-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathjunction_predictions_seabass.py
127 lines (109 loc) · 6.57 KB
/
junction_predictions_seabass.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
import os
import utils
import numpy as np
import pandas as pd
import tensorflow as tf
from data import load_data, label_and_filter_data, model_inputs, FEATURE_GROUPS, SCALAR_FEATS
from models import build_model, train_model, test_model
from normalization import get_normalization_object
# script arguments
parser = utils.common_parser_arguments()
parser.add_argument('--gencode_batch_size', type=int, default=10000, help='gencode prediction batch size')
parser.add_argument('--sat_quant_active', type=float, default=0.1, help='set sigmoid(LFC) = q for quantile(LFC, q)')
parser.add_argument('--sat_quant_inactive', type=float, default=0.9, help='set sigmoid(LFC) = q for quantile(LFC, q)')
parser.add_argument('--use_lfc', action='store_true', default=False, help='Use Day 21 LFC instead of sea-bass LMM slope')
args = utils.parse_common_arguments(parser)
assert args.dataset == 'junction'
assert args.pm_only
# save path
output_dir = os.path.join('predictions', 'junction', 'tiger-junc' if args.use_lfc else 'tiger-bass')
output_dir = os.path.join(output_dir, utils.data_directory(args.pm_only, args.indels, args.seq_only))
os.makedirs(output_dir, exist_ok=True)
# random seed
if args.seed is not None:
tf.config.experimental.enable_op_determinism()
tf.keras.utils.set_random_seed(args.seed)
# load training data and merge sea-bass slope estimates
data_junc = load_data(args.dataset, pm_only=args.pm_only, indels=args.indels, holdout=args.holdout)
data_junc = label_and_filter_data(*data_junc, args.nt_quantile, args.filter_method, args.min_active_ratio)
sea_bass = pd.read_pickle(os.path.join('data-processed', 'sea-bass.bz2'))
sea_bass_cols = list(set(sea_bass.columns) - set(data_junc.columns))
data_junc = data_junc.merge(sea_bass[['gene', 'guide_seq'] + sea_bass_cols], how='inner', on=['gene', 'guide_seq'])
if not args.use_lfc:
data_junc['observed_lfc'] = data_junc['slope']
normalizer = get_normalization_object(args.normalization)(df=data_junc, **args.normalization_kwargs)
data_junc = normalizer.normalize_targets(data_junc)
data_junc = data_junc.loc[~np.isnan(data_junc['observed_lfc']) & ~np.isinf(data_junc['observed_lfc'])]
# load qpcr data
data_qpcr = pd.read_pickle(os.path.join('data-processed', 'junction-qpcr.bz2'))
# load gencode data
data_gencode, _ = load_data(dataset='junction-all', pm_only=args.pm_only, indels=args.indels)
# determine available scalar feature set
junction_features = set(list(data_junc.columns) + list(data_qpcr.columns) + list(data_gencode.columns))
if args.seq_only:
available_features = set()
else:
available_features = set(SCALAR_FEATS).intersection(junction_features) - set(FEATURE_GROUPS['junction overlap'])
available_features = list(available_features)
available_features.sort()
print('Non-scalar features to be used:', available_features)
excluded_features = list((set(SCALAR_FEATS) - set(available_features)).intersection(set(data_junc.columns)))
excluded_features.sort()
print('Junction features excluded:', excluded_features)
# pre-validation
for i, fold in enumerate(set(data_junc['fold'].unique()) - {'training'}):
# assemble model inputs
train_data = model_inputs(data_junc[data_junc.fold != fold], args.context, scalar_feats=available_features)
valid_data = model_inputs(data_junc[data_junc.fold == fold], args.context, scalar_feats=available_features)
test_qpcr = data_qpcr.loc[~data_qpcr['guide_seq'].isin(data_junc.loc[data_junc.fold != fold, 'guide_seq'])]
test_qpcr = model_inputs(test_qpcr, args.context, scalar_feats=available_features)
# train model
model = build_model(name=args.model,
target_len=train_data['target_tokens'].shape[1],
context_5p=train_data['5p_tokens'].shape[1],
context_3p=train_data['3p_tokens'].shape[1],
use_guide_seq=args.use_guide_seq,
loss_fn=args.loss,
debug=args.debug,
output_fn=normalizer.output_fn,
**args.kwargs)
model = train_model(model, train_data, valid_data, args.batch_size)
# accumulate predictions
kwargs = dict(header=(i == 0), index=False, mode='w' if (i == 0) else 'a')
test_model(model, valid_data).to_csv(os.path.join(output_dir, 'predictions_junc_cv.csv'), **kwargs)
test_model(model, test_qpcr).to_csv(os.path.join(output_dir, 'predictions_qpcr_ensemble.csv'), **kwargs)
for j in range(0, len(data_gencode), args.gencode_batch_size):
j_stop = min(j + args.gencode_batch_size, len(data_gencode))
df = data_gencode.iloc[j:j_stop]
df = test_model(model, model_inputs(df, args.context, scalar_feats=available_features))
df.to_csv(os.path.join(output_dir, 'predictions_gencode_ensemble.csv'), **kwargs)
print('\rGencode percent complete: {:.2f}%'.format(100 * j_stop / len(data_gencode)), end='')
# free keras memory
tf.keras.backend.clear_session()
# load predictions
pred_junc = pd.read_csv(os.path.join(output_dir, 'predictions_junc_cv.csv'))
pred_qpcr = pd.read_csv(os.path.join(output_dir, 'predictions_qpcr_ensemble.csv'))
pred_gc = pd.read_csv(os.path.join(output_dir, 'predictions_gencode_ensemble.csv'))
# average ensembles
pred_qpcr = pred_qpcr.groupby(list(set(pred_qpcr.columns) - {'observed_lfc', 'predicted_lfc'})).mean().reset_index()
pred_gc = pred_gc.groupby(list(set(pred_gc.columns) - {'observed_lfc', 'predicted_lfc'})).mean().reset_index()
# fit sigmoid transformation to gencode predictions
lfc = pred_gc.loc[pred_gc['guide_type'] == 'PM', 'predicted_lfc']
x = np.array([[lfc.quantile(args.sat_quant_active), 1], [lfc.quantile(args.sat_quant_inactive), 1]])
y = np.log(np.array([[args.sat_quant_active], [args.sat_quant_inactive]]) ** -1 - 1)
a, b = np.squeeze(np.linalg.inv(x.T @ x) @ x.T @ y)
pred_gc['predicted_score'] = 1 - 1 / (1 + np.exp(a * pred_gc['predicted_lfc'] + b))
# score predictions and measure performance
for name, pred in [('junc', pred_junc), ('qpcr', pred_qpcr)]:
print('\n*** ' + name.upper() + ' ***')
print('LFC: ', end='')
utils.measure_performance(pred, pred_var='predicted_lfc')
pred['predicted_score'] = 1 - 1 / (1 + np.exp(a * pred['predicted_lfc'] + b))
print('Score: ', end='')
utils.measure_performance(pred, pred_var='predicted_score')
# denormalize junction observations and targets
pred_junc = normalizer.denormalize_targets_and_predictions(pred_junc)
# save results
pred_junc.to_csv(os.path.join(output_dir, 'predictions_junc.csv'), index=False)
pred_qpcr.to_csv(os.path.join(output_dir, 'predictions_qpcr.csv'), index=False)
pred_gc.to_csv(os.path.join(output_dir, 'predictions_gencode.csv'), index=False)