# # Calculate duration on food + duration in L1 diapause
# metadata = duration_on_food(metadata)
# metadata = duration_L1_diapause(metadata)
#%% Subset results
# Subset results (rows) to remove NA groups
metadata = metadata[~metadata[GROUPING_VAR].isna()]
features = features.reindex(metadata.index)
# Check case-sensitivity
assert len(metadata[GROUPING_VAR].unique())==len(metadata[GROUPING_VAR].str.upper().unique())
# Subset results (rows) to omit selected strains
if args.omit_strains is not None:
features, metadata = subset_results(features, metadata, GROUPING_VAR, args.omit_strains)
# Subset results (rows) for imaging dates of interest
if args.dates is not None:
features, metadata = subset_results(features, metadata, 'date_yyyymmdd', args.dates)
# Load Tierpsy Top256 feature set + subset (columns) for Top256 features only
if args.use_top256:
top256_path = AUX_DIR / 'top256_tierpsy_no_blob_no_eigen_only_abs_no_norm.csv'
top256 = load_top256(top256_path, add_bluelight=args.align_bluelight)
# Ensure results exist for features in featurelist
top256_feat_list = [feat for feat in top256 if feat in features.columns]
print("Dropped %d features in Top256 that are missing from results" %\
(len(top256)-len(top256_feat_list)))
def compare_strains_keio(features, metadata, args):
""" Compare Keio single-gene deletion mutants with wild-type BW25113 control and look to see if
they signfiicantly alter N2 C. elegans behaviour while feeding.
Subset results to omit selected strains (optional)
Inputs
------
features, metadata : pd.DataFrame
Matching features summaries and metadata
args : Object
Python object with the following attributes:
- drop_size_features : bool
- norm_features_only : bool
- percentile_to_use : str
- remove_outliers : bool
- omit_strains : list
- grouping_variable : str
- control_dict : dict
- collapse_control : bool
- n_top_feats : int
- tierpsy_top_feats_dir (if n_top_feats) : str
- test : str
- f_test : bool
- pval_threshold : float
- fdr_method : str
- n_sig_features : int
"""
assert set(features.index) == set(metadata.index)
# categorical variable to investigate, eg.'gene_name'
grouping_var = args.grouping_variable
n_strains = len(metadata[grouping_var].unique())
assert n_strains == len(
metadata[grouping_var].str.upper().unique()) # check case-sensitivity
print("\nInvestigating '%s' variation (%d samples)" %
(grouping_var, n_strains))
# Subset results (rows) to omit selected strains
if args.omit_strains is not None:
features, metadata = subset_results(features,
metadata,
column=grouping_var,
groups=args.omit_strains,
omit=True)
control = args.control_dict[grouping_var] # control strain to use
# Load Tierpsy Top feature set + subset (columns) for top feats only
if args.n_top_feats is not None:
top_feats_path = Path(
args.tierpsy_top_feats_dir) / "tierpsy_{}.csv".format(
str(args.n_top_feats))
topfeats = load_topfeats(top_feats_path,
add_bluelight=True,
remove_path_curvature=True,
header=None)
# Drop features that are not in results
top_feats_list = [
feat for feat in list(topfeats) if feat in features.columns
]
features = features[top_feats_list]
##### Control variation #####
control_metadata = metadata[metadata[grouping_var] == control]
control_features = features.reindex(control_metadata.index)
# Clean data after subset - to remove features with zero std
control_feat_clean, control_meta_clean = clean_summary_results(
control_features,
control_metadata,
max_value_cap=False,
imputeNaN=False)
if args.analyse_control:
control_variation(control_feat_clean,
control_meta_clean,
args,
variables=[
k for k in args.control_dict.keys()
if k != grouping_var
],
n_sig_features=10)
if args.collapse_control:
print("\nCollapsing control data (mean of each day)")
features, metadata = average_plate_control_data(features, metadata)
# Record mean sample size per group
mean_sample_size = int(
np.round(
metadata.join(features).groupby([grouping_var],
as_index=False).size().mean()))
print("Mean sample size: %d" % mean_sample_size)
save_dir = get_save_dir(args)
stats_dir = save_dir / grouping_var / "Stats" / args.fdr_method
plot_dir = save_dir / grouping_var / "Plots" / args.fdr_method
##### STATISTICS #####
# =============================================================================
# ##### Pairplot Tierpsy Features - Pairwise correlation matrix #####
# if args.n_top_feats == 16:
# g = sns.pairplot(features, height=1.5)
# for ax in g.axes.flatten():
# # rotate x and y axis labels
# ax.set_xlabel(ax.get_xlabel(), rotation = 90)
# ax.set_ylabel(ax.get_ylabel(), rotation = 0)
# plt.subplots_adjust(left=0.3, bottom=0.3)
# plt.show()
# =============================================================================
if not args.use_corrected_pvals:
anova_path = stats_dir / '{}_results_uncorrected.csv'.format(args.test)
else:
anova_path = stats_dir / '{}_results.csv'.format(args.test)
# load results + record significant features
print("\nLoading statistics results")
anova_table = pd.read_csv(anova_path, index_col=0)
pvals = anova_table.sort_values(
by='pvals', ascending=True)['pvals'] # rank features by p-value
fset = pvals[pvals < args.pval_threshold].index.to_list()
print(
"\n%d significant features found by %s (P<%.2f, %s)" %
(len(fset), args.test, args.pval_threshold,
('uncorrected' if not args.use_corrected_pvals else args.fdr_method)))
### k-significant features
if len(fset) > 0:
# Compare k sigfeat and ANOVA significant feature set overlap
if not args.use_corrected_pvals:
k_sigfeats_path = stats_dir / "k_significant_features_uncorrected.csv"
else:
k_sigfeats_path = stats_dir / "k_significant_features.csv"
ksig_table = pd.read_csv(k_sigfeats_path, index_col=0)
fset_ksig = ksig_table[
ksig_table['pvals'] < args.pval_threshold].index.to_list()
fset_overlap = set(fset).intersection(set(fset_ksig))
prop_overlap = len(fset_overlap) / len(fset)
print("%.1f%% overlap with k-significant features" %
(prop_overlap * 100))
if prop_overlap < 0.5 and len(fset) > 100:
print(
"WARNING: Inconsistency in statistics for feature set agreement between "
+ "%s and k significant features!" % args.test)
if args.use_k_sig_feats_overlap:
fset = list(ksig_table.loc[fset_overlap].sort_values(
by='pvals', ascending=True).index)
### t-test
t_test = 't-test' if args.test == 'ANOVA' else 'Mann-Whitney' # aka. Wilcoxon rank-sum
if not args.use_corrected_pvals:
ttest_path = stats_dir / '{}_results_uncorrected.csv'.format(
t_test)
else:
ttest_path = stats_dir / '{}_results.csv'.format(t_test)
# read t-test results + record significant features (NOT ORDERED)
ttest_table = pd.read_csv(ttest_path, index_col=0)
pvals_t = ttest_table[[c for c in ttest_table if "pvals_" in c]]
pvals_t.columns = [c.split('pvals_')[-1] for c in pvals_t.columns]
fset_ttest = pvals_t[(pvals_t < args.pval_threshold).sum(
axis=1) > 0].index.to_list()
print("%d significant features found by %s (P<%.2f, %s)" %
(len(fset_ttest), t_test, args.pval_threshold,
('uncorrected'
if not args.use_corrected_pvals else args.fdr_method)))
else:
print("No significant features found for %s by %s" %
(grouping_var, args.test))
##### PLOTTING #####
if len(fset) > 0:
# Rank strains by number of sigfeats by t-test
ranked_nsig = (pvals_t < args.pval_threshold).sum(axis=0).sort_values(
ascending=False)
# Select top hit strains by n sigfeats (select strains with > 5 sigfeats as hit strains?)
hit_strains_nsig = ranked_nsig[ranked_nsig > 0].index.to_list()
#hit_nuo = ranked_nsig[[i for i in ranked_nsig[ranked_nsig > 0].index if 'nuo' in i]]
# if no sigfaets, subset for top strains ranked by lowest p-value by t-test for any feature
print("%d significant strains (with 1 or more significant features)" %
len(hit_strains_nsig))
if len(hit_strains_nsig) > 0:
write_list_to_file(hit_strains_nsig, stats_dir / 'hit_strains.txt')
# Rank strains by lowest p-value for any feature
ranked_pval = pvals_t.min(axis=0).sort_values(ascending=True)
# Select top 100 hit strains by lowest p-value for any feature
hit_strains_pval = ranked_pval[
ranked_pval < args.pval_threshold].index.to_list()
hit_strains_pval = ranked_pval.index[:N_LOWEST_PVAL].to_list()
write_list_to_file(
hit_strains_pval,
stats_dir / 'lowest{}_pval.txt'.format(N_LOWEST_PVAL))
print("\nPlotting ranked strains by number of significant features")
ranked_nsig_path = plot_dir / (
'ranked_number_sigfeats' + '_' +
('uncorrected' if args.fdr_method is None else args.fdr_method) +
'.png')
plt.ioff()
plt.close('all')
fig, ax = plt.subplots(figsize=(20, 6))
ax.plot(ranked_nsig)
if len(ranked_nsig.index) > 250:
ax.set_xticklabels([])
else:
ax.set_xticklabels(ranked_nsig.index.to_list(),
rotation=90,
fontsize=5)
plt.xlabel("Strains (ranked)", fontsize=12, labelpad=10)
plt.ylabel("Number of significant features", fontsize=12, labelpad=10)
plt.subplots_adjust(left=0.08, right=0.98, bottom=0.15)
plt.savefig(ranked_nsig_path, dpi=600)
print("Plotting ranked strains by lowest p-value of any feature")
lowest_pval_path = plot_dir / (
'ranked_lowest_pval' + '_' +
('uncorrected' if args.fdr_method is None else args.fdr_method) +
'.png')
plt.close('all')
fig, ax = plt.subplots(figsize=(20, 6))
ax.plot(ranked_pval)
plt.axhline(y=args.pval_threshold, c='dimgray', ls='--')
if len(ranked_nsig.index) > 250:
ax.set_xticklabels([])
else:
ax.set_xticklabels(ranked_nsig.index.to_list(),
rotation=90,
fontsize=5)
plt.xlabel("Strains (ranked)", fontsize=12, labelpad=10)
plt.ylabel("Lowest p-value by t-test", fontsize=12, labelpad=10)
plt.subplots_adjust(left=0.08, right=0.98, bottom=0.15)
plt.savefig(lowest_pval_path, dpi=600)
plt.close()
print("\nMaking errorbar plots")
errorbar_sigfeats(features,
metadata,
group_by=grouping_var,
fset=fset,
control=control,
rank_by='mean',
max_feats2plt=args.n_sig_features,
figsize=[20, 10],
fontsize=5,
ms=8,
elinewidth=1.5,
fmt='.',
tight_layout=[0.01, 0.01, 0.99, 0.99],
saveDir=plot_dir / 'errorbar')
# =============================================================================
# print("Making boxplots")
# boxplots_grouped(feat_meta_df=metadata.join(features),
# group_by=grouping_var,
# control_group=control,
# test_pvalues_df=(pvals_t.T if len(fset) > 0 else None),
# feature_set=fset,
# max_feats2plt=args.n_sig_features,
# max_groups_plot_cap=None,
# p_value_threshold=args.pval_threshold,
# drop_insignificant=False,
# sns_colour_palette="tab10",
# figsize=[6,130],
# saveDir=plot_dir / ('boxplots' + '_' + (
# 'uncorrected' if args.fdr_method is None else args.fdr_method) +
# '.png'))
# =============================================================================
# If no sigfeats, subset for top strains ranked by lowest p-value by t-test for any feature
if len(hit_strains_nsig) == 0:
print(
"\Saving lowest %d strains ranked by p-value for any feature" %
N_LOWEST_PVAL)
write_list_to_file(hit_strains_pval,
stats_dir / 'Top100_lowest_pval.txt')
hit_strains = hit_strains_pval
elif len(hit_strains_nsig) > 0:
hit_strains = hit_strains_nsig
# Individual boxplots of significant features by pairwise t-test (each group vs control)
boxplots_sigfeats(
features,
y_class=metadata[grouping_var],
control=control,
pvals=pvals_t,
z_class=metadata['date_yyyymmdd'],
feature_set=None,
saveDir=plot_dir / 'paired_boxplots',
p_value_threshold=args.pval_threshold,
drop_insignificant=True if len(hit_strains) > 0 else False,
max_sig_feats=args.n_sig_features,
max_strains=N_LOWEST_PVAL if len(hit_strains_nsig) == 0 else None,
sns_colour_palette="tab10",
verbose=False)
if SUBSET_HIT_STRAINS:
strain_list = [control] + hit_strains[:TOP_N_HITS]
print("Subsetting for Top%d hit strains" % (len(strain_list) - 1))
features, metadata = subset_results(features,
metadata,
column=grouping_var,
groups=strain_list,
verbose=False)
else:
strain_list = list(metadata[grouping_var].unique())
# =============================================================================
# # NOT NECESSARY FOR ALL STRAINS - LOOK AT CONTROL ONLY FOR THIS
# # superplots of variation with respect to 'date_yyyymmdd'
# print("\nPlotting superplots of date variation for significant features")
# for feat in tqdm(fset[:args.n_sig_features]):
# # plot day variation
# superplot(features, metadata, feat,
# x1='date_yyyymmdd',
# x2=None,
# saveDir=plot_dir / 'superplots',
# figsize=[24,6],
# show_points=False,
# plot_means=True,
# dodge=False)
# # plot run number vs day variation
# superplot(features, metadata, feat,
# x1='date_yyyymmdd',
# x2='imaging_run_number',
# saveDir=plot_dir / 'superplots',
# figsize=[24,6],
# show_points=False,
# plot_means=True,
# dodge=True)
# # plot plate number variation
# superplot(features, metadata, feat,
# x1='date_yyyymmdd',
# x2='source_plate_id',
# saveDir=plot_dir / 'superplots',
# figsize=[24,6],
# show_points=False,
# plot_means=True,
# dodge=True)
# # plot instrument name variation
# superplot(features, metadata, feat,
# x1='date_yyyymmdd',
# x2='instrument_name',
# saveDir=plot_dir / 'superplots',
# figsize=[24,6],
# show_points=False,
# plot_means=True,
# dodge=True)
# =============================================================================
# from tierpsytools.analysis.significant_features import plot_feature_boxplots
# plot_feature_boxplots(feat_to_plot=features,
# y_class=metadata[grouping_var],
# scores=pvals_t.rank(axis=1),
# pvalues=np.asarray(pvals_t).flatten(),
# saveto=None,
# close_after_plotting=True)
##### Hierarchical Clustering Analysis #####
# Z-normalise control data
control_featZ = control_features.apply(zscore, axis=0)
#featZ = (features-features.mean())/features.std() # minus mean, divide by std
#from tierpsytools.preprocessing.scaling_class import scalingClass
#scaler = scalingClass(scaling='standardize')
#featZ = scaler.fit_transform(features)
### Control clustermap
# control data is clustered and feature order is stored and applied to full data
print("\nPlotting control clustermap")
control_clustermap_path = plot_dir / 'heatmaps' / 'date_clustermap.pdf'
cg = plot_clustermap(
control_featZ,
control_metadata,
group_by=([grouping_var] if grouping_var == 'date_yyyymmdd' else
[grouping_var, 'date_yyyymmdd']),
method=METHOD,
metric=METRIC,
figsize=[20, 6],
sub_adj={
'bottom': 0.05,
'left': 0,
'top': 1,
'right': 0.85
},
saveto=control_clustermap_path,
label_size=15,
show_xlabels=False)
# control clustermap with labels
if args.n_top_feats <= 256:
control_clustermap_path = plot_dir / 'heatmaps' / 'date_clustermap_label.pdf'
cg = plot_clustermap(
control_featZ,
control_metadata,
group_by=([grouping_var] if grouping_var == 'date_yyyymmdd' else
[grouping_var, 'date_yyyymmdd']),
method=METHOD,
metric=METRIC,
figsize=[20, 10],
sub_adj={
'bottom': 0.5,
'left': 0,
'top': 1,
'right': 0.85
},
saveto=control_clustermap_path,
label_size=(15, 15),
show_xlabels=True)
#col_linkage = cg.dendrogram_col.calculated_linkage
control_clustered_features = np.array(
control_featZ.columns)[cg.dendrogram_col.reordered_ind]
### Full clustermap
# Z-normalise data for all strains
featZ = features.apply(zscore, axis=0)
## Save z-normalised values
# z_stats = featZ.join(hit_metadata[grouping_var]).groupby(by=grouping_var).mean().T
# z_stats.columns = ['z-mean_' + v for v in z_stats.columns.to_list()]
# z_stats.to_csv(z_stats_path, header=True, index=None)
# Clustermap of full data
print("Plotting all strains clustermap")
full_clustermap_path = plot_dir / 'heatmaps' / (grouping_var +
'_clustermap.pdf')
fg = plot_clustermap(featZ,
metadata,
group_by=grouping_var,
row_colours=None,
method=METHOD,
metric=METRIC,
figsize=[20, 30],
sub_adj={
'bottom': 0.01,
'left': 0,
'top': 1,
'right': 0.95
},
saveto=full_clustermap_path,
label_size=8,
show_xlabels=False)
if args.n_top_feats <= 256:
full_clustermap_path = plot_dir / 'heatmaps' / (
grouping_var + '_clustermap_label.pdf')
fg = plot_clustermap(featZ,
metadata,
group_by=grouping_var,
row_colours=None,
method=METHOD,
metric=METRIC,
figsize=[20, 40],
sub_adj={
'bottom': 0.18,
'left': 0,
'top': 1,
'right': 0.95
},
saveto=full_clustermap_path,
label_size=(15, 10),
show_xlabels=True)
# clustered feature order for all strains
_ = np.array(featZ.columns)[fg.dendrogram_col.reordered_ind]
pvals_heatmap = anova_table.loc[control_clustered_features, 'pvals']
pvals_heatmap.name = 'P < {}'.format(args.pval_threshold)
assert all(f in featZ.columns for f in pvals_heatmap.index)
# Plot heatmap (averaged for each sample)
if len(metadata[grouping_var].unique()) < 250:
print("\nPlotting barcode heatmap")
heatmap_path = plot_dir / 'heatmaps' / (grouping_var + '_heatmap.pdf')
plot_barcode_heatmap(
featZ=featZ[control_clustered_features],
meta=metadata,
group_by=[grouping_var],
pvalues_series=pvals_heatmap,
p_value_threshold=args.pval_threshold,
selected_feats=None, # fset if len(fset) > 0 else None
saveto=heatmap_path,
figsize=[20, 30],
sns_colour_palette="Pastel1",
label_size=10)
##### Principal Components Analysis #####
pca_dir = plot_dir / 'PCA'
# remove outlier samples from PCA
if args.remove_outliers:
outlier_path = pca_dir / 'mahalanobis_outliers.pdf'
features, inds = remove_outliers_pca(df=features, saveto=outlier_path)
metadata = metadata.reindex(features.index) # reindex metadata
featZ = features.apply(zscore, axis=0) # re-normalise data
# Drop features with NaN values after normalising
n_cols = len(featZ.columns)
featZ.dropna(axis=1, inplace=True)
n_dropped = n_cols - len(featZ.columns)
if n_dropped > 0:
print("Dropped %d features after normalisation (NaN)" % n_dropped)
coloured_strains_pca = [control] + hit_strains[:15]
coloured_strains_pca = [
s for s in coloured_strains_pca
if s in metadata[grouping_var].unique()
]
#from tierpsytools.analysis.decomposition import plot_pca
_ = plot_pca(featZ,
metadata,
group_by=grouping_var,
control=control,
var_subset=coloured_strains_pca,
saveDir=pca_dir,
PCs_to_keep=10,
n_feats2print=10,
kde=False,
sns_colour_palette="plasma",
n_dims=2,
label_size=8,
sub_adj={
'bottom': 0.13,
'left': 0.13,
'top': 0.95,
'right': 0.88
},
legend_loc=[1.02, 0.6],
hypercolor=False)
# add details of COG category information to metadata
# (using hard-coded dict of info from Baba et al. 2006 paper)
metadata['COG_category'] = metadata['COG_category'].map(COG_category_dict)
# plot pca coloured by Keio COG category
_ = plot_pca(featZ,
metadata,
group_by='COG_category',
control=None,
var_subset=list(metadata['COG_category'].dropna().unique()),
saveDir=pca_dir / 'COG',
PCs_to_keep=10,
n_feats2print=10,
kde=False,
n_dims=2,
hypercolor=False,
label_size=8,
figsize=[12, 8],
sub_adj={
'bottom': 0.1,
'left': 0.1,
'top': 0.95,
'right': 0.7
},
legend_loc=[1.02, 0.6],
sns_colour_palette="plasma")
##### t-distributed Stochastic Neighbour Embedding #####
print("\nPerforming tSNE")
tsne_dir = plot_dir / 'tSNE'
perplexities = [mean_sample_size
] # NB: should be roughly equal to group size
_ = plot_tSNE(featZ,
metadata,
group_by=grouping_var,
var_subset=coloured_strains_pca,
saveDir=tsne_dir,
perplexities=perplexities,
figsize=[8, 8],
label_size=8,
marker_size=20,
sns_colour_palette="plasma")
print("\nPerforming tSNE")
tsne_dir = plot_dir / 'tSNE'
perplexities = [mean_sample_size
] # NB: should be roughly equal to group size
_ = plot_tSNE(featZ,
metadata,
group_by='COG_category',
var_subset=list(metadata['COG_category'].dropna().unique()),
saveDir=tsne_dir / 'COG_category',
perplexities=perplexities,
figsize=[8, 8],
label_size=8,
marker_size=20,
sns_colour_palette="plasma")
##### Uniform Manifold Projection #####
print("\nPerforming UMAP")
umap_dir = plot_dir / 'UMAP'
n_neighbours = [mean_sample_size
] # NB: should be roughly equal to group size
min_dist = 0.1 # Minimum distance parameter
_ = plot_umap(featZ,
metadata,
group_by=grouping_var,
var_subset=coloured_strains_pca,
saveDir=umap_dir,
n_neighbours=n_neighbours,
min_dist=min_dist,
figsize=[8, 8],
label_size=8,
marker_size=20,
sns_colour_palette="plasma")
dtype={
'comments': str,
'source_plate_id': str
})
# Subset for hit strains only (optional)
if HIT_STRAINS_ONLY:
# Read list from file
strain_list = read_list_from_file(HIT_STRAINS_256_BH_PATH)
print("%d strains found in: %s" %
(len(strain_list), HIT_STRAINS_256_BH_PATH))
# Subset for hit strains
print("Subsetting results for hit strains only")
features, metadata = subset_results(features,
metadata,
column='gene_name',
groups=strain_list)
else:
strain_list = list(metadata['gene_name'].unique())
# Load Tierpsy Top feature set + subset (columns) for top feats only
if N_TOP_FEATS is not None:
top_feats_path = Path(TOP_FEATS_DIR) / "tierpsy_{}.csv".format(
str(N_TOP_FEATS))
topfeats = load_topfeats(top_feats_path,
add_bluelight=True,
remove_path_curvature=True,
header=None)
# Drop features that are not in results
top_feats_list = [
def keio_stats(features, metadata, args):
""" Perform statistical analyses on Keio screen results:
- ANOVA tests for significant between strain variation among all strains for each feature
- t-tests for significant differences between each strain and control for each feature
- k-significant feature selection for agreement with ANOVA significant feature set
Inputs
------
features, metadata : pd.DataFrame
Clean feature summaries and accompanying metadata
args : Object
Python object with the following attributes:
- drop_size_features : bool
- norm_features_only : bool
- percentile_to_use : str
- remove_outliers : bool
- omit_strains : list
- grouping_variable : str
- control_dict : dict
- collapse_control : bool
- n_top_feats : int
- tierpsy_top_feats_dir (if n_top_feats) : str
- test : str
- f_test : bool
- pval_threshold : float
- fdr_method : str
- n_sig_features : int
"""
# categorical variable to investigate, eg.'gene_name'
grouping_var = args.grouping_variable
print("\nInvestigating '%s' variation" % grouping_var)
# assert there will be no errors duee to case-sensitivity
assert len(metadata[grouping_var].unique()) == len(metadata[grouping_var].str.upper().unique())
# Subset results (rows) to omit selected strains
if args.omit_strains is not None:
features, metadata = subset_results(features, metadata, grouping_var, args.omit_strains)
# Load Tierpsy Top feature set + subset (columns) for top feats only
if args.n_top_feats is not None:
top_feats_path = Path(args.tierpsy_top_feats_dir) / "tierpsy_{}.csv".format(str(args.n_top_feats))
topfeats = load_topfeats(top_feats_path, add_bluelight=args.align_bluelight,
remove_path_curvature=True, header=None)
# Drop features that are not in results
top_feats_list = [feat for feat in list(topfeats) if feat in features.columns]
features = features[top_feats_list]
assert not features.isna().any().any()
strain_list = list(metadata[grouping_var].unique())
control = args.control_dict[grouping_var] # control strain to use
assert control in strain_list
if args.collapse_control:
print("Collapsing control data (mean of each day)")
features, metadata = average_plate_control_data(features,
metadata,
control=control,
grouping_var=grouping_var,
plate_var='imaging_plate_id')
_ = df_summary_stats(metadata) # summary df # TODO: plot from?
# Record mean sample size per group
mean_sample_size = int(np.round(metadata.join(features).groupby([grouping_var],
as_index=False).size().mean()))
print("Mean sample size: %d" % mean_sample_size)
# construct save paths (args.save_dir / topfeats? etc)
save_dir = get_save_dir(args)
stats_dir = save_dir / grouping_var / "Stats" / args.fdr_method
plot_dir = save_dir / grouping_var / "Plots" / args.fdr_method
#%% F-test for equal variances
# Compare variance in samples with control (and correct for multiple comparisons)
# Sample size matters in that unequal variances don't pose a problem for a t-test with
# equal sample sizes. So as long as your sample sizes are equal, you don't have to worry about
# homogeneity of variances. If they are not equal, perform F-tests first to see if variance is
# equal before doing a t-test
if args.f_test:
levene_stats_path = stats_dir / 'levene_results.csv'
levene_stats = levene_f_test(features, metadata, grouping_var,
p_value_threshold=args.pval_threshold,
multitest_method=args.fdr_method,
saveto=levene_stats_path,
del_if_exists=False)
# if p < 0.05 then variances are not equal, and sample size matters
prop_eqvar = (levene_stats['pval'] > args.pval_threshold).sum() / len(levene_stats['pval'])
print("Percentage equal variance %.1f%%" % (prop_eqvar * 100))
#%% ANOVA / Kruskal-Wallis tests for significantly different features across groups
test_path_unncorrected = stats_dir / '{}_results_uncorrected.csv'.format(args.test)
test_path = stats_dir / '{}_results.csv'.format(args.test)
if not (test_path.exists() and test_path_unncorrected.exists()):
test_path.parent.mkdir(exist_ok=True, parents=True)
if (args.test == "ANOVA" or args.test == "Kruskal"):
if len(strain_list) > 2:
# perform ANOVA + record results before & after correcting for multiple comparisons
stats, pvals, reject = univariate_tests(X=features,
y=metadata[grouping_var],
control=control,
test=args.test,
comparison_type='multiclass',
multitest_correction=None, # uncorrected
alpha=args.pval_threshold,
n_permutation_test='all')
# get effect sizes
effect_sizes = get_effect_sizes(X=features,
y=metadata[grouping_var],
control=control,
effect_type=None,
linked_test=args.test)
# compile + save results (uncorrected)
test_results = pd.concat([stats, effect_sizes, pvals, reject], axis=1)
test_results.columns = ['stats','effect_size','pvals','reject']
test_results['significance'] = sig_asterix(test_results['pvals'])
test_results = test_results.sort_values(by=['pvals'], ascending=True) # rank pvals
test_results.to_csv(test_path_unncorrected, header=True, index=True)
# correct for multiple comparisons
reject_corrected, pvals_corrected = _multitest_correct(pvals,
multitest_method=args.fdr_method,
fdr=args.pval_threshold)
# compile + save results (corrected)
test_results = pd.concat([stats, effect_sizes, pvals_corrected, reject_corrected], axis=1)
test_results.columns = ['stats','effect_size','pvals','reject']
test_results['significance'] = sig_asterix(test_results['pvals'])
test_results = test_results.sort_values(by=['pvals'], ascending=True) # rank pvals
test_results.to_csv(test_path, header=True, index=True)
# use reject mask to find significant feature set
fset = pvals.loc[reject[args.test]].sort_values(by=args.test, ascending=True).index.to_list()
#assert set(fset) == set(anova_corrected['pvals'].index[np.where(anova_corrected['pvals'] <
#args.pval_threshold)[0]])
if len(fset) > 0:
print("%d significant features found by %s for '%s' (P<%.2f, %s)" % (len(fset),
args.test, grouping_var, args.pval_threshold, args.fdr_method))
anova_sigfeats_path = stats_dir / '{}_sigfeats.txt'.format(args.test)
write_list_to_file(fset, anova_sigfeats_path)
else:
fset = []
print("\nWARNING: Not enough groups for %s for '%s' (n=%d groups)" %\
(args.test, grouping_var, len(strain_list)))
#%% Linear Mixed Models (LMMs), accounting for day-to-day variation
# NB: Ideally report: parameter | beta | lower-95 | upper-95 | random effect (SD)
elif args.test == 'LMM':
with warnings.catch_warnings():
# Filter warnings as parameter is often on the boundary
warnings.filterwarnings("ignore")
#warnings.simplefilter("ignore", ConvergenceWarning)
(signif_effect, low_effect, error, mask, pvals
) = compounds_with_low_effect_univariate(feat=features,
drug_name=metadata[grouping_var],
drug_dose=None,
random_effect=metadata[args.lmm_random_effect],
control=control,
test=args.test,
comparison_type='multiclass',
multitest_method=args.fdr_method)
assert len(error) == 0
# save pvals
pvals.to_csv(test_path_unncorrected, header=True, index=True)
# save significant features -- if any strain significant for any feature
fset = pvals.columns[(pvals < args.pval_threshold).any()].to_list()
if len(fset) > 0:
lmm_sigfeats_path = stats_dir / '{}_sigfeats.txt'.format(args.test)
write_list_to_file(fset, lmm_sigfeats_path)
# save significant effect strains
if len(signif_effect) > 0:
print(("%d significant features found (%d significant %ss vs %s control, "\
% (len(fset), len(signif_effect), grouping_var.replace('_',' '),
control) if len(signif_effect) > 0 else\
"No significant differences found between %s "\
% grouping_var.replace('_',' '))
+ "after accounting for %s variation, %s, P<%.2f, %s)"\
% (args.lmm_random_effect.split('_yyyymmdd')[0], args.test,
args.pval_threshold, args.fdr_method))
signif_effect_path = stats_dir / '{}_signif_effect_strains.txt'.format(args.test)
write_list_to_file(signif_effect, signif_effect_path)
else:
raise IOError("Test '{}' not recognised".format(args.test))
#%% t-tests / Mann-Whitney tests
# t-test to use
t_test = 't-test' if args.test == 'ANOVA' else 'Mann-Whitney' # aka. Wilcoxon rank-sum
ttest_path_uncorrected = stats_dir / '{}_results_uncorrected.csv'.format(t_test)
ttest_path = stats_dir / '{}_results.csv'.format(t_test)
if not (ttest_path_uncorrected.exists() and ttest_path.exists()):
ttest_path.parent.mkdir(exist_ok=True, parents=True)
if len(fset) > 0 or len(strain_list) == 2:
# perform t-tests (without correction for multiple testing)
stats_t, pvals_t, reject_t = univariate_tests(X=features,
y=metadata[grouping_var],
control=control,
test=t_test,
comparison_type='binary_each_group',
multitest_correction=None,
alpha=0.05)
# get effect sizes for comparisons
effect_sizes_t = get_effect_sizes(X=features,
y=metadata[grouping_var],
control=control,
effect_type=None,
linked_test=t_test)
# compile + save t-test results (uncorrected)
stats_t.columns = ['stats_' + str(c) for c in stats_t.columns]
pvals_t.columns = ['pvals_' + str(c) for c in pvals_t.columns]
reject_t.columns = ['reject_' + str(c) for c in reject_t.columns]
effect_sizes_t.columns = ['effect_size_' + str(c) for c in effect_sizes_t.columns]
ttest_uncorrected = pd.concat([stats_t, effect_sizes_t, pvals_t, reject_t], axis=1)
ttest_uncorrected.to_csv(ttest_path_uncorrected, header=True, index=True)
# correct for multiple comparisons
pvals_t.columns = [c.split("_")[-1] for c in pvals_t.columns]
reject_t, pvals_t = _multitest_correct(pvals_t,
multitest_method=args.fdr_method,
fdr=args.pval_threshold)
# compile + save t-test results (corrected)
pvals_t.columns = ['pvals_' + str(c) for c in pvals_t.columns]
reject_t.columns = ['reject_' + str(c) for c in reject_t.columns]
ttest_corrected = pd.concat([stats_t, effect_sizes_t, pvals_t, reject_t], axis=1)
ttest_corrected.to_csv(ttest_path, header=True, index=True)
# record t-test significant features (not ordered)
fset_ttest = pvals_t[np.asmatrix(reject_t)].index.unique().to_list()
#assert set(fset_ttest) == set(pvals_t.index[(pvals_t < args.pval_threshold).sum(axis=1) > 0])
print("%d significant features found for any %s vs %s (%s, P<%.2f)" %\
(len(fset_ttest), grouping_var, control, t_test, args.pval_threshold))
if len(fset_ttest) > 0:
ttest_sigfeats_path = stats_dir / '{}_sigfeats.txt'.format(t_test)
write_list_to_file(fset_ttest, ttest_sigfeats_path)
#%% K significant features
ksig_uncorrected_path = stats_dir / 'k_significant_features_uncorrected.csv'
ksig_corrected_path = stats_dir / 'k_significant_features.csv'
if not (ksig_uncorrected_path.exists() and ksig_corrected_path.exists()):
ksig_corrected_path.parent.mkdir(exist_ok=True, parents=True)
fset_ksig, (scores, pvalues_ksig), support = k_significant_feat(feat=features,
y_class=metadata[grouping_var],
k=len(fset),
score_func='f_classif',
scale=None,
feat_names=None,
plot=False,
k_to_plot=None,
close_after_plotting=True,
saveto=None, #k_sigfeat_dir
figsize=None,
title=None,
xlabel=None)
# compile + save k-significant features (uncorrected)
ksig_table = pd.concat([pd.Series(scores), pd.Series(pvalues_ksig)], axis=1)
ksig_table.columns = ['scores','pvals']
ksig_table.index = fset_ksig
ksig_table.to_csv(ksig_uncorrected_path, header=True, index=True)
# Correct for multiple comparisons
_, ksig_table['pvals'] = _multitest_correct(ksig_table['pvals'],
multitest_method=args.fdr_method,
fdr=args.pval_threshold)
# save k-significant features (corrected)
ksig_table.to_csv(ksig_corrected_path, header=True, index=True)
#%% mRMR feature selection: minimum Redunduncy, Maximum Relevance #####
mrmr_dir = plot_dir / 'mrmr'
mrmr_dir.mkdir(exist_ok=True, parents=True)
mrmr_results_path = mrmr_dir / "mrmr_results.csv"
if not mrmr_results_path.exists():
estimator = Pipeline([('scaler', StandardScaler()), ('estimator', LogisticRegression())])
y = metadata[grouping_var].values
(mrmr_feat_set,
mrmr_scores,
mrmr_support) = mRMR_feature_selection(features, y_class=y, k=10,
redundancy_func='pearson_corr',
relevance_func='kruskal',
n_bins=10, mrmr_criterion='MID',
plot=True, k_to_plot=5,
close_after_plotting=True,
saveto=mrmr_dir, figsize=None)
# save results
mrmr_table = pd.concat([pd.Series(mrmr_feat_set), pd.Series(mrmr_scores)], axis=1)
mrmr_table.columns = ['feature','score']
mrmr_table.to_csv(mrmr_results_path, header=True, index=False)
n_cv = 5
cv_scores_mrmr = cross_val_score(estimator, features[mrmr_feat_set], y, cv=n_cv)
cv_scores_mrmr = pd.DataFrame(cv_scores_mrmr, columns=['cv_score'])
cv_scores_mrmr.to_csv(mrmr_dir / "cv_scores.csv", header=True, index=False)
print('MRMR CV Score: %f (n=%d)' % (np.mean(cv_scores_mrmr), n_cv))
else:
# load mrmr results
mrmr_table = pd.read_csv(mrmr_results_path)
mrmr_feat_set = mrmr_table['feature'].to_list()
print("\nTop %d features found by MRMR:" % len(mrmr_feat_set))
for feat in mrmr_feat_set:
print(feat)
parser.add_argument('--features_path', help="Path to feature summaries file",
default=FEATURES_PATH, type=str)
parser.add_argument('--metadata_path', help="Path to metadata file",
default=METADATA_PATH, type=str)
args = parser.parse_args()
args = load_json(args.json)
# Read feature summaries + metadata
features = pd.read_csv(FEATURES_PATH)
metadata = pd.read_csv(METADATA_PATH, dtype={'comments':str, 'source_plate_id':str})
# Subset for control data only
control_strain = args.control_dict[args.grouping_variable] # control strain to use
control_features, control_metadata = subset_results(features, metadata,
column=args.grouping_variable,
groups=[control_strain])
# Subset for imaging dates of interest
if args.dates is not None:
dates = [int(d) for d in args.dates]
control_features, control_metadata = subset_results(control_features, control_metadata,
column='date_yyyymmdd', groups=dates)
# Clean data after subset - to remove features with zero std
control_features, control_metadata = clean_summary_results(control_features,
control_metadata,
max_value_cap=False,
imputeNaN=False)
# Load Tierpsy Top feature set + subset (columns) for top feats only
if args.n_top_feats is not None:
