Add SVM classifier based separability scoring #10
Conversation
|
@duopeng we have automatic formatting with pre-commit. You can install by just running |
grassp/tools/scoring.py
Outdated
| Parameters | ||
| ---------- | ||
| data : DataFrame or AnnData |
There was a problem hiding this comment.
The docstrings should not have the data type (:str , etc). Rather the arguments themselves, e.g.:
label_col: str = "consensus_graph_annnotation",
grassp/tools/scoring.py
Outdated
| auc_clustermap = sns.clustermap(auc_mat, | ||
| square=True, | ||
| annot=True, | ||
| fmt=".2f", | ||
| cmap="rocket", | ||
| vmin=0.5, | ||
| vmax=1, | ||
| cbar_kws=dict(label=f"ROC-AUC ({auc_model.upper()})"), | ||
| figsize=(heatmap_size[0], heatmap_size[1])) | ||
| auc_clustermap.fig.suptitle("Label separability\nPair-wise classifier-AUC") | ||
| auc_clustermap.ax_heatmap.set_xticklabels( | ||
| auc_clustermap.ax_heatmap.get_xticklabels(), rotation=45, ha='right') | ||
| figures['auc_fig'] = auc_clustermap |
There was a problem hiding this comment.
Plotting should be handled separately from calculations in grassp.pl
grassp/tools/scoring.py
Outdated
| # Drop rows with missing coords or labels | ||
| df = df.dropna(subset=[label_col, *coord_cols]) | ||
|
|
||
| X_all = df[list(coord_cols)].values |
There was a problem hiding this comment.
You're converting a np.array (the .obsm or .X into a DataFrame and then back to a np.array. This seems inefficient
grassp/tools/scoring.py
Outdated
| if isinstance(data, ad.AnnData): | ||
| assert label_col in data.obs.columns, f"label_col {label_col} not in data.obs.columns" | ||
| X_all = sc.tools._utils._choose_representation(data, use_rep=use_rep, n_pcs=n_pcs) | ||
| df = pd.DataFrame(X_all) |
There was a problem hiding this comment.
You're still converting X_all into a dataFrame and then back into an array (line 316).
grassp/tools/scoring.py
Outdated
| if DataFrame, then use column name as label | ||
| Defaults to "consensus_graph_annnotation" | ||
| use_rep : str, optional | ||
| coordinates (X in the classifier) |
There was a problem hiding this comment.
X in the classifier might not mean much to users. Scanpy has Use the indicated representation. 'X' or any key for .obsm is valid.
grassp/tools/scoring.py
Outdated
| Defaults to "consensus_graph_annnotation" | ||
| use_rep : str, optional | ||
| coordinates (X in the classifier) | ||
| if AnnData, use .obsm[use_rep] if use_rep is a *str*, and .var[use_rep] if *list* |
There was a problem hiding this comment.
Don't need to support the .var[use_rep] case. Users can simply subset before!
grassp/tools/scoring.py
Outdated
| np.fill_diagonal(auc_mat.values, 0.5) | ||
|
|
||
| if inplace: | ||
| data.uns[f"separability ({label_col})"] = { |
There was a problem hiding this comment.
prefer no spaces or special characters in names: separability_{label_col}. don't forget to also fix in plotting
grassp/plotting/heatmaps.py
Outdated
|
|
||
|
|
||
| def sep_auc_heatmap( | ||
| data: np.ndarray | pd.DataFrame, |
There was a problem hiding this comment.
this should also be able to take an anndata object and look in .uns for entries with "separability_"
…satility to plotting function
|
I tested the code after Marika's cleanup, and it works nicely! we can merge this branch with main! |
No description provided.