[Response to Reviewers] Add Silhouette analysis #68
Conversation
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
| # In[2]: | ||
|
|
||
|
|
||
| np.random.seed(1234) |
There was a problem hiding this comment.
Recommend random seed 0 to be consistent with Way Lab standard in other projects.
| # For consistent Silhouette input space dimensionality | ||
| n_pca_components = 50 |
There was a problem hiding this comment.
Where does 50 come from? I know in UMAP we do 2 components, what is the difference when is comes to PCA?
There was a problem hiding this comment.
Also heads up, in the PR comment you say the number of components is 40 but in here it is 50, recommend confirming which one is correct/most appropriate.
There was a problem hiding this comment.
ah! Great catch.
In my experience, 50 is more than enough to capture the majority of the variance in the dataset, which is what we're aiming for. It's more or less an arbitrary number
| output_silhouette_results = pathlib.Path( | ||
| eval_path, "silhouette_score_results.tsv" | ||
| ) | ||
| output_silhouette_samples_results = pathlib.Path( | ||
| eval_path, "silhouette_score_results_per_sample.tsv" | ||
| ) |
There was a problem hiding this comment.
Recommend making these compressed TSVs for saving space plus that might be the standard convention in this repo.
There was a problem hiding this comment.
they are super tiny, will stick with tsv so they can be rendered on github
|
Thanks for the review @jenna-tomkinson - I caught a couple things too, which I addressed in the recent commits. Merging now! |
This PR is in response to the following reviewer comment:
We think this is a good idea, and therefore performed the following analysis:
n_components=50)We interpret the Silhouette scores how well cells of a given phenotype are clustered compared to other cells of the same phenotype. A positive score means cells of the same phenotype are more similar to other cells of the same phenotype (on average) compared to all other cells. A score of 1 indicates complete separation of similar phenotypes from other phenotypes.
New supplementary figure