A new computational pathology study reveals that deep learning models can identify critical biomarker status from routine tissue images to accelerate oncology diagnostics.
Tempus AI, Inc. presented findings at the European Society for Medical Oncology Congress 2026 in Madrid, detailing a computational approach to pre-screening advanced non-small cell lung cancer (NSCLC) patients for biomarker status. The study leveraged deep learning models to analyze routinely generated hematoxylin and eosin (H&E) pathology whole slide images, achieving greater than 80% sensitivity across 12 distinct biomarkers.
The research cohort comprised 60,459 advanced NSCLC cases derived from 56,969 individual patients, providing a massive dataset for model validation. By utilizing standard pathology slides, the model aims to optimize the diagnostic workflow for clinicians facing increasingly complex molecular testing requirements.
Sensitivity metrics for several of the 12 assessed biomarkers exceeded 90% in the multi-center analysis. This performance level suggests a high potential for the tool to serve as an effective triage mechanism in clinical settings where tissue availability is constrained.
The diagnostic framework is designed to prioritize subsequent molecular or immunohistochemistry testing rather than replacing these established standards. The initiative represents a collaboration between Tempus and Pfizer to mitigate systemic barriers to biomarker testing in oncology.
Ezra Cohen, MD, Chief Medical Officer of Oncology at Tempus, stated that the findings highlight the capacity for artificial intelligence to extract clinically relevant signals from existing diagnostic data. This methodology integrates directly into routine care pathways to provide actionable insights for precision medicine.
The underlying technology builds upon the company’s existing portfolio, including Paige Predict and Tempus Preview. These systems utilize similar multimodal data architectures to predict genomic and IHC biomarker status when traditional sequencing is either delayed or insufficient.
Processing H&E whole slide images requires sophisticated computer vision architectures capable of handling gigapixel-scale data. The model architecture likely employs convolutional neural networks or vision transformers to identify morphological patterns associated with specific genomic alterations, effectively mapping visual features to molecular phenotypes.
Training these models involves high-dimensional feature extraction where the algorithm learns to correlate subtle cellular arrangements with underlying genetic mutations. This process requires rigorous normalization of staining variations across different laboratory sites to ensure the model maintains consistent performance on diverse clinical inputs.
The computational pipeline must account for significant inter-laboratory variability in tissue preparation and scanning protocols. By implementing advanced color normalization and spatial feature aggregation techniques, the model minimizes the impact of non-biological noise, allowing the network to focus on the underlying tumor biology and microenvironment characteristics.
The integration of AI into pathology workflows addresses a critical bottleneck in oncology: the time-consuming nature of sequential diagnostic testing. By providing an early indicator of biomarker status, the model allows clinicians to allocate limited tumor tissue more efficiently toward the most likely therapeutic targets.
The reliance on H&E slides is significant because these images are generated as a matter of course in pathology departments. Extracting predictive genomic information from these images transforms a static diagnostic artifact into a dynamic data source for precision oncology.
The clinical utility of this pre-screening panel lies in its ability to accelerate access to personalized therapies. As the number of biomarker-driven treatments for NSCLC expands, the ability to rapidly identify patient cohorts becomes a primary determinant of successful therapeutic intervention.
Future development of the H&E pre-screening panel will likely focus on refining sensitivity across a broader range of tumor types and integrating the tool into clinical decision support systems. Researchers and clinicians will monitor how these predictive models perform in prospective trials compared to gold-standard molecular sequencing.
