def oned_svm(self):
solo_fns = os.listdir(self.solo_dp)
combo_fns = os.listdir(self.combo_dp)
df_dict = {'Label': [], 'Accuracy': []}
for combo_fn in combo_fns:
fpi = os.path.join(self.combo_dp, combo_fn)
combo_df = pd.read_csv(fpi, sep='\t', index_col=0)
labels = [
label for label in list(combo_df)
if label != 'Protein ID' and label != 'y'
]
for label in labels:
full_label = '{} {}'.format(combo_fn[5:-4], label)
df_dict['Label'].append(full_label)
norm_scores = combo_df[label]
X = np.array([norm_scores]).T
y = np.array(combo_df['y']).T
clf = svms.linear_svc(X, y)
print(full_label)
df_dict['Accuracy'].append(clf.score(X, y))
for solo_fn in solo_fns:
full_label = solo_fn[5:-4]
print(full_label)
df_dict['Label'].append(full_label)
fpi = os.path.join(self.solo_dp, solo_fn)
solo_df = pd.read_csv(fpi, sep='\t', index_col=0)
norm_scores = solo_df['Norm Scores']
X = np.array([norm_scores]).T
y = np.array(solo_df['y']).T
clf = svms.linear_svc(X, y)
df_dict['Accuracy'].append(clf.score(X, y))
df_out = pd.DataFrame(df_dict, columns=['Label', 'Accuracy'])
df_out.to_csv(self.oned_fpo, sep='\t')
def svm_len(self):
df_train = pd.read_csv(self.train_fp, sep='\t', index_col=0)
y = np.array(df_train['y']).T
X = np.array(df_train[['Length']])
lin_clf = svms.linear_svc(X, y)
print("The accuracy score for the length is {}".format(
lin_clf.score(X, y)))
def find_hyperplane(self):
"""Show that the hyperplane for the set intercept is close to 0"""
df = pd.read_csv(self.train_fpi, sep='\t', index_col=0)
seqs = list(df['Sequence'])
cs = NormScore()
lc_norm = cs.lc_norm_score(seqs)
X = np.array([lc_norm]).T
y = np.array(df['y']).T
clf = svms.linear_svc(X, y)
xs = np.arange(-2, 2, 0.01).reshape(1, -1).T
dists = list(clf.decision_function(xs))
for x, dist in zip(xs, dists):
if dist < 0:
print(x)
break
def check_one_charge(self):
"""
Result. If you remove K, R, E, your classification accuracy goes to 0.71
Hypothesis: it is the LCAs with K/R/E that matter the most for
classification. So what if we only count LCAs with a charged residue?
"""
df = pd.read_csv(self.train_fpi, sep='\t', index_col=0)
#df = df[df['y'] == 0]
seqs = list(df['Sequence'])
lca_counts = self.count_lca_charge(seqs)
#plt.hist(lca_counts, bins=20, range=(0, 70))
#plt.ylim([0, 900])
#plt.show()
X = np.array([lca_counts]).T
y = np.array(df['y']).T
clf = svms.linear_svc(X, y)
print(clf.score(X, y))
def raw_svm(self, fpi, fpo):
df_dict = {'SVM score': [], 'Label': []}
cols = ['Label', 'SVM score']
rem_cols = ['Protein ID', 'Length', 'y']
df_in = pd.read_csv(fpi, sep='\t', index_col=0)
k_lcs = [
lab for lab in df_in.columns.values.tolist() if lab not in rem_cols
]
for i, k_lc in enumerate(k_lcs):
print(i)
print(k_lc)
raw_scores = df_in[k_lc]
X = np.array([raw_scores]).T
y = np.array(df_in['y']).T
clf = svms.linear_svc(X, y)
df_dict['SVM score'].append(clf.score(X, y))
df_dict['Label'].append(k_lc)
df = pd.DataFrame(df_dict, columns=cols)
df.to_csv(fpo, sep='\t')
def svm_comp(self):
df_train = pd.read_csv(self.comp_fp, sep='\t', index_col=0)
y = np.array(df_train['y']).T
scores = []
for aa in self.aas:
cols = [aa]
X = np.array(df_train[cols])
lin_clf = svms.linear_svc(X, y)
score = lin_clf.score(X, y)
scores.append(score)
print("The mean accuracy is {} and standard deviation is {} for the "
"fraction of each amino acid used separately to "
"classify.".format(np.mean(scores), np.std(scores)))
all_cols = [aa for aa in self.aas]
X = np.array(df_train[all_cols])
rbf_clf = svms.smooth_rbf(X, y)
score = rbf_clf.score(X, y)
print("The accuracy score for the fraction of all amino acids used "
"to classify is {}".format(score))
def run_svm(self, motif_sum, y):
X = np.array([motif_sum]).T
clf = svms.linear_svc(X, y)
return clf.score(X, y)