The technology to determine whether to diagnose someone with cancer from single or multiplexed images and data types has made its way to the clinic. Pathologists and other disease experts require transparency and trust in applying computational and systems pathology, not black-box deep learning AI with biased heatmaps as substitutes for explanations. So, if a computer spits out an answer, is a clinician just supposed to accept it and prognosticate accordingly? What would a clinician need to be assured of the AI’s deduction?
Explainable AI is a set of tools and frameworks to help understand and interpret predictions made by machine learning models. This artificial intelligence has the potential to deconstruct why a particular recommendation is made in an understandable language for clinicians. Importantly, explainable AI can build trust and confidence in decisions recommended by the algorithms, guiding pathologists and disease experts who remain in complete control to make the final decisions.
Spun out of the University of Pittsburgh School of Medicine, SpIntellx guides clinicians and researchers with proprietary unbiased spatial analytics and explainable AI. The company aims to transform computational and systems pathology by offering software as a service (SaaS) for precision pathology applications harnessing unbiased spatial analytics. SpIntellx’s explainable AI also identifies novel targets, biomarkers, and previously unknown cell types or states that can aid drug discovery and advance companion diagnostics for radically improving prediction accuracies through deep insights into biological mechanisms of action. These explainable AI tools can optimize clinical trials for precision patient stratification and personalize therapeutic options for selecting optimal therapeutics based on insights into probable patient outcomes.
GEN Edge sat down with CEO Dusty Majumdar, PhD, and Chakra Chennubhotla, PhD, president, co-founder, CTO, and chief of AI, to learn about some of SpIntellx’s recent partnerships and the reach explainable AI has from R&D through the clinical journey.
GEN Edge: What was the founding mission behind SpIntellx?
Majumdar: As we look at the rapidly evolving space of spatial biology, there’s a tipping point that we are finding now in the spatial biology arena where a combination of transcriptomics and proteomics and other modalities and genomics are coming to fruition. The mission of SpIntellx is to ensure that with all this multimodal data coming together, we are getting actionable insights out of all the different tests and data sets that are coming out.
That’s a challenge because often you have to search for a needle in a haystack, and you have so much data, but you don’t know what to do with the data. Our approach of using microdomains in pathology images as guides to where to look for actionable information with transcriptomics and genomics differentiates us from everyone else. We call this guided exploration of transcriptomics and genomics within microdomains, where you can get insights that nobody else can get at this point.
Our mission and purpose are to unravel the heterogeneity in a tumor microenvironment. Today, the survival rate is still around 25% for the best immunotherapy treatment. In addition, less than 5% of the hidden circuitry of the tumor microenvironment is understood today. We believe that unraveling the heterogeneity and understanding how these microdomains and network biology works in the environment will improve survival rates of late-stage cancer.
Chennubhotla: SpIntellx is a spinoff from the University of Pittsburgh School of Medicine. I was a tenured faculty member of the computational and systems biology department, where I met D. Lansing Taylor. My background was in computer science with a focus on computer vision, machine learning, and AI. Lans had a cell biology and fluorescence imaging background and was coming back to academia with a long stint in industry building high-content screening systems.
One of the research areas that we were both interested in was understanding spatial tumor biology from pathology samples with spatial proteomics techniques. Together, we ran a large group of computational pathology researchers with support from federal grants including NCI. It was clear to us that there will be an explosion in the development of platforms for imaging spatial biology in the coming decade, but instead of worrying about reagent design and building another imaging platform, we decided to focus on the back-end AI and unbiased spatial analytics to be able to extract actionable insights from the massive datasets that the spatial biology platforms generate.
With the advancements we made, we were in a very good place to spin off SpIntellx as a company. Given the massive opportunity we have in this space, I’ve decided to move to the company full-time.
GEN Edge: Can you explain how SpIntellx’s precision pathology platform works?
Chennubhotla: Tumors are dynamic ecosystems. You want to let the imaging data tell you what the emergent biology is. All the algorithms we designed were shaped to extract emergent biology. Hence, we have developed an unbiased, hypothesis-free, data-driven spatial analysis. The explainable AI part came from our conversations with pathologists, clinicians, and oncologists because their comfort level with the AI system is higher if only the AI system could explain how it’s making a recommendation.
When we thought about starting this company, explainable AI (xAI) was right in the middle. We received the first explainable AI in computational pathology patent. The way our xAI interfaces are built is that there’s a ‘Why’ button at one place, wherein the end user can actually click on this button to ask the system for an explanation. The software pipeline remains in control of the end-user, they can either approve that recommendation or disagree with the recommendation so that they can let our xAI algorithms continue to learn what and how they perceive.
The actual mechanics of doing this is being very aware of the tissue architecture and the tissue heterogeneity in quantitative terms and speaking about these terms in a language that the clinicians, the pathologists, and the oncologist understand. That’s the core of our xAI technology.
Majumdar: The critical thing that AI companies have been missing out in the last decade is the clarity, explainability, and causality behind these algorithms. We see a lot of the radiology AI companies that came up five years ago are going belly up because, despite FDA approvals, they don’t have explainability—it’s a deep learning black box. Clinicians don’t trust the recommendations and are not really using them in their practices.
Our differentiation goes into three different levels. One is functional cell phenotyping, where we identify all the cell types and states by considering the spatial context around each cell. The biggest difference with our phenotyping approach is that we do not manually threshold the biomarker signals to define our cell phenotypes, our method is data-driven, hence hypothesis-free. The next thing is microdomain discovery, where, organically, we let the data discover repeated clusters of cells as microdomains across the tissue on a whole slide. Last, is the microdomain-specific network biology.
The fact that we can identify unique pathways and network biology from a multiplexed pathology image is something that I was very interested in when I first came into the company. I’ve seen it happen with various clinical trials and customers we have worked with. It’s very powerful, and it’s a straightforward tool. The lower overall cost of doing it from a pathology slide rather than using expensive equipment to delve into different omics is solid in terms of our value proposition.
GEN Edge: Tell us about your product pipeline and revenue stream.
Majumdar: Our target customers are biopharma and hospital systems. It’s bifurcated. We believe our first customers will be biopharma, not only in the clinical trial development space but also in the discovery and development of companion diagnostic tests.
We have two offerings: TumorMapr™ and HistoMapr™. The TumorMapr offering is primarily directed towards clinical trial development in identifying subpopulations of response. We use multiplex images for the TumorMapr.
Our first customer—not yet public—is a large pharma company who wants us to identify subpopulations of response in an immunotherapy clinical trial that they are undergoing right now on lymphoma. That’s a typical customer of ours where they want to identify the population that will respond to a drug versus populations that would not be based on our microdomain discovery from tissue samples. This customer also wants to develop a companion diagnostic test for the drug they’re working on.
We also have a few customers that want to improve their workflow. These include our partner, CellNetix, an extensive pathology network on the west coast. For them, it’s about efficiency and better concordance among different pathologists and ultimately lowering costs using automated analysis.
Chennubhotla: The traditional pathology practice is to take a biological tissue sample, put it on the slide, stain it with hematoxylin and eosin, and study the morphology of cells under a microscope for diagnosis. Now, you take a digital image of this slide using transmitted light. Our HistoMapr software is for these transmitted light applications. In the last 5–6 years, because of the interest in spatial biology, you have all these platforms which are doing imaging of the antibodies. So you have 5 to more than 500 biomarkers. TumorMapr is for the multiplex and highly-multiplex, either immunofluorescence or mass spectrometry. We are agnostic to any of the platforms.
Typically, biopharma customers take one tissue section for a digital image with transmitted light and then the other section for the multiplex. In this case, you can use both pieces of software. The core technology is the same, but the input is different. The knowledge you extract is slightly different, but it’s all integrated. Then, the customer uploads their data sets to the cloud, where we run our platforms and then deliver the knowledge background.
GEN Edge: What situations benefit from using SpIntellx’s explainable AI?
Chennubhotla: Because of explainable AI, the concordance between clinicians will go up. In our pilot studies, we have already shown that. With CellNetix, we are doing an extensive 10,000 patient case study to demonstrate that the inter pathology concordance improves because of explainable AI. That is a massive push for why you want to use explainability in the workflow.
The other concept we have a handle on is the notion of microdomains. These are distinct collections of the immune, stromal, and tumor cells. We are asking what is common across cancers about these microdomains. What would a library of these microdomains look like? That would be the true spatial knowledge universal to cancer.
Cell phenotypes and states are on a continuum. There’s evidence now that combining these pieces of information lets you dive deeper between proteomics and single-cell genomics. They are extending into these different cancers, finding those phenotypes that are universal whether they’re in the brain or the liver. The explainable AI side addresses the pathways and interactions coming into play to generate these microdomains.
For the discovery phase, wherein people are willing to put hundreds of biomarkers on the tissue. That’s not going to be practical if you want to go clinical. You would like it to be a very small number of biomarkers. Our tools can facilitate discovery among all these biomarkers which one is actionable. We have shown in studies explicitly that distinct spatial arrangements of biomarkers that one should be paying attention to. That’s the key insight—not the number of biomarkers per se but their spatial relationships. Of course, you need to have some, but spatial relationships have a lot more prognostic and predictive information.
GEN Edge: What are your major milestones and future trajectory plans?
Majumdar: We want to be the company that people think about when they think about explainable AI and spatial biology. There are many companies in that arena, but we want to be known for advanced spatial analytics and explainable AI that nobody else at this point really is focused on.
From a clinical trial perspective, we want to be the company that the CROs and biopharma companies want to work with. That’s our first target. We want to be in 5–10 clinical trials next year. Then, probably 100 by the year 2024–2025. We want to be the company that everyone thinks about and talks about when they think about incorporating explainability into their clinical trial development process. We also want to make forays into discovery and at least have a couple of companion diagnostics tests out by 2025 with biopharma, which takes a longer time with FDA approvals. Still, at least a couple is the aspiration by 2025–2026.
GEN Edge: How is SpIntellx growing, either internally or via partnerships?
Majumdar: We are looking at several partnerships. We announced a partnership with Inspirata that was all about reaching out to health systems. More recently, we announced a partnership with iCura—a CRO in the Philadelphia area. This is very interesting because they are working with biopharma on several clinical trials, including immunotherapy, to accelerate that penetration into clinical trial development. We feel that a partnership with a CRO like iCura would be beneficial.
Chennubhotla: We are actively executing on commercial contracts with biopharma and CROs. The initial contracts are all targeted to be around clinical trials, focused on identifying subpopulations of response to drugs. We are rapidly moving to identify drug targets with customers focused on drug discovery.
As I mentioned before, we are also actively working with our biopharma customers interested in building companion diagnostic tests, which we will develop in collaboration with the biopharma once the drug is on the market.