File size: 8,242 Bytes
3ac1d94 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 | import argparse
import os
import shutil
import time
import numpy as np
import torch
from torch_geometric.data import Batch
from torch_geometric.transforms import Compose
from torch_scatter import scatter_sum, scatter_mean
from tqdm.auto import tqdm
import onescience.utils.targetdiff.misc as misc
import onescience.utils.targetdiff.transforms as trans
from onescience.datapipes.targetdiff import get_dataset
from onescience.datapipes.targetdiff.pl_data import FOLLOW_BATCH
from models.molopt_score_model import ScorePosNet3D, log_sample_categorical
from onescience.utils.targetdiff.evaluation import atom_num
def unbatch_v_traj(ligand_v_traj, n_data, ligand_cum_atoms):
all_step_v = [[] for _ in range(n_data)]
for v in ligand_v_traj: # step_i
v_array = v.cpu().numpy()
for k in range(n_data):
all_step_v[k].append(v_array[ligand_cum_atoms[k]:ligand_cum_atoms[k + 1]])
all_step_v = [np.stack(step_v) for step_v in all_step_v] # num_samples * [num_steps, num_atoms_i]
return all_step_v
def sample_diffusion_ligand(model, data, num_samples, batch_size=16, device='cuda:0',
num_steps=None, pos_only=False, center_pos_mode='protein',
sample_num_atoms='prior'):
all_pred_pos, all_pred_v = [], []
all_pred_pos_traj, all_pred_v_traj = [], []
all_pred_v0_traj, all_pred_vt_traj = [], []
time_list = []
num_batch = int(np.ceil(num_samples / batch_size))
current_i = 0
for i in tqdm(range(num_batch)):
n_data = batch_size if i < num_batch - 1 else num_samples - batch_size * (num_batch - 1)
batch = Batch.from_data_list([data.clone() for _ in range(n_data)], follow_batch=FOLLOW_BATCH).to(device)
t1 = time.time()
with torch.no_grad():
batch_protein = batch.protein_element_batch
if sample_num_atoms == 'prior':
pocket_size = atom_num.get_space_size(data.protein_pos.detach().cpu().numpy())
ligand_num_atoms = [atom_num.sample_atom_num(pocket_size).astype(int) for _ in range(n_data)]
batch_ligand = torch.repeat_interleave(torch.arange(n_data), torch.tensor(ligand_num_atoms)).to(device)
elif sample_num_atoms == 'range':
ligand_num_atoms = list(range(current_i + 1, current_i + n_data + 1))
batch_ligand = torch.repeat_interleave(torch.arange(n_data), torch.tensor(ligand_num_atoms)).to(device)
elif sample_num_atoms == 'ref':
batch_ligand = batch.ligand_element_batch
ligand_num_atoms = scatter_sum(torch.ones_like(batch_ligand), batch_ligand, dim=0).tolist()
else:
raise ValueError
# init ligand pos
center_pos = scatter_mean(batch.protein_pos, batch_protein, dim=0)
batch_center_pos = center_pos[batch_ligand]
init_ligand_pos = batch_center_pos + torch.randn_like(batch_center_pos)
# init ligand v
if pos_only:
init_ligand_v = batch.ligand_atom_feature_full
else:
uniform_logits = torch.zeros(len(batch_ligand), model.num_classes).to(device)
init_ligand_v = log_sample_categorical(uniform_logits)
r = model.sample_diffusion(
protein_pos=batch.protein_pos,
protein_v=batch.protein_atom_feature.float(),
batch_protein=batch_protein,
init_ligand_pos=init_ligand_pos,
init_ligand_v=init_ligand_v,
batch_ligand=batch_ligand,
num_steps=num_steps,
pos_only=pos_only,
center_pos_mode=center_pos_mode
)
ligand_pos, ligand_v, ligand_pos_traj, ligand_v_traj = r['pos'], r['v'], r['pos_traj'], r['v_traj']
ligand_v0_traj, ligand_vt_traj = r['v0_traj'], r['vt_traj']
# unbatch pos
ligand_cum_atoms = np.cumsum([0] + ligand_num_atoms)
ligand_pos_array = ligand_pos.cpu().numpy().astype(np.float64)
all_pred_pos += [ligand_pos_array[ligand_cum_atoms[k]:ligand_cum_atoms[k + 1]] for k in
range(n_data)] # num_samples * [num_atoms_i, 3]
all_step_pos = [[] for _ in range(n_data)]
for p in ligand_pos_traj: # step_i
p_array = p.cpu().numpy().astype(np.float64)
for k in range(n_data):
all_step_pos[k].append(p_array[ligand_cum_atoms[k]:ligand_cum_atoms[k + 1]])
all_step_pos = [np.stack(step_pos) for step_pos in
all_step_pos] # num_samples * [num_steps, num_atoms_i, 3]
all_pred_pos_traj += [p for p in all_step_pos]
# unbatch v
ligand_v_array = ligand_v.cpu().numpy()
all_pred_v += [ligand_v_array[ligand_cum_atoms[k]:ligand_cum_atoms[k + 1]] for k in range(n_data)]
all_step_v = unbatch_v_traj(ligand_v_traj, n_data, ligand_cum_atoms)
all_pred_v_traj += [v for v in all_step_v]
if not pos_only:
all_step_v0 = unbatch_v_traj(ligand_v0_traj, n_data, ligand_cum_atoms)
all_pred_v0_traj += [v for v in all_step_v0]
all_step_vt = unbatch_v_traj(ligand_vt_traj, n_data, ligand_cum_atoms)
all_pred_vt_traj += [v for v in all_step_vt]
t2 = time.time()
time_list.append(t2 - t1)
current_i += n_data
return all_pred_pos, all_pred_v, all_pred_pos_traj, all_pred_v_traj, all_pred_v0_traj, all_pred_vt_traj, time_list
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('config', type=str)
parser.add_argument('-i', '--data_id', type=int)
parser.add_argument('--device', type=str, default='cuda:0')
parser.add_argument('--batch_size', type=int, default=100)
parser.add_argument('--result_path', type=str, default='./outputs')
args = parser.parse_args()
logger = misc.get_logger('sampling')
# Load config
config = misc.load_config(args.config)
logger.info(config)
misc.seed_all(config.sample.seed)
# Load checkpoint
ckpt = torch.load(config.model.checkpoint, map_location=args.device)
logger.info(f"Training Config: {ckpt['config']}")
# Transforms
protein_featurizer = trans.FeaturizeProteinAtom()
ligand_atom_mode = ckpt['config'].data.transform.ligand_atom_mode
ligand_featurizer = trans.FeaturizeLigandAtom(ligand_atom_mode)
transform = Compose([
protein_featurizer,
ligand_featurizer,
trans.FeaturizeLigandBond(),
])
# Load dataset
dataset, subsets = get_dataset(
config=ckpt['config'].data,
transform=transform
)
train_set, test_set = subsets['train'], subsets['test']
logger.info(f'Successfully load the dataset (size: {len(test_set)})!')
# Load model
model = ScorePosNet3D(
ckpt['config'].model,
protein_atom_feature_dim=protein_featurizer.feature_dim,
ligand_atom_feature_dim=ligand_featurizer.feature_dim
).to(args.device)
model.load_state_dict(ckpt['model'])
logger.info(f'Successfully load the model! {config.model.checkpoint}')
data = test_set[args.data_id]
pred_pos, pred_v, pred_pos_traj, pred_v_traj, pred_v0_traj, pred_vt_traj, time_list = sample_diffusion_ligand(
model, data, config.sample.num_samples,
batch_size=args.batch_size, device=args.device,
num_steps=config.sample.num_steps,
pos_only=config.sample.pos_only,
center_pos_mode=config.sample.center_pos_mode,
sample_num_atoms=config.sample.sample_num_atoms
)
result = {
'data': data,
'pred_ligand_pos': pred_pos,
'pred_ligand_v': pred_v,
'pred_ligand_pos_traj': pred_pos_traj,
'pred_ligand_v_traj': pred_v_traj,
'time': time_list
}
logger.info('Sample done!')
result_path = args.result_path
os.makedirs(result_path, exist_ok=True)
shutil.copyfile(args.config, os.path.join(result_path, 'sample.yml'))
torch.save(result, os.path.join(result_path, f'result_{args.data_id}.pt'))
|