def _evo_op(pwn, evo_arr, conv_layer=pairwise_conv_layer):
"""Performs Template network operation
Args:
pwn (tf.Tensor): pssm of shape (1, l, 21)
evo_arr (tf.Tensor): evfold/ccmpred matrix of shape (1, l, l, 1)
conv_layer (function): conv layer for deep conv
Returns:
tf.Tensor: contact map embedding of shape (1, l, l, 1)
"""
shp = pwn.get_shape().as_list()[-1]
with tf.variable_scope('evo_op', reuse=tf.AUTO_REUSE):
aa_proj_ref = tf.get_variable(
"pwn_projection",
shape=(shp, PROJECTION_DIM),
initializer=tf.contrib.layers.xavier_initializer())
# aa_proj_ref = _print_max_min(aa_proj_ref, 'aa_proj_ref')
s = tf.matmul(pwn, aa_proj_ref) # (1, l, PROJECTION_DIM)
s_pair = outer_concat(s)
c = tf.concat([s_pair, evo_arr], axis=3)
logits = deep_conv_op(c,
input_shape=PROJECTION_DIM * 2 + 1,
num_bins=2,
name_prefix='evo_op_conv',
residual=True,
conv_layer=conv_layer)
logits = logits[..., 1:] # * tf.constant(float(2))
# logits = _print_max_min(logits, "logits evo op")
return logits
def _template_op(seq, dm, conv_layer=pairwise_conv_layer):
"""Performs Template network operation
Args:
seq (tf.Tensor): reference sequence of shape (1, l, 22)
dm (tf.Tensor): reference distance matrix of shape (1, l, l, 1)
conv_layer (function): the conv layer to use in deep conv
Returns:
tf.Tensor: contact map embedding of shape (1, l, l, 1)
"""
with tf.variable_scope('template_op', reuse=tf.AUTO_REUSE):
seq_dim = int(seq.shape[-1])
aa_proj_ref = tf.get_variable(
"ref_projection",
shape=(seq_dim, PROJECTION_DIM),
initializer=tf.contrib.layers.xavier_initializer())
# aa_proj_ref = _print_max_min(aa_proj_ref, "aa_proj_ref template op")
s = tf.matmul(seq, aa_proj_ref) # (1, l, PROJECTION_DIM)
s_pair = outer_concat(s)
c = tf.concat([s_pair, dm], axis=3)
logits = deep_conv_op(c,
input_shape=PROJECTION_DIM * 2 + 1,
num_bins=2,
name_prefix='template_op_conv',
residual=True,
conv_layer=conv_layer)
logits = logits[..., 1:]
# logits = _print_max_min(logits, "logits template op")
return logits
def test_outer_concat(self):
seq1 = np.random.random((1, 4, 7))
seq1_tf = tf.constant(seq1)
outer_concat_seq_tf = outer_concat(seq1_tf)
with tf.Session():
outer_concat_seq = outer_concat_seq_tf.eval()
assert outer_concat_seq[0, 2, 3, 4] == seq1[0, 2, 4]
assert outer_concat_seq[0, 2, 3, 8] == seq1[0, 3, 1]
def _weighting_op_template(seq_template,
seq_target,
conv_layer=pairwise_conv_layer):
"""Computes weights for template
Args:
seq_template (tf.Tensor): template sequence/pwm of shape (1, l, 22/21)
seq_target (tf.Tensor): target sequence/conservation of shape (1, l, 22/21)
conv_layer (function): conv layer to use for deep conv
Returns:
tf.Tensor: weights of shape (1, l, l, 1)
"""
with tf.variable_scope('weight_template_op', reuse=tf.AUTO_REUSE):
seq_template = seq_template[..., 0:21]
aa_dim = tf.shape(seq_template)[-1]
aa_dim_int = int(seq_template.shape[-1])
l = tf.shape(seq_template)[1]
one = tf.shape(seq_template)[0]
name_w = 'template'
O = tf.einsum('ijk,ijl->ijkl', seq_template, seq_target)
O_s = tf.reshape(O, shape=(one, l, aa_dim * aa_dim))
conv0_filt_shape = [10, aa_dim_int**2, PROJECTION_DIM]
# initialise weights and bias for the filter
weights0 = tf.get_variable(
f"{name_w}_conv0_w",
shape=conv0_filt_shape,
initializer=tf.contrib.layers.xavier_initializer())
# weights0 = _print_max_min(weights0, "weights0 template")
# O_s = _print_max_min(O_s, "O_s template")
O_smooth = tf.nn.tanh(
tf.nn.conv1d(input=O_s, filters=weights0, padding='SAME'))
# O_smooth = _print_max_min(O_smooth, "O_smooth template")
conv0_filt_shape = [10, PROJECTION_DIM, 1]
# initialise weights and bias for the filter
weights1 = tf.get_variable(
f"{name_w}_conv1_w",
shape=conv0_filt_shape,
initializer=tf.contrib.layers.xavier_initializer())
# weights1 = _print_max_min(weights1, "weights1 template")
W_1d = tf.nn.conv1d(input=O_smooth, filters=weights1, padding='SAME')
W = outer_concat(W_1d)
input_shape = 2
W_smooth = deep_conv_op(W,
input_shape=input_shape,
name_prefix=name_w,
num_bins=1,
residual=True,
num_layers=2,
conv_layer=conv_layer)
return W_smooth
def _target_property_op(seq_target,
properties_target,
conv_layer=pairwise_conv_layer_2):
"""Computes weights for target property
Args:
seq_target (tf.Tensor): target sequence of shape (1, l, 22/21)
properties_target (tf.Tensor): target properties of shape (1, l, 13)
conv_layer (function): conv layer to use for deep conv
Returns:
tf.Tensor: weights of shape (1, l, l, 1)
"""
with tf.variable_scope('target_property_op', reuse=tf.AUTO_REUSE):
seq_target = seq_target[..., 0:21]
aa_dim_int = int(seq_target.shape[-1]) + int(
properties_target.shape[-1])
name_w = 'target_property'
O_s = tf.concat([seq_target, properties_target], axis=2)
conv0_filt_shape = [10, aa_dim_int, PROJECTION_DIM]
# initialise weights and bias for the filter
weights0 = tf.get_variable(
f"{name_w}_conv0_w",
shape=conv0_filt_shape,
initializer=tf.contrib.layers.xavier_initializer())
# weights0 = _print_max_min(weights0, "weights0 template")
# O_s = print_max_min(O_s, "O_s property")
O_smooth = tf.nn.tanh(
tf.nn.conv1d(input=O_s, filters=weights0, padding='SAME'))
# O_smooth = print_max_min(O_smooth, "O_smooth property")
conv0_filt_shape = [10, PROJECTION_DIM, 1]
# initialise weights and bias for the filter
weights1 = tf.get_variable(
f"{name_w}_conv1_w",
shape=conv0_filt_shape,
initializer=tf.contrib.layers.xavier_initializer())
# weights1 = print_max_min(weights1, "weights1 property")
W_1d = tf.nn.conv1d(input=O_smooth, filters=weights1, padding='SAME')
W = outer_concat(W_1d)
# W = print_max_min(W, "W property")
input_shape = 2
W_smooth = deep_conv_op(W,
input_shape=input_shape,
name_prefix=name_w,
num_bins=1,
residual=True,
num_layers=2,
conv_layer=conv_layer)
# W_smooth = print_max_min(W_smooth, "W_smooth property")
return W_smooth
def _weighting_op_evo(conservation,
evo_arr,
beff,
name,
conv_layer=pairwise_conv_layer):
"""Computes weights for evo
Args:
conservation (tf.Tensor): conservation score of shape (1, l, 1)
evo_arr (tf.Tensor): ccmpred/evfold energy matrix of shape (1, l, l, 1)
beff (tf.Tensor): Effective number of sequences within alignment Beff
name (str): layer name
conv_layer (function): conv layer for deep conv
Returns:
tf.Tensor: weights of shape (1, l, l, 1)
"""
with tf.variable_scope('weight_evo_op', reuse=tf.AUTO_REUSE):
conservation_outer = outer_concat(conservation)
# conservation_outer = _print_max_min(conservation_outer, "conservation_outer ")
shape_w = tf.shape(conservation_outer)
beff_arr = tf.ones(shape=(shape_w[0], shape_w[1], shape_w[2],
shape_w[0])) * tf.log(beff)
# beff_arr = _print_max_min(beff_arr, "beff_arr ")
W = tf.concat([beff_arr, evo_arr], axis=3)
# W = _print_max_min(W, f"W {name} ")
W_smooth = deep_conv_op(W,
input_shape=2,
name_prefix=name,
num_bins=1,
residual=True,
num_layers=2,
conv_layer=conv_layer)
# W_smooth = _print_max_min(W_smooth, f"W_smooth evo {name}")
return W_smooth