-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathtrain_gtex.py
165 lines (136 loc) · 6.57 KB
/
train_gtex.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
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
"""
Trains the model on GTEx data
"""
import torch
import numpy as np
import pandas as pd
import wandb
import argparse
from torch.utils.data import Dataset, DataLoader
from src.hnn import HypergraphNeuralNet
from src.data import Data
from src.dataset import HypergraphDataset
from src.data_utils import *
from src.eval_utils import *
from src.train_utils import train
import scanpy as sc
np.random.seed(0)
num_workers = 4
GTEX_FILE = 'data/GTEX_data.csv'
METADATA_FILE = 'data/GTEx_Analysis_v8_Annotations_SubjectPhenotypesDS.txt'
def GTEx(file=GTEX_FILE):
"""
Loads processed GTEx data
:param file: path of the CSV file
:return: Returns
- data: numpy array of shape=(nb_samples, nb_genes)
- gene_symbols: numpy array with gene symbols. Shape=(nb_genes,)
- sampl_ids: numpy array with sample IDs (GTEx IDs of individuals, e.g. GTEX-1117F). Shape=(nb_samples,)
- tissues: numpy array indicating the tissue of each sample. Shape=(nb_samples,)
"""
# Load data
df = pd.read_csv(file, index_col=0) # .sample(frac=1, random_state=random_seed)
tissues = df['tissue'].values
sampl_ids = df.index.values
del df['tissue']
data = np.float32(df.values)
gene_symbols = df.columns.values
return data, gene_symbols, sampl_ids, tissues
def GTEx_metadata(file=METADATA_FILE):
"""
Loads metadata DataFrame with information about individuals
:param file: path of the file
:return: Pandas DataFrame with subjects' information
"""
df = pd.read_csv(file, delimiter='\t')
df = df.set_index('SUBJID')
return df
def GTEx_v8_normalised_adata(file=GTEX_FILE):
data, gene_symbols, sampl_ids, tissues = GTEx(file=file)
metadata_df = GTEx_metadata()
adata = sc.AnnData(data)
adata.var['Symbol'] = gene_symbols
adata.obs['Participant ID'] = sampl_ids
adata.obs['Tissue'] = tissues
# Delete participants with only one measured tissue
adata = adata[adata.obs['Participant ID'].duplicated(keep=False)]
# Static keys
adata.obs['Tissue_idx'], tissue_dict = map_to_ids(adata.obs['Tissue'].values)
adata.uns['Tissue_dict'] = tissue_dict
# del adata.obs['Tissue']
# Dynamic keys
adata.obs['Participant ID_dyn'] = adata.obs['Participant ID']
# Populate participant features
adata.obs['Age'] = [float(a[:2]) for a in metadata_df.loc[adata.obs['Participant ID']]['AGE'].values]
adata.obs['Sex'] = metadata_df.loc[adata.obs['Participant ID']]['SEX'].values-1
donor_age = adata.obs['Age'] / 100
donor_sex, donor_sex_dict = map_to_ids(adata.obs['Sex'])
adata.obsm['Participant ID_feat'] = np.stack((donor_age, donor_sex), axis=-1)
adata.uns['Sex_dict'] = donor_sex_dict
# Put gene expression in layer
adata.layers['x'] = adata.X
# Set up tissue colors
colors = ['#ffaa56', '#cdad22', '#8fbc8f', '#8b1c62', '#ee6a50', '#ff0000', '#eeee00', '#eeee00', '#eeee00',
'#eeee00', '#eeee00', '#eeee00', '#eeee00', '#eeee00', '#eeee00', '#eeee00', '#eeee00', '#eeee00',
'#eeee00', '#00cdcd', '#9ac0cd', '#ee82ee', '#cdb79e', '#eec591', '#8b7355', '#8b7355', '#cdaa7d',
'#b452cd', '#7a378b', '#cdb79e', '#cdb79e', '#9acd32', '#cdb79e', '#7A67EE', '#FFD700', '#FFB6C1',
'#CD9B1D', '#B4EEB4', '#D9D9D9', '#3A5FCD', '#1E90FF', '#CDB79E', '#CDB79E', '#FFD39B', '#A6A6A6',
'#008B45', '#EED5D2', '#EED5D2', '#FF00FF']
adata.uns['Tissue_colors'] = colors
return adata
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--config', dest='config', default='configs/default.yaml', type=str)
args, unknown = parser.parse_known_args()
# Initialise wandb
wandb.init(project='multitissue_imputation', entity="multitissue_imputation_project", config=args.config)
config = wandb.config
print(config)
# Load data
adata = GTEx_v8_normalised_adata()
# Dictionaries
_, tissue_dict = map_to_ids(adata.obs['Tissue'])
tissue_dict_inv = {v: k for k, v in tissue_dict.items()}
# Split train/val/test
donors = adata.obs['Participant ID'].values
train_donors = np.loadtxt('data/splits/gtex_train.txt', delimiter=',', dtype=str)
val_donors = np.loadtxt('data/splits/gtex_val.txt', delimiter=',', dtype=str)
test_donors = np.loadtxt('data/splits/gtex_test.txt', delimiter=',', dtype=str)
train_mask = np.isin(donors, train_donors)
test_mask = np.isin(donors, test_donors)
val_mask = np.isin(donors, val_donors)
print(len(train_donors), len(val_donors), len(test_donors))
collate_fn = Data.from_datalist
dtype = torch.float32 # torch.double
target_tissues = ['Lung', 'Pancreas', 'Heart_Atrial', 'Esophagus_Muscularis']
source_tissues = [t for t in adata.obs['Tissue'].unique() if t not in target_tissues] # All tissues except targets
train_dataset = HypergraphDataset(adata[train_mask], dtype=dtype, disjoint=True, static=False)
val_dataset = HypergraphDataset(adata[val_mask], dtype=dtype, disjoint=False, static=True, obs_source={'Tissue': source_tissues}, obs_target={'Tissue': target_tissues})
# test_dataset = HypergraphDataset(adata[test_mask], dtype=dtype, static=True)
train_loader = DataLoader(train_dataset, batch_size=config.batch_size, collate_fn=collate_fn, shuffle=True, num_workers=num_workers)
val_loader = DataLoader(val_dataset, batch_size=config.batch_size, collate_fn=collate_fn, shuffle=False, num_workers=num_workers)
# test_loader = DataLoader(test_dataset, batch_size=config.batch_size, collate_fn=collate_fn, shuffle=False, num_workers=num_workers)
# device = torch.device("cpu")
# Use certain GPU
device = torch.device("cuda:{}".format(config.gpu) if torch.cuda.is_available() else "cpu")
# Select dynamic/static node types
config.static_node_types = {'Tissue': (len(adata.obs['Tissue_idx'].unique()), config.d_tissue),
'metagenes': (config.meta_G, config.d_gene)}
config.dynamic_node_types = {'Participant ID': (len(adata.obs['Participant ID'].unique()), config.d_patient)}
# Model
config.G = adata.shape[-1]
model = HypergraphNeuralNet(config).to(device) # .double()
# Train
def rho(x, out):
x_pred = out['px_rate'].detach().cpu().numpy()
return np.mean(pearson_correlation_score(x, x_pred, sample_corr=True))
metric_fns = [rho]
train(config,
model=model,
loader=train_loader,
val_loader=val_loader,
device=device,
preprocess_fn=None,
compute_metrics_train=False,
metric_fns=metric_fns)
torch.save(model.state_dict(), 'data/model.pth')