Summary
Researchers at Shandong University and BGI Research describe DiffGSP, a computational framework that models mRNA diffusion to reconstruct spatial gene-expression patterns. The work is a bioRxiv preprint evaluated across diverse spatial-transcriptomics datasets.
Researchers at Shandong University and BGI Research have introduced DiffGSP, a computational framework designed to correct for mRNA diffusion in spatial transcriptomics. The bioRxiv preprint, posted on September 15, 2026, describes a method that combines a physical model of diffusion with graph-based signal processing to reconstruct gene-expression patterns more faithfully within tissue.
Spatial transcriptomics measures gene activity while retaining information about where transcripts were located in an intact tissue. That spatial context can reveal how cells are arranged and how disease-associated regions form. During tissue processing, however, mRNA can move from one location to nearby locations. A sequencing spot may therefore contain transcripts originating from adjacent cells, creating a mismatch between the measured signal and its original position.
How DiffGSP models the problem
DiffGSP treats diffusion as a source of distortion rather than an unavoidable feature of the final measurement. The framework incorporates Fick’s law, which describes how particles spread through a medium, to model the movement of mRNA between neighbouring tissue locations. It also uses graph signal processing, a set of methods for analysing data distributed across connected points, to represent relationships between spatial measurements.
Together, these components allow the method to estimate the underlying gene-expression landscape from measurements that have been altered by diffusion. The approach is described as physics-informed because its computational reconstruction is constrained by a physical description of how diffusion occurs, rather than relying only on statistical patterns in the data.
The authors say DiffGSP is broadly applicable to sequencing-based spatial-transcriptomics technologies. It is a computational correction method, not a new sequencing instrument or a biological treatment.
Findings reported across tissues
The preprint describes benchmarking across diverse datasets and evaluation metrics, with the authors reporting consistent restoration of spatial gene-expression patterns. The reconstructed data were used to identify several types of biological organisation:
- fine anatomical structures in the mouse brain;
- spatially organised gene modules in the kidney;
- intratumoral heterogeneity in colorectal cancer; and
- tertiary lymphoid structures in lung adenocarcinoma.
These examples represent different analytical challenges. Brain anatomy depends on distinguishing closely positioned structures, while tumours can contain several molecularly distinct regions within the same mass. Tertiary lymphoid structures are organised immune-cell formations that can be studied through the spatial arrangement of gene-expression signals.
More faithful spatial localisation could make it easier to interpret cell neighbourhoods, tissue organisation and disease microenvironments from sequencing-based measurements. In this study, that potential comes from computationally reconstructing the spatial signal before downstream biological interpretation.
The evidence is currently presented in a bioRxiv preprint and consists of computational benchmarking and reconstructed patterns across tissue datasets. The supplied record does not provide numerical performance values, dataset sample sizes or details of experimental confirmation, so the magnitude of the reported improvement and its biological validation require assessment of the full study and subsequent peer-reviewed work.