File size: 10,194 Bytes
3f0529e |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import typing as T
import numpy as np
import torch
import torch.nn.functional as F
from einops import rearrange, repeat
from torch import nn
from openfold.np import residue_constants
from openfold.np.protein import Protein as OFProtein
from openfold.np.protein import to_pdb
from openfold.utils.feats import atom14_to_atom37
def encode_sequence(
seq: str,
residue_index_offset: T.Optional[int] = 512,
chain_linker: T.Optional[str] = "G" * 25,
) -> T.Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
if chain_linker is None:
chain_linker = ""
if residue_index_offset is None:
residue_index_offset = 0
chains = seq.split(":")
seq = chain_linker.join(chains)
unk_idx = residue_constants.restype_order_with_x["X"]
encoded = torch.tensor(
[residue_constants.restype_order_with_x.get(aa, unk_idx) for aa in seq]
)
residx = torch.arange(len(encoded))
if residue_index_offset > 0:
start = 0
for i, chain in enumerate(chains):
residx[start : start + len(chain) + len(chain_linker)] += (
i * residue_index_offset
)
start += len(chain) + len(chain_linker)
linker_mask = torch.ones_like(encoded, dtype=torch.float32)
chain_index = []
offset = 0
for i, chain in enumerate(chains):
if i > 0:
chain_index.extend([i - 1] * len(chain_linker))
chain_index.extend([i] * len(chain))
offset += len(chain)
linker_mask[offset : offset + len(chain_linker)] = 0
offset += len(chain_linker)
chain_index = torch.tensor(chain_index, dtype=torch.int64)
return encoded, residx, linker_mask, chain_index
def batch_encode_sequences(
sequences: T.Sequence[str],
residue_index_offset: T.Optional[int] = 512,
chain_linker: T.Optional[str] = "G" * 25,
) -> T.Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
aatype_list = []
residx_list = []
linker_mask_list = []
chain_index_list = []
for seq in sequences:
aatype_seq, residx_seq, linker_mask_seq, chain_index_seq = encode_sequence(
seq,
residue_index_offset=residue_index_offset,
chain_linker=chain_linker,
)
aatype_list.append(aatype_seq)
residx_list.append(residx_seq)
linker_mask_list.append(linker_mask_seq)
chain_index_list.append(chain_index_seq)
aatype = collate_dense_tensors(aatype_list)
mask = collate_dense_tensors(
[aatype.new_ones(len(aatype_seq)) for aatype_seq in aatype_list]
)
residx = collate_dense_tensors(residx_list)
linker_mask = collate_dense_tensors(linker_mask_list)
chain_index_list = collate_dense_tensors(chain_index_list, -1)
return aatype, mask, residx, linker_mask, chain_index_list
def output_to_pdb(output: T.Dict) -> T.List[str]:
"""Returns the pbd (file) string from the model given the model output."""
# atom14_to_atom37 must be called first, as it fails on latest numpy if the
# input is a numpy array. It will work if the input is a torch tensor.
final_atom_positions = atom14_to_atom37(output["positions"][-1], output)
output = {k: v.to("cpu").numpy() for k, v in output.items()}
final_atom_positions = final_atom_positions.cpu().numpy()
final_atom_mask = output["atom37_atom_exists"]
pdbs = []
for i in range(output["aatype"].shape[0]):
aa = output["aatype"][i]
pred_pos = final_atom_positions[i]
mask = final_atom_mask[i]
resid = output["residue_index"][i] + 1
pred = OFProtein(
aatype=aa,
atom_positions=pred_pos,
atom_mask=mask,
residue_index=resid,
b_factors=output["plddt"][i],
chain_index=output["chain_index"][i] if "chain_index" in output else None,
)
pdbs.append(to_pdb(pred))
return pdbs
def collate_dense_tensors(
samples: T.List[torch.Tensor], pad_v: float = 0
) -> torch.Tensor:
"""
Takes a list of tensors with the following dimensions:
[(d_11, ..., d_1K),
(d_21, ..., d_2K),
...,
(d_N1, ..., d_NK)]
and stack + pads them into a single tensor of:
(N, max_i=1,N { d_i1 }, ..., max_i=1,N {diK})
"""
if len(samples) == 0:
return torch.Tensor()
if len(set(x.dim() for x in samples)) != 1:
raise RuntimeError(
f"Samples has varying dimensions: {[x.dim() for x in samples]}"
)
(device,) = tuple(set(x.device for x in samples)) # assumes all on same device
max_shape = [max(lst) for lst in zip(*[x.shape for x in samples])]
result = torch.empty(
len(samples), *max_shape, dtype=samples[0].dtype, device=device
)
result.fill_(pad_v)
for i in range(len(samples)):
result_i = result[i]
t = samples[i]
result_i[tuple(slice(0, k) for k in t.shape)] = t
return result
class Attention(nn.Module):
def __init__(self, embed_dim, num_heads, head_width, gated=False):
super().__init__()
assert embed_dim == num_heads * head_width
self.embed_dim = embed_dim
self.num_heads = num_heads
self.head_width = head_width
self.proj = nn.Linear(embed_dim, embed_dim * 3, bias=False)
self.o_proj = nn.Linear(embed_dim, embed_dim, bias=True)
self.gated = gated
if gated:
self.g_proj = nn.Linear(embed_dim, embed_dim)
torch.nn.init.zeros_(self.g_proj.weight)
torch.nn.init.ones_(self.g_proj.bias)
self.rescale_factor = self.head_width**-0.5
torch.nn.init.zeros_(self.o_proj.bias)
def forward(self, x, mask=None, bias=None, indices=None):
"""
Basic self attention with optional mask and external pairwise bias.
To handle sequences of different lengths, use mask.
Inputs:
x: batch of input sequneces (.. x L x C)
mask: batch of boolean masks where 1=valid, 0=padding position (.. x L_k). optional.
bias: batch of scalar pairwise attention biases (.. x Lq x Lk x num_heads). optional.
Outputs:
sequence projection (B x L x embed_dim), attention maps (B x L x L x num_heads)
"""
t = rearrange(self.proj(x), "... l (h c) -> ... h l c", h=self.num_heads)
q, k, v = t.chunk(3, dim=-1)
q = self.rescale_factor * q
a = torch.einsum("...qc,...kc->...qk", q, k)
# Add external attention bias.
if bias is not None:
a = a + rearrange(bias, "... lq lk h -> ... h lq lk")
# Do not attend to padding tokens.
if mask is not None:
mask = repeat(
mask, "... lk -> ... h lq lk", h=self.num_heads, lq=q.shape[-2]
)
a = a.masked_fill(mask == False, -np.inf)
a = F.softmax(a, dim=-1)
y = torch.einsum("...hqk,...hkc->...qhc", a, v)
y = rearrange(y, "... h c -> ... (h c)", h=self.num_heads)
if self.gated:
y = self.g_proj(x).sigmoid() * y
y = self.o_proj(y)
return y, rearrange(a, "... lq lk h -> ... h lq lk")
class Dropout(nn.Module):
"""
Implementation of dropout with the ability to share the dropout mask
along a particular dimension.
"""
def __init__(self, r: float, batch_dim: T.Union[int, T.List[int]]):
super(Dropout, self).__init__()
self.r = r
if type(batch_dim) == int:
batch_dim = [batch_dim]
self.batch_dim = batch_dim
self.dropout = nn.Dropout(self.r)
def forward(self, x: torch.Tensor) -> torch.Tensor:
shape = list(x.shape)
if self.batch_dim is not None:
for bd in self.batch_dim:
shape[bd] = 1
return x * self.dropout(x.new_ones(shape))
class SequenceToPair(nn.Module):
def __init__(self, sequence_state_dim, inner_dim, pairwise_state_dim):
super().__init__()
self.layernorm = nn.LayerNorm(sequence_state_dim)
self.proj = nn.Linear(sequence_state_dim, inner_dim * 2, bias=True)
self.o_proj = nn.Linear(2 * inner_dim, pairwise_state_dim, bias=True)
torch.nn.init.zeros_(self.proj.bias)
torch.nn.init.zeros_(self.o_proj.bias)
def forward(self, sequence_state):
"""
Inputs:
sequence_state: B x L x sequence_state_dim
Output:
pairwise_state: B x L x L x pairwise_state_dim
Intermediate state:
B x L x L x 2*inner_dim
"""
assert len(sequence_state.shape) == 3
s = self.layernorm(sequence_state)
s = self.proj(s)
q, k = s.chunk(2, dim=-1)
prod = q[:, None, :, :] * k[:, :, None, :]
diff = q[:, None, :, :] - k[:, :, None, :]
x = torch.cat([prod, diff], dim=-1)
x = self.o_proj(x)
return x
class PairToSequence(nn.Module):
def __init__(self, pairwise_state_dim, num_heads):
super().__init__()
self.layernorm = nn.LayerNorm(pairwise_state_dim)
self.linear = nn.Linear(pairwise_state_dim, num_heads, bias=False)
def forward(self, pairwise_state):
"""
Inputs:
pairwise_state: B x L x L x pairwise_state_dim
Output:
pairwise_bias: B x L x L x num_heads
"""
assert len(pairwise_state.shape) == 4
z = self.layernorm(pairwise_state)
pairwise_bias = self.linear(z)
return pairwise_bias
class ResidueMLP(nn.Module):
def __init__(self, embed_dim, inner_dim, norm=nn.LayerNorm, dropout=0):
super().__init__()
self.mlp = nn.Sequential(
norm(embed_dim),
nn.Linear(embed_dim, inner_dim),
nn.ReLU(),
nn.Linear(inner_dim, embed_dim),
nn.Dropout(dropout),
)
def forward(self, x):
return x + self.mlp(x)
|