TIGER: AI-Powered Histology-Based Gene Expression Signature Prediction for Breast Cancer
SUMMARY
TIGER is a transformer-based computational pathology pipeline that enables accurate prediction of breast cancer gene-expression signatures and neoadjuvant therapy response from routine H&E-stained histology images, facilitating rapid, cost-effective prognostic and treatment stratification without the need for RNA sequencing or microarray assays
The Unmet Need: Current transcriptomic testing for cancer prognosis and therapy response is costly and slow with significant tissue and quality constraints
- Breast cancer treatment selection and prognosis often rely on gene-expression profiling, which requires expensive molecular assays such as RNA sequencing or microarrays; these assays are limited by high cost, tissue preservation issues including RNA degradation, and long turnaround times, which hinder widespread clinical use and rapid decision-making.
-
Advances in computational pathology and artificial intelligence have increasingly focused on leveraging digital whole-slide histology images to extract molecular and genomic information non-invasively, thus creating opportunities to improve clinical workflow efficiency and accessibility of molecular biomarkers.
The Proposed Solution: TIGER leverages a transformer-based multiple-instance learning model to infer gene-expression signatures directly from digitized H&E histology whole-slide images
- The faculty inventor developed the TIGER system employing a self-normalizing neural network architecture combined with transformer-based feature aggregation adapted for histology images, trained on The Cancer Genome Atlas and validated across multiple clinical trials including ISPY2 and CALGB cohorts. It reconstructs curated breast cancer gene-expression signatures—such as immune infiltration, proliferation, and hormone signaling pathways—allowing prediction of pathological complete response with demonstrated AUC up to 0.794, surpassing clinical-feature-based models. Unlike previous approaches predicting individual genes, TIGER emphasizes biologically validated gene-expression signatures, enhancing interpretability and robustness. The model's outputs include response prediction visualizations and probability stratifications, supporting treatment selection decisions, including HER2-targeted neoadjuvant therapies.
FIGURE
ADVANTAGES
-
Rapid gene-expression signature inference from routine histology
-
Cost-effective alternative to molecular assays
-
High predictive accuracy for neoadjuvant therapy response
-
Robust across multiple breast cancer subtypes and cohorts
-
Integrates multiple biological pathways for enhanced interpretability
-
Compatible with standard FFPE and frozen tissue specimens
APPLICATIONS
-
Prediction of neoadjuvant chemotherapy response in HER2-negative breast cancer
-
Stratification of HER2-positive breast cancer for targeted neoadjuvant therapies
-
Prognostic assessment using gene-expression signatures in breast cancer clinical management