Spatial biology has promised a more complete view of how cells behave inside tissues, but that promise has often come with a steep technical price. Researchers who want to measure both gene activity and protein abundance while preserving each molecule’s location must typically combine specialized instruments, costly reagents and complex computational workflows. A new method called NicheTrans aims to make that challenge more manageable by predicting one molecular layer from another while preserving information about the cellular neighborhood in which those measurements were made.
Reported by Wang, Zou, Lin and colleagues in Nature Methods, NicheTrans is designed as a spatially aware cross-omics translation method. Its central goal is to infer inaccessible or missing molecular information from more readily obtained single-omics measurements, such as spatial transcriptomics data. Rather than treating each cell as an isolated unit, the method incorporates both the cell’s molecular profile and the characteristics of its surrounding niche. That distinction is important because cells with similar gene-expression signatures can behave very differently depending on the neighboring cells, extracellular signals and tissue structures around them.
Cross-omics translation is not an entirely new idea. Machine-learning models have previously been used to predict protein abundance from RNA measurements or to estimate one type of single-cell assay from another. However, many of these approaches focus primarily on the internal molecular state of individual cells. In spatial tissues, that can leave out a crucial source of biological information: location. Immune cells, neurons, astrocytes and other cell types are influenced by physical proximity, cell-cell communication and local tissue architecture. NicheTrans was developed to capture those relationships alongside the multimodal measurements themselves.
The framework uses a flexible Transformer-based architecture, a class of models known for learning complex relationships among multiple features. In this context, the model can process molecular signals from individual cells together with information describing their local microenvironment. Spatial relationships can include neighboring cell identities, local expression patterns and broader tissue organization. By integrating these inputs, NicheTrans learns associations between molecular modalities while retaining the spatial context that gives those associations biological meaning. The result is not simply a list of predicted markers, but a translated molecular map that can be examined across tissue regions.
The researchers evaluated NicheTrans across diverse biological settings to test whether its predictions were reliable beyond a single tissue or experimental platform. Their analyses indicated that incorporating neighborhood information improved cross-omics translation compared with approaches based only on single-cell molecular profiles. The model was also able to reveal spatial multiomics domains that were not readily visible when either omics layer was analyzed alone. These domains represent tissue regions defined by combinations of molecular and spatial properties, potentially exposing biological states that would otherwise remain hidden in separate datasets.
One of the study’s notable features is its use of model interpretation to connect predictions with biological mechanisms. Rather than treating the Transformer as an opaque prediction engine, the researchers examined molecular relationships that contributed to its output. This analysis identified gene programs associated with dopamine metabolism, a process central to neuronal signaling and brain function. The result suggests that NicheTrans can do more than fill in missing measurements: it may also help researchers identify coordinated molecular programs linking different omics layers within specific tissue niches.
The model likewise highlighted molecular patterns connected with amyloid β-associated cell states. Amyloid β accumulation is a defining feature of Alzheimer’s disease, but the effects of the protein are not uniform across the brain. Different glial and neuronal populations can respond in distinct ways, and those responses may depend on where the cells are located and which neighboring populations surround them. By linking predicted protein markers with spatial gene-expression patterns, NicheTrans helped characterize these localized states and offered a way to investigate how disease-associated cellular programs are organized in tissue.
The Alzheimer’s disease analysis also demonstrated a practical use of translated data. Protein markers predicted by NicheTrans were used as spatial landmarks to quantify the organization of key glial subtypes. Glial cells, including astrocytes and microglia, are increasingly recognized as active participants in neurodegeneration rather than passive support cells. Their distribution, clustering and proximity to pathological features can provide clues about disease progression. When direct spatial protein measurements are unavailable or limited, translated markers could provide an additional route for mapping these cell populations and comparing their organization across healthy and diseased samples.
The authors present NicheTrans as a way to broaden access to spatial multiomics rather than as a replacement for direct measurement. Predictions remain dependent on the quality and diversity of the data used to train and evaluate the model, and computational translation cannot eliminate the need for experimental validation. Nevertheless, the framework addresses a major bottleneck in spatial biology: researchers may already possess valuable spatial transcriptomic datasets but lack matching protein or other molecular measurements. By using the information embedded in cellular neighborhoods, NicheTrans could help transform those partial datasets into more comprehensive maps of tissue biology.
As spatial technologies continue to expand, methods that combine experimental accessibility with computational depth may become increasingly influential. NicheTrans points toward a model of analysis in which molecular modalities are not interpreted independently and cells are not treated as isolated points. Instead, genes, proteins, cells and neighborhoods are analyzed as parts of an interconnected system. If validated across additional tissues, diseases and experimental platforms, the approach could help researchers uncover molecular relationships and disease-associated spatial patterns from datasets that were never originally designed to support full multiomics analysis.
Subject of Research: Spatially aware cross-omics translation and Transformer-based multimodal analysis for spatial multiomics.
Article Title: NicheTrans: spatial-aware cross-omics translation
Article References: Wang, Z., Zou, Q., Lin, S. et al. NicheTrans: spatial-aware cross-omics translation. Nature Methods 23, 1528–1539 (2026). https://doi.org/10.1038/s41592-026-03153-3
Image Credits: AI Generated
DOI: 10.1038/s41592-026-03153-3
Keywords: NicheTrans, spatial multiomics, cross-omics translation, spatial transcriptomics, Transformer model, multimodal learning, cellular microenvironment, Alzheimer’s disease, glial cells, dopamine metabolism.
