acceptodds
Under review as a conference paper at ICLR 2027

DeSpotX: Identifiability-Based Decontamination for Spatial Transcriptomics

Abstract

Spatial transcriptomics (ST) at single-cell resolution profiles gene expression in its native spatial context, but a substantial fraction of transcripts contaminate neighboring cells, compromising downstream biological analyses. Existing decontamination methods either ignore the spatial structure of contamination or aggregate over neighbors without constraining how contamination is separated from native expression, leaving the decomposition ambiguous. To resolve this ambiguity, we introduce DeSpotX, a deep generative model that uses anchor genes, defined as genes not natively expressed in a given cell cluster, to constrain the contamination decomposition. DeSpotX further uses spatial information to estimate contamination locally through a cluster-masked, distance-weighted average over neighboring cells. On spike-in simulations across five datasets and four ST platforms, DeSpotX attains the highest AUROC in 27 of 30 benchmark settings, exceeding the best baseline by up to 0.12, and remains robust to inaccuracies in the cell-cluster annotation and in anchor gene construction. Using matched scRNA-seq as an independent reference for cell-type-specific expression, DeSpotX retains 94% of within-type signal while improving marker-gene specificity. Decontaminated expression also shows greater spatial coherence and yields cell-cell communication networks consistent with known biology. Finally, iterating between decontamination and cell-cluster annotation refines these outcomes, reassigning ligand-receptor signaling to expected source cells in mouse brain and breast cancer tissues.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
spatial transcriptomics, decontamination, identifiability, deep generative models, single-cell genomics

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over one review, above 85% of the others.
Reviewer_fSmoOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this study, the authors introduced DeSpotX, a deep generative model to address contamination in single cell resolution spatial transcriptomics. It integrated anchor genes with spatially informed contamination estimation to distinguish native gene expression from the transcripts from neighboring cells. The authors performed extensive benchmarking across several datasets and technologies, and showed it outperforms existing methods. The authors further show that decontamination improves downstream analysis including marker gene specificity, spatial coherence and inferred cell-cell communication.

Critical strengths

The method is well designed and novel. Most of the claims are well supported by the experiments.

Critical weaknesses
  1. The deep investigation of the anchor gene selection is lacking.
  2. There is potential circularity in validation.
  3. The spatial validation in the higher resolution is missing.
  4. There are some caveats in cell-cell communication analysis.
  5. The quantitative evaluation for clustering needs more metrics.
Questions for authors
  1. About the anchor genes, I wonder if the number of anchor genes should be adjusted by the size of the clusters. Or whether these two are highly correlated? Whether the cluster with a higher number of cells should have a higher number of anchor genes?
  2. Whether the anchor genes used in the model are overlapped with marker genes which were shown in Figure 3? If overallped, it’s a circularity about the argument. The validation or the conclusion drawn in Figure 3 should be independent of the anchor genes. Or the anchor genes used in the model should be removed for all downstream analysis.
  3. For clustering shown by UMAP in Figure 3, I wonder if there is a quantitative way to show the separation of clusters after decontamination is better than before.
  4. Table 10, I wonder if the bold column should be the “anchor” one?
  5. For Figure 4, the authors showed the expression is spatially coherent, but some panels are full of signals, for example, panels b and d, which makes it hard to visually see. I wonder if the high-resolution cell type or subtype level analysis will be clear to showcase the spatial coherence. For instance, in MERFISH data, there is subtype level annotation for all cell types, the bottom panel could be replaced by some sparse cell subtypes. Overall, the evaluation on the subtypes or high-resolution level annotations would significantly improve the study.
  6. About the cell-cell communication analysis, I wonder if CellChat analysis used spatial information. If not, I suggest the authors use a spatial-aware method to do the analysis. Additionally, the results of changes between iterations are not convincing to make any conclusions since there is no ground truth to support which one is better or real. For example, in Figure 5e, the interactions on non-neurons are reduced after decontamination, which may be biased by the remaining signals. The authors should provide more analysis to show which interactions are shifted and which interactions keep left in the decontaminated dataset, and look for literature evidence to support the findings if possible. Otherwise, it’s hard to make a solid conclusion from the predicted CCC shift.
  7. It would be helpful to have some explanation for Figure 1 in the caption.
5 comments
AuthorsOpenReview ↗

Response to Reviewer fSmo (Part 1/2)

Thank you for the many concrete suggestions! We are encouraged that you found the method well designed and novel, and most of the claims well supported by the experiments. We address each question below, together with the weakness it relates to.

1. Anchor gene selection and cluster size (Weakness 1; Question 1)

Thank you for this question. The number of anchor genes does not need to be adjusted by cluster size, since the two are not correlated and larger clusters do not benefit from more anchors. We added three analyses to show this.

First, the number of anchors does not depend on cluster size. Over the 103 clusters with at least 20 cells in the five real datasets, the Spearman correlation between the two is 0.02 (p = 0.86). Cluster sizes vary up to 1,056-fold within a dataset, but anchor counts only 1.6-4.5-fold, since an anchor is selected by the fraction of a cluster's cells that detect a gene, not by their number.

Second, a larger cluster does not need more anchors. We scored the paper's runs with one anchor per cluster (Figure 2c) separately for each cluster. Compared with all anchors, AUROC changes by at most 0.004 in both large and small clusters, including the largest ones, since Lemma 2 needs, for each cell, only one anchor that receives contamination, whatever the size of its cluster.

Third, giving larger clusters more anchors lowers accuracy. To test this directly, we reassigned the paper's anchor counts so that larger clusters receive more anchors and smaller clusters fewer, with the same total, and retrained DeSpotX. At 20% contamination, relMAE rises from 0.494 to 0.526 on Xenium_Breast and from 0.555 to 0.847 on MERFISH_Brain. The extra anchors are genes these clusters express, so they remove native expression.

Details are in the revised Appendix I.

2. Overlap of anchor genes with the Figure 3 markers (Weakness 2; Question 2)

Thank you for raising this possible circularity. We added analyses that leave the anchors out, and the conclusions hold.

First, the overlap is partial. DeSpotX selects anchors from the data as genes that a cell type rarely detects, without using the markers, so some off-target markers are anchors, but no canonical marker in the red boxes is. Among the off-target dots, 58 of 78 are anchors on Xenium_Breast, 11 of 45 on MERFISH_Brain, 38 of 45 on CosMx_NSCLC, 27 of 40 on Stereo-seq_E14 and 46 of 50 on Xenium5k_Breast.

Second, we left the anchors out and repeated the analysis on the off-target dots that are not anchors. DeSpotX still removes a median of 31-55% of their marker counts on the four datasets with enough such dots, against 2-7% for the canonical markers in the red boxes (Table R1).

Table R1. Median percentage of marker counts that DeSpotX removes per Figure 3a dot.

Xenium_BreastMERFISH_BrainCosMx_NSCLCStereo-seq_E14
Off-target dots that are not anchors33.6% (n = 20)44.8% (n = 34)55.1% (n = 7)30.9% (n = 13)
Canonical markers in the red boxes2.6%2.1%7.1%3.9%

Third, we also left the anchor entries out of the spike-in scoring on all five datasets. DeSpotX still has the highest AUROC on every dataset (Table R2).

Table R2. AUROC ↑ on non-anchor entries, local cross-cluster spike-in, mean ± std across the five contamination levels.

Xenium_BreastMERFISH_BrainCosMx_NSCLCXenium5k_BreastStereo-seq_E14
DeSpotX0.930 ± 0.0130.932 ± 0.0110.965 ± 0.0030.996 ± 0.0000.976 ± 0.001
Best baseline0.793 ± 0.011 (ResolVI)0.802 ± 0.018 (SpaceBender)0.854 ± 0.011 (ResolVI)0.976 ± 0.002 (ResolVI)0.910 ± 0.007 (ResolVI)

Fourth, we also validated the downstream cell-cell communication analysis independently of the anchors. Using the raw counts and dissociated atlases, we found that the ligands behind the lost OPIOID and PTN interactions, such as Penk from non-neuronal cells in Figure 5e, are consistent with contamination.

Details are in the revised Section 4.3.1 and Appendices G, M (Figure 17) and S.

AuthorsOpenReview ↗

Response to Reviewer fSmo (Part 2/2)

3. Quantitative evaluation of cluster separation (Weakness 5; Question 3)

Thank you for this suggestion. Yes. In addition to the kNN purity shown on the UMAPs (Figure 18), we added five metrics of cluster separation, and all of them improve after DeSpotX.

First, DeSpotX improves every metric on every dataset. We computed kNN purity in PCA space, the silhouette score, the Davies-Bouldin and Calinski-Harabasz indices, and the NMI between Leiden clusters and the published labels (Table R3). DeSpotX is also the best of all methods in 24 of the 25 combinations.

Table R3. Raw counts → DeSpotX with the published labels.

MetricXenium_BreastMERFISH_BrainCosMx_NSCLCXenium5k_BreastStereo-seq_E14
kNN purity, PCA ↑0.772 → 0.9430.927 → 0.9920.501 → 0.8280.365 → 0.8140.742 → 0.936
Silhouette ↑0.027 → 0.1790.058 → 0.244−0.094 → −0.003−0.045 → −0.0360.047 → 0.168
Davies-Bouldin ↓2.574 → 1.6672.034 → 1.4274.427 → 2.5546.975 → 2.6592.561 → 1.665
Calinski-Harabasz ↑1,473 → 3,0681,034 → 2,377549 → 1,359260 → 1,2471,636 → 2,821
Leiden NMI ↑0.596 → 0.7950.683 → 0.8060.437 → 0.7390.079 → 0.4320.581 → 0.780

Second, to check that the gain does not come from the labels given to DeSpotX, we added an analysis with finer subtypes that no method receives. kNN purity in PCA space still rises on every dataset, from 0.232-0.832 to 0.429-0.910, and is the highest among all methods.

Details are in the revised Appendix N (Table 17).

4. Bold column of Table 10 (Question 4)

Thank you for catching this. We now bold the anchor column of Table 10 (Table 11 in the revised paper), which has the highest AUROC in every row, as in our other tables. We had bolded the non-anchor column because it tests whether the AUROC could come from the anchor mask alone, and the caption now states this instead.

Details are in the revised Table 11.

5. High-resolution spatial validation in Figure 4 (Weakness 3; Question 5)

Thank you for this suggestion. We replaced Figure 4a-e with subtype-level maps. Each panel now shows three subtypes, each with its marker, so the spatial coherence is much easier to see. The subtypes are the Allen subclasses on MERFISH_Brain, types transferred from reference atlases on CosMx_NSCLC, Xenium5k_Breast and Stereo-seq_E14, and subclusters of the raw counts on Xenium_Breast.

First, DeSpotX clears the marker from the surrounding cells and keeps it in the subtype. The share of the surrounding cells that detect the marker falls from 5.9-35.4% to 0.2-8.8%, while the subtypes keep 85.4-99.5% of their marker counts.

Second, DeSpotX is given only the paper's labels, not these subtypes, so this is an independent validation.

The new Figure 4 is in Section 4.3.2, and the values of each subtype are in Appendix Q (Table 19).

6. Caveats in the cell-cell communication analysis (Weakness 4; Question 6)

Thank you for these suggestions. We replaced CellChat with the spatial-aware CellChat v2 [1], and the conclusions hold (Figure 5d, 5e).

First, we listed which interactions are shifted and which are kept. On MERFISH_Brain, most lost interactions (14 of 18) are OPIOID interactions sent by non-neuronal cells, while neurons keep all of theirs. On Xenium_Breast, PTN interactions from cancer epithelial and endothelial cells are lost, and those from myoepithelial cells are kept.

Second, to see whether these shifts are real, we checked them against dissociated single-cell atlases and literature evidence. The senders that lose interactions barely express the ligand in the atlases, whereas those that keep them express it highly, consistent with neurons being the main source of enkephalin, the product of Penk [2].

Third, we tested whether the reduced interactions on non-neurons in Figure 5e are biased by the remaining signals. In the raw counts alone, their Penk is 17-26 times higher next to striatal neurons than farther away, consistent with contamination from these neurons.

Details are in the revised Section 4.3.3 and Appendix S (Figure 21).

7. Explanation of Figure 1 in the caption (Question 7)

Thank you for this suggestion. The caption of Figure 1 now explains the encoder, the decoder and the output, each with the equation that defines it.

Thank you again for the detailed comments, which have helped us improve the paper. We would be happy to answer any further questions.

Sincerely,

Authors

References

[1] Jin, Suoqin, Maksim V. Plikus, and Qing Nie. "CellChat for systematic analysis of cell–cell communication from single-cell transcriptomics." Nature protocols (2025).

[2] Le Merrer, Julie, et al. "Reward processing by the opioid system in the brain." Physiological reviews (2009).

AuthorsOpenReview ↗

Happy to clarify any remaining points

Dear Reviewer fSmo,

Thank you again for your detailed and helpful questions, which have improved the paper. We have responded to each of them and revised the paper accordingly.

We understand that this is a busy time. If you have a moment before the discussion period closes, we would be grateful to hear whether our responses address your concerns, and we would be glad to clarify any remaining points.

Best regards,

Authors

Comments on the paper

9 comments
AuthorsOpenReview ↗

Response to Reviewer joxp (Part 1/3)

Thank you for the thorough review and the helpful questions! We appreciate that you found the rationale for the anchor constraints clear and the inclusion of off-model contamination mechanisms valuable. We address each comment below.

1. Contributions and the identification principle (Weakness 1)

Thank you for this suggestion. The revised Section 2 now distinguishes our contributions from the identification principle. Our main contribution is to apply this principle to each cell, combining anchor constraints with a spatial contamination model and learning the per-cell decomposition from data. To show what the spatial formulation and the learned estimation add, we replaced each part of DeSpotX on the Xenium_Breast and MERFISH_Brain spike-ins (Table R1).

First, the spatial formulation helps both contamination detection and removal magnitude. Replacing the local contamination profile with a global one lowers AUROC and raises relMAE.

Second, the learned estimation is needed for the removal magnitude. Replacing the network with the plug-in estimator of Lemma 2 or with EM estimates keeps a similar AUROC but greatly raises relMAE.

Table R1. DeSpotX with one part replaced, mean ± std.

MethodXenium_Breast AUROC ↑Xenium_Breast relMAE ↓MERFISH_Brain AUROC ↑MERFISH_Brain relMAE ↓
DeSpotX0.946 ± 0.0140.536 ± 0.0620.939 ± 0.0150.642 ± 0.201
Global contamination profile0.935 ± 0.0070.636 ± 0.0270.918 ± 0.0070.661 ± 0.024
EM estimates0.942 ± 0.0140.771 ± 0.1610.934 ± 0.0141.212 ± 0.380
Plug-in estimator of Lemma 20.940 ± 0.0130.806 ± 0.1850.936 ± 0.0131.292 ± 0.473

Third, we also added our local profile to SoupX and our anchors to ResolVI and SpaceBender. Some of them improve, but all remain well below DeSpotX (AUROC at most 0.854, against 0.939-0.946 for DeSpotX), so the gain comes from integrating both into one per-cell model.

Details are in the revised Section 4.2.2 and Appendix H (Tables 12 and 13).

2. Dependence on cell-cluster labels (Weakness 2)

Thank you for raising this point. We quantified these failure modes by giving DeSpotX six kinds of wrong labels on the Xenium_Breast and MERFISH_Brain spike-ins.

  • Novel state (M1). One cell type is left unannotated and absorbed into its most similar type.
  • Ambiguous states (M2-M4). Cell types are merged to half their number (M2), cells are clustered without supervision into about 60 clusters (M3), or 30% of cells are relabeled as their most similar type (M4).
  • Severely wrong initial labels (M5, M6). Unsupervised clusters are named by their majority type, leaving 17-38% of cells wrong (M5), or 50% of labels are random and re-annotated after one iteration of the loop (M6).

First, label errors reduce performance only slightly. AUROC drops by at most 0.031 and relMAE rises by at most 0.109 (Table R2). Even in the worst case, relMAE stays below that of SpaceBender, the best baseline without labels (0.642 and 0.965), and that of DecontX even with correct labels (0.749-0.808).

Second, the iterative loop corrects severely wrong initial labels. Starting from 50% random labels, one iteration raises the label accuracy from 50% to 77-94% and lowers relMAE from 0.62-0.64 to 0.54-0.56, close to 0.49-0.56 with correct labels (M6, 20% contamination).

Table R2. DeSpotX with correct and wrong labels at 20% contamination (40% is similar).

LabelsXenium_Breast AUROC ↑Xenium_Breast relMAE ↓MERFISH_Brain AUROC ↑MERFISH_Brain relMAE ↓
Correct labels0.9570.4940.9500.555
M10.9550.5070.9480.552
M20.9320.6030.9210.586
M30.9460.5280.9370.590
M40.9490.5340.9430.641
M50.9260.5970.9340.566
M60.9400.5410.9480.562

Details are in the revised Section 4.2.2 and Appendix I (Table 15).

AuthorsOpenReview ↗

Response to Reviewer joxp (Part 2/3)

3. Sparse or noisy anchors under shared markers or pervasive low-level expression (Weakness 3)

Thank you for raising these two cases. We added analyses for both, and in all of them every cluster keeps anchors and AUROC on non-anchor entries drops by at most 0.026.

First, we tested shared markers on real sections and spike-ins. We relabeled four real sections into finer subtypes that share markers. Every cluster still keeps at least 132 anchors, and more than 99.8% of cells have an anchor that receives contamination, as Lemma 2 requires. We also built spike-ins in which cells share markers with their neighbors, either by taking up to 50% of their expression from the neighboring type or by natively expressing the genes that contaminate them. After DeSpotX derives its anchors again, every cluster keeps at least 46 anchors. This is because sharing markers with a neighbor removes only those few genes from a cluster's anchors, and most of its anchors are markers of unrelated lineages, which it still does not express.

Second, we built spike-ins in which every gene gains pervasive low-level expression, and let DeSpotX derive its anchors again. Native expression on the new anchors does not increase, the cluster with the fewest anchors keeps 11-64 of them across settings, and DeSpotX keeps 76-86% of the added expression. This is because the per-cluster detection threshold is relative to each gene's highest-expressing cluster, so a gene expressed at a similar low level in every cluster does not become an anchor.

Details are in the revised Section 4.2.2 and Appendix I (Table 14).

4. Identification vs recovery by the learned estimator (Question 1)

Thank you for this question. The revised Introduction and Section 3.2 now distinguish the two. Lemma 2 guarantees identification when the contamination profile is known and the anchor zeros are exact. Lemma 4 then gives a conditional bound for the learned estimator with data-derived anchors, in terms of the deviations from these two conditions and the model fit. We measured these deviations where possible, and their effect on performance is small.

First, we measured the terms of Lemma 4 on the Xenium_Breast and MERFISH_Brain spike-ins at 20% contamination and on the real sections. The measured deviations are small (Table R3).

Table R3. Terms of Lemma 4 on the trained model.

Term of Lemma 4Measured ↓Xenium_Breast, spike-inXenium_Breast, realMERFISH_Brain, spike-inMERFISH_Brain, real
Profile error ΔError of the contamination mass on the anchors, relative to the true mass5.1%–4.9%–
Anchor zeros δFitted native mass on the anchors0.024%0.055%0.016%0.034%
Anchor mass χ̂_ACells with χ̂_A < 0.011.0%0.83%2.3%0.045%
Native anchor expression ηCounts on the anchors, an upper bound–4.6%–3.2%

Second, the remaining gap to the two conditions of Lemma 2 has little effect on performance. We supplied the true contamination profile or enforced the anchor zeros with a 10 times stronger penalty, and AUROC changes by at most 0.004 and relMAE by at most 0.042.

Details are in the revised Introduction, Section 3.2 and Appendix B.4.

5. Contamination mass on the anchor set and the single-anchor experiment (Question 2)

Thank you for this question. With the default anchors, almost every cell receives contamination on its anchors. A stress test with one anchor per cluster leaves a subset of cells without contamination on their anchor. Even for these cells, DeSpotX's per-cell estimates track the injected contamination, not only aggregate AUROC.

First, with the default anchors, 99.76-100% of cells in the five real datasets have an anchor that receives contamination, with a median of 1.1-42.4 counts per cell. Even where a cell carries few such counts, its cluster shares one native profile, and every cluster with at least 10 cells carries at least 307 counts on its anchors.

Second, we built new spike-ins on Xenium_Breast and MERFISH_Brain in which we let the contamination fraction vary across cells, as clean and contaminated patches or as a gradient from 0 to 0.4, and ran DeSpotX with one random anchor per cluster. Here 58-68% of cells receive no contamination on their anchor. For these cells, the estimated injected fraction, the fraction of counts DeSpotX removes beyond what it removes before injection, correlates with the true one (Pearson 0.73-0.90) as well as for cells with contamination on their anchor (0.70-0.88).

Third, in the same spike-ins, DeSpotX correctly estimates near-zero contamination for the uncontaminated cells in the clean patches. 89-90% of them have an estimated injected fraction below 0.05, and this estimate separates clean from contaminated cells with AUROC 0.94-0.98.

Details are in the revised Appendices B.4 and I.

AuthorsOpenReview ↗

Response to Reviewer joxp (Part 3/3)

6. Treatment quality of individual cells within homogeneous clusters (Question 3)

Thank you for this question. Yes. The residual affects every cell that receives within-cluster contamination, in proportion to that contamination as Lemma 3 predicts, but it is below the sampling noise of the cell's own counts in 92-100% of cells.

First, we measured each cell's residual in the current simulation. In its spike-in where 70% of each cell's contamination comes from its own cluster, we ran DeSpotX twice, once as is and once with this within-cluster part removed. The difference between the two outputs, the within-cluster contamination left in each cell, matches the residual that Lemma 3 predicts for that cell (median ratio 0.99 and 1.01, Spearman 0.97 and 0.92 on Xenium_Breast and MERFISH_Brain).

Second, we built a spike-in in which each cell's contamination comes from its 12 nearest neighbors regardless of type, so cells inside homogeneous clusters receive only within-cluster contamination. For these 1,437 and 1,291 cells, the residual is below sampling noise in 98.5% and 93.0% of them.

Details are in the revised Appendix K.

7. Presence or absence, 94% retention and macrophage-state retention (Question 4)

Thank you for asking for these details. We added all three to the revised paper.

First, presence or absence. For each cell type and gene in the matched scRNA-seq, detection is the fraction of the type's cells with at least one count. The gene is present in the type if detection is above 50% and absent if it is below 1%, and genes in between are not used.

Second, the 94% retention means that DeSpotX keeps 94% of the raw counts of present genes, counting all cells and genes together. For each Xenium cell, we add up the counts of the genes present in its cell type, and divide the total over all cells after DeSpotX by that in the raw data.

Third, macrophage-state retention. The polarization of each macrophage is the mean z-score of eight M2 markers minus that of eight M1 markers. Strongly polarized cells are those whose raw polarization is more than one standard deviation from the mean. Retention is the mean absolute polarization of these cells after decontamination, divided by that of the same cells in the raw data.

Details are in the revised Section 4.3.1 and Appendix P.

8. Independent measures of annotation accuracy (Question 5)

Thank you for this suggestion. We checked the labels against independent references that the iterations never use, and the labels improve against all of them.

First, we compared the labels with the expert labels published with the two datasets [1,2]. From iteration 0 to iteration 5, many more cells change from a wrong to a correct label than the reverse, 2,658 against 438 on MERFISH_Brain and 1,904 against 932 on Xenium_Breast.

Second, we checked the labels with a held-out annotator and with cell types that the brain section should lack (Table R4). We annotated the counts at iteration 0 and iteration 5 with a CellTypist model that the iterations never use, and its labels agree better with the expert labels at iteration 5. The MERFISH_Brain section is registered to the Allen brain coordinates and contains no cerebellum, pons or medulla, so Bergmann glia and glycinergic neurons should be almost absent, and calls of both fall by iteration 5. Removing the same amount of counts at random gains much less, so the improvement comes from which counts DeSpotX removes.

Table R4. Labels at iteration 0 and iteration 5 checked against references outside the iterations.

CheckIteration 0Iteration 5Random removal
Held-out annotator agreement, Xenium_Breast ↑0.5000.5470.511-0.515
Held-out annotator agreement, MERFISH_Brain ↑0.2490.2700.251-0.253
Glycinergic neuron calls, MERFISH_Brain (45 in expert labels) ↓1,2588831,104-1,195
Bergmann glia calls, MERFISH_Brain (0 in expert labels) ↓1,004897975-1,004

Third, we used spike-ins with known true labels and contamination. Starting from 50% random labels, one iteration of the loop raises the label accuracy from 50% to 77-94% and lowers relMAE from 0.62-0.64 to 0.54-0.56, close to 0.49-0.56 with correct labels.

Details are in the revised Section 4.3.3 and Appendices S (Table 20) and I (Table 15).

Thank you again for the thoughtful comments, which have helped us improve the paper. We would be happy to answer any further questions.

Sincerely,

Authors

References

[1] Yao, Zizhen, et al. "A high-resolution transcriptomic and spatial atlas of cell types in the whole mouse brain." Nature (2023).

[2] Janesick, Amanda, et al. "High resolution mapping of the tumor microenvironment using integrated single-cell, spatial and in situ analysis." Nature communications (2023).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.