SPRINT enables spatial proteomic profile reconstruction through multimodal fusion of histology and transcriptomics

Tissues are not merely static collections of cells; they are dynamic, organized cities where different neighborhoods perform specific jobs. To understand how these cities function, especially in health and disease, scientists need to see not just which cells are present, but what they are actually doing. For decades, the most detailed maps of tissue activity have relied on measuring RNA, the molecular instructions cells use to build proteins. However, RNA is only the blueprint; the proteins themselves are the workers, the machines, and the signals that actually carry out the work. While we can now map where genes are active in a tissue sample, measuring the actual proteins in those same precise locations has been difficult, expensive, and slow. This gap leaves researchers with a partial picture, knowing the plans but missing the construction site.

A new computational tool called SPRINT offers a way to fill in these missing details. By combining two types of data that are already commonly available—microscope images of tissue and maps of gene activity—SPRINT can predict where specific proteins are located with high accuracy. The researchers behind this work, based at the Harbin Institute of Technology, developed a system that learns the relationship between what a tissue looks like under a microscope, what genes are turned on, and where the resulting proteins end up. Once trained, this system can reconstruct detailed protein maps for tissue sections where proteins were never directly measured, effectively turning standard microscope slides and gene maps into comprehensive protein atlases.

The challenge of mapping proteins in space is significant because proteins behave differently than genes. Some proteins are tightly linked to the physical shape of the tissue, such as those that form the walls of blood vessels or the outer layer of skin, making them easy to spot in a microscope image. Others, however, are involved in complex chemical signaling or immune responses that do not leave a clear visual trace. Traditional methods that rely only on microscope images often miss these invisible proteins, while methods that rely only on gene maps can be inaccurate because the amount of a protein in a cell does not always match the amount of its corresponding gene instructions. The researchers found that relying on just one of these sources limits the ability to see the full picture.

To solve this, the team created a framework that treats the microscope image and the gene map as two complementary languages describing the same biological reality. The system first looks at the tissue image to understand the physical structure, such as the arrangement of cells and the texture of the tissue. It then looks at the gene expression data to understand the molecular state of those cells. By fusing these two streams of information, the model learns to predict protein locations even when the protein is not obvious in the image or when the gene instructions are a poor predictor of the final protein amount. The system is designed to be flexible, handling both high-resolution images that show individual cell details and lower-resolution images that capture broader tissue patterns, ensuring it works across different types of laboratory equipment.

The researchers tested this approach on a variety of biological samples to see if it could generalize beyond the specific data it was trained on. They began with human tonsil tissue, training the model on one section and then asking it to predict protein locations in a completely different section from a different person. The system successfully reconstructed the spatial patterns of dozens of proteins, including those involved in immune defense and tissue structure. It performed better than existing methods that relied on either images or genes alone, accurately capturing both the broad organization of the tissue and the fine details of local protein distribution. This success held true even when the researchers tested the model on mouse spleen tissue, which was imaged at a much lower resolution, proving that the system could adapt to different image qualities without losing its predictive power.

Perhaps the most rigorous test involved a human brain sample, where the researchers split a single tissue section diagonally. They trained the model on one half and asked it to predict the protein landscape of the other half, a region it had never seen before and which contained different types of cells in different proportions. Despite this spatial separation and the shift in cellular composition, the model accurately reconstructed the protein patterns across the unseen region. It correctly identified complex biological programs, such as areas of immune activation and tissue repair, demonstrating that it had learned the fundamental rules connecting tissue structure, gene activity, and protein location, rather than simply memorizing the specific layout of the training sample.

The utility of this approach extends beyond just filling in missing data. When the researchers used the reconstructed protein maps to help identify distinct regions within the tissue, the results were significantly clearer than when using gene data or images alone. In the mouse spleen and brain samples, adding the predicted protein information helped the computer distinguish between anatomical zones that were previously blurred together, such as the marginal zone of the spleen. This suggests that the protein maps generated by the system provide a level of biological detail that is essential for understanding how tissues are organized and how they respond to disease.

The researchers also explored whether this method could work across species and for molecules other than proteins. They trained the model on human breast tissue and applied it to mouse breast tissue, successfully predicting protein locations in the mouse sample without ever seeing mouse protein data during training. This indicates that the relationships between tissue structure, genes, and proteins are conserved enough to be transferred between humans and mice. Furthermore, they tested the framework on metabolites, the small molecules that cells use for energy and signaling. By training the system directly on metabolite data, they were able to reconstruct spatial maps of these molecules in a mouse brain, showing that the underlying strategy is not limited to proteins but can be adapted to reconstruct various types of molecular landscapes.

This work establishes a new way to extract maximum value from existing biological data. Many research labs already possess vast archives of tissue images and gene maps, but the protein data for these samples is often missing or incomplete. SPRINT provides a general framework to complete these spatial molecular profiles, allowing scientists to infer the presence and location of proteins, and potentially other molecules, across diverse tissues, species, and imaging conditions. By bridging the gap between what we can easily measure and what we need to know, this tool opens the door to a more complete understanding of the molecular architecture of life, turning partial snapshots into comprehensive views of how tissues are built and function.