Using GLM_supervised mode

[ryangossart]

I am trying to use GLM_supervised mode on a 10X sample and I noticed a few errors in the documentation for the metadata file

1. The file format currently needs to be .csv instead of .tsv as in the example
2. The program will fail if the file has a header line. The example file has a header line.
3. The order of columns needs to be cell type, cell barcode, sample name instead of sample name, cell barcode, cell type.

The program will run if these modifications are made, but does not produce the output directory or the files mentioned in the documentation. Helmert decomposition effect size bin values and p values are present in result.after_correction.scores.tsv, but I am unsure how to interpret these values.

Is there any guidance available on how to interpret these results. Also a nice feature to have when exploring these results would be the counts for each anchor/extender pair in each metadata group.

Thanks in advance,
Ryan G

[marekkokot]

Hi,

Thanks for letting us know. The confusion arises because we have several options for supervised runs, depending on the data type (non-10X vs. 10X, GML vs. Helmert-based). In the case of 10X, we were using GLM, and to run it, the --supervised_test_samplesheet flag should be used; for this, the switch format is as described in the docs. In the docs, I have added a direct example usage at the top to avoid future confusion: https://github.com/refresh-bio/SPLASH/wiki/GLM_supervised

I also realised that this example file is not part of our tar.gz releases, so I updated the latest release archived by adding this file:
https://github.com/refresh-bio/SPLASH/releases.

I hope this is helpful (let me know), and thanks again for reporting this!

[ryangossart]

Thanks for the quick response! Now I can run the tool on both my data and the example. However, in both cases, the GLM_supervised_anchors.tsv just one blank line and the /plots directory is empty. Not sure why this is happening, but any guidance would be appreciated.

Also, you may want to add the --supervised_test_samplesheet flag on the configuration page. I noticed it wasn't there.

Thanks for the help,
Ryan G

[marekkokot]

Hi, it's there in the docs https://github.com/refresh-bio/SPLASH/wiki/Configuration
Do you have any suggestions regarding where we could mention this parameter beyond this?

@roozbehdn Any thoughts regarding the empty plots directory?

[roozbehdn]

Hi @ryangossart,
Glad to hear you’ve been able to run sc-SPLASH on your data. For 10x analysis, we recommend the GLM-based approach (enabled with the --supervised_test_samplesheet flag, which is the one you have also been able to run.

When you enable the GLM-based approach, the *_supervised_metadata folder will contain a selected_anchors_GLMnet.txt file. This file lists all anchors called by sc-SPLASH that are selected as input for the GLM-based supervised analysis. Anchors are selected based on the --supervised_test_anchor_sample_fraction_cutoff input parameter, which sets the minimum fraction of samples (or cells, for 10x data) in which an anchor must be present to be included in the GLM analysis (default = 0.4). If you see few anchors in selected_anchors_GLMnet.txt, try lowering this cutoff (e.g., 0.2 or 0.1).

The GLM_supervised_anchors.tsv file contains only anchors with non-zero GLM coefficients (i.e., meta-dependent or cell-type-specific). If none are found, this file and the plots folder (which only includes plots for these anchors) will be empty.

[roozbehdn]

Also @ryangossart, regarding your following comment:

Also a nice feature to have when exploring these results would be the counts for each anchor/extender pair in each metadata group.

Currently, we don’t output anchor/extendor counts for each metadata group, but I like the idea and we’ll address it in the next SPLASH release. As a quick workaround, you can merge the per-cell anchor/target counts from the satcOut*.tsv files in the *_supervised_metadata folder with your metadata file to get counts per metadata group.