83 Quick summary: Researchers in China have developed an artificial intelligence approach called Deep Visual Spatial Transcriptomics and Proteomics that successfully maps the three dimensional inner molecular complexity of colorectal tumours with 94% accuracy. By combining standard pathology images with spatial data, the technology revealed that specific protein variations directly create an immunosuppressive microenvironment, which helps explain why certain regions of a tumour manage to evade medical detection and treatment. This breakthrough carries immense value for clinical healthcare practice and public health policy, providing oncology teams with an accessible predictive tool to decode rich molecular information from routine biopsies and design highly tailored therapies that improve patient survival outcomes and mental well-being. Researchers in China have developed an artificial intelligence approach that combines detailed pathology images with spatial molecular data to map the inner complexity of colorectal tumours in three dimensions, offering new insights into what drives the disease and how it may be treated more precisely. The technique, known as Deep Visual Spatial Transcriptomics and Proteomics (DVSTP), addresses a long-standing limitation in cancer research: conventional sequencing methods that homogenise tumour tissue lose the spatial context essential for understanding cellular interactions. Colorectal tumours are not uniform masses but complex ecosystems composed of diverse cell populations with distinct behaviours, and variation in gene expression and protein levels across different regions plays a critical role in how cancers progress and respond to therapy. “Bulk sequencing gives us an average picture, like blending an entire landscape into a single colour,” explains the research team from Union Hospital, Tongji Medical College, Huazhong University of Science and Technology. “We needed a way to see both the molecular details and where they occur in the tissue.” DVSTP combines three data layers: high-resolution histopathology images, spatial transcriptomics, and mass spectrometry-based spatial proteomics. Using tissue samples from 123 colorectal cancer patients, the team trained a deep learning model to identify cell types from H&E-stained images, achieving 94% accuracy in distinguishing malignant cells, immune cells, and other stromal components. The researchers applied this approach to serial sections from 2 distinct sites within a single stage II colorectal tumour, analysing 380 slices and reconstructing its three-dimensional architecture. “This allowed us to see how different cell populations are organised in space and how they interact across the entire tumour volume,” says the lead author. Comparing spatial transcriptomics and proteomics data, the correlation between mRNA and protein levels was surprisingly modest, with an average Spearman correlation of just 0.37. “The mRNA–protein relationship is not a simple linear translation,” the researchers note. “This is why integrating direct protein measurements is so crucial for understanding functional cell states.” The spatial proteomics analysis identified 2,805 proteins and revealed four distinct tumour subtypes, each with unique molecular signatures. One of the most promising findings was that computational analysis of standard pathology images alone could predict molecular profiles. The deep learning model achieved an area under the curve (AUC) of 0.718 for predicting protein expression from H&E images, improving to 0.755 when combining images with transcriptomics data. “This suggests that the morphological patterns pathologists have used for decades contain rich molecular information that can now be decoded using artificial intelligence,” the researchers explain. The approach was validated using data from The Cancer Genome Atlas, confirming its generalisability across independent patient cohorts. Among the heterogeneous protein patterns identified, Serine/Arginine-Rich Splicing Factor 6 (SRSF6) stood out for its pronounced spatial variation and strong association with immune exclusion. SRSF6-high areas showed significantly lower infiltration of CD4+ and CD8+ T cells, suggesting the protein contributes to an immunosuppressive tumour microenvironment. Experiments in cell lines and mouse models confirmed that SRSF6 overexpression promoted cancer cell migration and tumour growth while reducing T cell infiltration, with knockdown producing the opposite effects. In human patient samples, high SRSF6 expression correlated with poorer overall survival. “SRSF6 appears to be a master regulator that shapes both tumour cell behaviour and the immune landscape,” the team concludes. “Its spatial heterogeneity within tumours may explain why some regions evade immune detection.” The researchers acknowledge that current spatial transcriptomics and proteomics technologies remain limited by resolution and cost. DVSTP addresses this by using computational deconvolution to enhance effective resolution while integrating widely available pathology images with spatial omics data. Its capacity for three-dimensional tumour reconstruction also reveals spatial infiltration patterns that are invisible in two-dimensional sections, which could help clinicians identify aggressive tumour regions and predict therapeutic responses. The findings, published in Science Bulletin, highlight how combining accessible pathology resources with deep learning can bridge the gap between research and clinical application in precision oncology.