Predicting the Future: How AI is Unlocking Secrets in Tissue Images

                                                                  Image credit: https://openai.com/index/dall-e/ 

Imagine looking at a simple picture of a tissue sample and being able to tell which genes are active inside it. That's the promise of a new area of research called spatial gene expression prediction, and it could revolutionize how we understand and treat diseases like cancer.

What is Spatial gene expression? Think of your body as a complex city, with each cell acting as a building. Genes are like the blueprints that tell each building (cell) what to do. Spatial transcriptomics (SRT) is a way of mapping out which genes are switched on in different parts of the tissue. This is super useful because it shows us how cells organize themselves and interact, which is key to understanding how diseases develop.
The problem? SRT is still quite expensive. A more common and cheaper method involves staining tissue samples with haematoxylin and eosin (H&E). H&E staining has been around for ages and is used in pretty much every clinic to help doctors diagnose diseases by looking at tissue samples under a microscope.

Enter artificial intelligence: Now, what if we could use AI to predict the spatial gene expression from those H&E images? That’s the big idea. If it works, we could get a wealth of information from standard tissue samples, helping us to:

• Find new biomarkers to diagnose diseases earlier

• Discover new therapeutic targets for drug development

• Examine how genes vary across large populations

Benchmarking the methods: Because this field is so new, lots of different AI methods are popping up to tackle this challenge. A recent study has put eleven of these methods to the test, comparing how well they predict spatial gene expression from H&E images. This study looked at everything from how accurate the predictions were to how easy the methods were to use and how well they could be applied to real-world clinical data.

How the methods work: Most of these methods use ConvolutionalNeural Networks (CNNs) and Transformers. These are types of AI that are good at recognizing patterns in images. They learn to associate certain visual features in the H&E images with specific patterns of gene activity. Some methods also use fancy techniques like Graph Neural Networks (GNNs) to understand how gene expression varies between neighboring spots in the tissue.

What the study found: The study used five different spatial transcriptomics datasets to train and test the AI methods. The researchers then checked how well the methods could: First, predict gene expression within the same image, Second, generalize to new images from different datasets and third, predict patient survival outcomes. They also looked at how easy each method was to use and how efficiently it used computer resources.

Here’s what they discovered:

• No single method was the best at everything. Some were good at predicting gene expression in specific tissues, while others were better at generalizing to new datasets.

• Simpler methods can be better. The most complex AI architectures didn't always perform the best. Methods that focused on capturing both local and global features within single spots in the image often outperformed those that tried to incorporate information from neighboring spots.

• Some methods are more robust to image quality. HisToGene and Hist2ST were less affected by variations in H&E image quality, which is important for real-world clinical applications.

• Usability is a major issue. Many of the methods were difficult to use, lacking proper documentation, testing, and user-friendly interfaces.

Why this matters: This study is a big step forward because it gives us a clear picture of the current state of spatial gene expression prediction. It highlights the strengths and weaknesses of different methods, helping researchers choose the right tool for the job. It also points out areas where more work is needed, such as improving the usability of these methods and making them more robust to variations in image quality and resolution.
The future of AI in pathology: AI-powered spatial gene expression prediction has the potential to transform how we diagnose and treat diseases. By unlocking the wealth of information hidden within standard tissue images, we can gain a deeper understanding of disease mechanisms and develop more effective therapies.
While there’s still work to be done, this study shows that we’re on the right track. As AI methods continue to improve, we can expect to see even more exciting applications of this technology in the years to come.
 
Additional information: Benchmarking the translational potential of spatial gene expression prediction from histology. Nature communications (2025). https://doi.org/10.1038/s41467-025-56618-y

Journal information: https://www.nature.com/ncomms/