‍

Key Takeaways:

• Generative and agentic AI offer exciting opportunities for digital pathology and clinical research, with the potential to streamline pathologist workloads and accelerate biomarker research.

• Regulatory approval and financial reimbursement are top barriers to clinical deployment of agentic AI today. Non-explainable biomarkers validated by therapy responses could be transformative.

• Data diversity is more valuable than data quantity for foundational models; AI-based image analysis could be a strong alternative to traditional computational models for rare indications with less available data.

Hi Jill! To get us started, I'd love to hear about your background and what you've done in your career that has brought you to what you're working on now.

I have a PhD in molecular medicine and molecular pathology in veterinary medicine. I managed a veterinary diagnostic lab for several years before making a significant career shift. I focused on assisting diagnostic companies in advancing various molecular tests into clinical settings. Predominantly, I worked with technologies like PCR and the early stages of next-generation sequencing, at a time when many doubted their clinical viability. I dedicated myself to identifying early adopters and helping to transform that space.

I am now at Modella AI. We are very focused on medical imaging and generative and agentic AI, with our initial efforts largely concentrated in pathology. We believe there are significant opportunities to leverage generative and agentic AI in clinical use cases, as well as in biomedical research and computational pathology.

Could you start by describing what you mean by generative and agentic AI and what they mean within the context of imaging, pathology, and clinical research?

Generative AI is a type of artificial intelligence that creates content. In pathology, for example, users can highlight a region of interest within an image and ask the AI to describe what it sees. The AI then generates medical text detailing the morphology. Conversely, users can prompt the AI to identify areas where a tumor is present. This functionality varies based on the training of multimodal foundation models, which use image caption prompts and undergo extensive pretraining.

Agentic AI, on the other hand, functions as an intelligent agent capable of reasoning, performing tasks, and interacting with humans. It can devise research plans and recommend steps for training AI classifiers. Initially focused on computational pathology research, we have recently developed agentic AI for clinical use. After a diagnosis, this AI can autonomously consult medical guidelines to determine necessary tests or identify relevant clinical trials for patients, guiding the next steps in their healthcare journey.

‍

What role do you think generative and agentic AI will play in the clinical setting? What are some of the biggest barriers to their clinical deployment?

I've been thinking a lot about how historically, many companies have focused on one disease at a time due to the limitations of AI and data access. This approach has been challenging for academics and for-profit organizations to adopt because it's often seen as a "nice to have" rather than a necessity. Pathologists can perform many of these tasks themselves, so the return on investment must be considered.

In my view, a minimal viable product in this space should be pan-disease. There's a significant shortage of pathologists, and their workloads are increasing. Many tasks in anatomical pathology are mundane and could potentially be automated by AI. A product that addresses all pathological diseases, even beyond oncology to include infectious and inflammatory diseases, could be highly beneficial. Most cases or slides are often normal with no disease present. A tool that allows pathologists to focus primarily on suspicious disease areas across various diseases could greatly impact their workflow. They could spend most of their time on disease-positive images and quickly handle the negative cases.

Another barrier is the financial aspect—how these tools will be reimbursed, who will pay for them, and how much they will pay. No group wants to receive less reimbursement for using these tools, but this remains a significant challenge.

It's also important for systems to handle various types of images, including H&E and IHC, which can be quite subjective. The future looks promising with developments like quantitative continuous scoring and antibody drug conjugates. These advancements could be transformative, providing a strong return on investment for going digital, as they might be necessary for determining patient treatment options.

How are these tools validated by the FDA? The Breakthrough Device Designation (BDD) that you received suggests a willingness from the FDA to collaborate, but there still isn't a clear regulatory pathway. What do you think this process could look like?

Yes, that's a very good and challenging question. It's outstanding that the FDA has approved both the Paige Prostate and the Ibex Prostate tests. There is definitely some predicate success here, with the ability to analytically and clinically validate on a per disease basis.

When you move beyond using these as clinical assistive tools, where the pathologist still signs out, and start to explore generative and eventually agentic applications, a similar process will likely be necessary. This might involve more prospective work and a significant study to prove that the underlying large language model is adequately trained. This includes extensive human preference testing before reaching a design lock.

Continuous training presents a regulatory challenge, as agencies typically require that reagents or kits be GMP manufactured and locked. Any changes would necessitate revalidation. There are examples of how other technologies have navigated these challenges, but it will require significant education and alignment from the beginning to establish what needs to be proven and to confirm regulatory agreement.

What do you think are the most exciting promises for generative AI within preclinical and translational into clinical research, that are doing things traditional computational models aren't well suited for?

I think it starts with the use of foundation models, particularly multimodal foundation models. These models are heavily pretrained with various data types such as H&E and IHC. They might also integrate a vision and language encoder, or combine H&E with whole transcriptome or extensive IHC data. One of the advantages is that when you extract embeddings using a foundation model, you can develop an MIL classifier with much less data. Historically, MIL without foundation models required so much data that the size of a phase one clinical trial was too small to confidently detect any significant signals from the images. Now, with these models, you can analyze a phase one trial and extract meaningful insights, training AI MIL classifiers from that data.

Additionally, having access to generative and agentic tools can allow this process to occur autonomously in the background, supporting various downstream tasks. For example, determining if there is enough tumor in a biopsy for use in a clinical trial can be automatically flagged as soon as the biopsy is uploaded. This addresses the common discrepancy between what is requested and what is actually received in clinical protocols.

What do you think is the opportunity to utilize these algorithms in rare indications with smaller patient populations and less available data?

One key lesson is that the diversity of data in a foundation model is more crucial than the quantity of data. For instance, having numerous cases of breast cancer with the same morphology won't significantly enhance the model's performance. However, including rare morphologies can lead to foundation models that outperform larger models. Therefore, ensuring the inclusion of diverse data is incredibly important for the effectiveness of downstream tasks and few-shot learning capabilities of these tools.

One thing that is important, is the need to evaluate how those models and tools work in specific ethnicities. For instance, the FDA has certain requirements for what needs to be represented in your analytical and clinical validation. The same is true in Europe and essentially across the world, including the Asia Pacific region. I don't think anyone should assume that their models will always work perfectly. You probably need to fine-tune the model with a representation of what you're trying to achieve in terms of those claims. We have quite a few opportunities to collaborate with different countries to start to see how much, if any, additional training we would need, but we are still in the early stages there.

How do you drive the adoption of these novel tools among pathologists?

I think it's important to build tools that address the pain points pathologists face. Pathologists need to be involved in the process, using tools that enhance their efficiency and allow them to focus on critical aspects of their work. Initially, I anticipated negative reactions, but that hasn't been the case. I've received many positive messages and emails.

It's also important to consider the needs of future pathologists. Many new pathologists seek additional education and prefer having another year with a supervising pathologist to review their work. A pathology-trained copilot could provide this support, allowing them to consult with peers while continuously educating themselves with these tools.

Slide search capabilities where generative tools can search an entire archive, not just within a single case, are exciting. These tools could identify the top cases with specific morphologies or create heat maps to highlight particular cells in an image.

Providing these tools can mitigate potential pushback by demonstrating their value in improving diagnostic efficiency. Groups that adopt these technologies may see increased volumes due to faster and more efficient diagnoses, which is crucial for treating patients. Additionally, accessibility to residents, fellows, and others in training is important to get them excited about the possibilities these technologies offer for their future practice.

Shifting gears a bit, let's consider the overall landscape, which is quite fragmented. We're beginning to see some early CDx partnerships that are likely to shape the future of this landscape. How do you see this evolving?

It is very fragmented. At the front end, you have the scanners and hardware, primarily manufactured by large diagnostic companies. Then, there are the IMS and viewers, which in some cases are provided by the scanner companies themselves, or by other companies focused solely on that aspect. Additionally, there are companies building AI models and companies like ours that are more generative and agentic. It's important, especially for clinical sign-out, as more sites are going digital, which is outstanding, though it has taken longer than most had hoped due to various reasons. It's crucial to be agnostic and understand where you fit in the full workflow.

I believe we will see more strategic partnering within this landscape and expect mergers and acquisitions in this space. Regarding foundation models, if you have a lot of data, you can build them. This raises the question of whether to build your own foundation model or use one that's available. As we move towards regulated products, we might have to be less agnostic since regulators will require specific certifications like 510(k) or IVDR. From an R&D perspective, flexibility is key as it's still to be determined which scanners and IMS will dominate. The market will likely remain segmented, which could be challenging.

Looking forward, what key events, milestones, or technologies do you see revolutionizing this space?

I'm particularly excited about the development of multimodal fusion classifiers, where we can use various approaches to analyze not only a patient's pathology and outcome data but also radiology and other clinical inputs. This comprehensive analysis could transform precision medicine by building models that predict conditions much earlier. For instance, despite standard screening ages, we could predict higher risks based on a patient's comprehensive data profile, allowing for more proactive and personalized care.

Another exciting aspect is translating technologies like multiplexed IF and spatial transcriptomics. Diseases are incredibly diverse, and sometimes it just comes down to which part of the tissue was sampled. Getting the diagnosis right can truly enable the treatment of the patient for the best possible outcome.

Personally, my motivation is the impact on patients' lives. If my work can help extend someone's life, that is incredibly rewarding, whether the impact is direct or indirect through the products and technologies I work on.

Thank you so much, Jill. I really appreciate your time today.

‍

‍