def preprocessing_function(nucleotide_sequence: str) -> np.ndarray:
"""One-hot-encode (a single) the sequence.
The kipoi deepbind model does not accept a string of nucleotides.
Args:
nucleotide_sequence (str): defined to be of lenght 101, though other
lenghts might be accepted.
Returns:
np.ndarray: of shape `[1, len(nucleotide_sequence), 4]`
"""
return encodeDNA([nucleotide_sequence])
def onehot_dinucl_shuffle(seqs):
"""Di-nucleotide shuffle the sequences
"""
return encodeDNA(
[dinuc_shuffle(s) for s in one_hot2string(seqs, vocab=DNA)])
def encode_dna(seq):
return encodeDNA(seq)
def data(n_bases=10,
spline_order=3,
pos_class_weight=1.0,
truncate=True,
encode_splines=True,
cache=True,
minmax_scale=False):
dtw_train = pd.read_csv(BR_DATA + "/train/wide_data.csv")
dtw_test = pd.read_csv(BR_DATA + "/test/wide_data.csv")
# replace some columns names
dtw_train.rename(columns={
"dist.1": "dist1",
"dist.2": "dist2"
},
inplace=True)
dtw_test.rename(columns={
"dist.1": "dist1",
"dist.2": "dist2"
},
inplace=True)
y_train = col2num_array(dtw_train["is_branchpoint"], -1)
x_train_pos = {col: col2num_array(dtw_train[col]) for col in pos_columns}
x_train_seq = encodeDNA(dtw_train["seq"])
y_test = col2num_array(dtw_test["is_branchpoint"])
x_test_pos = {col: col2num_array(dtw_test[col]) for col in pos_columns}
x_test_seq = encodeDNA(dtw_test["seq"])
# get secondary structure - use a larger sequence window to precompute it
# to get the larger sequence window first
assert dtw_train["seq_wide"][0][100:-100] == dtw_train["seq"][
0] # correct values
x_train_struct = get_structure(dtw_train["seq_wide"],
BR_DATA + "/train/rna_structure.npy", cache)
x_train_struct = x_train_struct[:, 100:-100]
x_test_struct = get_structure(dtw_test["seq_wide"],
BR_DATA + "/test/rna_structure.npy", cache)
x_test_struct = x_train_struct[:, 100:-100]
# convert all np.nans into -1's for y's and 0's for x's
if truncate:
x_train_pos = {
k: truncate_values(k, v)
for k, v in x_train_pos.items()
}
x_test_pos = {k: truncate_values(k, v) for k, v in x_test_pos.items()}
# Ranges computed only on the trainin set
position_stats = get_minmax_range(x_train_pos, x_train_pos)
minmax_scalers = None
if encode_splines:
x_train_pos_spl = encodeSplines_common(x_train_pos, position_stats,
n_bases, spline_order)
x_test_pos_spl = encodeSplines_common(x_test_pos, position_stats,
n_bases, spline_order)
else:
x_train_pos_spl = {
k: fill_nan(v)[:, :, np.newaxis]
for k, v in x_train_pos.items()
}
x_test_pos_spl = {
k: fill_nan(v)[:, :, np.newaxis]
for k, v in x_test_pos.items()
}
if minmax_scale:
minmax_scalers = {
k: MinMaxScaler().fit(v.reshape(-1, 1))
for k, v in x_train_pos_spl.items()
}
x_train_pos_spl = {
k: minmax_scalers[k].transform(v.reshape(
(-1, 1))).reshape(v.shape)
for k, v in x_train_pos_spl.items()
}
x_test_pos_spl = {
k: minmax_scalers[k].transform(v.reshape(
(-1, 1))).reshape(v.shape)
for k, v in x_test_pos_spl.items()
}
x_train = merge_dicts(x_train_pos_spl, {
"seq": x_train_seq,
"struct": x_train_struct
})
x_test = merge_dicts(x_test_pos_spl, {
"seq": x_test_seq,
"struct": x_test_struct
})
y_train = fill_nan(y_train, -1)[:, :, np.newaxis]
y_test = fill_nan(y_test, -1)[:, :, np.newaxis]
# todo update return
sample_weight = np.squeeze(np.where(y_train == 1, pos_class_weight, 1), -1)
return (x_train, y_train, sample_weight, pos_columns,
get_branchpoint_pwm_list(), position_stats,
minmax_scalers), (x_test, y_test)