hear from regulatory consultants and everybody who's helping companies do these uh clinical trials or for devices that is document document from the get-go because you cannot go back to redocument.

00:26:02
>> It's very challenging to go back right. I mean, you need to document where your your training images came from, the different versions, you know, the iterations that you might be doing throughout development, making sure you have the right, you know, target demographics in your training data for what the the clinical trial is going to be. And the sooner you do that and the more data you collect, of course, the the stronger and the better the model you'll have, but ultimately you'll be

00:26:23
farther along when you move towards that regulatory process. >> Definitely. And and what about the multimodel? Does the development look different? I think the development maybe is similar, but it just adds complexity if you've got more than one measurement. >> Uh, and an example I like to use is in the ADC space, there's something called the bystander effect. >> Um, so the bystander effect when the the ADC drug hits the cell, gets internalized, the toxin gets released

00:26:51
and kills that tumor cell, you end up with free toxin. So that cell, you know, that cell dies and you get free toxin. So if you have a neighboring tumor cell next to it, that toxin might end up killing that and you'd get a bystander kill, which would be a positive effect, right? So being able to quantify that can add some value to that that scoring metric. >> And we do that through what we call spatial analysis. So you're measuring the expression level of those neighboring cells and its proximity to

00:27:17
the target positive cell >> to get a score about if there's more >> if there's more tumor cells close to those tumor uh target positive cells then you're more likely to have that bystander effect and you're going to be more efficacious for that that therapeutic. But it is a different metric. It's a different scoring metric. It is one that's very challenging for a pathologist, right? Trying to capture the the spatial relationship of all the tumor cells. So it uh bodess well for

00:27:43
image analysis to bring that modality in as well. And ultimately it's just going to be part of that scoring metric right that scoring that predictive score is going to be a combination of membrane expression cytoplasmic expression and the maybe the proximity as I remember the story of immunoscore where um they were measuring the was it CD8 and >> the CD8 the the this the um cytotoxic tea cells that were close to that tumor margin. >> Exactly. So that was that that's already something you cannot do visually and it

00:28:18
ended up being a image analysis driven score for colurectal cancer prognosis. Yeah, I mean that was really exciting I think for the image analysis world too, right? Something that you can't capture, you know, visually or very hard to capture visually and from an image analysis perspective, >> it's a relatively simple um uh calculation, right? You're looking at all the immune cells, you define the tumor border and then you just measure that that infiltration region. >> You know, one of the challenges is the

00:28:44
fact that you're going through the regulatory process with the therapeutic, with the AI algorithm, with the actual cutoff that you're working with, right? trying to understand what that cutoff is. You know, all of those are going through the regulatory process in parallel. >> Any other considerations for regulatory? What else is challenging? >> I think with ADCs and companion diagnostics, the biggest challenge is that the the diagnostic, right, the biomarker assay needs to go through that

00:29:10
regulatory process in parallel with a therapeutic. So you know therapeutic you're looking at uh adverse side effects and efficacy and meanwhile you're you're going through the biomarker process where you're developing the assay you're validating the assay um both analytical and clinical validation and then going through that regulatory process in parallel and somewhat at risk right >> um so it's just a lot more documentation needs a lot more communications back and

00:29:36
forth between you know your diagnostic partner the pharma company and the regulatory agency so you know That relationship is obviously extremely critical uh for any biomarker development but especially companion diagnostics. >> Yes, the relationship aspect another keyword I sometimes like highlight these keywords the relationship because you have a compounding of complexity right there is already enough complexity in filing for a drug. Now you have the test, you have people from the pharma

00:30:06
company, you have people who are uh maybe software developers in one way or another, image analysis specialists, the lab people. So a a very complex landscape that needs to be coordinated enough to later be able to uh file the necessary documentation. So definitely important. >> Yeah. And I think finding that partner, >> the people you want to work with, >> right? I mean both from a a project relationship, right? need to have trust that they're going to be able to execute on the technical aspects but then also

00:30:35
the regulatory aspects. >> And it is one area we see our partnership with Leica being one of our strengths is you know we have that experience not only from a technical aspect from end to end but from the regulatory aspect as well and even in within our AI diagnostics group we do have experience >> going through that regulatory process. We work closely with our regulatory group. Um, so we have a CE marked algorithm. We have a 510K cleared Halo AP which I mentioned earlier. So we've

00:31:02
already gone through that kind of regulatory process on a couple of our products. >> Um, and obviously looking forward to doing that with a CDX. >> Yes. And I think it's I mean it's a process. It has like definite steps. So anybody who wants to do that can go through the process. But the experience of having done it is invaluable. >> It is. And on both sides, uh, both companies have that experience, right, for for different things. This is like a different way of thinking about your

00:31:30
products once you're in the regulatory environment and you're trying to get something cleared, something uh maybe approved. >> Yeah. And I think ultimately, you know, when you're earlier in the process, you're trying to decide between, you know, a a ma manual visual read versus, you know, a single plex biomarker assay or maybe a more complex measurement. So, you kind of want to go through the process in parallel with that as well because if it's a simple onoff type biomarker, then there's, you

00:32:01
know, there's probably little reason to do the quantitative analysis. Um, but if it if the quantitative analysis adds value, then obviously you want to take care of that, too. uh and then having a platform that you can deploy that you know at a diagnostic lab is is important as well. >> Definitely I think the availability of options doesn't mean that you need to take advantage of all the options and and what I mean is like okay if you can throw all the markers and do a multiplex it's going to be challenging to deploy.

00:32:29
So, it's it's kind of a balance. The least complicated, the least complex that gives you the most benefit. And sometimes it's going to be a simple marker for visual assessment and sometimes it's going to be spatial relationships between different um immune cell populations, right? So, >> yeah, I think that's one of the biggest challenges is what what that added value is and making sure you can come up with that that decision and story. Mhm. And I think that makes image analysis now it

00:32:57
takes it >> from the just like specific application level into the level of this is an acknowledge the method scientific method that should be applied and then the scientists um define okay what's going to be the application but but but image analysis is something that needs to be leveraged because you're not going to find it out any other way Right. >> Yep. >> Which makes me excited. >> It's very exciting. >> I come from the image analysis background as well.

00:33:29
>> I mean um for me obviously being able to impact patient care is is ultimate goal of everyone that's working in this field. >> Um but personally the my kind of career path where I kind of started off in basic research and then now I'm going to have you know potentially have an assay that is impacting like a therapeutic. That's super exciting. >> I know. >> Yeah. Especially with digital image analysis and microscopy with all the technologies that I've grown up with.

00:33:52
>> I know it's like full circle moment. >> Definitely. >> You know what else is important that ties back to the user group meetings u meeting that you have today is the image analysis literacy because now that it is becoming this established method. Um it it it used to be like cutting edge who liked the scientists that liked it would get involved in it but now it's similar to molecular pathology. When you become a pathologist there's no option to like not learn molecular. This is part of the

00:34:26
patient care. >> It's a fundamental tool that [clears throat] everyone needs to be using >> and and I see image analysis like that right now as well. And we were when we were coming uh to this podcast, when we were walking and we were talking about okay the when I was doing my image analysis work at the image analysis company that um I was working at the crucial thing that we had to start with was was the communication between pathologists and image analysis scientists. The different image analysis

00:34:54
concepts that the pathologists needed to learn. So like the segmentation, classification and all these computer vision high high level computer vision ter terms just to be able to communicate with your scientists and basically do a joint work of creating image analysis solution that actually matches the tissue. >> Yeah. And I think that's where the um we call them application scientists, but they're the application level people. These are typically people that are really experts in image analysis or

00:35:25
experts, you know, they've been in science labs and, you know, biology degrees, but they're that kind of middle person, right? They can understand what is needed in the industry, you know, what what algorithms are needed, what tools are needed, and they can convey that back to our development team so they can develop those tools. >> And that's one of the really values of our user group meeting is we get users together. Um, we hear, you know, of course we're presenting to them some of

00:35:47
the new features of our software and giving them some information about our software, but we're also having some users present to see what they're doing. Other users in the audience can understand how other people are addressing certain challenges. And then, of course, all of that, all of those ideas are feeding back into our development team so that we can increase, you know, improve upon our product or continually improve our product. >> And I looked at the names uh at the registration table. There was people

00:36:11
from pharmaceutical industry, academia, some private research places. I don't know. Basically a a big breath of a big variety of >> image application spaces. >> Yeah. I mean, Boston, Cambridge is like a hot spot for that. You know, you've got bio, you know, dozens, hundreds of biotech companies, large pharma, you have CRO's that are actually running the the facility or running the the the laboratories. Um, so it's a really kind of a hot pot of of technology and excitement,

00:36:42
>> good place to host the user group meeting, >> right? Yep. >> So what do you think is going to be the future? What's relevant for the future for people who are starting in this field? Because now that this is an established method that we have proof of concepts of image analysis being used in the companion diagnostic space, there are still more people that have not done it than those who have done it. there's going to be a lot of beginners and how do we deal with that?

00:37:10
>> Yeah, I think the education of of the next class of scientists is always extremely important. We see that partly because we have a lot of our software in academic labs. So, a lot of these research labs are using our software um to answer a lot of these questions in the early early research. Our user group meetings are kind of like an educational opportunity as well because if you're if you're not even a user, it's not just a customer, it might be a prospective customer. they can come and listen to

00:37:34
the talks too. So I think it's extremely important. I think it's more and more common um people in graduate school even undergraduate school are getting exposure to image analysis algorithms earlier and earlier. I think you know there's a lot of other organizations there your there's your podcast right an excellent resource um for people learning about the different technologies the buzzwords just understanding what the buzzwords mean. >> Um there's groups like the digital

00:37:57
pathology association that puts out white papers and educational papers. Um, I think all of that just continues to grow. >> Anything specific that you're hopeful for in the future? >> Well, I think first would be a digital CDX. That would be super exciting. >> I'm waiting for that as well. >> I think there is so much value in the multiplex technologies, whether it's multiplex chromogenic or multiplex fluoresence. I mean, in a simple case, if you have a bspecific ADC, you know,

00:38:23
you got two targets. So, you need to have two assays that do that. in the immunoncology space just trying to understand the immune micro environment and the relationship between all these um different immune populations relative to the tumor micro environment you know that work is going to continue to to expand and I think as soon as they start identifying these multiparametric signatures on these multiplex panels it's inevitable that it's going to move into the clinic >> um you know as you mentioned there are a

00:38:51
number of challenges >> you know pre-analytical variables the wet lab work a lot of challenges to overcome but you know the technologies there. >> Mhm. >> Yeah. >> Yeah. And the challenges maybe they will disappear but for the nearest future probably not. So we have to figure out how to overcome them to get to the the insights from the biology that are behind. >> I think from an AI model perspective you know there's new AI models continually coming out. Um you know as a company

00:39:19
we're assessing different models. There's foundation models where you've developed a >> you know with a lot of training data you've developed like a core model that's works for pathology. >> Um what we've seen is that that is a nice starting point to customize that training. So if you're going into a specific indication instead of starting from scratch we can start with some kind of base pre-trained model and then add training to that. >> Um so I think the models are going to

00:39:45
continue to evolve. the training sets, you know, getting access to training sets is one of the big limitations. Um, we're going to continue to build relationships with pharma companies because a lot of this data is sitting in a pharma pharma space. So, collaborating with them to continue to improve these models. [snorts] >> Thank you so much. Thank you so much for the insights into AI powered companion diagnostics. And if somebody's interested in Indica exploring what you guys are doing, where can they find you?

00:40:13
Uh yeah, look on the web indicab.com. Uh obviously the website has a contact form. We have info@indicolab.com. Um and obviously reach out to any of us at any time. We have application scientists spread across the globe. We'd be glad to answer any questions. >> Thank you so much, Doug. Thank you. And uh I hope you have a great day, great user group meeting. I'm going to be there as well. And I talk to you in the next episode.