def test_from_biadjacency_weight(self):
M = sparse.csc_matrix([[1,2],[0,3]])
B = bipartite.from_biadjacency_matrix(M)
assert_edges_equal(B.edges(),[(0,2),(0,3),(1,3)])
B = bipartite.from_biadjacency_matrix(M, edge_attribute='weight')
e = [(0,2,{'weight':1}),(0,3,{'weight':2}),(1,3,{'weight':3})]
assert_edges_equal(B.edges(data=True),e)
def test_from_biadjacency_weight(self):
M = sparse.csc_matrix([[1,2],[0,3]])
B = bipartite.from_biadjacency_matrix(M)
assert_edges_equal(B.edges(),[(0,2),(0,3),(1,3)])
B = bipartite.from_biadjacency_matrix(M, edge_attribute='weight')
e = [(0,2,{'weight':1}),(0,3,{'weight':2}),(1,3,{'weight':3})]
assert_edges_equal(B.edges(data=True),e)
def load(path):
biadjacency = sparse.load_npz(path)
bipartite = nxb.from_biadjacency_matrix(biadjacency)
print(f'{bipartite.number_of_edges():,} edges in the bipartite graph')
print(f'connected: {nx.is_connected(bipartite)}')
# nx.write_graphml(bipartite, 's2_2_bipartite_graph/paper_author.graphml')
return bipartite
def Draw_Bipartite_Graph(user, seq_item, attn_value, name, save=True):
if True:
fig = plt.figure(name)
entity2ids, index_user, index_item = dict(), 0, 0
for i in range(len(user)):
if user[i] not in entity2ids:
entity2ids[user[i]] = index_user
index_user += 1
for i in range(len(seq_item)):
if seq_item[i] not in entity2ids:
entity2ids[seq_item[i]] = index_item
index_item += 1
row = [entity2ids[user[i]] for i in range(len(user))]
col = [entity2ids[seq_item[i]] for i in range(len(seq_item))]
X_name = [i for i in range(len(set(row)))]
Y_name = [i + len(set(row)) for i in range(len(set(col)))]
a_matrix = coo_matrix((attn_value, (row, col))).toarray()
a_matrix = coo_matrix(a_matrix)
G = bipartite.from_biadjacency_matrix(a_matrix,
create_using=None,
edge_attribute='weight')
pos = dict()
Y_len = int((len(Y_name) - 1) * 10)
X_unit_len = int(Y_len / (len(X_name) + 1))
pos.update(
(n, (0, (i + 1) * X_unit_len)) for i, n in enumerate(X_name))
pos.update((n, (0.5, i * 10)) for i, n in enumerate(Y_name))
num_edges = G.number_of_edges()
num_nodes = G.number_of_nodes()
color_map = []
for node in G:
if node < len(set(user)):
color_map.append('xkcd:red')
else:
color_map.append('xkcd:blue')
nx.draw(
G,
pos=pos, #with_labels=True,
edge_color=attn_value,
edge_cmap=plt.get_cmap('rainbow'),
node_color=color_map,
cmap=plt.get_cmap('Reds'))
plt.savefig('/home/hsucheng/DRS/code/RS_2/graph/draw_test-' +
str(name) + '.png')
plt.close(name)
def birkhoff_von_neumann(Y, tol=0.0001):
if Y.shape[0] != Y.shape[1]:
raise ValueError('Y.shape[0] != Y.shape[1]')
if np.any(Y < -tol):
raise ValueError('np.any(Y < -tol)')
Y = np.where(Y < tol, 0, Y)
m = Y.shape[0]
lambdas = []
perms = []
residuals = Y > tol
while np.any(residuals):
adj = residuals.astype(int)
adj = sparse.csr_matrix(adj)
G = bp.from_biadjacency_matrix(adj)
M = bp.maximum_matching(G)
M_ = [(kk, v - m) for kk, v in M.items() if kk < m]
if len(
M_
) < m: # this can happen due to numerical stability issues TODO add test
break
M_ = sorted(
M_, key=itemgetter(0)
) # if tuples sorted by rows, then the columns are the permutation
rows, columns = zip(*M_)
perm = np.array(columns)
assert perm.shape == (m, )
lambda_ = np.min(Y[rows, columns])
P = np.zeros((m, m), dtype=float)
P[rows, columns] = 1.
lambdas.append(lambda_)
perms.append(perm)
Y -= lambda_ * P
residuals = Y > tol
return np.array(lambdas), np.array(perms)
def matching(idxA, idxB):
labelsA = np.unique(idxA)
m = len(labelsA)
labelsB = np.unique(idxB)
n = len(labelsB)
W = np.zeros((m, n))
for j in range(n):
for i in range(m):
W[i, j] = -sum(idxA[idxB == labelsB[j]] == labelsA[i])
G = bipartite.from_biadjacency_matrix(scipy.sparse.coo_matrix(W))
top_nodes = {n for n, d in G.nodes(data=True) if d["bipartite"] == 0}
match = bipartite.minimum_weight_full_matching(G, top_nodes)
new_labels = dict()
for i in range(n):
new_labels[labelsB[i]] = labelsA[match[i+n]]
matched = np.array([new_labels[idxB[i]] for i in range(len(idxB))])
return matched
def bipartite_graph_from_matrix(matrix,
labels1,
labels2,
threshold=0,
match=False):
if match:
matrix, permutation = sort_matrix(matrix)
labels1 = [labels1[int(permutation[i])] for i in range(len(labels1))]
label_dict = {i: l for i, l in enumerate(labels1 + labels2)}
print("Topic labels:", label_dict)
if not isinstance(matrix, np.ndarray):
matrix = np.array(matrix)
matrix[matrix < threshold] = 0
sp_matrix = sparse.coo_matrix(matrix)
g = bipartite.from_biadjacency_matrix(sp_matrix)
nx.set_node_attributes(g, label_dict, name='label')
return g
else:
numero += i
f.close()
#matrix.pop()
matrix.pop()
numpy_matrix = np.matrix(matrix)
return numpy_matrix
matrix = leer_matrix("Code_CPP/matrix.csv")
adjacency = scipy.sparse.csc_matrix(matrix)
G = bi.from_biadjacency_matrix(adjacency)
infected = leer_matrix("Code_CPP/Datos/infectados.csv")
color = bi.color(G)
'''
pos=nx.spring_layout(G)
w = csv.writer(open("output.csv", "w"))
for key,val in pos.items():
w.writerow([key,val[0],val[1]])
'''
pos = {}
for i in range(10):
pos[i] = np.array([0, 1 - (0.2 * i)])
for i in range(9):
pos[10 + i] = np.array([0.2, 0.9 - (0.2 * i)])
#https://stackoverflow.com/questions/35392342/how-to-change-colours-of-nodes-and-edges-of-bipartite-graph-in-networkx
import numpy as np
import networkx as nx
import matplotlib.pyplot as plt
from networkx.algorithms import bipartite
import scipy.sparse as sparse
a_matrix = sparse.rand(5, 10, format='coo', density=0.8)
G = bipartite.from_biadjacency_matrix(a_matrix,
create_using=None,
edge_attribute='weight')
X, Y = bipartite.sets(G)
pos = dict()
pos.update((n, (-1, i * 11)) for i, n in enumerate(X))
pos.update((n, (0.5, i * 5)) for i, n in enumerate(Y))
num_edges = G.number_of_edges()
num_nodes = G.number_of_nodes()
#nx.draw(G, pos=pos, with_labels=True,edge_color=np.random.random(num_edges), edge_cmap=plt.get_cmap('Blues'), node_color=np.random.random(num_nodes), cmap=plt.get_cmap('Reds'))
nx.draw(G,
pos=pos,
with_labels=True,
edge_color=np.random.random(num_edges),
edge_cmap=plt.get_cmap('Reds'),
node_color=range(num_nodes),
node_size=1400,
cmap=plt.cm.Reds)
#C = nx.connected_component_subgraphs(G)
#for g in C:
def test_from_biadjacency_multigraph(self):
M = sparse.csc_matrix([[1, 2], [0, 3]])
B = bipartite.from_biadjacency_matrix(M, create_using=nx.MultiGraph())
assert_edges_equal(B.edges(), [(0, 2), (0, 3), (0, 3), (1, 3), (1, 3),
(1, 3)])
def test_from_biadjacency_roundtrip(self):
B1 = nx.path_graph(5)
M = bipartite.biadjacency_matrix(B1, [0, 2, 4])
B2 = bipartite.from_biadjacency_matrix(M)
assert_true(nx.is_isomorphic(B1, B2))
def aco_preprocessing(path_expr,
path_ppi,
col,
val1,
val2,
patients_info,
log2,
gene_list=None,
size=None,
HI_calc="exact",
sample=None):
#gene_list - preselected genes
#th if genes are not preselected specify threshold for standard deviation selection
# HI calculation is either exact or corelation based (for big datasets)
expr = pd.read_csv(path_expr, sep="\t")
expr = expr.set_index("Unnamed: 0")
group1_true = list(expr[expr[col] == val1].index)
group2_true = list(expr[expr[col] == val2].index)
patients_new = group1_true + group2_true
if sample != None:
idx = list(expr.index)
new_idx = np.random.choice(idx, int(sample * len(idx)), False)
expr = expr.loc[new_idx]
group1_true = list(expr[expr[col] == val1].index)
group2_true = list(expr[expr[col] == val2].index)
patients_new = group1_true + group2_true
expr = expr.loc[patients_new]
net = pd.read_csv(path_ppi, sep="\t", header=None)
nodes_ppi = set(net[0]).union(set(net[1]))
genes_ge = list(set(expr.columns) - set(patients_info))
new_genes = [int(x) for x in genes_ge]
intersec_genes = set.intersection(set(new_genes), set(nodes_ppi))
genes_for_expr = [str(x) for x in list(intersec_genes)]
expr = expr[genes_for_expr]
#20188 genes
if log2:
expr = np.log2(expr)
z_scores = stats.zscore(expr)
z_scores = pd.DataFrame(z_scores, columns=expr.columns, index=expr.index)
if gene_list != None and size == None: # gene list is given
new_genes = [str(gene) for gene in gene_list]
elif gene_list == None and size != None: #std selection
std_genes = expr[genes_for_expr].std()
std_genes, genes_for_expr = zip(
*sorted(zip(std_genes, genes_for_expr)))
genes_for_expr = genes_for_expr[len(std_genes) - size:]
new_genes = list(genes_for_expr)
elif gene_list == None and size == None: #all genes
new_genes = genes_for_expr
else:
print(
"please specify gene selection method: predifined list, standart deviation filtering or none of them"
)
return ()
expr = expr[new_genes]
z_scores = z_scores[new_genes].values
labels_B = dict()
rev_labels_B = dict()
node = 0
#nodes = set(deg_nodes + genes_aco)
for g in new_genes:
labels_B[node] = g
rev_labels_B[g] = node
node = node + 1
for p in patients_new:
labels_B[node] = p
rev_labels_B[p] = node
node = node + 1
#scaler = preprocessing.MinMaxScaler(feature_range=(0, 1))
#sim = scaler.fit_transform(expr)
data_aco = pd.DataFrame(z_scores, columns=new_genes, index=patients_new)
data_aco = data_aco.T
n, m = data_aco.shape
GE = pd.DataFrame(data_aco.values,
index=np.arange(n),
columns=np.arange(n, n + m))
t = 2
b = np.matrix(data_aco > t)
b_sp = csr_matrix(b)
B = bipartite.from_biadjacency_matrix(b_sp)
G = nx.Graph()
G.add_nodes_from(np.arange(n))
for row in net.itertuples():
node1 = str(row[1])
node2 = str(row[2])
if node1 in set(new_genes) and node2 in set(new_genes):
G.add_edge(rev_labels_B[node1], rev_labels_B[node2])
A_new = nx.adj_matrix(G).todense()
A_j = joined_net(B, G)
if HI_calc == "exact":
H = hi(A_j, n, m)
if HI_calc == "corr":
H = HI_big(data_aco, gtg_weight=1, gtp_weight=1, ptp_weight=1)
group1_true_ids = [rev_labels_B[x] for x in group1_true]
group2_true_ids = [rev_labels_B[x] for x in group2_true]
return B, G, H, n, m, GE, A_new, group1_true_ids, group2_true_ids, labels_B, rev_labels_B
def sim_data(genes1, genes2, background, patients1, patients2, dens):
n = genes1 + genes2 + background
m = patients1 + patients2
genes = np.arange(n)
groups_genes = list(np.ones(genes1)) + list(np.ones(genes2) * 2) + list(
np.ones(background) * 3)
groups_p = [1 if node < patients1 else 2 for node in range(m)]
to_sparce = 0.3 #to sparcify bipartite
to_mix = 0.99 # to mix edges berween groups
b = np.zeros((n, m))
ge = np.random.normal(0, 1, n * m).reshape(n, m)
for patient in range(m):
for gene in range(n):
p_gr = groups_p[patient]
g_gr = groups_genes[gene]
if p_gr == 1 and g_gr == 1: #all up
ge[gene, patient] = np.random.normal(1, 0.35, 1)
elif p_gr == 2 and g_gr == 2:
ge[gene, patient] = np.random.normal(1, 0.35, 1) #also up
elif p_gr == 1 and g_gr == 2:
ge[gene, patient] = np.random.normal(-1, 0.35, 1) #down
elif p_gr == 2 and g_gr == 1:
ge[gene, patient] = np.random.normal(-1, 0.35, 1) #down
for patient in range(m):
for gene in range(genes1 + genes2):
prob = np.random.uniform(0, 1)
if prob > 0.9:
ge[gene, patient] = np.random.normal(0, 1, 1)
for gene in range(genes1 + genes2, n):
prob = np.random.uniform(0, 1)
if prob < 0.05:
for patient in range(m):
if groups_p[patient] == 1: #all up
ge[gene, patient] = np.random.normal(0.3, 0.35, 1)
else:
ge[gene, patient] = np.random.normal(-0.3, 0.35, 1)
if prob > 0.05 and prob < 0.1:
for patient in range(m):
if groups_p[patient] == 1: #all up
ge[gene, patient] = np.random.normal(-0.3, 0.35, 1)
else:
ge[gene, patient] = np.random.normal(0.3, 0.35, 1)
g1 = nx.barabasi_albert_graph(genes1, 1)
g2 = nx.barabasi_albert_graph(genes2, 1)
g3 = nx.barabasi_albert_graph(background, 1)
G = nx.disjoint_union(g1, g2)
G = nx.disjoint_union(G, g3)
for _ in range(int(dens * n)):
node1 = np.random.randint(0, genes1)
node2 = np.random.randint(genes1, genes1 + genes2)
node3_1 = np.random.randint(genes1 + genes2, n)
node3_2 = np.random.randint(genes1 + genes2, n)
G.add_edges_from([(node1, node3_1), (node2, node3_2)])
d = nx.density(G)
count = 0
while d > 0.002 and count < 10:
node3_1 = np.random.randint(genes1 + genes2, n)
node3_2 = np.random.randint(genes1 + genes2, n)
count = count + 1
if G.has_edge(node3_1, node3_2):
G.remove_edge(node3_1, node3_2)
d = nx.density(G)
#A_g = nx.adj_matrix(G).todense() *1
b_sp = csr_matrix(b) #sparse matrix for making bipartite graph
B = bipartite.from_biadjacency_matrix(b_sp)
GE = pd.DataFrame(ge, index=np.arange(n), columns=np.arange(n, n + m))
H = HI_big(GE, 1, 1, 1)
return (B, GE, G, H, d, n, m)
def aco_preprocessing_strings(expr_str,
ppi_str,
col,
log2,
gene_list=None,
size=None,
sample=None):
# path_expr - path for gene expression
# path_ppi - path for ppi
# col - split variable name (ONLY TWO CLASSES)
# log2 - log2 transform
#gene_list - preselected genes (if any)
#size - if genes are not preselected specify size of the gene set for standard deviation selection
# sample = None - all patients, otherwise specify fraction of patients taken
EXPRDATA = StringIO(expr_str)
expr = pd.read_csv(EXPRDATA, sep="\t")
expr = expr.set_index("Unnamed: 0")
#TODO: check if column 'prognosis' or 'cancer type' exists, set column based on this info
if ('cancer_type' in list(expr)):
col = 'cancer_type'
else:
col = 'prognosis'
val1, val2 = list(set(expr[col]))
group1_true = list(expr[expr[col] == val1].index)
group2_true = list(expr[expr[col] == val2].index)
patients_new = group1_true + group2_true
if sample != None:
idx = list(expr.index)
new_idx = np.random.choice(idx, int(sample * len(idx)), False)
expr = expr.loc[new_idx]
group1_true = list(expr[expr[col] == val1].index)
group2_true = list(expr[expr[col] == val2].index)
patients_new = group1_true + group2_true
expr = expr.loc[patients_new]
PPIDATA = StringIO(ppi_str)
net = pd.read_csv(PPIDATA, sep="\t", header=None)
nodes_ppi = set(net[0]).union(set(net[1]))
genes_ge = list(set(expr.columns) - set([col]))
new_genes = [int(x) for x in genes_ge]
intersec_genes = set.intersection(set(new_genes), set(nodes_ppi))
genes_for_expr = [str(x) for x in list(intersec_genes)]
expr = expr[genes_for_expr]
#20188 genes
if log2:
expr = np.log2(expr)
z_scores = stats.zscore(expr)
z_scores = pd.DataFrame(z_scores, columns=expr.columns, index=expr.index)
if gene_list != None and size == None: # gene list is given
new_genes = [str(gene) for gene in gene_list]
elif gene_list == None and size != None: #std selection
std_genes = expr[genes_for_expr].std()
std_genes, genes_for_expr = zip(
*sorted(zip(std_genes, genes_for_expr)))
genes_for_expr = genes_for_expr[len(std_genes) - size:]
new_genes = list(genes_for_expr)
elif gene_list == None and size == None: #all genes
new_genes = genes_for_expr
else:
print(
"please specify gene selection method: predifined list, standart deviation filtering or none of them"
)
return ()
expr = expr[new_genes]
z_scores = z_scores[new_genes].values
labels_B = dict()
rev_labels_B = dict()
node = 0
#nodes = set(deg_nodes + genes_aco)
for g in new_genes:
labels_B[node] = g
rev_labels_B[g] = node
node = node + 1
for p in patients_new:
labels_B[node] = p
rev_labels_B[p] = node
node = node + 1
#scaler = preprocessing.MinMaxScaler(feature_range=(0, 1))
#sim = scaler.fit_transform(expr)
data_aco = pd.DataFrame(z_scores, columns=new_genes, index=patients_new)
data_aco = data_aco.T
n, m = data_aco.shape
GE = pd.DataFrame(data_aco.values,
index=np.arange(n),
columns=np.arange(n, n + m))
t = 2
b = np.matrix(data_aco > t)
b_sp = csr_matrix(b)
B = bipartite.from_biadjacency_matrix(b_sp)
G = nx.Graph()
G.add_nodes_from(np.arange(n))
for row in net.itertuples():
node1 = str(row[1])
node2 = str(row[2])
if node1 in set(new_genes) and node2 in set(new_genes):
G.add_edge(rev_labels_B[node1], rev_labels_B[node2])
A_new = nx.adj_matrix(G).todense()
H = HI_big(data_aco, gtg_weight=1, gtp_weight=1, ptp_weight=1)
group1_true_ids = [rev_labels_B[x] for x in group1_true]
group2_true_ids = [rev_labels_B[x] for x in group2_true]
#print(group1_true + "babaaba")
return B, G, H, n, m, GE, A_new, group1_true_ids, group2_true_ids, labels_B, rev_labels_B, val1, val2, group1_true, group2_true
if tempCol[j] > r:
closeSegments = 2 * closeSegments
elif tempCol[j] < (-r):
closeSegments = 2 * closeSegments + 1
else:
closeSegments = np.concatenate(
(2 * closeSegments, 2 * closeSegments + 1))
#closeTargets = np.reshape(targetPoints[closeSegments], (closeSegments.shape[0] * k, d))
for j in range(closeSegments.shape[0]):
norms = np.linalg.norm(targetPoints[closeSegments[j]] -
latentPoints[i],
axis=-1)
smallNorms = norms < r
data = norms[smallNorms]
rows = np.repeat(i, data.shape[0])
cols = np.arange(k)[smallNorms] + (closeSegments[j] * k)
B = B + coo_matrix(
(data, (rows, cols)), shape=(n, (2**d) * k), dtype=np.float16)
#closeTargets = targetPoints[closeSegments]
#norms = np.linalg.norm(closeTargets - latentPoints[i], axis=-1)
#for j in range(containments.shape[0]):
# closeLatentPoints[containments[j]].extend([i])
#create the biadjacency sparse matrix of the desired bipartate graph
#for i in range(n):
# segmentPointIndices = np.asarray(closeLatentPoints[i], dtype=int)
# segmentPoints = latentPoints[segmentPointIndices]
G = bipartite.from_biadjacency_matrix(B)
pass
else:
adjmatrix[i][x-1]=1
counter+=1
counter
import networkx as nx # define the graph
from networkx.algorithms import bipartite
G=nx.Graph()
import numpy, scipy.sparse # define the adj. matrix as scipy matrix for input graph
A = numpy.array(adjmatrix)
Asp = scipy.sparse.csr_matrix(A)
from networkx.algorithms.bipartite import from_biadjacency_matrix
G = from_biadjacency_matrix(Asp, create_using=None, edge_attribute=None)#scipy sparse matrix
X, Y = bipartite.sets(G)
XX = list(X)
YY = list(Y)
import matplotlib.pyplot as plt
import numpy as np
nx.draw(G)
plt.savefig("simple_path.png") # save as png
plt.show() # display
# Define I_tem dictionary
tem = {}
for i in range(1, aMAXX+1):
for j in range(1, bMAXX+1):
def test_from_biadjacency_multigraph(self):
M = sparse.csc_matrix([[1,2],[0,3]])
B = bipartite.from_biadjacency_matrix(M, create_using=nx.MultiGraph())
assert_edges_equal(B.edges(),[(0,2),(0,3),(0,3),(1,3),(1,3),(1,3)])
def test_from_biadjacency_roundtrip(self):
B1 = nx.path_graph(5)
M = bipartite.biadjacency_matrix(B1, [0,2,4])
B2 = bipartite.from_biadjacency_matrix(M)
assert_true(nx.is_isomorphic(B1,B2))