def cartesian_gaussian():
"""Generate a 3D Cartesian Gaussian with random mean and covariance."""
rotation = stats.special_ortho_group(3).rvs()
eigenvalues = stats.gamma(2).rvs(3)
cov = rotation @ np.diag(eigenvalues) @ rotation.T
mean = stats.norm.rvs(size=3)
return stats.multivariate_normal(mean, cov)
def test_frozen_matrix(self):
dim = 7
frozen = special_ortho_group(dim)
rvs1 = frozen.rvs(random_state=1234)
rvs2 = special_ortho_group.rvs(dim, random_state=1234)
assert_equal(rvs1, rvs2)
def test_frozen_matrix(self):
dim = 7
frozen = special_ortho_group(dim)
rvs1 = frozen.rvs(random_state=1234)
rvs2 = special_ortho_group.rvs(dim, random_state=1234)
assert_equal(rvs1, rvs2)
def build_moment_index(self, name, mesh, grid_size):
print(f'Building HuIndex for {name}')
n_samples = 1000
for i in range(n_samples):
print(f'{i}/{n_samples}')
R = special_ortho_group(3).rvs()
mask, offset, vt = self.mesh_to_mask(mesh, R, grid_size)
hu = cv2.HuMoments(cv2.moments(np.float32(mask)))
lower, upper = hu - abs(hu * 0.3), hu + abs(hu * 0.3)
self.Index.insert(i, np.concatenate((lower, upper), axis = 0), obj = (R, offset, vt))
def pair_model(model, FLAGS, node_embed):
# Generate a dataset where two atoms are very close to each other and everything else is very far
# Indices for atoms
atom_names = ["X", "C", "N", "O", "S"]
residue_names = [
"ALA",
"ARG",
"ASN",
"ASP",
"CYS",
"GLU",
"GLN",
"GLY",
"HIS",
"ILE",
"LEU",
"LYS",
"MET",
"PHE",
"PRO",
"SER",
"THR",
"TRP",
"TYR",
"VAL",
]
energies_output_dict = {}
def make_key(n_rotations, residue_name1, residue_name2, atom_name1,
atom_name2):
return f"{n_rotations}_{residue_name1}_{residue_name2}_{atom_name1}_{atom_name2}"
# Save a copy of the node embed
node_embed = node_embed[0]
node_embed_orig = node_embed.clone()
# Try different combinations
for n_rotations in [5]:
# Rotations
so3 = special_ortho_group(3)
rot_matrix_neg = so3.rvs(
n_rotations) # number of random rotations to average
residue_names_proc = ["ALA", "TYR", "LEU"]
atom_names_proc = ["C", "N", "O"]
for residue_name1, residue_name2 in itertools.product(
residue_names_proc, repeat=2):
for atom_name1, atom_name2 in itertools.product(atom_names_proc,
repeat=2):
eps = []
energies = []
residue_index1 = residue_names.index(residue_name1)
residue_index2 = residue_names.index(residue_name2)
atom_index1 = atom_names.index(atom_name1)
atom_index2 = atom_names.index(atom_name2)
for i in np.linspace(0.1, 1.0, 100):
node_embed = node_embed_orig.clone()
node_embed[-2, -3:] = torch.Tensor([1.0, 0.5, 0.5])
node_embed[-1, -3:] = torch.Tensor([1.0 + i, 0.5, 0.5])
node_embed[-1, 0] = residue_index1
node_embed[-2, 0] = residue_index2
node_embed[-1, 1] = atom_index1
node_embed[-2, 1] = atom_index2
node_embed[-1, 2] = 6 # res_counter
node_embed[-2, 2] = 6 # res_counter
node_embed = np.tile(node_embed[None, :, :],
(n_rotations, 1, 1))
node_embed[:, :, -3:] = np.matmul(node_embed[:, :, -3:],
rot_matrix_neg)
node_embed_feed = torch.Tensor(node_embed).cuda()
node_embed_feed[:, :,
-3:] = node_embed_feed[:, :,
-3:] - node_embed_feed[:, :, -3:].mean(
dim=1,
keepdim=True)
energy = model.forward(node_embed_feed) #
energy = energy.mean()
eps.append(i * 10)
energies.append(energy.item())
key = make_key(n_rotations, residue_name1, residue_name2,
atom_name1, atom_name2)
energies_output_dict[key] = (eps, energies)
# Optionally make plots here -- potentially add conditions to avoid making too many
plt.plot(eps, energies)
plt.xlabel("Atom Distance")
plt.ylabel("Energy")
plt.title(
f"{n_rotations} rots: {atom_name1}, {atom_name2} distance in {residue_name1}/{residue_name2}"
)
plt.savefig(
f"distance_plots/{n_rotations}_{atom_name1}_{atom_name2}_in_{residue_name1}_{residue_name2}_distance.png"
)
plt.clf()
# Back to outside
output_path = osp.join(FLAGS.outdir, "atom_distances.p")
pickle.dump(energies_output_dict, open(output_path, "wb"))
def new_model(model, FLAGS, node_embed):
BATCH_SIZE = 120
pdb_name = FLAGS.pdb_name #'6mdw.0'
pickle_file = f"/private/home/yilundu/dataset/mmcif/mmCIF/{pdb_name[1:3]}/{pdb_name}.p"
(node_embed, ) = pickle.load(open(pickle_file, "rb"))
par, child, pos, pos_exist, res, chis_valid = parse_dense_format(
node_embed)
angles = compute_dihedral(par, child, pos, pos_exist)
chis_target_initial = angles[:, 4:8].copy(
) # dihedral for backbone (:4); dihedral for sidechain (4:8)
NUM_RES = len(res)
all_energies = np.empty(
(NUM_RES, 4, 360)) # 4 is number of possible chi angles
surface_core_type = []
for idx in range(NUM_RES):
dist = np.sqrt(np.square(pos[idx:idx + 1, 2] - pos[:, 2]).sum(axis=1))
neighbors = (dist < 10).sum()
if neighbors >= 24:
surface_core_type.append("core")
elif neighbors <= 16:
surface_core_type.append("surface")
else:
surface_core_type.append("unlabeled")
for idx in tqdm(range(NUM_RES)):
for chi_num in range(4):
if not chis_valid[idx, chi_num]:
continue
# init_angle = chis_target[idx, chi_num]
for angle_deltas in batch(range(-180, 180, 3), BATCH_SIZE):
pre_rot_node_embed_short = []
for angle_delta in angle_deltas:
chis_target = chis_target_initial.copy(
) # make a local copy
# modify the angle by angle_delta amount. rotate to chis_target
chis_target[
idx,
chi_num] += angle_delta # Set the specific chi angle to be the sampled value
# pos_new is n residues x 20 atoms x 3 (xyz)
pos_new = rotate_dihedral_fast(angles, par, child, pos,
pos_exist, chis_target,
chis_valid, idx)
node_neg_embed = reencode_dense_format(
node_embed, pos_new, pos_exist)
# sort the atoms by how far away they are
# sort key is the first atom on the sidechain
pos_chosen = pos_new[idx, 4]
close_idx = np.argsort(
np.square(node_neg_embed[:, -3:] -
pos_chosen).sum(axis=1))
# Grab the 64 closest atoms
node_embed_short = node_neg_embed[
close_idx[:FLAGS.max_size]].copy()
# Normalize each coordinate of node_embed to have x, y, z coordinate to be equal 0
node_embed_short[:,
-3:] = node_embed_short[:, -3:] - np.mean(
node_embed_short[:, -3:], axis=0)
node_embed_short = torch.from_numpy(
node_embed_short).float().cuda()
pre_rot_node_embed_short.append(
node_embed_short.unsqueeze(0))
pre_rot_node_embed_short = torch.stack(
pre_rot_node_embed_short)
# Now rotate all elements
n_rotations = 100
so3 = special_ortho_group(3)
rot_matrix = so3.rvs(n_rotations) # n x 3 x 3
node_embed_short = pre_rot_node_embed_short.repeat(
1, n_rotations, 1, 1)
rot_matrix = torch.from_numpy(rot_matrix).float().cuda()
node_embed_short[:, :, :, -3:] = torch.matmul(
node_embed_short[:, :, :, -3:],
rot_matrix) # (batch_size, n_rotations, 64, 20)
# Compute the energies for the n_rotations * batch_size for this window of 64 atoms.
# Batch the first two dimensions, then pull them apart aftewrads.
node_embed_short = node_embed_short.reshape(
node_embed_short.shape[0] * node_embed_short.shape[1],
*node_embed_short.shape[2:],
)
energies = model.forward(node_embed_short) # (12000, 1)
# divide the batch dimension by the 10 things we just did
energies = energies.reshape(BATCH_SIZE, -1) # (10, 200)
# Average the energy across the n_rotations, but keeping batch-wise seperate
energies = energies.mean(1) # (10, 1)
# Save the result
all_energies[idx, chi_num,
angle_deltas] = energies.cpu().numpy()
# Can use these for processing later.
avg_chi_angle_energy = (
all_energies * chis_valid[:NUM_RES, :4, None]).sum(0) / np.expand_dims(
chis_valid[:NUM_RES, :4].sum(0),
1) # normalize by how many times each chi angle occurs
output = {
"all_energies": all_energies,
"chis_valid": chis_valid,
"chis_target_initial": chis_target_initial,
"avg_chi_angle_energy":
avg_chi_angle_energy, # make four plots from this (4, 360),
"res": res,
"surface_core_type": surface_core_type,
}
# Dump the output
output_path = osp.join(FLAGS.outdir, f"{pdb_name}_rot_energies.p")
if not osp.exists(FLAGS.outdir):
os.makedirs(FLAGS.outdir)
pickle.dump(output, open(output_path, "wb"))
def make_tsne(model, FLAGS, node_embed):
"""
grab representations for each of the residues in a pdb
"""
pdb_name = FLAGS.pdb_name
pickle_file = MMCIF_PATH + f"/mmCIF/{pdb_name[1:3]}/{pdb_name}.p"
(node_embed, ) = pickle.load(open(pickle_file, "rb"))
par, child, pos, pos_exist, res, chis_valid = parse_dense_format(
node_embed)
angles = compute_dihedral(par, child, pos, pos_exist)
all_hiddens = []
all_energies = []
n_rotations = 2
so3 = special_ortho_group(3)
rot_matrix = so3.rvs(n_rotations)
rot_matrix = torch.from_numpy(rot_matrix).float().cuda()
for idx in range(len(res)):
# sort the atoms by how far away they are
# sort key is the first atom on the sidechain
pos_chosen = pos[idx, 4]
close_idx = np.argsort(
np.square(node_embed[:, -3:] - pos_chosen).sum(axis=1))
# Grab the 64 closest atoms
node_embed_short = node_embed[close_idx[:FLAGS.max_size]].copy()
# Normalize each coordinate of node_embed to have x, y, z coordinate to be equal 0
node_embed_short[:, -3:] = node_embed_short[:, -3:] - np.mean(
node_embed_short[:, -3:], axis=0)
node_embed_short = torch.from_numpy(node_embed_short).float().cuda()
node_embed_short = node_embed_short[None, :, :].repeat(
n_rotations, 1, 1)
node_embed_short[:, :, -3:] = torch.matmul(node_embed_short[:, :, -3:],
rot_matrix)
# Compute the energies for the n_rotations * batch_size for this window of 64 atoms.
# Batch the first two dimensions, then pull them apart aftewrads.
# node_embed_short = node_embed_short.reshape(node_embed_short.shape[0] * node_embed_short.shape[1], *node_embed_short.shape[2:])
energies, hidden = model.forward(node_embed_short,
return_hidden=True) # (12000, 1)
# all_hiddens.append(hidden.mean(0)) # mean over the rotations
all_hiddens.append(hidden[0]) # take first rotation
all_energies.append(energies[0])
surface_core_type = []
for idx in range(len(res)):
# >16 c-beta neighbors within 10A of its own c-beta (or c-alpha for gly).
hacked_pos = np.copy(pos)
swap_hacked_pos = np.swapaxes(hacked_pos, 0, 1) # (20, 59, 3)
idxs_to_change = swap_hacked_pos[4] == [0, 0, 0] # (59, 3)
swap_hacked_pos[4][idxs_to_change] = swap_hacked_pos[3][idxs_to_change]
hacked_pos_final = np.swapaxes(swap_hacked_pos, 0, 1)
dist = np.sqrt(
np.square(hacked_pos_final[idx:idx + 1, 4] -
hacked_pos_final[:, 4]).sum(axis=1))
neighbors = (dist < 10).sum()
if neighbors >= 24:
surface_core_type.append("core")
elif neighbors <= 16:
surface_core_type.append("surface")
else:
surface_core_type.append("unlabeled")
output = {
"res": res,
"surface_core_type": surface_core_type,
"all_hiddens": torch.stack(all_hiddens).cpu().numpy(),
"all_energies": torch.stack(all_energies).cpu().numpy(),
}
# Dump the output
output_path = osp.join(FLAGS.outdir, f"{pdb_name}_representations.p")
if not osp.exists(FLAGS.outdir):
os.makedirs(FLAGS.outdir)
pickle.dump(output, open(output_path, "wb"))
def rotamer_trials(model, FLAGS, test_dataset):
test_files = test_dataset.files
random.shuffle(test_files)
db = load_rotamor_library()
so3 = special_ortho_group(3)
node_embed_evals = []
nminibatch = 4
if FLAGS.ensemble > 1:
models = model
# The three different sampling methods are weighted_gauss, gmm, rosetta
rotamer_scores_total = []
surface_scores_total = []
buried_scores_total = []
amino_recovery_total = {}
for k, v in kvs.items():
amino_recovery_total[k.lower()] = []
counter = 0
rotations = FLAGS.rotations
for test_file in tqdm(test_files):
(node_embed, ) = pickle.load(open(test_file, "rb"))
node_embed_original = node_embed
par, child, pos, pos_exist, res, chis_valid = parse_dense_format(
node_embed)
angles = compute_dihedral(par, child, pos, pos_exist)
amino_recovery = {}
for k, v in kvs.items():
amino_recovery[k.lower()] = []
if node_embed is None:
continue
rotamer_scores = []
surface_scores = []
buried_scores = []
types = []
gt_chis = []
node_embed_evals = []
neg_chis = []
valid_chi_idxs = []
res_names = []
neg_sample = FLAGS.neg_sample
n_amino = pos.shape[0]
amino_recovery_curr = {}
for idx in range(1, n_amino - 1):
res_name = res[idx]
if res_name == "gly" or res_name == "ala":
continue
res_names.append(res_name)
gt_chis.append(angles[idx, 4:8])
valid_chi_idxs.append(chis_valid[idx, :4])
hacked_pos = np.copy(pos)
swap_hacked_pos = np.swapaxes(hacked_pos, 0, 1) # (20, 59, 3)
idxs_to_change = swap_hacked_pos[4] == [0, 0, 0] # (59, 3)
swap_hacked_pos[4][idxs_to_change] = swap_hacked_pos[3][
idxs_to_change]
hacked_pos_final = np.swapaxes(swap_hacked_pos, 0, 1)
neighbors = np.linalg.norm(
pos[idx:idx + 1, 4] - hacked_pos_final[:, 4], axis=1) < 10
neighbors = neighbors.astype(np.int32).sum()
if neighbors >= 24:
types.append("buried")
elif neighbors < 16:
types.append("surface")
else:
types.append("neutral")
if neighbors >= 24:
tresh = 0.98
else:
tresh = 0.95
if FLAGS.sample_mode == "weighted_gauss":
chis_list = interpolated_sample_normal(db,
angles[idx, 1],
angles[idx, 2],
res[idx],
neg_sample,
uniform=False)
elif FLAGS.sample_mode == "gmm":
chis_list = mixture_sample_normal(db,
angles[idx, 1],
angles[idx, 2],
res[idx],
neg_sample,
uniform=False)
elif FLAGS.sample_mode == "rosetta":
chis_list = exhaustive_sample(db,
angles[idx, 1],
angles[idx, 2],
res[idx],
tresh=tresh)
neg_chis.append(chis_list)
node_neg_embeds = []
length_chis = len(chis_list)
for i in range(neg_sample):
chis_target = angles[:, 4:8].copy()
if i >= len(chis_list):
node_neg_embed_copy = node_neg_embed.copy()
node_neg_embeds.append(node_neg_embeds[i % length_chis])
neg_chis[-1].append(chis_list[i % length_chis])
continue
chis = chis_list[i]
chis_target[idx] = (
chis * chis_valid[idx, :4] +
(1 - chis_valid[idx, :4]) * chis_target[idx])
pos_new = rotate_dihedral_fast(angles, par, child, pos,
pos_exist, chis_target,
chis_valid, idx)
node_neg_embed = reencode_dense_format(node_embed, pos_new,
pos_exist)
node_neg_embeds.append(node_neg_embed)
node_neg_embeds = np.array(node_neg_embeds)
dist = np.square(node_neg_embeds[:, :, -3:] -
pos[idx:idx + 1, 4:5, :]).sum(axis=2)
close_idx = np.argsort(dist)
node_neg_embeds = np.take_along_axis(node_neg_embeds,
close_idx[:, :64, None],
axis=1)
node_neg_embeds[:, :, -3:] = node_neg_embeds[:, :, -3:] / 10.0
node_neg_embeds[:, :, -3:] = node_neg_embeds[:, :, -3:] - np.mean(
node_neg_embeds[:, :, -3:], axis=1, keepdims=True)
node_embed_evals.append(node_neg_embeds)
if len(node_embed_evals) == nminibatch or idx == (n_amino - 2):
n_entries = len(node_embed_evals)
node_embed_evals = np.concatenate(node_embed_evals)
s = node_embed_evals.shape
# For sample rotations per batch
node_embed_evals = np.tile(node_embed_evals[:, None, :, :],
(1, rotations, 1, 1))
rot_matrix = so3.rvs(rotations)
if rotations == 1:
rot_matrix = rot_matrix[None, :, :]
node_embed_evals[:, :, :, -3:] = np.matmul(
node_embed_evals[:, :, :, -3:], rot_matrix[None, :, :, :])
node_embed_evals = node_embed_evals.reshape((-1, *s[1:]))
node_embed_feed = torch.from_numpy(
node_embed_evals).float().cuda()
with torch.no_grad():
energy = 0
if FLAGS.ensemble > 1:
for model in models:
energy_tmp = model.forward(node_embed_feed)
energy = energy + energy_tmp
else:
energy = model.forward(node_embed_feed)
energy = energy.view(n_entries, -1, rotations).mean(dim=2)
select_idx = torch.argmin(energy, dim=1).cpu().numpy()
for i in range(n_entries):
select_idx_i = select_idx[i]
valid_chi_idx = valid_chi_idxs[i]
rotamer_score, _ = compute_rotamer_score_planar(
gt_chis[i], neg_chis[i][select_idx_i],
valid_chi_idx[:4], res_names[i])
rotamer_scores.append(rotamer_score)
amino_recovery[str(res_names[i])] = amino_recovery[str(
res_names[i])] + [rotamer_score]
if types[i] == "buried":
buried_scores.append(rotamer_score)
elif types[i] == "surface":
surface_scores.append(rotamer_score)
gt_chis = []
node_embed_evals = []
neg_chis = []
valid_chi_idxs = []
res_names = []
types = []
counter += 1
rotamer_scores_total.append(np.mean(rotamer_scores))
if len(buried_scores) > 0:
buried_scores_total.append(np.mean(buried_scores))
surface_scores_total.append(np.mean(surface_scores))
for k, v in amino_recovery.items():
if len(v) > 0:
amino_recovery_total[k] = amino_recovery_total[k] + [
np.mean(v)
]
print(
"Obtained a rotamer recovery score of ",
np.mean(rotamer_scores_total),
np.std(rotamer_scores_total) / len(rotamer_scores_total)**0.5,
)
print(
"Obtained a buried recovery score of ",
np.mean(buried_scores_total),
np.std(buried_scores_total) / len(buried_scores_total)**0.5,
)
print(
"Obtained a surface recovery score of ",
np.mean(surface_scores_total),
np.std(surface_scores_total) / len(surface_scores_total)**0.5,
)
for k, v in amino_recovery_total.items():
print(
"per amino acid recovery of {} score of ".format(k),
np.mean(v),
np.std(v) / len(v)**0.5,
)
def __init__(
self,
FLAGS,
mmcif_path="./mmcif",
split="train",
rank_idx=0,
world_size=1,
uniform=True,
weighted_gauss=False,
gmm=False,
chi_mean=False,
valid=False,
):
files = []
dirs = os.listdir(osp.join(mmcif_path, "mmCIF"))
self.split = split
self.so3 = special_ortho_group(3)
self.chi_mean = chi_mean
self.weighted_gauss = weighted_gauss
self.gmm = gmm
self.uniform = uniform
# Filter out proteins in test dataset
for d in tqdm(dirs):
directory = osp.join(mmcif_path, "mmCIF", d)
d_files = os.listdir(directory)
files_tmp = [osp.join(directory, d_file) for d_file in d_files if ".p" in d_file]
for f in files_tmp:
name = f.split("/")[-1]
name = name.split(".")[0]
if name in test_rotamers and self.split == "test":
files.append(f)
elif name not in test_rotamers and self.split in ["train", "val"]:
files.append(f)
self.files = files
if split in ["train", "val"]:
duplicate_seqs = set()
# Remove proteins in the train dataset that are too similar to the test dataset
with open(osp.join(mmcif_path, "duplicate_sequences.txt"), "r") as f:
for line in f:
duplicate_seqs.add(line.strip())
fids = set()
# Remove low resolution proteins
with open(
osp.join(mmcif_path, "cullpdb_pc90_res1.8_R0.25_d190807_chains14857"), "r"
) as f:
i = 0
for line in f:
if i is not 0:
fid = line.split()[0]
if fid not in duplicate_seqs:
fids.add(fid)
i += 1
files_new = []
alphabet = []
for letter in range(65, 91):
alphabet.append(chr(letter))
for f in files:
tup = (f.split("/")[-1]).split(".")
if int(tup[1]) >= len(alphabet):
continue
seq_id = tup[0].upper() + alphabet[int(tup[1])]
if seq_id in fids:
files_new.append(f)
self.files = files_new
elif split == "test":
fids = set()
# Remove low resolution proteins
with open(
osp.join(mmcif_path, "cullpdb_pc90_res1.8_R0.25_d190807_chains14857"), "r"
) as f:
i = 0
for line in f:
if i is not 0:
fid = line.split()[0]
fids.add(fid)
i += 1
files_new = []
alphabet = []
for letter in range(65, 91):
alphabet.append(chr(letter))
for f in files:
tup = (f.split("/")[-1]).split(".")
if int(tup[1]) >= len(alphabet):
continue
seq_id = tup[0].upper() + alphabet[int(tup[1])]
if seq_id in fids:
files_new.append(f)
self.files = files_new
chunksize = len(self.files) // world_size
n = len(self.files)
# Set up a validation dataset
if split == "train":
n = self.files[int(0.95 * n) :]
elif split == "val":
n = self.files[: int(0.95 * n)]
self.FLAGS = FLAGS
self.db = load_rotamor_library()
print(f"Loaded {len(self.files)} files for {split} dataset split")
self.split = split