Lab meetingLiterature reviewPart 2 of 2 · Downstream experimentsDraft · 27 October 2026
Spatial multiomics: downstream experiments after integration.
382 downstream experiments from 60 papers first published in 2025 and 2026, plus 163 self-tests: what each one did, on which data, with which metric, and where its number sits.
Thanh Bui · University of Central Florida · Fall 2026Press → to begin · ? for controls
01DataSection and spot
Two units: the section, and the spot inside it.
Every modality count in this deck is per spot, within one section.
Section (slice)
One thin slice cut from a tissue block and placed on one slide. A dataset is one section or a series of them; "slice" means the same.
Spot (pixel)
One capture site on that section: 100 µm in ST, 55 µm in Visium, 20 or 10 µm in DBiT-based assays. One spot is one row of every count matrix.
Modalities per spot
What one spot of one section reads: 1 in ST; 2 in most of our data (RNA + protein, or RNA + ATAC); 3 in spatial tri-omics; at most 5, in Spatial-Mux-seq.
Across sections
The P22 brain has ATAC and three histone marks, but each section pairs only one of them with RNA: 2 per spot, never 5. Joining sections that share no spot is the mosaic problem.
Read the next slides asmodalities in one spot of one section. Counts across a series of sections are not added up.
01DataOne spot, more layers
One spot, more layers.
The most modalities read in the same spot of one section went from 1 in 2016 to 2 in 2020 and 5 in 2025.
Most modalities in one spot
1
2016 · RNA only
2
2020 · RNA + 22 proteins
5
2025 · ATAC, H3K27me3, H3K27ac, RNA, 122 proteins
2023: ATAC or one histone mark joins RNA in the same 20 µm pixels of the P22 brain: 2 per spot, one pair per section.
Sep 2025: DNA methylation joins RNA in the same pixels, down to 10 µm: 2 per spot. The stack adds up assays; none reads all six.
Nov 2025: 3 per spot: RNA, about 150 proteins, and ATAC or H3K27me3, depending on the section; includes LPC lesions.
Ståhl et al., Science 2016 · Liu et al., Cell 2020 (DBiT-seq) · Zhang et al., Nature 2023 · Jiang et al., Nat. Methods 2023 (MISAR-seq) · Guo et al., Nat. Methods 2025 (Spatial-Mux-seq) · Lee et al., Nature 2025 (Spatial-DMT) · Zhang et al., Nature 2025 (spatial tri-omics). Full entries on the last slide of Part 1.
01DataSpot size
Smaller spots, while layers were added.
Spot or pixel size fell from 100 µm, about 3–30 cells, to 10 µm, about one cell.
100 µm spots · ST, 2016
55 µm spots · Visium · lymph node RNA + protein data
Left: a 1 mm × 1 mm field to scale; spacing drawn at twice the size (ST is 200 µm centre to centre, Visium 100 µm). Right: one point per assay; marker size = modalities on one section. Hover for details.
Sizes: Rao et al., Nature 2021 (ST 100 µm, Visium 55 µm); Vickovic et al., Nat. Methods 2019 ("100 µm (3–30 cells)"); assay papers on slide 3, plus Rodriques et al., Science 2019 (Slide-seq), Deng et al., Nature 2022 and Science 2022, Long et al., Nat. Methods 2024, Enninful et al., Nat. Methods 2026 (DBiTplus).
01DataSeries of sections
From one section to series of sections.
Datasets became time courses, lesion series and patient cohorts, so a method now meets many sections, not one.
MISAR-seq · embryonic mouse brain, RNA + ATAC · 2 sections per stage in our copy
E11.0
E13.5
E15.5
E18.5
Spatial tri-omics · mouse brain, ATAC or H3K27me3 + RNA + proteins · development, then LPC lesions
P0
P2
P5
P7
P10
P21
LPC 5 dpl
10 dpl
21 dpl
HTAN · 131 tumour sections from 78 cases, 6 cancer types · Visium, with snRNA-seq and CODEX on other material
Key points
Time: MISAR-seq covers E11.0, E13.5, E15.5 and E18.5, with 1,263–2,248 grids per section after filtering.
Lesions: tri-omics adds P0–P21 brains and LPC lesions at 5, 10 and 21 days after injection.
Cohorts: HTAN joins 131 sections with 48 snRNA-seq and 22 CODEX samples that do not share spots.
SoIntegration now has to line up sections across time, lesions and patients, not only fuse the layers of one spot. That is what the multi-section experiments in this deck test.
01DataLedger
Seven co-profiling assays; four feed our benchmark.
What one spot of one section reads, at what size, in which tissue. Sizes are pixel widths or spot diameters.
Assay · first online
In one spot
Per spot
Spot / pixel
Tissue
In our benchmark
DBiT-seq · 2020
RNA + 22 proteins
2
10–50 µm
mouse embryo, E10–E12
no
ATAC–RNA / CUT&Tag–RNA · 2023
RNA + ATAC, or RNA + one histone mark
2
20–50 µm
P22 mouse brain, E13 embryo, human hippocampus
P22: ATAC–RNA (labelled), 3 histone–RNA sections
MISAR-seq · 2023
RNA + ATAC
2
50 µm
embryonic mouse brain, E11.0–E18.5
8 slices, 2 per stage
Visium CytAssist RNA + protein · 2024
RNA + 35-antibody panel (4 isotype controls)
2
55 µm
human lymph node, sections A1 and D1
A1, D1 (labelled)
Spatial-Mux-seq · 2025
ATAC + 2 histone marks + RNA + 122 proteins
5
20–50 µm
E13 embryo, P21 hippocampus, EAE brain
no
Spatial-DMT · 2025
DNA methylation + RNA
2
10–50 µm
E11 and E13 embryo, P21 brain
no
Spatial tri-omics · 2025
ATAC or H3K27me3 + RNA + about 150 proteins
3
15–20 µm
developing mouse and human brain, LPC lesions
all slices, no labels
02CorpusWhat was read, and how
60 papers, 382 downstream experiments, each with its figure.
The data section showed what is now measured. The rest is the downstream experiments the papers run after integration: which ones, on which data, scored how.
Key points
Corpus: 60 papers from our catalogue first online in 2026 (42) or 2025 (18); 2024 papers are left out.
Access: 47 full texts from PubMed Central, 8 preprints, 4 open-access publisher pages and 1 from figure legends only.
Unit: one downstream experiment is one analysis of the integrated output on named data, with its metric, number and figure. Self-tests (simulation, ablation, robustness, runtime, uncertainty) are counted apart: 163.
Checks: 379 of 380 extracted quotes match the saved full text word for word; every number in a result appears there too, except 2 SMOPCA numbers that sit in unsaved supplementary tables.
How the deck runs
Data, corpus. How the data grew; what gets tested, on which data.
Recipes. Seven downstream experiments: a recipe, then its metrics and which way is better.
More tasks, evaluation. Same format: what they do, what it tells you, metrics.
Applications, review. Benchmarks, what is new, a checklist.
Not yet in the counts29 more papers from 2025 and 2026 are still to be read: mostly bioRxiv preprints (rate-limited) and paywalled journals. The counts will move once they are in.
02CorpusWhat gets tested
Domains first, then the papers test themselves.
Papers (of 60) that run each downstream experiment at least once, by first-online year. Below the line: evaluation experiments, which are nearly as common.
Read the chart
Domains: 45 of 60 papers cluster the joint embedding and score it (98 experiments).
Rare: trajectories 8, cell–cell communication 9, uncertainty 3 papers.
02CorpusWhich data carry the evidence
The same few tissues carry most of the evidence.
Papers (of 60) using each dataset family, by keyword match on the dataset labels in the records (blue 2026, orange 2025). Families overlap; one paper can use several.
Dataset family
Papers
Mouse brain, spatial ATAC or CUT&Tag with RNA (P21/P22 and others)
29
Spleen (SPOTS and others)
29
Mouse embryo sections (MISAR-seq, Stereo-seq and others)
27
Human lymph node (SpatialGlue A1, D1)
26
Simulated or semi-synthetic data
26
Human tonsil
20
Mouse thymus (Stereo-CITE-seq)
13
Human DLPFC (12 sections)
5
Why it matters
Comparable: shared sections let a reader set numbers side by side, if preprocessing and K match.
Over-used: the same lymph node, brain and embryo sections tune and test most methods (inference).
Reference rows: MultiSpaceNet adds a naive concatenation (mean ARI 0.422) and an untrained network (0.310) next to its model (0.489), Table 2.
03Recipe 1 of 745 of 60 papers · 32 from 2026
Spatial domains: cluster, map back, score.
The most common experiment. Cluster the joint embedding, colour the tissue by cluster, then score the map with labels and without them.
Recipe
Cluster. mclust on the joint embedding (SpatialGlue, MultiSP) or Leiden (COSMOS). MultiSP sets K to the number of annotated regions; SpatialGlue chose K by hand.
Map back. Colour each spot by its cluster at its (x, y).
Score with labels. ARI and NMI against the manual annotation; report the mean over seeds.
Score without labels. Moran's I of each cluster's map (SpatialGlue, via Squidpy): do neighbouring spots agree?
Worked example COSMOS on P22: ARI 0.63, against 0.43 for SpatialGlue (Fig. 3c). SpatialGlue on lymph node A1: median Moran's I 0.62, with 6 clusters against 10 annotated categories (Fig. 2e, g).
Pitfall The annotation is one view of the tissue: in lymph node A1 the paracortex "more resembled the RNA-specific and protein-specific modalities than the H&E annotation" (SpatialGlue).
Long et al., Nat. Methods 2024 (SpatialGlue), Fig. 2e, g, Results and Supplementary Information · Zhou et al., Nat. Commun. 2025 (COSMOS), Fig. 3c · Mo C, Zou X, Jin S. Cell Genomics 6, 101141 (2026) (MultiSP), doi:10.1016/j.xgen.2026.101141 · Hubert & Arabie, J. Classif. 1985 (ARI) · Moran, Biometrika 1950 · Scrucca et al., R J. 2016 (mclust).
03Recipe 1 · Spatial domainsHow 2025–26 papers score them
ARI still leads; label-free scores are spreading.
Scores named in the domain experiments of the 45 papers that run them, by keyword. One paper can use several.
Score
Papers
Better when
ARI
33
Higher; 1 = labels, 0 = chance
NMI
26
Higher; 1 = labels
AMI
18
Higher; 0 = chance
Homogeneity
13
Higher: clusters are pure
Moran's I
11
Higher, short of over-smoothing
V-measure
10
Higher
FMI
5
Higher
Silhouette / ASW
5
Higher: compact clusters
Completeness
4
Higher: labels not split
Beyond one ARI
Reference rows: MultiSpaceNet scores a naive concatenation (mean ARI 0.422) and an untrained network (0.310) against its trained model (0.489), Table 2.
Seeds: SCIGMA reports pairwise ARI over 20 seeds; on brain it is 7% above SpatialGlue (Fig. 2E).
K matters: PRAGA's ARI is 28.28% with the true number of clusters and 24.19% at C = 8 (Table 3).
No labels: smiMAD uses MMSCS and NMMI; SpaHDSRL uses five coherence scores such as Moran's I and Geary's C.
Caveats: CoMo reads a top Moran's I as over-smoothing; in LATTICE, ARI against an RNA-based reference fell as epigenomic channels were added (Section 4.2).
Suggested practice (inference)Give ARI with K and its choice stated, the spread over seeds, one label-free score, and a check for over-smoothing.
03Recipe 2 of 719 of 60 papers · 14 from 2026
Modality weights: which layer led each spot.
Read the per-spot weights the model learned, map them, average them per domain, and test what they claim.
Recipe
Extract. Per-spot weights: SpatialGlue's between-modality attention is a softmax, so the weights sum to 1; COSMOS uses WNN weights.
Map. Colour each spot by its RNA weight against the other modality's.
Summarise. Per domain, as split violins (SpatialGlue Fig. 3e) or mean bars; a domain led by one modality is a claim to check.
Check. Rerun without that modality (MISO reruns every combination), or map private factors (SpaMV).
Worked example SpatialGlue, P22 ATAC–RNA: RNA weighs more in the corpus callosum, ATAC in cortex and lateral ventricle (Fig. 3e). SpaMV, E13 embryo: the H3K27ac-private topic sits in the spinal cord, the H3K27me3-private topic in jaw and nose (Fig. 4e).
Pitfall SpatialGlue's weights "are calculated based on the latent embeddings of each modality, rather than raw feature matrices", so some discrepancy with the raw data is expected. MISO "cannot directly quantify the contribution of each modality".
Long et al., Nat. Methods 2024 (SpatialGlue), Fig. 3e and Results · Zhou et al., Nat. Commun. 2025 (COSMOS), Fig. 3d · Liu et al., Nat. Commun. 2026 (SpaMV), Fig. 4e · Coleman et al., Nat. Methods 2025 (MISO), Discussion, Extended Data Figs 8–9.
Modality weights are a claim until a modality is broken.
12 of the 22 experiments report a number; the rest show maps or plots only.
What they do
Read the weight or attention the model gives each modality per spot or cluster, map it on the tissue, and compare it with where each modality carries signal. A few papers corrupt or drop one modality and check that its weight falls.
What it tells you
Whether the model knows which modality to trust where. A weight map is the model's own claim; only a perturbation tests it.
Metric
Papers
Better when
Weight maps and distributions
9
Not scored; read by eye
Correlation with modality signal (PCC, Spearman)
6
Higher: weight follows signal
Clustering with and without a modality (ARI, NMI)
2
Larger drop when the high-weight modality is removed
Paired tests on weights (Wilcoxon)
2
Lower p: weights differ where expected
ExampleMultiSpaceNet corrupted one modality inside a region; the attention weights did not follow it (AUC 0.511 for RNA corruption; Section 3.4, Table S25).
03Recipe 3 of 733 of 60 papers · 22 from 2026
Domain markers: test, rank, map back.
How a domain map gets biological names. Run it for every modality: genes, peaks, proteins.
Recipe
Contrast. One domain against all the others.
Test every feature. Wilcoxon rank-sum: Seurat FindAllMarkers for genes, ArchR for peaks (SpatialGlue); adjust for multiple testing.
Rank. Top markers per domain by fold change, shown as a dot plot or heat map.
Map back. The top marker should outline its domain and match known anatomy.
Worked example SpatialGlue, P22 H3K27ac–RNA: Tspan2, Cldn11 and Ugt8a mark the corpus callosum and anterior commissure; Olfm1, Cux2 and Rorb mark cortical layers (Fig. 3j).
Pitfall Double dipping: when the groups come from clustering the same data, a classical test "yields an extremely inflated type I error rate" (Gao et al. 2024); use p-values to rank, not to prove. Test every modality: in SpaMV's thymus, RNA-only markers would mislabel a protein-defined cluster as T-cell-enriched (Supp. Fig. 19).
Long et al., Nat. Methods 2024 (SpatialGlue), Fig. 3j and Supplementary Information · Liu et al., Nat. Commun. 2026 (SpaMV), Supp. Fig. 19 · Gao LL, Bien J, Witten D. Selective inference for hierarchical clustering. J. Am. Stat. Assoc. 119, 332–342 (2024), doi:10.1080/01621459.2022.2116331, abstract.
Markers name each domain with what it over-expresses.
19 of the 39 experiments report a number; the rest show maps or plots only.
What they do
Test each domain against the rest for genes, peaks, proteins or metabolites, map the top features on the tissue, and run pathway enrichment.
What it tells you
Turns cluster numbers into biology, and shows whether the second modality adds markers the first one misses.
Metric
Papers
Better when
Differential test (p, FDR, Mann–Whitney)
7
Lower p after correction
Fold change (log2FC)
4
Larger, in the expected direction
Pathway or GO enrichment
3
Significant and matches known tissue function
Spatial autocorrelation (Moran's I)
3
Higher: the marker forms a region
Correlation across modalities
2
Higher: gene and protein agree
ExampleIn rheumatoid arthritis synovium, mean local Moran's I of RSPO3+ fibroblasts is 64.12 against 9.75 in healthy tissue (Fig. 7B).
03Recipe 4 of 713 of 60 papers · 10 from 2026
Cross-modal imputation: hide, predict, compare.
Hold out a modality where it was measured, predict it from the rest, and score the prediction against the measurement.
Recipe
Hold out. Remove protein or ATAC from one section and train on the others; SpaMosaic uses each of 3 sections as the test set once.
Predict. From RNA, through nearest neighbours in the shared latent space (SpaMosaic) or a regressor on an RNA-only embedding (MultiSpaceNet).
Compare. Pearson correlation per feature and per spot; AUROC on binarised peaks for ATAC (SpaMosaic).
Baseline. Beat a simple predictor on the same split, such as a Ridge regression on raw RNA.
Worked example MultiSpaceNet, spleen replicate 2, five-fold cross-validation: mean Pearson r 0.651, against 0.547 for SpatialGlue and 0.582 for a Ridge regression on raw RNA (Table 3).
Pitfall Leakage: an embedding fused with protein has "already consumed the protein it would be asked to predict" (MultiSpaceNet). The ground truth matters too: smoothing sparse ATAC raised SpaMosaic's spot-level PCC by 34.4% (preprint Supp. Fig. S34b).
Imputation predicts the hidden modality; leakage decides if it counts.
14 of the 26 experiments report a number; the rest show maps or plots only.
What they do
Hide one modality on held-out spots, sections or features, predict it from the others through the shared space, and compare with what was measured.
What it tells you
The joint space carries real cross-modal information, which would let a cheap assay stand in for an expensive one. It counts only if the hidden data never entered training.
Metric
Papers
Better when
Pearson correlation (PCC)
10
Higher; 1 = exact up to scale
Error (RMSE)
4
Lower
Rank or angle (Spearman, cosine)
3
Higher
Spatial PCC (SPCC)
3
Higher: the map matches, not only the levels
Domains from imputed data (ARI, NMI)
1
Higher: imputed data still separates domains
ExampleLeakage audit: an early-fusion embedding "predicts" protein at 99.4% of a Ridge model that is given the protein itself (MultiSpaceNet, Section 3.6).
03Recipe 5 of 721 of 60 papers · 16 from 2026
Multi-section: mix the sections, keep the domains.
Score two axes at once: are the sections mixed, and are the domains still apart? Then check that matching regions line up.
Recipe
Embed. Integrate every section into one embedding.
Mixing. iLISI, kBET or batch ASW, with points labelled by section.
Biology kept. ARI, NMI or cLISI, with points labelled by domain; scIB-style scores weight the two 40/60 (PRESENT).
Align. Check that matching regions line up across sections, by registration or optimal transport.
Worked example SpaMosaic, embryonic brain, 3 sections: MIDAS ranked first on iLISI and SpaMosaic fourth (preprint Fig. 3g), yet MIDAS's clusters "showed substantial overlapping".
Pitfall Mixing alone is not success. Scales differ too: SMART normalises iLISI to 0–1 on graph distances, so values do not compare across papers; and single-section ARI "does not, by itself, test integration" (MultiSpaceNet).
Yan et al., Nat. Genet. 2026 (SpaMosaic), preprint v3 Fig. 3g · Li Z et al. Brief. Bioinform. 27, bbag214 (2026) (PRESENT), doi:10.1093/bib/bbag214 · Du Z et al., Nat. Commun. 17, 2876 (2026) (SMART) · Zhang et al. 2026 (MultiSpaceNet) · Korsunsky et al., Nat. Methods 2019 (LISI) · Büttner et al., Nat. Methods 2019 (kBET) · Luecken et al., Nat. Methods 2022 (scIB).
Multi-section integration: mix the batches, keep the biology.
25 of the 46 experiments report a number; the rest show maps or plots only.
What they do
Integrate several sections into one embedding, or register sections and modalities into one coordinate frame. Then score batch mixing, kept biology and alignment error.
What it tells you
Atlases and time courses need one space across sections. Mixing and conservation pull against each other, so a method has to report both.
Metric
Papers
Better when
Biology kept (ARI, NMI, cLISI, domain ASW)
10
Higher
Batch mixing (iLISI, kBET, batch ASW)
8
Higher, unless biology is lost
Overlap and accuracy (Dice, accuracy, AUROC)
7
Higher
Correlation of matched features (PCC, cosine)
4
Higher
Registration error (TRE, Hausdorff, landmarks)
3
Lower
ExampleMGW scores metabolite-to-Visium alignment by whether dopamine lands on dopamine-high spots: AUROC up to 0.995 (Fig. 6b).
03Recipe 6 of 78 of 60 papers · 5 from 2026
Trajectories: order spots in time and space.
Pseudotime orders spots within one section; optimal transport links sections from successive stages.
Recipe
Root. Pick the start from biology: "the first cell of a given cell type" (COSMOS), or progenitor states (CoMo).
Pseudotime. Diffusion pseudotime (scanpy DPT) on the joint embedding (COSMOS, CoMo), or PAGA (SpatialCOC); map it on the tissue.
Check. Genes correlated with pseudotime should be known markers; COSMOS found 577 on P22 (Fig. 3g–i).
Across stages. Couple spots of successive time points by optimal transport on expression and spatial structure (moscot).
Worked example CoMo, E13 embryo ATAC–RNA: pseudotime runs from the ventricular zone to subpallium to pallium (Fig. 4e). moscot, mouse embryo E9.5–E16.5: annotation transitions 5% and 13% more accurate than moscot.time and TOME (Fig. 4b).
Pitfall On one section, pseudotime is a spatial gradient, not elapsed time, and neither COSMOS nor CoMo scores it with a number. moscot's couplings "probabilistically link cells": they are model output, not lineage.
Zhou et al., Nat. Commun. 2025 (COSMOS), Methods, Fig. 3g–i · Gui Y et al. Brief. Bioinform. 27, bbag192 (2026) (CoMo), doi:10.1093/bib/bbag192, Fig. 4e · Li M et al., Nat. Commun. 17, 5268 (2026) (SpatialCOC) · Klein D et al. Mapping cells through time and space with moscot. Nature 638, 1065–1075 (2025), doi:10.1038/s41586-024-08453-2, Fig. 4b, Methods · Haghverdi et al., Nat. Methods 2016 (DPT).
Trajectories order states in time; the best papers test a driver.
8 of the 15 experiments report a number; the rest show maps or plots only.
What they do
Infer pseudotime inside a section, or map cells across time points with optimal transport, and read which cell types turn into which.
What it tells you
Whether the integrated space follows development or disease progression. A predicted driver tested in the lab is the strongest evidence.
Metric
Papers
Better when
Maps only
4
Not scored; read by eye
Correlation with known time or genes
2
Higher
Matching scores (cosine, cell-type matching)
2
Higher
Transition or descendancy scores
2
Higher: known lineages recovered
Examplemoscot's NEUROD2 prediction was tested by knockout in iPS-cell-derived islets: fewer ghrelin-expressing cells (Fig. 5h,i).
03Recipe 7 of 716 of 60 papers · 11 from 2026
Regulatory links: peaks to genes, motifs to domains.
Needs chromatin and RNA on the same spots, or an imputed pair. Correlate, compare with a matched null, then validate outside the data.
Recipe
Link. Correlate each peak's accessibility with a nearby gene's expression across spots (Signac LinkPeaks in SWITCH), or learn the links by attention within 150 kb (MultiGATE).
Null. Compare each correlation with the one expected for peaks of the same GC content, accessibility and length (Signac).
Motifs. TF motif deviations per spot with chromVAR (SWITCH), averaged per domain.
Validate. Against outside data: eQTL, enhancer atlases, ChIP-seq (MultiGATE).
Worked example MultiGATE, human hippocampus ATAC–RNA: its peak–gene links recover GTEx eQTL with AUROC 0.703, against 0.530 for Cicero and 0.515 for Spearman correlation (Fig. 2d).
Pitfall Two features that both mark one domain correlate with no regulatory link between them. SWITCH's authors concede motif plus correlation "may not allow unambiguous assignment of regulatory roles" (Peer Review File), and its links use imputed ATAC.
Miao J et al., Nat. Commun. 16, 9403 (2025) (MultiGATE), doi:10.1038/s41467-025-63418-x, Fig. 2d, Methods · Li et al., Nat. Comput. Sci. 2025 (SWITCH), Supplementary Information and Peer Review File · Stuart et al., Nat. Methods 2021 (Signac; LinkPeaks documentation) · Schep et al., Nat. Methods 2017 (chromVAR).
Denoising gives cleaner maps; confirm the finding in raw data.
3 of the 11 experiments report a number; the rest show maps or plots only.
What they do
Denoise or sharpen the measured features through the joint model and show maps or domains before and after.
What it tells you
Can rescue sparse protein or accessibility signal, but a smoother map is also what over-smoothing looks like.
Metric
Papers
Better when
Share of locations above a threshold
2
Higher
Spatial autocorrelation (Moran's I)
2
Higher, but too high means over-smoothed
Differential tests after denoising
1
Lower p, then confirm in raw data
Domains after denoising (ARI)
1
Higher
ExampleSOTMGF: CD117 and Cpt1a become differential only after denoising, and both associate with survival (p = 0.016 and 0.0032, Fig. 2F). Check such findings in raw data.
Cell types are scored against an independent reference.
15 of the 18 experiments report a number; the rest show maps or plots only.
What they do
Assign cell types, map a single-cell reference onto the tissue, or deconvolve spots. Then compare with expert labels, matched single-cell data or a pathologist.
What it tells you
Most later biology rests on the labels. A second modality should fix calls that RNA alone gets wrong.
Metric
Papers
Better when
Agreement of proportions (correlation)
4
Higher
Label accuracy (F1, precision, recall)
2
Higher
Proportion error (RMSE, JSD)
2
Lower
Clustering agreement (NMI, ARI, homogeneity)
1
Higher
ExampleTACIT on the same cells: protein and RNA cell-type calls agree on about 34% of cells with full panels, 81% with shared markers (Fig. 7).
Cell–cell communication: few papers and no ground truth.
4 of the 10 experiments report a number; the rest show maps or plots only.
What they do
Run ligand–receptor tools such as CellChat on the integrated labels and neighbourhoods; some take the receptor from measured protein.
What it tells you
Proposes which neighbouring cells signal to each other. With no ground truth, papers lean on permutation tests and the literature.
Metric
Papers
Better when
Interaction strength, number, probability
3
Descriptive; larger means more signalling
Permutation test
2
Lower p
Spatial or neighbourhood correlation
2
Higher
Receptor measured as protein
1
Higher protein support for the pair
ExampleMultiSP takes the ligand from RNA and the receptor from measured protein for CellChat; receptor levels are higher than with RNA alone (Fig. 6E,H).
Everything else: platform checks and lab validation.
15 of the 29 experiments report a number; the rest show maps or plots only.
What they do
Cross-platform agreement on the same tissue, a metabolite correlated with gene programmes, knockouts of predicted drivers, and spatial methods run on non-spatial data.
What it tells you
The checks that make the rest believable: platforms agree before biology is read, and predictions are tested outside the computer.
Metric
Papers
Better when
Correlation (Pearson, Spearman)
7
Higher
Clustering agreement (ARI, NMI, AMI)
5
Higher
Wet-lab readout (knockout, mouse model)
2
Effect in the predicted direction
Matching (accuracy, FOSCTTM)
1
Higher accuracy, lower FOSCTTM
ExampleThree platforms on the same synovium agree on disease fold changes at Pearson r of 0.75 or higher before any biology is read (Methods 2.3.6).
05EvaluationHow papers test their own models
Most papers stress-test; almost none quantify doubt.
Self-evaluation experiments across the 60 papers, and what the 60 robustness experiments vary (keyword classes, first match wins).
Kind
Of 60 papers
Simulation
22 papers · 26 exp.
Ablation
38 papers · 45 exp.
Robustness, sensitivity
39 papers · 60 exp.
Runtime, memory
29 papers · 29 exp.
Uncertainty
3 papers · 3 exp.
Robustness: what is varied
Exp.
Hyper-parameters, K, graph size
28
Noise, dropout, corruption
12
Wrong input: rotation, references, preprocessing
10
Missing modality or data
5
Seeds, repeats, bootstrap
3
Other
2
Gap3 papers map uncertainty, and none of the 545 downstream or evaluation experiments tests whether it is calibrated.
Robustness: how fast the score falls under stress.
28 of the 60 experiments report a number; the rest show maps or plots only.
What they do
Vary hyper-parameters, add noise or dropout, mask a modality, change seeds or feed a deliberately wrong input, and track the score.
What it tells you
Whether a result survives settings a user would not tune. A flat curve matters more than the best point.
Metric
Papers
Better when
Clustering under stress (ARI, NMI)
23
Flatter curve, higher retention
Task metric under stress (PCC, RMSE, accuracy)
8
Stays near the clean value
Spread across runs (variance, IQR, pairwise ARI)
7
Lower spread
ExampleSubPoE reports retention: under the strongest noise it keeps 71.8% of its ARI (Fig. 7a–c).
05Evaluation · Runtime and memory29 evaluation experiments · 29 papers · 16 from 2026
Runtime and memory: report the largest section that runs.
17 of the 29 experiments report a number; the rest show maps or plots only.
What they do
Time and peak memory on growing numbers of spots or cells, often against the same baselines.
What it tells you
Whether the method fits atlas-scale and subcellular data on ordinary hardware.
Metric
Papers
Better when
Time (seconds, minutes)
27
Lower; flat growth with size
Peak memory (GB, RAM, VRAM)
17
Lower
Largest size that runs
1
Higher
ExampleSCIGMA was the only method to run on 500K–1M locations; the others failed within 48 h (Fig. 5).
05Evaluation · UncertaintyNew · 3 papers, all 2026
Uncertainty: mapped in three papers, checked in none.
A map shows where the model doubts. Only a calibration test shows whether to believe it.
What they do
Next to each spot's label, the model reports how sure it is, and the paper paints that on the tissue.
Where the doubt sits
BaySC (arXiv) derives it from posterior co-membership. It is high in the lymph-node sinus and trabeculae and on brain-region boundaries at E15.5, and near zero in cortex and follicles (Fig. 4a-b).
SCIGMA (Nature Genetics) finds it highest where RNA and ATAC disagree. SubPoE splits it into molecular and spatial experts on tonsil (Fig. 6a–d).
What it tells you
That doubt falls where the biology is ambiguous. It does not tell you whether a doubtful label is more often wrong, which is what a user needs before trusting the map.
Missing recipe (inference)Sort spots by uncertainty and plot the error rate against expert labels, as classifiers report calibration. None of the 545 experiments does this.
06ApplicationsTissue studies with a second modality
Applications check the platforms, then read the biology.
Four 2025–26 studies where the experiments are the point: cross-platform checks, same-section metabolites, a negative control.
RA synovium (Front. Immunol. 2026): Xenium 5K, CosMx 1K and Visium agree on cell-type proportions (rho 0.89–0.92, Supp. Fig. 2B); immune cells are about 54% of RA cells and macrophages rise 11-fold (Fig. 2D).
Diabetic PDAC (Cancer Res. 2026): MALDI-MSI then Visium on the same section; cholesterol is higher in diabetic tumours (P = 0.0128, Fig. 1C–I), followed up with Mendelian randomisation and mouse perturbations.
MAGPIE (Nat. Commun. 2026): co-registers same-section Visium and MSI with accuracy at least 0.9 (Fig. 2f); the factor most correlated with an inhaled drug is enriched for immune and defence responses (Fig. 4b–d).
MIIT (GigaScience 2025): in prostate, dropping spots whose tissue type disagrees raises the citrate correlation from 0.302 to 0.437; a 180°-rotated section serves as a deliberately wrong integration (Figs 5–6).
PatternThe applied studies spend their first figures proving the modalities line up; most method papers skip that step.
06Applications · Disease and clinical27 downstream experiments · 22 papers · 17 from 2026
Disease: separate the conditions, then predict the outcome.
16 of the 27 experiments report a number; the rest show maps or plots only.
What they do
Compare disease with healthy tissue, score niches or programmes, and test them against survival in outside cohorts such as TCGA.
What it tells you
The clinical case for integration: a niche or pathway that only the combined modalities reveal.
Metric
Papers
Better when
Pathway enrichment (NES)
6
Larger |NES|, significant
Group tests (Wilcoxon, t, Cauchy)
5
Lower p
Survival (log-rank)
4
Lower p: groups differ in outcome
Status recovery (AUC, F1, ARI, NMI)
3
Higher
ExampleSpaMTP finds 539 pathways in lesioned striatum that neither genes nor metabolites show alone (Fig. 2j–n).
06ApplicationsBenchmarks and platform comparisons
Benchmarks: no single winner, and platforms differ first.
Two 2025–26 benchmarks in the corpus. Four more (scMultiSim, SPATCH, SCMMIB, Giotto Suite) are not read yet.