AI Predicts Cancer Gene Expression from Routine Tissue Slides with 30% Better Accuracy
New deep learning framework Path2Omics accurately predicts molecular signatures from histopathology slides, potentially revolutionizing precision oncology.
Summary
Researchers developed Path2Omics, an AI framework that predicts gene expression and DNA methylation patterns from routine cancer tissue slides. The system achieved 30% better accuracy by combining predictions from both standard diagnostic slides and fresh frozen tissue samples. When tested across seven external datasets, the AI-predicted molecular data matched actual laboratory measurements well enough to accurately predict patient survival and treatment responses, suggesting this approach could make precision cancer medicine more accessible.
Detailed Summary
Precision oncology relies heavily on molecular profiling to guide treatment decisions, but obtaining comprehensive genetic data remains expensive and time-consuming. Researchers at the National Cancer Institute have developed Path2Omics, a deep learning framework that can predict gene expression and DNA methylation patterns directly from routine histopathology slides.
The team trained their AI system on data from 30 cancer types in The Cancer Genome Atlas, using over 22,000 slides from 8,637 patients. Uniquely, they developed two complementary models: one trained on standard formalin-fixed paraffin-embedded (FFPE) slides used in clinical practice, and another on fresh frozen (FF) slides that more closely match the tissue used for molecular sequencing.
When validated across seven independent datasets, the integrated approach achieved 30% better performance than using FFPE slides alone, successfully predicting approximately 4,400 out of 18,000 genes. Remarkably, the AI-predicted gene expression data performed nearly as well as actual laboratory measurements when used to predict patient survival outcomes and treatment responses.
The framework demonstrated particular strength in breast cancer applications, accurately classifying molecular subtypes and predicting responses to neoadjuvant chemotherapy. This suggests that routine pathology slides, already collected for every cancer patient, could provide much of the molecular information currently requiring expensive specialized testing.
While the approach shows promise for democratizing precision oncology, the study was limited to retrospective analysis and will require prospective validation before clinical implementation.
Key Findings
- AI framework predicts gene expression from histopathology with 30% improved accuracy using integrated approach
- Fresh frozen slides outperformed standard diagnostic slides for molecular prediction across cancer types
- Predicted gene expression matched actual measurements for survival and treatment response prediction
- Successfully classified breast cancer subtypes and predicted chemotherapy responses from routine slides
- Framework validated across seven independent datasets spanning multiple cancer types
Methodology
Researchers used deep learning with CTransPath feature extraction and multi-layer perceptron regression, training on 22,369 slides across 30 cancer types with 5×5 nested cross-validation. The integrated model combined predictions from both FFPE and fresh frozen slide models.
Study Limitations
The study was retrospective and limited to specific cancer types in TCGA. Clinical implementation requires prospective validation, and performance may vary with different staining protocols or imaging systems. The approach depends on high-quality digitized pathology slides.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.