def get_error_table_using_percentile_positives_new(err_df, target_scores, num_null):
""" transfer error statistics in err_df for many target scores and given
number of estimated null hypothesises 'num_null' """
num_total = len(target_scores)
num_alternative = num_total - num_null
target_scores = np.sort(to_one_dim_array(target_scores)) # ascending
# optimized
num_positives = count_num_positives(target_scores.astype(np.float64))
num_negatives = num_total - num_positives
# the last coertion is needed because depending on the scale of num_total
# numpy switched to 64 bit floats
pp = (num_positives.astype(np.float32) / num_total).astype(np.float32)
# find best matching row in err_df for each percentile_positive in pp:
imax = find_nearest_matches(err_df.percentile_positive.values, pp)
qvalues = err_df.qvalue.iloc[imax].values
svalues = err_df.svalue.iloc[imax].values
pvalues = err_df.pvalue.iloc[imax].values
fdr = err_df.FDR.iloc[imax].values
fdr[fdr < 0.0] = 0.0
fdr[fdr > 1.0] = 1.0
fdr[num_positives == 0] = 0.0
fp = np.round(fdr * num_positives)
tp = num_positives - fp
tn = num_null - fp
fn = num_negatives - tn
sens = tp / num_alternative
if num_alternative == 0:
sens = np.zeros_like(tp)
sens[sens < 0.0] = 0.0
sens[sens > 1.0] = 1.0
df_error = pd.DataFrame(
dict(qvalue=qvalues,
svalue=svalues,
pvalue=pvalues,
TP=tp,
FP=fp,
TN=tn,
FN=fn,
FDR=fdr,
sens=sens,
cutoff=target_scores),
columns="qvalue svalue pvalue TP FP TN FN FDR sens cutoff".split(),
)
return df_error
def get_error_table_using_percentile_positives_new(err_df, target_scores,
num_null):
""" transfer error statistics in err_df for many target scores and given
number of estimated null hypothesises 'num_null' """
num_total = len(target_scores)
num_alternative = num_total - num_null
target_scores = np.sort(to_one_dim_array(target_scores)) # ascending
# optimized
num_positives = count_num_positives(target_scores.astype(np.float64))
num_negatives = num_total - num_positives
# the last coertion is needed because depending on the scale of num_total
# numpy switched to 64 bit floats
pp = (num_positives.astype(np.float32) / num_total).astype(np.float32)
# find best matching row in err_df for each percentile_positive in pp:
imax = find_nearest_matches(err_df.percentile_positive.values, pp)
qvalues = err_df.qvalue.iloc[imax].values
svalues = err_df.svalue.iloc[imax].values
fdr = err_df.FDR.iloc[imax].values
fdr[fdr < 0.0] = 0.0
fdr[fdr > 1.0] = 1.0
fdr[num_positives == 0] = 0.0
fp = np.round(fdr * num_positives)
tp = num_positives - fp
tn = num_null - fp
fn = num_negatives - tn
sens = tp / num_alternative
if num_alternative == 0:
sens = np.zeros_like(tp)
sens[sens < 0.0] = 0.0
sens[sens > 1.0] = 1.0
df_error = pd.DataFrame(
dict(qvalue=qvalues,
svalue=svalues,
TP=tp,
FP=fp,
TN=tn,
FN=fn,
FDR=fdr,
sens=sens,
cutoff=target_scores),
columns="qvalue svalue TP FP TN FN FDR sens cutoff".split(),
)
return df_error
def get_error_table_from_pvalues_new(p_values, lambda_=0.4, use_pfdr=False):
""" estimate error table from p_values with method of storey for estimating fdrs and q-values
"""
# sort descending:
p_values = np.sort(to_one_dim_array(p_values))[::-1]
# estimate FDR with storeys method:
num_null = 1.0 / (1.0 - lambda_) * (p_values >= lambda_).sum()
num_total = len(p_values)
# optimized with numpys broadcasting: comparing column vector with row
# vector yields a matrix with pairwise comparison results. sum(axis=0)
# sums up each column:
num_positives = count_num_positives(p_values)
num_negatives = num_total - num_positives
pp = 1.0 * num_positives / num_total
tp = num_positives - num_null * p_values
fp = num_null * p_values
tn = num_null * (1.0 - p_values)
fn = num_negatives - num_null * (1.0 - p_values)
fdr = fp / num_positives
# storey published pFDR as an improvement over FDR,
# see http://www.genomine.org/papers/directfdr.pdf:
if use_pfdr:
fac = 1.0 - (1.0 - p_values)**num_total
fdr /= fac
# if we take the limit p->1 we achieve the following factor:
fdr[p_values == 0] = 1.0 / num_total
# cut off values to range 0..1
fdr[fdr < 0.0] = 0.0
fdr[fdr > 1.0] = 1.0
# estimate false non-discovery rate
fnr = fn / num_negatives
# storey published pFDR as an improvement over FDR,
# see http://www.genomine.org/papers/directfdr.pdf:
if use_pfdr:
fac = 1.0 - p_values**num_total
fnr /= fac
# if we take the limit p->1 we achieve the following factor:
fnr[p_values == 0] = 1.0 / num_total
# cut off values to range 0..1
fnr[fnr < 0.0] = 0.0
fnr[fnr > 1.0] = 1.0
sens = tp / (num_total - num_null)
# cut off values to range 0..1
sens[sens < 0.0] = 0.0
sens[sens > 1.0] = 1.0
if num_null:
fpr = fp / num_null
else:
fpr = 0.0 * fp
# assemble statistics as data frame
df = pd.DataFrame(
dict(
pvalue=p_values.flatten().astype(np.float32),
percentile_positive=pp.flatten().astype(np.float32),
positive=num_positives.flatten().astype(np.float32),
negative=num_negatives.flatten().astype(np.float32),
TP=tp.flatten().astype(np.float32),
FP=fp.flatten().astype(np.float32),
TN=tn.flatten().astype(np.float32),
FN=fn.flatten().astype(np.float32),
FDR=fdr.flatten().astype(np.float32),
FNR=fnr.flatten().astype(np.float32),
sens=sens.flatten().astype(np.float32),
FPR=fpr.flatten().astype(np.float32),
),
columns="""pvalue percentile_positive positive negative TP FP
TN FN FDR FNR sens FPR""".split(),
)
# cummin/cummax not available in numpy, so we create them from dataframe
# here:
df["qvalue"] = df.FDR.cummin()
df["svalue"] = df.sens[::-1].cummax()[::-1]
return ErrorStatistics(df, num_null, num_total)
def get_error_table_from_pvalues_new(p_values, lambda_=0.4):
""" estimate error table from p_values with method of storey for estimating fdrs and q-values
"""
# sort descending:
p_values = np.sort(to_one_dim_array(p_values))[::-1]
# estimate FDR with storeys method:
num_null = 1.0 / (1.0 - lambda_) * (p_values >= lambda_).sum()
num = len(p_values)
# p_values = p_values[:,None]
# optimized with numpys broadcasting: comparing column vector with row
# vector yields a matrix with pairwise comparison results. sum(axis=0)
# sums up each column:
num_positives = count_num_positives(p_values)
num_negatives = num - num_positives
pp = 1.0 * num_positives / num
tp = num_positives - num_null * p_values
fp = num_null * p_values
tn = num_null * (1.0 - p_values)
fn = num_negatives - num_null * (1.0 - p_values)
fdr = fp / num_positives
# cut off values to range 0..1
fdr[fdr < 0.0] = 0.0
fdr[fdr > 1.0] = 1.0
sens = tp / (num - num_null)
# cut off values to range 0..1
sens[sens < 0.0] = 0.0
sens[sens > 1.0] = 1.0
if num_null:
fpr = fp / num_null
else:
fpr = 0.0 * fp
# assemble statistics as data frame
error_stat = pd.DataFrame(
dict(pvalue=p_values.flatten(),
percentile_positive=pp.flatten(),
positive=num_positives.flatten(),
negative=num_negatives.flatten(),
TP=tp.flatten(),
FP=fp.flatten(),
TN=tn.flatten(),
FN=fn.flatten(),
FDR=fdr.flatten(),
sens=sens.flatten(),
FPR=fpr.flatten()),
columns="""pvalue percentile_positive positive negative TP FP
TN FN FDR sens FPR""".split()
)
# cummin/cummax not available in numpy, so we create them from dataframe
# here:
error_stat["qvalue"] = error_stat.FDR.cummin()
error_stat["svalue"] = error_stat.sens[::-1].cummax()[::-1]
return error_stat, num_null, num
def get_error_table_from_pvalues_new(p_values, lambda_=0.4, use_pfdr=False):
""" estimate error table from p_values with method of storey for estimating fdrs and q-values
"""
# sort descending:
p_values = np.sort(to_one_dim_array(p_values))[::-1]
# estimate FDR with storeys method:
num_null = 1.0 / (1.0 - lambda_) * (p_values >= lambda_).sum()
num_total = len(p_values)
# optimized with numpys broadcasting: comparing column vector with row
# vector yields a matrix with pairwise comparison results. sum(axis=0)
# sums up each column:
num_positives = count_num_positives(p_values)
num_negatives = num_total - num_positives
pp = 1.0 * num_positives / num_total
tp = num_positives - num_null * p_values
fp = num_null * p_values
tn = num_null * (1.0 - p_values)
fn = num_negatives - num_null * (1.0 - p_values)
fdr = fp / num_positives
# storey published pFDR as an improvement over FDR,
# see http://www.genomine.org/papers/directfdr.pdf:
if use_pfdr:
fac = 1.0 - (1.0 - p_values) ** num_total
fdr /= fac
# if we take the limit p->1 we achieve the following factor:
fdr[p_values == 0] = 1.0 / num_total
# cut off values to range 0..1
fdr[fdr < 0.0] = 0.0
fdr[fdr > 1.0] = 1.0
sens = tp / (num_total - num_null)
# cut off values to range 0..1
sens[sens < 0.0] = 0.0
sens[sens > 1.0] = 1.0
if num_null:
fpr = fp / num_null
else:
fpr = 0.0 * fp
# assemble statistics as data frame
df = pd.DataFrame(
dict(pvalue=p_values.flatten().astype(np.float32),
percentile_positive=pp.flatten().astype(np.float32),
positive=num_positives.flatten().astype(np.float32),
negative=num_negatives.flatten().astype(np.float32),
TP=tp.flatten().astype(np.float32),
FP=fp.flatten().astype(np.float32),
TN=tn.flatten().astype(np.float32),
FN=fn.flatten().astype(np.float32),
FDR=fdr.flatten().astype(np.float32),
sens=sens.flatten().astype(np.float32),
FPR=fpr.flatten().astype(np.float32),
),
columns="""pvalue percentile_positive positive negative TP FP
TN FN FDR sens FPR""".split(),
)
# cummin/cummax not available in numpy, so we create them from dataframe
# here:
df["qvalue"] = df.FDR.cummin()
df["svalue"] = df.sens[::-1].cummax()[::-1]
return ErrorStatistics(df, num_null, num_total)
