Seeing is Believing
Recent advances in spatial multi-omics technologies have enabled scientists to study tissue architecture with detailed molecular measurements such as gene expression, protein abundance, and metabolite levels, complemented by morphological context from H&E-stained images. In our AI in Genomics Lab, we have developed several widely used tools, including SpaGCN, TESLA, and MISO, to analyze these complex multimodal datasets. While these methods excel at tasks like cell type identification and tissue region segmentation, their "black-box" nature makes it difficult to interpret how they connect morphological features to molecular changes. Uncovering these connections is critical. A fundamental question remains: how exactly do molecular features shape tissue architecture? Quantitative measurement of the relationship is helpful, but not sufficient—we also need to visualize it. As humans are naturally visual thinkers, seeing how tissue structure shifts with molecular activity can be more intuitive and insightful for many biologists than simply reviewing statistical abstractions like p-values.
MorphLink make AI-Driven Spatial Biology Transparent
To address these challenges, we developed MorphLink—an interpretable AI tool that visually bridges tissue morphology and molecular signals. By opening the "black box" of spatial biology, MorphLink uncovers and explains how tissue structure connects to molecular activity, enabling new biological insights. True to its name, MorphLink provides statistically robust and visually intuitive linkages between molecular and morphological changes. It achieves this by solving three key challenges:
1) Interpretable Morphological Feature Extraction: While deep learning models excel at tissue segmentation, their features often lack biological interpretability. MorphLink extracts comprehensive, label-free morphological measurements that are both biologically meaningful and human-understandable.
2) Quantifying Spatial Relationships: Current methods lack robust metrics to correlate morphology with molecular data in spatial contexts. MorphLink introduces CPSI (Curve-based Pattern Similarity Index), a novel metric that efficiently quantifies these relationships.
3) Visual Integration of Multi-modal Data: MorphLink provides intuitive visualizations that simultaneously track how cellular behavior evolves from both structural and molecular perspectives.
An Example of How MorphLink Reveals Tumor Heterogeneity
We demonstrate how MorphLink helps researchers uncover the connection between tissue appearance and molecular activity using a bladder cancer sample. The first example focuses on nuclei, the most prominent structures in tumors. MorphLink detects that genes involved in immune signalling, such as CD74, are strongly associated with variations in chromatin density, as measured by nuclear solidity in H&E images. As shown in Fig. 1a, increased CD74 expression corresponds to a morphological transition in the nuclei, from dark, dense chromatin (heterochromatin) to a more clumped and dispersed state (euchromatin), characterized by inconspicuous nucleoli and minimal cytoplasm. This loosening and opening of chromatin are typically associated with increased transcriptional activity, reflecting activation of genes involved in antigen presentation.
Beyond nuclei, MorphLink also examines other tissue structures. For example, it reveals a link between genes related to tumor proliferation, such as MYCL, and the rapid development of cancer-associated fibroblasts (CAFs). Identifying CAFs in H&E images is challenging because their nuclei are smaller, less dense, and more irregular than those of tumor cells. Additionally, they are often embedded within the stroma, a dense tissue background that can obscure their features. Despite these challenges, MorphLink successfully captures CAF regions. As shown in Fig. 1b, it demonstrates that higher MYCL expression is associated with larger CAF areas (highlighted in green) surrounding tumor cells. These CAFs support tumor growth and invasion, making them key contributors to cancer progression.
MorphLink makes it easier to find and understand how different tissue features seen under the microscope relate to specific molecular signals. These connections help us better understand how cells behave, especially when cells of the same type show different morphology. This is important because changes in cell morphology often reflect changes in their functions. Since spatial omics technologies are still too expensive or hard to access for routine use, we believe the interpretable and biologically relevant morphology features identified by MorphLink will catalyze the advancement of molecularly informed histopathology analysis, paving the way for developing accurate predictive and prognostic risk models based on histology images.