Feature/exe 1769 dsb implementation #147
Conversation
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## dev #147 +/- ##
==========================================
+ Coverage 81.80% 81.83% +0.02%
==========================================
Files 118 119 +1
Lines 6568 6605 +37
==========================================
+ Hits 5373 5405 +32
- Misses 1195 1200 +5 ☔ View full report in Codecov by Sentry. |
There was a problem hiding this comment.
Excellent! I have focuses on the python aspect here. I think @maxkarlsson is more qualified to comment on the actual implementation of the method.
Very nice and clear code. I had some minor suggestions.
| def dsb_normalize( | ||
| raw_expression: pd.DataFrame, isotype_controls: Union[List, None] = None | ||
| ): | ||
| """empty-droplet-free method as implemented in the dsb package. |
There was a problem hiding this comment.
I think we should add a reference to the original package here.
There was a problem hiding this comment.
Mulè, M.P., Martins, A.J. & Tsang, J.S. Normalizing and denoising protein expression data from droplet-based single cell profiling. Nat Commun 13, 2099 (2022).
https://doi.org/10.1038/s41467-022-29356-8
…randomizations for the same seed across platforms.
Changed terminology somewhat. expression --> abundance limma --> _regress_out_confounder baseline --> background
There was a problem hiding this comment.
Looks great! I have some comments and a control question for the PCA step. I have a little trouble understanding whether the linear regression is equivalent, mostly because I am not used to the syntax. I would suggest that we generate data for both versions of the normalization and verify that the result is close to exactly the same. Great and quick work!
I also made some commits with terminology changes. Please hava a look and see if you agree with the changes.
| current_axis = dataframe.loc[i, :] if axis == 0 else dataframe.loc[:, i] | ||
| gmm = gmm.fit(current_axis.to_frame()) | ||
| background[i] = np.min(gmm.means_) | ||
| scores[i] = np.abs(gmm.means_[1] - gmm.means_[0]) / np.sum(gmm.covariances_) |
There was a problem hiding this comment.
What is done at this line? Scores for a marker is calculated as how many standard deviations the positive population mean is from the negative population mean?
There was a problem hiding this comment.
Yes! right now the score is not used anywhere as we do the baseline subtraction for all markers, but could later be used as a check for existence of negative population.
|
|
||
| def _get_background_abundance(dataframe: pd.DataFrame, axis=0): | ||
| """Fit a double gaussian distribution to the abundance data and return the mean | ||
| of the first gaussian as an estimation of the background level.""" |
There was a problem hiding this comment.
We also return some scores, right?
| log_abundance = log_abundance - marker_background | ||
| component_background, _ = _get_background_abundance(log_abundance, axis=0) | ||
|
|
||
| if isotype_controls is not None: |
There was a problem hiding this comment.
I think we should force the user to input isotypes, since dsb relies on those.
| if isotype_controls is not None: | ||
| control_signals = log_abundance.loc[:, isotype_controls] | ||
| control_signals["component_background"] = component_background | ||
| pheno = PCA(n_components=1).fit_transform(control_signals) |
There was a problem hiding this comment.
Are variables in control_signals scaled to unit variance before PCA?
…dsb_normalize to be called with at least one isotype.
Description
Added dsb_normalization as part of the new normalization package in the analysis stage. dsb normalization has been shown to effectively compensate for noise effects in expression data.
Fixes: #(issue number)
Type of change
Please delete options that are not relevant.
How Has This Been Tested?
It has been tested against a similar function that is already implemented in R and the results are almost identical (A slight variation is expected given that the gaussian mixture model stage is indeterministic).
A corresponding test function is implemented to compare against future changes to the function.
PR checklist: