HistoTME is an AI software that predicts the tumor microenvironment’s molecular makeup from standard pathology images, helping identify lung cancer patients likely to benefit from immunotherapy without needing costly molecular tests.
The tumor microenvironment (TME) plays a crucial role in the progression and treatment response of cancers, particularly non-small cell lung cancer (NSCLC). As immunotherapies such as immune checkpoint inhibitors (ICIs) become increasingly central to cancer care, the ability to accurately characterize the TME is essential for predicting which patients will benefit from these treatments. Traditionally, the assessment of TME relies on molecular assays, such as RNA sequencing or immunohistochemistry, which provide insights into the immune landscape of tumors. However, these tests are often expensive, require specialized equipment and expertise, and may not be accessible in all clinical settings. The growing demand for precision oncology has highlighted the need for more accessible, cost-effective, and scalable methods to evaluate the TME and guide immunotherapy decisions. Current approaches to TME characterization face several significant limitations. Biomarkers like PD-L1 expression, while widely used, are not always reliable predictors of response to ICIs, especially in patients with low PD-L1 levels. Molecular assays, though informative, are resource-intensive and may not be feasible in many healthcare environments due to cost, infrastructure, or insurance coverage constraints. Additionally, these methods often require fresh or high-quality tissue samples, which are not always available, and involve lengthy turnaround times that can delay treatment decisions. Digital pathology offers a potential alternative, but existing computational tools typically depend on expert pathologist annotation or lack the ability to directly infer molecular features from routine histopathology slides, limiting their clinical utility and scalability.
HistoTME is a software-based artificial intelligence tool designed to predict the molecular composition of the tumor microenvironment (TME) directly from standard hematoxylin and eosin (H&E)-stained histopathology images. The system operates through a sophisticated two-step pipeline: first, it uses the UNI foundation model to extract detailed image features, which are then processed by a multitask-attention-based multiple instance learning (AB-MIL) framework to generate prediction scores and heatmaps. In the second step, HistoTME applies clustering analysis and a two-step random forest classification algorithm to categorize patients into clinically relevant TME subtypes—immune-desert or immune-inflamed. This approach enables the prediction of patient response to immune checkpoint inhibitor (ICI) therapy, using only digital pathology slides, without the need for expert annotation or costly molecular assays. What differentiates HistoTME is its ability to serve as a digital biomarker, offering a scalable, cost-effective alternative to traditional molecular testing for immunotherapy stratification. Unlike conventional biomarkers such as PD-L1 expression, which often lack predictive power in patients with low expression levels, HistoTME leverages routinely available pathology images and advanced AI to infer molecularly defined TME subtypes. The tool was trained and validated on a large, multi-modal dataset of over 650 lung cancer patients with matched histopathology and RNA sequencing data, ensuring robust performance and clinical relevance. Its design eliminates the need for specialized molecular infrastructure, making it particularly valuable for resource-limited settings. Additionally, the integration of interpretability features, such as spatial heatmaps, enhances clinical trust and usability, positioning HistoTME as a transformative solution in the field of computational pathology and precision oncology.
• Enables prediction of tumor microenvironment molecular composition directly from standard H&E-stained histopathology images without requiring molecular testing.
• Improves identification of non-small cell lung cancer (NSCLC) patients likely to benefit from immune checkpoint inhibitor (ICI) therapy, enhancing treatment personalization.
• Provides a cost-effective and accessible alternative to expensive molecular assays, suitable for resource-limited medical centers.
• Utilizes advanced AI techniques combining foundation models and multiple instance learning for accurate and interpretable predictions with spatial heatmaps.
• Classifies patients into clinically relevant immune-inflamed or immune-desert TME subtypes, aiding clinical decision-making.
• Validated on a large multi-modal dataset with matched histopathology and RNA sequencing data, ensuring robust performance and clinical relevance.
• Does not require expert pathologist annotation, facilitating scalable deployment in diverse clinical settings.
• Immunotherapy response prediction
• Digital pathology workflow integration
• Cost-effective biomarker development
• Resource-limited cancer diagnostics
Patent Pending
TRL 5
This technology is available for licensing.