Can Measuring Just One Molecular Layer Still Fill In the Full Tissue Picture? NicheTrans Uses Spatial Information to Infer Missing Data

Spatial multi-omics aims not only to determine which genes cells express, but also to identify where they are located and which cells surround them. However, simultaneously measuring RNA, proteins, metabolites, or chromatin states in the same tissue section often requires expensive equipment and complex experiments. NicheTrans, developed by teams at Monash University and Fudan University, seeks to infer one unmeasured molecular signal from another, more readily obtainable type of molecular data while preserving the tissue’s spatial structure.

Published in *Nature Methods*, the model uses a Transformer architecture. When learning correspondences between different omics layers, it does not treat each cell as an isolated data point, but incorporates information about surrounding cells and the microenvironment. For example, it can infer protein markers from spatial gene expression and then use the inferred results to analyze how cell populations are arranged within tissue.

The research team validated the model on multiple existing datasets spanning breast cancer, lymph nodes, brain diseases, and different spatial multi-omics platforms. According to the paper, NicheTrans identified spatial regions that single-omics analysis could not distinguish. Model attribution analysis also linked its findings to dopamine metabolism-related gene programs and cell states associated with amyloid beta. These results suggest that incorporating information about which cells are adjacent to one another may restore biological context lost when only molecular concentrations are considered.

Using Alzheimer’s disease brain tissue data, the team further treated the protein markers inferred by the model as spatial landmarks to quantify the distribution of different glial cell subtypes. This type of analysis could eventually be used to identify cellular neighborhoods associated with disease progression, treatment response, or recurrence. In tumor research, researchers also hope to use it to identify microenvironmental features linked to treatment resistance or recurrence. However, the current findings are limited to research data analysis and have not demonstrated that the model can be used for clinical diagnosis or to select treatments for patients.

The implementation, training, and visualization workflows for NicheTrans have been made public, along with documentation for examples involving Parkinson’s disease, Alzheimer’s disease, breast cancer, lymph nodes, and MISAR-seq. Open-source code will help other teams reproduce and compare the results. It may also allow laboratories without advanced multi-omics equipment to first generate hypotheses from existing single-omics data and then decide which molecular signals warrant experimental validation.

However, the data it “fills in” are ultimately model predictions, not direct measurements of the sample. Reliability is affected by the training data, experimental platform, tissue quality, and differences among disease populations. When the model encounters cell states that are rare in the training data, it may also produce results that appear plausible but are inaccurate. Before the approach can be used for patient stratification or biomarker development, it must still be compared item by item with experimentally measured multi-omics results in independent, prospective clinical samples, and standards must be established for uncertainty assessment and cross-platform validation.