Dr. Suzanne K. Coberly, MD, MS, Senior Director of Translational Pathology at Bristol Myers Squibb, brings more than 20 years of experience across biotech, companion diagnostics, IVD, and translational pathology. Her work focuses on discovery and translational pathology in oncology and immuno-oncology, including biomarker and predictive assay development.
At BMS, Dr. Coberly works within a translational pathology group supporting programmes from discovery through first-in-human studies. Her team focuses on specialised, exploratory pathology work, particularly novel assays that help research teams understand mechanism of action, pharmacodynamics, efficacy signals, and potential toxicity. This includes work across protein degradation, immuno-oncology, and CAR-T programmes.
For Dr. Coberly, computational pathology is not a sudden revolution, but the result of years of progress in digital pathology, image analysis, and AI-enabled tissue interpretation. The digitisation of glass slides enabled pathologists and researchers to work with whole-slide images, quantify expression, and analyse complex tissue features more consistently. As AI tools have improved, they have expanded what can be measured, including multiplex expression, cell-type relationships, and spatial patterns within the tumour microenvironment.
One of the most exciting areas is virtual IHC: the ability to use H&E slides, which are routinely available for most patients, to predict protein expression or other molecular features. This could be particularly valuable in early-phase oncology trials, where tissue is often limited and repeated testing is not always feasible. By extracting more information from standard histology, computational pathology may allow teams to explore more biomarkers, reduce tissue burden, and identify signals that would otherwise be missed.
AI also has the potential to improve translational decision-making by integrating pathology with clinical, genomic, and molecular data. Foundation models may help identify new biomarker hypotheses across complex datasets, while virtual IHC or virtual genomics could support more targeted, patient-level exploration. This creates opportunities to better understand whether a therapy is acting through its intended mechanism and which biomarkers may predict response.
However, Dr. Coberly emphasises that strong tissue biomarkers still require a stepwise evidence pathway. Before moving from exploratory research toward clinical development or patient stratification, a biomarker must be supported by strong preclinical work, assay validation, relevant model systems, and analysis of patient samples. Researchers need to understand expression patterns, biological relevance, association with mechanism of action, and potential links to efficacy. Clinical data can then help determine whether a biomarker is strong enough to progress toward a companion diagnostic or stratification strategy.
Computational pathology can reveal biology that traditional pathology may not easily capture, including spatial relationships between tumour and immune cells, heterogeneity across tissue regions, and microenvironment features that are difficult to quantify manually. These insights are especially relevant in oncology and immuno-oncology, where response may depend not only on whether a biomarker is present, but where it is located, how it interacts with neighbouring cells, and how consistently it appears across a tumour.
The challenge is trust. AI-enabled pathology biomarkers must be validated in ways that satisfy scientists, clinicians, development teams, and regulators. Black-box models remain a concern, particularly when they identify predictive features without a clear explanation of the underlying biology. For biomarkers to be adopted confidently, teams need to show that AI-derived signals are reproducible, biologically meaningful, clinically relevant, and robust across patient populations.
Dr. Coberly’s advice for teams working in computational pathology is clear: do not treat AI as a shortcut around biology. Use it to expand discovery, generate hypotheses, and make better use of limited tissue, but always connect findings back to mechanism, assay performance, and clinical relevance. The most predictive and meaningful biomarkers will come from combining computational power with rigorous pathology expertise, translational validation, and a clear understanding of the disease biology.
