MM: Hi Joe, thanks for taking the time today to discuss recent innovation and approaches in AI-powered digital pathology, especially in the context of your experience at Bristol Myers Squibb. To start, can you talk about your role at BMS, the role of your organization, and how digital pathology fits into that organization?
JS: Sure, I lead a computational research organization that’s called Translational Bioinformatics. We are a team with diverse computational expertise, and we partner with the translational medicine organization, as well as clinical development teams to perform exploratory biomarker analyses to understand disease, disease variability, mechanisms of our drugs, etc., ultimately to identify which patients are most likely to benefit from medicines in our portfolio.
Within this organization are teams focused on specific therapeutic areas, like solid tumor oncology. We have also built teams with deep technical expertise, including emerging areas such as digital pathology [DP]. Our DP team establishes the infrastructure to handle, process, and analyze DP images and drives the development and deployment of cutting-edge AI-based techniques for automated image analysis.
MM: Does your team work across the entire development process or on specific aspects?
JS: We are part of a centralized organization that works across the entire drug development lifecycle. While our team focuses primarily on translational research for mid- to late-stage assets, we also take signals or insights generated from those analyses and feed those insights back into earlier parts of the pipeline. We call this reverse translation.
MM: BMS has clearly established their expertise in digital pathology. Let’s take a step back – how would you describe or define AI-powered digital pathology?
JS: AI applications in DP involve training computers and algorithms to analyze scans of pathology slides in a manner consistent with what human pathologists would do. We can then deploy these algorithms at a scale not always possible with humans.AI computational techniques are often more sensitive, more accurate, more precise, and more consistent than read-outs from human pathologists.
MM: One question we hear a lot is, if we train and validate algorithms using human pathologists, how or why would AI-based algorithms perform better than those humans?
JS: We collaborate very closely with pathologists to develop data sets that have hundreds of slide images and human annotations. We then use images annotated by multiple pathologists, rather than any one individual pathologist input, to train algorithms. This approach mitigates intra-pathologist variability. Algorithms trained in this manner tend to perform better than a single reader – and because they are computer driven, they are consistent and reproducible. We also look for opportunities to contextualize and validate AI approaches using orthogonal information such as genomic data and retrospective or prospective clinical outcome data.
MM: What value have you realized from utilizing AI approaches in pathology?
JS: One success story has been digital PD-L1 read-outs. PD-L1 is an established biomarker for many indications to identify patients likely to respond to I-O therapy. Digital PD-L1 read-outs have better overall performance and consistency than manual read-outs. Consequently, digital read-outs tend to identify more PD-L1 positive patients at any given threshold (e.g., positivity at 1%, 5%, etc.) than manual scoring. Now, what is really exciting is that we see similar rates of clinical response in PD-L1 positive patients identified by digital and manual readouts.1 This suggests that digital pathology readouts have the potential to identify more patients who are likely to benefit from specific therapies.
MM: This is fantastic and such a great result for patients. PD-L1 is an interesting example of improving upon the performance of manual pathology. What are emerging use cases of DP that go beyond traditional pathology?
JS: With AI analysis of slide images, we can extract many more features than what’s feasible by human pathologists. We can detect, characterize, and quantify tumor cells, normal cells, the boundaries between those cells, the expression or prevalence of biomarkers, etc. But what’s unique to pathology is we also have spatial information. We can interrogate various signals along with their spatial relationships to identify more complex biological patterns that may be relevant to disease status or patient response.
Take tumor classification as an example. We classify “hot” tumors as those inflamed by many immune cells, “cold” tumors as those with very few immune cells present, and there are versions in between – one of which is an “excluded” condition, where immune cells are present but have not invaded the tumor itself. This is a pattern uniquely identifiable though spatial computational analysis, and in partnership with PathAI we have built automated and scalable approaches to characterize and quantify the features necessary to classify a tumor as hot, cold, or excluded.2
This example shows us quantifying a spatial relationship between two categories of cell types – immune cells versus tumor cells. But with AI methods we can go even further and analyze multiple categories of spatial relationships simultaneously and rapidly.
One such use case examines the spatial relationships between receptors and ligands. LAG3 is a protein that tends to be expressed on the surface of exhausted T cells. It interacts with major histocompatibility complex (MHC) molecules expressed on tumor cells or antigen-presenting cells (APCs). Because we can detect and quantify so much information from the slide image, we can detect, quantify, and interrogate the spatial relationships between LAG3 and MHC expressing cells, and identify patterns that are linked to disease state or patient response to therapy. This integrated analysis of molecular, cellular, and spatial information allows us to explore axes of biology that are not accessible through other means.
MM: There is certainly opportunity for AI and DP in the translational space. How do you move these insights further down the development process, ultimately to positively impact patients?
JS: Yes, up until recently, we have been using AI computational approaches in a retrospective and exploratory space. We recognize the opportunity to move these deployments and impact both upstream and downstream in the development cycle. This is an important pivot.
An example of downstream impact is if we could deploy digital PD-L1 read-outs as a complementary or companion diagnostic. This could be game-changing in providing consistent and improved identification of patients eligible for treatment.
MM: For emerging approaches like DP and AI, how do you develop and build these capabilities?
JS: The problems we are attacking with DP approaches are really complicated. To be successful, we need to put together teams that have the diversity of expertise across computer science and machine learning, the disease area, clinical development, and even regulatory. We need to build those teams with complementary external expertise that can take the form in relationships with technology-focused companies, such as PathAI, or more traditional contract research organizations (CROs) to deploy these technologies.
MM: Yes, DP reminds me of the earlier days of Next Generation Sequencing (NGS) in that we need partner across the ecosystem, but you still need to have enough internal expertise to know what you are doing.
JS: Exactly, you need expertise and knowledge internally to make good choices about strategy and external partnerships.
Another similarity between NGS and DP is they both can generate oceans of data. This enables a few approaches. In some cases, you are casting a wide net to screen and identify novel biological signals that no one has used before. In other cases, you’re testing specific biological hypotheses against the data sets generated by NGS and DP. The DP community is learning how to balance between these two approaches, and when to pivot from one to another. If you’re doing a screening exercise, how do you recognize a hot lead that has the potential for broad impact and how do you transition that to a more focused, hypothesis-driven approach?
MM: Right, and then get it validated and evaluated across cohorts. Joe, thanks for the fascinating discussion and it’s exciting to both see and be part of AI evolution in digital pathology.