We’ve long said that the next major therapeutic breakthrough may lie in pathology data that already exists. At this year’s AACR Annual Meeting, it became clear the biopharma community is arriving at the same conclusion.
Across more than 7,000 abstracts presented in San Diego, multiple independent groups demonstrated that AI applied to routine H&E whole slide images can predict genetic mutations, biomarker expression, and treatment response with accuracy that rivals, and in some cases exceeds, molecular testing.
Pathology data offers the most detailed and direct profile of disease. What AACR 2026 showed is that the field is finally learning how to unlock it, and the implications for biomarker strategy, patient identification, and companion diagnostic development are immediate.
The Slide’s Potential to Challenge Molecular Testing
Several abstracts illustrate how far pathology AI has progressed in capturing molecular signals directly from tissue.
Natera presented a deep learning model trained on more than 45,000 colorectal cancer patients that predicts MSI status from H&E images with 0.98 accuracy and BRAF V600E mutation status with 0.93 accuracy. Importantly, the model learns from paired molecular and histopathology data, but operates on the slide alone at inference time.
For R&D teams, this changes the constraints around biomarker development. Genetic and molecular characteristics that typically require sequencing or IHC may now be gleaned from slides that already exist in trial archives.
MD Anderson Cancer Center’s Path-IO was validated across more than 1,000 patients from multiple institutions and countries. The model predicts immunotherapy response in non-small cell lung cancer directly from whole slide images and pathology data, and outperforms PD-L1 across validation datasets.
This has direct consequences for companion diagnostics. If an AI-derived biomarker consistently outperforms established approaches, the case for building and deploying these models becomes difficult to ignore.
In a third example, researchers from the National Cancer Institute, NIH, Cedars-Sinai, Harvard Medical School, Dana-Farber, and Yale presented on TIME_ACT, a model that predicts immunotherapy response directly from pathology slides using a framework called Path2Omics. The system infers gene expression patterns from H&E images, identifying 66 genes associated with tumor immune activation, without sequencing, additional staining, or molecular testing.
These three abstracts may all take different approaches, but they all support the same underlying conclusion. Signals that have historically required dedicated molecular workflows are accessible from tissue that already exists.
Why Pathology Data? Why Now?
The case for pathology data in drug development is not simply that it is cheaper or faster than molecular testing, though for clinical teams those advantages are real. For biopharma R&D teams, the argument is more fundamental.
Retrospective access
Sequencing archived samples is often impractical due to cost, degradation, or tissue limitations. Slides can be digitized and reanalyzed, unlocking historical cohorts for biomarker discovery and patient stratification that molecular testing never could.
Spatial context
Molecular assays describe what is expressed. Pathology reveals where within tissue architecture and in relation to other cell types. This spatial dimension is increasingly where predictive signals are found.
Tissue preservation
Molecular testing consumes tissue. AI-based slide analysis does not, making it especially valuable for small biopsies and limited samples.
Scale
AI enables simultaneous analysis of thousands of slides, expanding what is possible in cohort design, enrichment, and biomarker discovery.
The Work Starts Here
In our experience, biopharma teams don’t need to be convinced that pathology AI is valuable. Instead, they need to overcome two challenges that consistently stand in the way of fully capitalizing on it.
The first is data. Unlike molecular data, which has to be newly created, pathology data already exists in study archives, CRO partnerships, and biobanks across indications of interest. Every patient who participates in an oncology trial has at least one associated slide. But existing data is not the same as usable data. Slides must be digitized, annotated, and curated into datasets that are comprehensive, diverse, and structured enough to train or fine-tune a model against. That work is unavoidable.
The second is deployment. A model that lives in a research environment is not a tool. It becomes one only when it is integrated into routine R&D workflows at scale, with the compute resources and interoperability to support it. For this reason, the platform where the model runs is more important than the model itself.
Neither challenge is insurmountable, and the window for leading the field is still open. The organizations that invest in the right data foundation, the right platform, and the right partner will define the next era of precision oncology, turning what AACR revealed into clinical and commercial advantage.
The rest will be working to catch up.
