SpaGCN: Integrating gene expression, spatial location and histology to identify spatial domains and spatially variable genes by graph convolutional network

Classified as probably not a substantial claim. May combine two claims.

Previous studies have shown that KLK6 and MBP expression is restricted to oligodendrocytes, while ATP1B1 is mainly expressed in neurons and astrocytes.

“Previous studies have shown that KLK6 and MBP expression is restricted to oligodendrocytes, while ATP1B1 is mainly expressed in neurons and astrocytes 19 .”

• cites [19] Fu, X. et al. Continuous polony gels for tissue mapping with high resolution and RNA capture efficiency. Preprint at bioRxiv https://doi.org/10.1101/2021.03.17.435795 (2021). sourceCould not verify

  The passages on mouse brain expression patterns (Abstract, page unavailable) and the marker-gene list in Figure 3 (PIXEL-seq data analysis, page unavailable) discuss other genes, not the claimed cell-type distributions. Searches for KLK6, MBP, ATP1B1, and oligodendrocytes found no passage in our copy addressing the claim. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

KRT17 functions as a tumor promoter and regulates proliferation in pancreatic cancer, and MMP11 has been found to be a prognostic biomarker for pancreatic cancer.

“KRT17 functions as a tumor promoter and regulates proliferation in pancreatic cancer 21 , and MMP11 has been found to be a prognostic biomarker for pancreatic cancer 22 .”

• cites [21] Liu, Y. et al. High-spatial-resolution multi-omics sequencing via deterministic barcoding in tissue. Cell 183, 1665–1681 e1618 (2020). sourceCould not verify

  The Results sections on whole mouse embryos and internal-organ development discuss spatial gene-expression mapping, not pancreatic cancer. Searches for KRT17 as a tumor promoter and MMP11 as a prognostic biomarker found no passage addressing either claim in our copy. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [22] Cho, C. S. et al. Microscopic examination of spatial transcriptome using Seq-Scope. Cell 184, 3559–3572 e3522 (2021). sourceCould not verify

  The Results passages describe Seq-Scope findings in liver and colon tissue, including Mmp7 expression in hepatic progenitor cells, but do not address pancreatic cancer. Searches for KRT17 and MMP11 found no passage supporting either the tumor-promoter/proliferation claim or the prognostic-biomarker claim. (LLM, conf=0.75)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The unsupervised iterative clustering algorithm has been previously utilized for scRNA-seq analysis and showed superior performance over Louvain's method.

“This unsupervised iterative clustering algorithm has been previously utilized for scRNA-seq analysis and showed superior performance over Louvain's method 25,26 .”

• cites [25] Sun, S., Zhu, J. & Zhou, X. Statistical analysis of spatial expression patterns for spatially resolved transcriptomic studies. Nat. Methods 17, 193–200 (2020). sourceCould not verify

  The Results (Simulations; page unavailable) compare methods for detecting spatially expressed genes, not clustering algorithms. Searches for iterative clustering, scRNA-seq clustering, and Louvain found no passage in our copy addressing the claimed prior use or performance comparison. (LLM, conf=0.74)

  Experimental — accuracy not yet measured
• cites [26] Xie, J., Girshick, R. & Farhadi, A. Unsupervised deep embedding for clustering analysis. In Proc. 33rd International Conference on Machine Learning Vol. 48 (JMLR: W&CP, 2016). sourceNot checked — no full text of the source

May combine two claims.

SpaGCN detected 12 SVGs in pancreatic cancer dataset, with 3, 8, and 1 SVGs for domains 0, 1, and 2, respectively.

“In total, SpaGCN detected 12 SVGs, with 3, 8, and 1 SVGs for domains 0, 1, and 2, respectively (Fig. 5b; Supplementary Fig. 21).”

May combine two claims.

Using SpaGCN, spatial domains detected agreed well with annotated hypothalamic nuclei, with domain 2 corresponding to ACA, domain 3 to PS, and domain 7 to MnPo.

“Using this approach, SpaGCN detected spatial domains that agreed well with the annotated hypothalamic nuclei (Fig. 6a), with domain 2 corresponding to ACA, domain 3 corresponding to PS, and domain 7 corresponding to MnPo.”

• AssumptionRests on an assumption: In MERFISH data, the captured tissue area is much smaller and fewer genes are measured, making it difficult to detect spatial domains because cells within such a small area are more similar to each other.

May combine two claims.

PVALB is enriched in domain 1, and TRM62 is enriched in domain 8.

“As shown in Fig. 3c, PVALB is enriched in domain 1, and TRM62 is enriched in domain 8.”

May combine two claims.

Louvain's method identified 5 main spatial domains in the MOB dataset, but domains 2, 3, and 4 have blurred boundaries and more outliers than SpaGCN.

“Louvain's method identified 5 main spatial domains.However, since it does not consider spatial and histology information, domains 2, 3, and 4 have blurred boundaries and more outliers than SpaGCN.”

Probably not a substantial claim: may not pass S4 (based).

SpaGCN is applicable to both sequencing-based and smFISH-based data.

“SpaGCN is applicable to both sequencing-based and smFISH-based data.”

May combine two claims.

When using top 18 SVGs, ARI is 0.07 for SpatialDE and 0.05 for SPARK; using 806 most significant SpatialDE SVGs, ARI is 0.14; using 1,114 most significant SPARK SVGs, ARI is 0.15.

“When only using their top 18 SVGs, the ARI is only 0.07 for SpatialDE and 0.05 for SPARK.Even when using the 806 most significant SpatialDE detected SVGs, the ARI is only 0.14.When using the 1,114 most significant SPARK detected SVGs, the ARI is 0.15 (Fig. 4e).”

May combine two claims.

SpaGCN detected 61 SVGs in LIBD slice 151673, with 53 specific to domain 4 (white matter).

“In total, SpaGCN detected 61 SVGs, with 53 of them specific to domain 4, which corresponds to the white matter region (Supplementary Fig. 14).”

May combine two claims.

SPARK and SpatialDE detected 203 and 163 SVGs, respectively, in pancreatic cancer dataset.

“SPARK and SpatialDE detected 203 and 163 SVGs, respectively (Supplementary Fig. 22).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

SpaGCN detected SVGs and meta genes are transferrable and can be utilized for downstream analyses in independent tissue sections.

“Additionally, the SpaGCN detected SVGs and meta genes are transferrable and can be utilized for downstream analyses in independent tissue sections.”

May combine two claims.

SpaGCN detected meta genes for domains 2, 7, 8 and 9, and these meta genes are transferrable to an adjacent tissue slice.

“Using this meta gene detection procedure, we also detected meta genes for domains 2, 7, 8 and 9, and found that these meta genes are transferrable to an adjacent tissue slice (Supplementary Fig. 13).”

May combine two claims.

SVGs detected by SpaGCN in mouse posterior brain were enriched in specific spatial domains and their expression patterns are transferable to an adjacent tissue slice.

“By contrast, the SVGs detected by SpaGCN were enriched in specific spatial domains (Supplementary Fig. 11) and their expression patterns are transferable to an adjacent tissue slice in the mouse posterior brain (Supplementary Fig. 12).”

May combine two claims.

K-means only identified 3 main spatial domains in the MOB dataset, with few spots assigned to domains 1 and 3.

“K-means only identified 3 main spatial domains, with only few spots assigned to domains 1 and 3.”

May combine two claims.

SVGs detected by SpaGCN have more coherent expression patterns and better biological interpretability than those detected by SPARK and SpatialDE.

“We also compared SVGs detected by SpaGCN with those detected by SPARK 14 and SpatialDE 13 , and found that the SVGs detected by SpaGCN have more coherent expression patterns and better biological interpretability than the other two methods.”

May combine two claims.

SPARK results indicate 274 genes have FDR-adjusted p-values less than 0.00001, with 14 having the smallest identical FDR-adjusted p-value of 4.42e-13.

“However, we found that the SPARK results indicate that 274 genes have FDR-adjusted p-values less than 0.00001 with 14 of them having the smallest identical FDR-adjusted p-value of 4.42e-13.”

May combine two claims.

Although all SpaGCN detected SVGs have clear spatial patterns, their Moran's I values are not significantly higher than SVGs detected by SPARK and SpatialDE (median 0.20 vs 0.18 and 0.25).

“Although all SpaGCN detected SVGs have clear spatial patterns, their Moran's I values are not significantly higher than the SVGs detected by SPARK and SpatialDE (median of 0.20 for SpaGCN against 0.18 for SPARK and 0.25 for SpatialDE).”

May combine two claims.

The Adjusted Rand Indexes for SpaGCN, K-means, and Louvain identified domains are 0.42, 0.24, and 0.33, respectively, suggesting SpaGCN results better agree with manually curated layer structure.

“The Adjusted Rand Indexes (ARIs) for the SpaGCN, K-means, and Louvain identified domains are 0.42, 0.24, and 0.33, respectively, suggesting that the SpaGCN results better agree with the manually curated layer structure reported in the original study.”

May combine two claims.

In mouse posterior brain data, Louvain's clustering is similar to SpaGCN, but SpaGCN's spatial domains are more spatially contiguous.

“Fig. 3a shows that Louvain's clustering is similar to SpaGCN, but the spatial domains detected by SpaGCN are more spatially contiguous than Louvain's results.”

May combine two claims.

SPARK detected 772 genes in MOB, with 49 overlapping with SpaGCN.

“SPARK detected 772 genes, with 49 overlapping with SapGCN (Supplementary Fig. 2).”

May combine two claims.

SpaGCN detected 523 SVGs for the 10 spatial domains in mouse posterior brain, while SPARK and SpatialDE detected 9,678 and 12,676 SVGs, respectively.

“SpaGCN detected 523 SVGs for the 10 spatial domains while SPARK and SpatialDE detected 9,678 and 12,676 SVGs, respectively (Supplementary Fig. 7).”

May combine two claims.

K-means and Louvain failed to separate the LIBD DLPFC tissue into layers with clear boundary, while SpaGCN successfully identified layer structures with clear boundaries.

“As shown in Fig. 4a, K-means and Louvain failed to separate the tissue into layers with clear boundary.By contrast, SpaGCN successfully identified layer structures with clear boundaries.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The spatial domain-specific DE analysis typically detects SVGs with enriched expression for the majority of the domains.

“The spatial domain-specific DE analysis described above typically detects SVGs with enriched expression for the majority of the domains.”

May combine two claims.

SPARK detected 3,187 SVGs, with 1,131 having FDR-adjusted p-values equal to 0, most of which only marked the white matter region.

“SPARK detected 3,187 SVGs with 1,131 of them having FDR-adjusted p-values equal to 0, most of which only marked the white matter region.”

May combine two claims.

SpaGCN completed spatial domain and SVG detection in less than one minute, whereas SpatialDE took approximately 13 minutes and SPARK more than 18 hours.

“As shown in Supplementary Fig. 24, SpaGCN completed spatial domain and SVG detection in less than one minute, whereas the computing time is ~13 minutes for SpatialDE and more than 18 hours for SPARK.”

May combine two claims.

SpaGCN required 1.3 GB of memory, whereas SpatialDE and SPARK required more than 3.1 GB and 7.2 GB, respectively.

“Furthermore, SpaGCN only required 1.3 GB of memory, whereas SpatialDE and SPARK required more than 3.1 GB and 7.2 GB of memory, respectively.”

May combine two claims.

SpaGCN is computationally fast and memory efficient.

“SpaGCN is computationally fast and memory efficient.”

May combine two claims.

Although it is difficult to identify single genes to mark the other neuronal layers, SpaGCN was able to find layer-specific meta genes.

“Although it is difficult to identify single genes to mark the other neuronal layers, SpaGCN was able to find layer-specific meta genes.”

May combine two claims.

SpatialDE detected 3,654 SVGs, with 806 having q-values equal to 0, but these genes do not necessarily show better spatial pattern than genes with larger q-values.

“SpatialDE detected 3,654 SVGs with 806 of them having q-values equal to 0, but these genes do not necessarily show better spatial pattern than genes with larger q-values (Supplementary Fig. 16).”

May combine two claims.

Although SPARK and SpatialDE detected much larger numbers of SVGs than SpaGCN, the genes detected by these two methods lack ability to distinguish different degrees of spatial variability in expression as their p-value and q-value distributions are highly skewed toward 0.

“Although SPARK and SpatialDE detected much larger numbers of SVGs than SpaGCN (Supplementary Fig. 17), the genes detected by these two methods lack ability to distinguish different degrees of spatial variability in expression as their p-value and q-value distributions are highly skewed toward 0 (Supplementary Fig. 18).”

May combine two claims.

FDR-adjusted p-value from SPARK and q-value from SpatialDE are not informative; genes with similar p-values/q-values do not necessarily show similar spatial pattern and a smaller p-value/q-value does not guarantee a better spatial pattern.

“Another limitation of these two methods is that the FDR-adjusted p-value from SPARK and q-value from SpatialDE are not informative.Genes with similar p-values/q-values do not necessarily show similar spatial pattern and a smaller p-value/q-value does not guarantee a better spatial pattern (Supplementary Fig. 8 and Supplementary Fig. 9).”

May combine two claims.

High expression of NRGN in domains 5 and 7 indicates these two domains are neuroanatomically similar, both consisting of cortex and the pyramidal layer of the hippocampus.

“The high expression of NRGN in domains 5 and 7 also indicate that these two domains are neuroanatomically similar -both consisting of cortex and the pyramidal layer of the hippocampus.”

May combine two claims.

Domain 2 was separated into 5 sub-domains located either in the center or different boundary regions of domain 2, suggesting that differences in neighborhoods of spots contribute to within-domain heterogeneity.

“As shown in Fig. 3e, domain 2 was separated into 5 sub-domains which are located either in the center or different boundary regions of domain 2, suggesting that differences in the neighborhoods of spots contribute to within-domain heterogeneity.”

• AssumptionRests on an assumption: The expression profile and biological function of a spot is heavily influenced by its neighboring spots.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

By varying characteristic length scale l, users can get spatial domain separations with different patterns in which a higher l results in spatial domains with higher contiguity.

“By varying 𝑙𝑙, users can get spatial domain separations with different patterns in which a higher 𝑙𝑙 will result in spatial domains with higher contiguity.”

May combine two claims.

Gene expression profiles of spots from white matter are distinct from spots in neuronal layers, while gene expression differences among the six neuronal layers are much smaller and more difficult to distinguish.

“These results indicate that gene expression profiles of spots from white matter are distinct from spots in the neuronal layers, while gene expression differences among the six neuronal layers are much smaller and more difficult to distinguish”

May combine two claims.

ARIs of both SpatialDE and SPARK are much lower than SpaGCN, even though both used many more SVGs than SpaGCN, confirming lack of spatial expression specificity for genes detected by these methods.

“The ARIs of both SpatialDE and SPARK are much lower than SpaGCN, even though both used many more SVGs than SpaGCN, which further confirmed the lack of spatial expression specificity for genes detected by these methods.”

May combine two claims.

Moran's I values for SPARK and SpatialDE SVGs in pancreatic cancer are much lower than those detected by SpaGCN, suggesting lack of spatial expression patterns.

“However, the Moran's I values for their SVGs are much lower than those detected by SpaGCN, suggesting their lack of spatial expression patterns (Fig. 5d).”

May combine two claims.

Multiple domain adaptive filtering criteria implemented in SpaGCN allow it to eliminate false positive SVGs and ensure all detected SVGs have clear spatial expression patterns.

“Further, multiple domain adaptive filtering criteria implemented in SpaGCN allow it to eliminate false positive SVGs and ensure all detected SVGs have clear spatial expression patterns.”

May combine two claims.

Results consistently showed SpaGCN can identify spatial domains with coherent gene expression and histology and detect SVGs and meta genes that have much clearer spatial expression patterns and biological interpretations than genes detected by SPARK and SpatialDE.

“The results consistently showed that SpaGCN can identify spatial domains with coherent gene expression and histology and detect SVGs and meta genes that have much clearer spatial expression patterns and biological interpretations than genes detected by SPARK and SpatialDE.”

Probably not a substantial claim: may not pass S4 (based).

Results from SpaGCN will enable researchers to accurately identify spatial domains and SVGs in their studies.

“Results from SpaGCN will enable researchers to accurately identify spatial domains and SVGs in their studies.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The graph convolutional layer was implemented based on Kipf and Welling, where backpropagation is operated via a localized first-order approximation of spectral graph convolution.

“The graph convolutional layer was implemented based on Kipf and Welling 23 , where the backpropagation is operated via a localized first-order approximation of spectral graph convolution.”

• cites Kipf and Welling 23Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Genes expressed in multiple but disconnected domains, though not uniquely expressed in a particular domain, are useful for understanding spatial variation of gene expression and can be used to form meta genes uniquely expressed in a specific domain.

“Although they are not uniquely expressed in a particular domain, these genes are still useful for understanding spatial variation of gene expression and can be used to form meta genes that are uniquely expressed in a specific domain.”

• AssumptionRests on an assumption: Some genes may be expressed in multiple but disconnected domains.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The top 50 principal components work well for all datasets analyzed in this paper.

“The top 50 principal components are used as input, which work well for all datasets analyzed in this paper.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A similar function has been employed in SpatialDE.

“A similar function has been employed in SpatialDE 13 .”

• cites [13] Moncada, R. et al. Integrating microarray-based spatial transcriptomics and single-cell RNA-seq reveals tissue architecture in pancreatic ductal adenocarcinomas. Nat. Biotechnol. 38, 333–342 (2020). sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Since layer 1 only has few spots, it is difficult to find a highly enriched gene.

“Since layer 1 only has few spots, it is difficult to find a highly enriched gene.”

May combine two claims.

The auxiliary target distribution upweights spots assigned with high confidence, and normalizes the contribution of each centroid to the overall loss function to prevent large clusters from distorting the hidden feature space.

“which upweights spots assigned with high confidence, and normalizes the contribution of each centroid to the overall loss function to prevent large clusters from distorting the hidden feature space.”