timoast
commented
4 years ago Hi Maria, thanks for your questions, I will try to answer them below:
Integration I think of as finding peaks that intersect at least x bp, taking the coordinates of these peaks, and introducing a new peak coordinate based on the intersecting peaks (after e.g. centering)
I think this is a description of one way dataset merging can work for scATAC-seq data. This is similar to what MergeWithRegions does, only that there is no centering of peaks in that function (for convenience we keep the peak coordinates of one of the input datasets). The integration workflow is slightly different in that it will adjust the data to remove dataset-specific differences in chromatin accessibility. However the integration needs common features across datasets, which is why we have the MergeWithRegions step in the integration vignette on the Signac website. You can read about how the integration algorithm works here: https://doi.org/10.1016/j.cell.2019.05.031
MergeWithRegions - merge datasets in a coordinate-aware manner. Does this mean that no shifts in peak coordinates were introduced (meaning no integration)?
Answered above
When running FindTopFeatures (min.cutoff = "q75") on this "unintegrated" object, I do get some 24K peaks - can I expect these peaks to contribute to the technical variance then (since UMAP generates 2 different clusters)?
This will return peaks that were accessible in more than 75% of the cells, I suppose if you expect any differences between the two groups to be purely technical, and there is differential accessibility between the groups for these peaks (which there must be if they're separating on the UMAP), then they would contribute to the technical variance.
So I continued with GetIntersectingFeatures which gives me some 97K peaks out of a total of 267K across 2 samples. Peaks are with original coordinates. Here I have a uniform peak set that I could also use for my objective 1) - see how well they agree with DNase footprints. Correct?
Yes, you could compare the intersecting peaks across objects with DHS peaks. You might also try taking the union of peaks, rather than the intersection.
I can now continue with subsampling some peaks and running FeatureMatrix, and add a new assay with these subsampled peaks to one of my samples (seurat objects) "# create a new assay and add it to the 10x dataset". Will that seurat object subset be used by FindIntegrationAnchors or IntegrateData, in any other way than the peaks.use subset directly that I specify for the FindIntegrationAnchors? After making the integrated object, I have there 2 samples with subsampled as well as all peaks across both samples.
The subset of peaks will only be used for the integration. After integration, you will have the integrated assay, which will only contain the subset of peaks that were corrected, and the original data for both datasets as a merged "peaks" assay. Be aware that any peaks that were not shared across the datasets will be given zero counts in the dataset where it was not detected, but this does not necessarily mean that there were no fragments in that region of the genome for those cells. For this reason, it might be a good idea to take the union of peaks across the datasets and quantify the new peak set in both datasets using FeatureMatrix.
Can I now use FindTopFeatures to compare the top peaks between the two samples in the integrated object (objective 2)? If so, is this correct: integrated <- RunTFIDF(integrated, assay = 'peaks') integrated <- FindTopFeatures(integrated, assay = 'peaks', min.cutoff = "q75")
A better way to compare peaks across datasets would be to perform a differential accessibility test, as this can take the total number of reads sequenced in each cell into account as a latent variable. For example, you can use the LR test: FindMarkers(object, ident.1 = dataset1, ident.2 = dataset2, test.use = 'LR', latent.vars = 'nCount_peaks'). However the approach you suggest of comparing the percentage of cells with the peak accessible across datasets is not wrong, and will be faster, but the LR test could give slightly better results and doesn't depend on setting a cutoff (eg, 75% cells with the peak).
Alternatively, I guess I could use FeatureMatrix -> CreateGeneActivityMatrix for each sample separately using the same set of peaks and compare the peak-associated genes (quantify gene expression ‘activity’ from ATAC-seq reads)?
Yes you could also do this if you are particularly interested in gene accessibilty.
mliiv
Dragonmasterx87
Hello,
I would like to understand the concepts and functions for scATAC-seq integration vs merging in signac a bit better. My data is very low-dimensional, same cell type, couple of treatments, run on same 10X Chromium run on different lanes (one lane per treatment), sequenced separately and then run through cellranger v3 separately. I would like to 1) Verify that my top peaks (peaks represented in e.g. 50% cells) overlap some existing DNase footprints (little quality control),
2) Compare top peaks between treatment and control. For this I would need to first identify peak regions that were discovered across the treatment and control samples, so let's say I want to get peaks intersect between 2 samples for simplicity here. I have followed the tutorial "scATAC-seq data integration". Integration I think of as finding peaks that intersect at least x bp, taking the coordinates of these peaks, and introducing a new peak coordinate based on the intersecting peaks (after e.g. centering).
MergeWithRegions - merge datasets in a coordinate-aware manner. Does this mean that no shifts in peak coordinates were introduced (meaning no integration)? I first tried this but when looking at the UMAP the two samples clearly separated in 2 clusters, which I do not expect from the biology. I was a bit unexpected, though (since I should have very low tech variation). When running FindTopFeatures (min.cutoff = "q75") on this "unintegrated" object, I do get some 24K peaks - can I expect these peaks to contribute to the technical variance then (since UMAP generates 2 different clusters)?
So I continued with GetIntersectingFeatures which gives me some 97K peaks out of a total of 267K across 2 samples. Peaks are with original coordinates. Here I have a uniform peak set that I could also use for my objective 1) - see how well they agree with DNase footprints. Correct?
I can now continue with subsampling some peaks and running FeatureMatrix, and add a new assay with these subsampled peaks to one of my samples (seurat objects) "# create a new assay and add it to the 10x dataset". Will that seurat object subset be used by FindIntegrationAnchors or IntegrateData, in any other way than the peaks.use subset directly that I specify for the FindIntegrationAnchors? After making the integrated object, I have there 2 samples with subsampled as well as all peaks across both samples.
Can I now use FindTopFeatures to compare the top peaks between the two samples in the integrated object (objective 2)? If so, is this correct: integrated <- RunTFIDF(integrated, assay = 'peaks') integrated <- FindTopFeatures(integrated, assay = 'peaks', min.cutoff = "q75")
Since I am not sure it's comparing the peaks between my two samples (two sets of cells with different treatments) and rather across all the cells.
Alternatively, I guess I could use FeatureMatrix -> CreateGeneActivityMatrix for each sample separately using the same set of peaks and compare the peak-associated genes (quantify gene expression ‘activity’ from ATAC-seq reads)?
I hope you can understand my questions and thank you in advance, Maria