Imaging-based spatial transcriptomics enables molecule-resolved profiling of gene expression within the spatial context of tissue. While promising for resolving tissue organization and cell states, these technologies are subject to technical artifacts. Specifically, segmentation errors, transcript spillover between adjacent cells and three-dimensional cell overlap can introduce misassigned transcripts into cell-level expression profiles. Such contamination can compromise biological interpretation and mask genuine spatial patterns.
Existing decontamination approaches take different strategies: some remove suspect expression, which reduces false positives but sacrifices true biological signal; others lack a principled, biologically grounded rule for deciding where a transcript truly belongs. These limitations motivated the development of a method that can make transcript-level decisions that are both accurate and traceable.
The authors present CellDot, a decontamination method based on an optimal transport formulation. CellDot operates at the molecule level and evaluates the plausible fate of each observed transcript. For every transcript, the method considers three possible outcomes: retain the transcript in its reported host cell, reassign it to a neighboring cell that is a better expression match, or remove it as background noise.
CellDot frames these choices within an optimal-transport problem that balances expression compatibility with spatial proximity and data-adaptive constraints. By treating transcripts as units that can be transported (retained, reassigned or removed), the framework yields a global solution that assigns transcripts in a way that is consistent with the observed spatial layout and expression references.
A core feature of CellDot is the integration of reference-guided expression compatibility with spatial information. The method leverages expression profiles or reference-guided signals to assess whether a transcript's gene identity is compatible with a candidate cell's expression program. Spatial proximity and local tissue geometry are incorporated to limit assignments to plausible neighboring cells and to penalize unlikely transports.
Data-adaptive constraints are applied so that the model conforms to the quality and properties of each dataset. This combination of expression-guided compatibility and spatial constraints produces molecule-level corrections that are both biologically informed and spatially coherent.
CellDot's output is a traceable decision for each transcript: keep the transcript in the original host cell, move it to a neighboring cell that fits expression and spatial constraints better, or delete it as background contamination. This fine-grained approach differs from methods that simply subtract an estimated contamination profile or discard entire categories of expression. The molecule-level decisions enable preservation of biologically meaningful variation while addressing contaminated measurements.
By providing explicit reassignments and removals, CellDot allows downstream users to inspect and audit the decontamination decisions, improving interpretability and reproducibility compared with more opaque approaches.
The authors evaluated CellDot across multiple human tumor datasets. According to the preprint, CellDot achieved superior performance compared with existing decontamination methods in restoring spatial expression patterns. The restored patterns were reported to match independent cross-platform measurements, indicating that CellDot's corrections recovered signals consistent with orthogonal data.
The manuscript emphasizes that CellDot enhanced recovery of biologically relevant signals in these tumor datasets, although specific quantitative metrics, dataset names and numerical results are reported in the source manuscript and not repeated here.
Beyond correcting individual transcript assignments, CellDot reportedly improved downstream biological analyses. The authors describe significant enhancement in the recovery of cellular states, intercellular communication signals and spatial niche programs after applying CellDot. These improvements suggest that accurate molecule-level decontamination can materially affect biological interpretation derived from spatial transcriptomics data.
The ability to recover refined cellular states and communication patterns is particularly relevant in complex tissues such as tumors, where microenvironmental interactions and spatial niches drive disease-relevant biology.
The authors demonstrate CellDot's scalability on real datasets and highlight that it is the only method, per their report, applicable to a whole-transcriptome Atera dataset among the methods they considered. This point underscores CellDot's computational design for handling large, high-resolution spatial transcriptomics datasets without the compromises required by less scalable approaches.
Detail on algorithmic complexity, runtime benchmarks and hardware used are presented in the full manuscript; readers should consult the preprint for those specifics.
This work is presented as a preprint on bioRxiv and has not been peer reviewed. The authors declare no competing interests. As with any preprint, conclusions and performance claims should be interpreted in the context of pending peer review and independent validation. For full methodological details, quantitative results, and dataset-specific evaluations, refer to the original manuscript.