Stanford Medicine researchers have developed an artificial intelligence (AI) program that can predict gene activity in tumor cells using standard microscope images of biopsy samples. This tool, described in Nature Communications on November 14, promises to offer an affordable and faster alternative to the costly and time-consuming genetic sequencing currently used in cancer treatment planning.
The AI program, named SEQUOIA (Slide-based Expression Quantification Using Linearized Attention), was trained on data from more than 7,000 tumor samples across 16 cancer types. By analyzing the standard microscope images, SEQUOIA can predict the expression of over 15,000 genes in tumor cells. In some cancer types, the AI’s predictions matched real genetic data with over 80% accuracy. The more data SEQUOIA had to learn from, the better it performed.
Clinicians traditionally determine cancer treatments based on a combination of a tumor’s location and its genetic activity. This helps them understand how aggressive the cancer might be, whether it is likely to spread, and how it might respond to treatments such as chemotherapy, immunotherapy, or hormone therapy. However, obtaining this genetic information usually requires expensive genomic sequencing.
The research team, led by Stanford’s Dr. Gevaert and graduate students Marija Pizuria, Yuanning Zheng, and Francisco Perez, recognized that gene activity could subtly change the appearance of tumor cells under a microscope. With AI, they were able to detect these changes. SEQUOIA’s ability to accurately predict gene activity is particularly useful for identifying large gene programs, such as those related to inflammation or cell growth, which are often key to cancer progression.
To make the results user-friendly, SEQUOIA visually maps the genetic activity onto the tumor image, allowing scientists and clinicians to easily interpret how gene expression varies across different parts of a tumor.
For validation, the team tested SEQUOIA using breast cancer gene signatures already known to be associated with treatment outcomes. The AI tool was able to predict the same risk scores as existing genomic tests, such as the FDA-approved MammaPrint test, which helps assess the risk of cancer recurrence. The results showed that patients flagged as high-risk by SEQUOIA had worse outcomes, including higher recurrence rates and shorter times before relapse.
While SEQUOIA is not yet ready for clinical use, it could eventually offer a cheaper, quicker alternative to genomic sequencing. The team plans to refine the model and explore its potential for a variety of cancers and genetic signatures. With further development, SEQUOIA could revolutionize cancer treatment by providing valuable genetic insights from standard biopsy images alone.
This research was supported by the National Cancer Institute and other funding organizations, with contributions from Roche Diagnostics.
Related topics:
- French Researchers Develop MassiveFold: A Faster, More Efficient Protein Structure Prediction Tool
- Study Investigates Genetic Factors Linked to Alzheimer’s Disease Using Cerebrospinal Fluid Proteins
- NIH Awards $3.96 Million to Combat Antibiotic-Resistant Superbugs with Innovative Research at UH
