AI-Driven Biomarker Assessment Could Shift How Pathologists Work

In oncology, the path from drug discovery to clinical approval is fraught with challenges, and the reliance on traditional, manually-scored biomarkers is a significant contributor to trial failure.

While there are various categories of biomarkers, our focus is on IHC. This laboratory technique stains tissue samples to visualize specific antigens (proteins), and it remains the gold standard for cancer diagnosis, grading, and prognosis.

With the rise of antibody-drug conjugates—a class of targeted therapies that deliver cytotoxic payloads directly to cells expressing specific antigens—IHC has become more critical than ever. It is the primary tool used to determine antigen density, which directly dictates whether a drug will be effective for a specific patient.

However, the traditional method of scoring IHC stains presents a major hurdle: visual interpretation.

Pathologists are often tasked with classifying hundreds of thousands of cells into complex categories based on subtle color gradients. This process is inherently subjective. Studies consistently show high inter-observer variability (disagreement between different pathologists) and intra-observer variability (disagreement by the same pathologist at different times).

In the high-stakes environment of a clinical trial, this subjectivity introduces significant risk. If a biomarker is scored inconsistently, patients who might have benefited from a drug may be excluded, or the drug’s perceived efficacy may be skewed, ultimately leading to a failure to meet clinical endpoints.

AI-driven interpretation is fundamentally transforming the development and commercialization of IHC-based biomarkers in three distinct ways:

• Transcending human cognitive limits: The human eye and mind are naturally limited in their ability to quantify the subtle nuances of IHC stains. Historically, scoring systems were designed to be 'cognitively feasible' for pathologists—meaning they had to be simple enough for a human to estimate visually. AI removes these constraints, enabling more complex and granular scoring systems. A prime example is the Trop2 Quantitative Continuous Score (QCS), an AI-based scheme developed to predict treatment response in non-small cell lung cancer. QCS involves calculating complex ratios that are simply impossible for a human to perform manually, yet they provide potentially more accurate predictions of drug efficacy.
• Standardization across multi-site trials: Implementing an AI-based algorithm within a clinical trial ensures that biomarker scoring is synchronized across dozens of global sites. Whether the AI scores autonomously or assists a pathologist, it could provide a 'digital gold standard.' This could drastically reduce the noise generated by subjective interpretation, ensuring that the data used to evaluate a drug’s success is consistent and reproducible.
• Accelerating clinical rollout: Once a drug and its associated IHC test receive regulatory clearance, AI enables a rapid, standardized rollout into the clinical setting. Instead of retraining thousands of pathologists worldwide on a new, complex scoring manual, the AI algorithm could provide immediate, objective support. This level of standardization ensures that no matter where a patient is treated, they are correctly identified as a candidate for life-saving therapy.

To transition from a 'promising' research tool to a regulated medical device, AI-based scoring tools must undergo a rigorous, two-tiered validation process. This ensures the technology is not only accurate in a lab setting but also safe and effective for real-world clinical decision-making.

The two critical levels of validation include:

• Analytical validation: This is the foundation. It involves demonstrating that the AI can accurately and reproducibly identify the biomarker (in this case, HER2) across a wide range of variables. Developers must prove the algorithm's robustness across different tissue preparation methods, staining protocols, and digital scanners. For our HER2 tool, this included demonstrating high sensitivity and specificity—particularly in the challenging HER2-low and HER2-ultralow categories, where manual discordance is highest.
• Clinical validation: Here, we must show that the AI’s output aligns with—or exceeds—the 'ground truth' established by expert pathologists. This is typically achieved through multi-reader multi-case studies. In a recent multi-reader study, the Ibex HER2 AI tool improved inter-observer agreement among pathologists for most HER2 IHC scoring categories, particularly for HER2 0, 1+, and 3+ cases, demonstrating the potential of AI to enhance scoring consistency. This level of consistency is the benchmark regulators look for to ensure the tool provides a 'digital gold standard'.

Ibex recently received IVDR certification for our HER2 breast cancer biomarker scoring solution. This is the certification required to confirm that a tool meets the highest global standards for routine clinical performance and use in clinical trials. 

The practice of medicine is in a constant state of evolution. As new technologies emerge, certain manual skills naturally become obsolete, while more sophisticated ones are developed in their place. Pathology is currently undergoing this exact transformation.

One of our pathologist users summed it up perfectly: “AI allows me to better use my cognitive fuel.”

In the traditional workflow, pathologists spend a significant portion of their day on 'needle-in-a-haystack' tasks—such as searching for minute clusters of stained cells—or on highly repetitive, labor-intensive quantification. These routine tasks are not only exhausting but are arguably a suboptimal use of a specialist's high-level training.

By integrating AI, we aren't replacing the pathologist; we are reallocating their expertise. AI handles the heavy lifting of quantification and screening, liberating the pathologist to devote their 'cognitive fuel' to where it is needed most: navigating complex interpretations, resolving borderline cases, and diagnosing rare pathologies. Ultimately, AI transforms the role from one of manual counting and searching to one of high-level clinical decision-making and data-driven consultation.

The introduction to this interview includes text that has been created with the assistance of generative AI and has undergone editorial review before publishing. Technology Networks' AI policy can be found here.