PRISM Targets Two Artifacts That Distort Spatial Transcriptomics at Once
Spatial transcriptomics has made it possible to read gene expression in tissue while preserving anatomy, but the data rarely arrive cleanly. Two artifacts frequently travel together: ambient RNA that leaks into nearby spots or cells, and doublets created when segmentation captures more than one cell in a single measurement. PRISM was built to address both problems in a single probabilistic framework rather than forcing analysts to solve them one after the other.
The method uses an Expectation-Maximization approach to assign each cell a probabilistic cell-type identity, estimate a per-cell contamination fraction α, and detect doublets through log-likelihood ratio testing of two-component mixtures. In practical terms, the model is trying to answer three questions at once: what cell type is most likely present, how much foreign RNA has entered the profile, and whether the profile really belongs to one cell or to a merged pair. In tissue regions where cells are tightly packed, especially in brain layers with dense architecture, those questions are inseparable. A correction strategy that ignores one artifact can easily mistake contamination for biology or misread mixed profiles as rare cell states.
Simulated Benchmarks Show Exact Recovery Across Broad Noise and Depth Ranges
Performance was first tested in silico on 40,800 simulated cells. Across contamination levels ranging from 0.05 to 0.50 and UMI depths from 500 to 5,000, PRISM recovered α with Pearson r=1.000. The simulation also produced zero doublet false positives and 100% singlet accuracy.
Those results matter because the benchmark spans the kinds of sequencing depth and contamination burdens that routinely complicate spatial experiments. A decontamination tool that performs well only at one depth or under low background noise can look elegant in a narrow test set while failing on real tissue. PRISM’s behavior across a wide range of conditions suggests that its joint modeling strategy can keep contamination estimates stable even when expression counts fluctuate. The perfect singlet classification in simulation also indicates that the doublet test is not simply overcalling mixed profiles to improve apparent cleanup.
A single-pass model changes the error structure. Once ambient RNA and doublets are estimated together, one artifact no longer has to be absorbed as unexplained variance by the other.
Cerebellar Tissue from an MLD Model Exposes the Cost of Uncorrected Background Signal
The most clinically relevant demonstration came from a Visium HD mouse cerebellum dataset drawn from a metachromatic leukodystrophy MLD model. The dataset included four samples, 125,289 cells, and 17 reference cell types. PRISM identified a median α=0.60, indicating a substantial contamination burden across the tissue profiles. It also flagged 67.4% doublets, driven by the densely packed cerebellar granule layer.
The high doublet rate in that region is biologically plausible. Granule cells are tightly arranged, and dense cellular neighborhoods increase the likelihood that segmentation boundaries will split poorly or merge neighboring cells into a single expression profile. In a disease setting such as MLD, where myelin biology is central to interpretation, dense cellular architecture can distort the apparent contribution of glial populations if ambient RNA and mixed-cell profiles are not corrected before downstream analysis. PRISM’s spatially aware framework is designed for exactly that kind of tissue-level complication.
Myelin Signal Sharpened After Decontamination
The most concrete effect of correction was seen in lineage-specific signal cleanup. PRISM eliminated spurious myelin signal from microglia, reducing it from 0.82% to 0.00%. It also doubled oligodendrocyte myelin purity from 2.8% to 5.6%.
Those shifts are small in absolute percentage terms but large in interpretive value. Microglia should not present as meaningful myelin-bearing cells, so any apparent myelin signature in that compartment risks misclassifying background RNA as cell identity. Oligodendrocytes, by contrast, are the cell type where myelin-related expression should be concentrated, so increased purity suggests a more faithful biological profile after correction. In tissue studies of demyelinating disease, false myelin signal in the wrong compartment can flatten cell-type boundaries and obscure which glial population is truly changing. Cleaning the matrix improves not just signal-to-noise ratio, but the logic of the cell atlas itself.
Practical Speed Keeps the Method Usable on Routine Hardware
PRISM completes its workflow on commodity CPUs in under 15 minutes and returns decontaminated integer count matrices for every measured gene. That computational profile matters because spatial transcriptomics analysis often involves large tissue sections and multiple samples, and methods that require specialized hardware can be difficult to adopt in routine pipelines. A fast CPU-based approach lowers the barrier for laboratories that need a direct output matrix they can pass into standard downstream analysis without rebuilding their workflow around a separate correction engine.
The integer-count output is also important. Downstream differential expression, clustering, and cell-state inference methods often assume count-based data structures. By producing a decontaminated matrix rather than only summary scores, PRISM fits into existing analysis habits while altering the underlying signal that drives them.
Joint Modeling Fits the Biology Better Than Sequential Cleanup
The appeal of PRISM is not only technical efficiency. It matches the structure of the problem. Ambient RNA and doublets are not independent nuisances; they are intertwined sources of error that can reinforce each other in densely packed tissue. A cell profile contaminated with nearby RNA may appear to express a hybrid state, while a genuine doublet can exaggerate marker overlap and inflate cross-lineage contamination estimates. Joint probabilistic modeling reduces the chance that one artifact is misread as the other.
The method still needs validation beyond a single preprint and beyond the mouse cerebellum setting used here. The data presented do not establish how PRISM behaves across other organs, other disease models, or samples with different tissue densities and segmentation quality. Even so, the current results define a clear use case: spatial transcriptomics datasets in which contamination is substantial, cell density is high, and lineage boundaries are biologically important enough that mixed profiles would distort interpretation. For those cases, simultaneous decontamination and doublet correction is not a luxury feature. It is the difference between a map that reflects tissue biology and one that still carries the noise of the measurement process.