In our latest of our digital pathology interview series, we spoke with Dusty Majumdar and Chakra Chennubhotla, CEO and CTO at SpIntellx to understand how their technology is driving the next generation of computational pathology.
- SpIntellx focuses on computational systems biology as an extension of computational pathology to try to get at the “why” and help unravel the complexities of cancer. Tying pathological features to specific pathways implicated in disease has the ability to improve patient outcomes and drive the next generation of companion diagnostics.
- Biopharma is starting to venture into advanced DP algorithms, including computational systems pathology, as a means to better identify responders to therapy and characterize the progression of disease. These technologies also allow pharma to prioritize ROIs for their downstream genomics and transcriptomics studies by identifying “microdomains” in the tumor microenvironment.
- “Explainable” AI is key for addressing pathologists' concerns with black box machine learning algorithms. The language that pathologists already know and use on a daily basis is a powerful tool; when the AI is able to communicate in those same terms, it can ease the path to adoption of AI-based computational pathology tools.
Could you both share a bit of your background as it relates to AI and spatial biology?
I’m the CEO of SpIntellx. I have over two decades of leadership experience in healthcare and biopharma. In the last 10 years, I have primarily focused on oncology and deployments of AI solutions across clinical decision support, clinical trials, and drug discovery. I have worked at companies like GE Healthcare, Exact Sciences, and IBM in their AI efforts in a range of leadership positions. I also spent time at ASCO as an executive consultant, helping bring CancerLinQ to the market. Also, I’ve been involved in genomics and transcriptomics in cancer since the early days. Now at SpIntellx, we certainly have been delving into the various aspects of spatial biology.
I’m the co-founder, chief technology officer and chief of AI at SpIntellx. I started as an electrical engineer by training, then I studied computer science in graduate school at University of Toronto, where I did computer vision, machine learning and AI. From there, I moved into computational biology and did a lot of work on computational biophysics, understanding intrinsic protein motions and how it impacts drug discovery. I then started working in computational pathology, which I realized should really be computational and systems pathology. All of the work I did there was pushing AI to be more pathologist and oncologist friendly. SpIntellx was a spin-off from the University of Pittsburgh. I left my tenured faculty position at school to be at SpIntellx full-time.
How would you characterize the utilization of AI-based tools in biopharma and health systems today?
We are making strides every day in both biopharma and health systems. AI is making a difference in many workflow related solutions across radiology and cardiology and it is beginning to creep into pathology as well. In biopharma, AI is being deployed in clinical trials, as we speak, to select patients for clinical trials. It's also making its way into discovery in biopharma, but that's going to take a little bit more time.
One of the areas where I feel AI can make a huge difference – and you're just beginning to see this – is the unraveling of the heterogeneity of the microenvironment of the tumor. Today, we understand less than 5% of the hidden circuitry of the tumor microenvironment. It's this heterogeneity in the tumor microenvironment that fuels resistance in patients with advanced cancer. We feel that AI, especially Explainable AI, which is an evolution of machine learning, is getting to the bottom of the “why” behind the disease, what we call “computational systems pathology”. This will be instrumental in establishing the causal relationships, which today are largely not understood in complex diseases, like cancer. That's where I feel that AI is headed down the line.
To explain where computational systems pathology fits in the digital pathology ecosystem, you have to go back to the origin of these words. Digital pathology is where the glass slide becomes a digital image. Computational pathology is when you can run an algorithm on the digital image to do tasks, such as counting of cells, in an automated way. Computational systems pathology is when you are asking a deeper question of why. Why am I seeing this in the tissue? Why should this be a marker of disease progression?
So, the question of “Why?” really comes from thinking about the systems aspect. What I learned from pathologists is that when they are looking at the tissue, there are three questions that they're always asking: “What might have happened before? What is it that I'm seeing now? What will happen next?”. They're always asking the question of how the structure of this tissue is going to change. Now you have technologies that allow you to probe in much greater depth such as multiplexing spatial proteomics and genomics technologies. Hence you can ask the question, “What pathways might be indicated in the formation or progress of disease?”. This line of thinking led us to formulate and discover microdomains in the tissue samples and then understand intra-tumor heterogeneity from this idea of microdomains. I think biopharma appreciates that in a serious way. Explainable AI, where AI meets systems biology, will be the key for making further progress.
If you think of a natural progression from digital pathology to computational pathology to computational systems pathology, do you see computational systems pathology as the final frontier? Or do you have a sense of what else is to come as the landscape evolves?
I think the computational systems pathology is a natural evolution of where computational pathology is today. As we get more into understanding the network biology from multiplex pathology images, the next obvious step would be to come up with digital twins of the tumor microenvironment, for example, that can lead to in silico trial arms simulating the virtual patient. If you can use the knowledge that we're gleaning from computational systems pathology on the network biology, we might be able to ultimately try virtual interventions on the patient.
This is all really related to the pain points we have in the clinical trials arena today. I'll expand a little bit on that because I think that probably should come first in terms of why we're doing this. It takes close to $1.8 billion to develop a drug; it takes 10 to 15 years to bring the drug to the market, especially in oncology; and 15% of phase one drugs make it to the market today in cancer. For the last 10 years, the survival rate for patients with advanced cancers with immunotherapy has pretty much stayed stagnant at 24%. If you're going to make a dent in all of this, we'll have to do trials faster, more economically, and have higher hit rates of the targets that we started with.
Ultimately, if you can come up with a digital twin of the tumor microenvironment with the multimodal data that we are bringing together from pathology, genomics, transcriptomics, lab data, EMR data, that's where this is headed. That's where I really believe it needs to go to resolve the huge pain points today that exist in the discovery and clinical trial continuum.
When you think about this, over what timeline do you expect this to have an impact on patients? We're just seeing digital pathology, let alone computational and computational systems pathology, beginning to gain a foothold in the clinical space (though it's a bit further along in the research and pharma space).
In 2017, the FDA approved the scanner technology, and there are now multiple companies’ scanners in the market and that has opened up the floodgates. In the last couple of years, we saw a huge uptick in biopharma’s use of algorithms to analyze and predict who will respond to therapy or the expected prognosis. I think that this work comes with a timeline. There are many studies where companies like us are participating in phase II and phase III trials. And folks like us can help biopharma drive companion diagnostic tests. In the next five years, I would expect to see a huge adoption of computational pathology algorithms, along with systems biology.
In terms of biopharma, we have a customer today that is using our solution to identify subpopulations of response with immunotherapy to certain cancers. Once you identify the subpopulations in response to the clinical trial, the next logical step is when the drug is in the market, what is the companion diagnostic test that goes along with the drug? So, we're in conversations with our customers around that as well.
The vision that I talked about of in-silico simulation or in-silico clinical trial arms, that's the convergence of several different things coming together. Not only just computational pathology and the network biology that we can extract from that, but also advancements in spatial biology, single cell genomics, transcriptomics, and integration of other multimodal data, such as diagnostic imaging data, lab data, EMR data. All of this needs to come together. It's a matter of constructing networks of disease. I think for all of this, “Explainable AI” or “Causal AI” is going to be critical. If you are going to establish a network of disease, which is going to be fairly complicated for cancer, you need to know what drives what. You need to know which genomic mutations are driving which gene expressions and which gene expressions are driving which proteins. I think it's critical that Explainable AI play a vital role in all of this.
Absolutely. Do you have any specific examples of how biopharma is using computational systems pathology to support translational or clinical trial research efforts today?
With our approach at SpIntellx, using advanced spatial analytics and Explainable AI, we are able to get into cell-cell communication. At this point, I think we are the only company that does that. Through the understanding of cell-cell communication, we also endeavor to have some glimpse at the network biology. Moreover, we are also discovering tumor microdomains that drive disease progression. Once you do that, you are able to, for example, predict the progression of disease way better than you could with simple spatial analytics and rudimentary genomic exploration that many companies use today.
We published studies in Nature Communications and Cell Reports predicting the five-year risk of recurrence of colorectal cancer patients from colon cancer samples across various stages. We had 400+ patient cohorts across stages I, II and III. This was a retrospective study, and we were able to predict, with 93% accuracy, the progression of colorectal cancer. Once you are able to do that - identify which populations will progress after treatment and which population where cancer will be arrested - you're able to identify clinical trial subpopulations much more effectively.
We have a very logical framework; there's a discovery phase, a translational phase, and then the clinical phase. Typically, in the discovery phase, you image tissue with lots of biomarkers (perhaps more than what you think is necessary) and then run the advanced spatial analytics.
Because we take an unbiased approach, we can point out the unexpected interactions between pathways that have emerged in this tissue. The idea that these networks can change depending on where you're looking at the tissue is really another way of saying there's heterogeneity. That allows you to now start with any number of biomarkers and construct this network and make the network suggest to you what additional biomarkers you might want to bring in. By the time it goes to a clinical application, then you say, “No, we have found the minimum number of biomarkers.” We actually have done that in a GI study. We have shown that you can actually reduce the number of biomarkers by 40% and increase the accuracy by ~30%, just by being very clever about how you think about the emergent network biology.
How do you see computational pathology increasing the efficiency of other types of analysis?
In the spatial biology arena, all these companies talk about transcriptomics and trying to predict the progression of cancer or recommend treatments based on that. We have a little different approach. We feel that this unguided spatial exploration often leads to a lot of waste of money, waste of time. I feel that there is an opportunity to do more guided exploration. We believe that microdomains could be a fantastic guiding tool to explore genomics and transcriptomics. In other words, instead of looking willy nilly across the tumor microenvironment, why not use these microdomains as guides, which are these repeated patterns of clusters of cells, which we have seen replicated across the tumor microenvironment. If you look in there, you see actionable mutations and actionable genomic expressions that accurately determine the disease progression and what treatments will and won’t work; and we’re extracting network biology from that.
There's increasing support for microdomains with a number of independent studies appearing recently. In some of these papers, a pathologist looks at the tissue and realizes that there is a very special histological pattern that they are observing. They then decided to do a microdissection of the tissue and determine where to do the genomics study based on those patterns.
Cancer cells typically engage in multiple tasks, which is to survive and flourish, and for which they want to manipulate the environment around them, so they don't get sucked in by the immune cells. What we realized is that some of the microdomains we found in the study of colorectal cancer are tumor promoting and some of them are tumor arresting. The whole idea of computational systems pathology is that we will not only identify these microdomains automatically, but also say why these microdomains come into existence. We want to be able to say that this pathway has an additive effect, and this pathway here has a suppressive effect on this other pathway, and so on. This opens up an entirely new way of thinking because if you know that a patient is not responding to therapy, and you know that it involves a particular kind of interaction of pathways, and you know which microdomains are tumor arresting and what the underlying pathway interactions are, you could change the microenvironment for this patient so that all the microdomains that you find are now tumor arresting, if you had the right combination of drugs. Also, once you can find these microdomains, you can perform microdomain-specific genomics and transcriptomics in those sections because you increase your chance of finding actionable insights.
There is a term that you have been mentioning throughout our conversation: that the AI is “explainable”. This addresses something we've heard pushback from in the past, that pathologists don't want to use blackbox algorithms. They don't trust it, they're not familiar with it, and there's going to be a huge hurdle to adoption in the clinical setting. How does Explainable AI address these concerns?
I've been involved in AI ever since the inception of AI in healthcare and life sciences, starting with GE and then going into IBM and a plethora of other companies. Chakra did his PhD at the University of Toronto, which has been a premier institute for computer vision, machine learning including deep learning and AI. So, we're both very familiar with how this field has evolved.
Not just in pathology, but across the board, there have been plenty of companies coming up with blackbox deep learning AI over the last decade or so. Many of the radiology solutions have achieved FDA approval five or six years ago. If you look at how many of these companies exist now, a lot of them are no longer there (either folded or acquired). The reason for that is the lack of credibility that these outfits generated with clinicians as a result of the ‘black-box’ type solution. Explainability is important because we want to know why. Why is this cancerous? Pathologists have expressed very early in their journey that explainability is critical.
With deep learning, you can probably identify subpopulations of disease progression; you can probably identify which patients will respond and which will not respond to an intervention just using a blackbox convolutional neural network. What we will not be able to do is answer questions related to why this is happening or what causes a patient not to respond, because you have no insight into the network of disease. That is what Explainable AI allows us to do.
What we do is completely hypothesis free; we let the data provide us the answers organically, instead of imposing a hypothesis on it and then seeing if the data satisfies that hypothesis or not. In the area of spatial biology, the fact that we do unbiased spatial omics sets us apart from the others out there who essentially do transcriptomics but don't have this access to microdomains.
Something that really stood out for me is that a pathologist’s reading of a slide is the ground truth, the final thing, right? But then, when these pathologists have a machine learning software that tells them something is atypical and does not explain why, especially in cases where the discordance between the colleagues is so high, it is not helpful! It is in these cases with a huge amount of discordance, where the AI has to be most helpful. The only way it will be helpful is if you can explain the recommendations made by the AI software in a language that the pathologist understands. And in fact, when we did market research speaking to pathologists, it was very clear that explainability is going to be the game changer in their adoption.
We know that the knowledge that pathologists have accumulated in describing any tissue is very important. For example, how they describe the shape of a lumen, or the tissue architecture, and so forth. What we decided to do is to actually build analytical or quantitative models for all the qualitative descriptions or vocabulary that they use. Now, when you give our software a new piece of tissue to analyze, we can use the same language that they would have used to describe this tissue. We give them the option to either agree or disagree with the AI, and thus help the AI systems to learn. This is very, very different from the deep learning blackbox algorithms which think that they are ‘explaining’ by showing a heat map of the pixels implicated in the decision made by the network. But if you ask a pathologist, those heat maps are like ink blots, they don't show you “Why?”. Having that language built into the analytical models allows the AI to use their same language back and explain results back to them in terms they have seen. We built our multiplex AI software to be about network biology because everybody understands pathways. We show them the pathways and how the pathways interact. We provide an explanation as to why. This is encouraging trust and transparency and reducing the bias of AI.
This makes sense. As you go into higher plex, you're entering a territory where you don’t always have language to describe what you’re seeing because it's something that can't be done manually. You have to bring the pathways into it because while they maybe don't have familiarity with the technology, they do have this systems biology background.
We touched on partnering really briefly earlier; I'd love to hear your thoughts about how important partnering is to build out this spatial biology and computational pathology ecosystem?
Our target, right from the beginning, was to be at the back end, doing spatial analytics, building Explainable AI. But it is very clear that you need a digital technology workflow solution provider to reach the customer. That's how we entered into a collaboration with Inspirata. It makes a lot of sense for us to focus on what we are good at, which is the Explainable AI and unbiased spatial analytics, and Inspirata to focus on providing the digital pathology workflow solution.
Our partnership with iCura is our newest partnership. They are experts at understanding multiplex technology and working with biopharma. We believe that through partnering with iCura, we will be able to really implement this microdomain-based discovery approach. We believe that more targeted explorations of genomics and transcriptomics within the microdomains will enable biopharma to have better optimized trials and more efficiently offer companion diagnostic tests.
Thank you so much for your insights!