File size: 22,684 Bytes
c87881a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 | """Set Transformer using only sequence embeddings and relative position."""
from __future__ import annotations
from dataclasses import asdict, dataclass
import torch
from torch import nn
from torch.nn import functional as F
@dataclass(frozen=True)
class ModelConfig:
architecture: str = "setnet"
esm_dimension: int = 1280
pfam_dimension: int = 64
pfam_vocab_size: int = 0
hidden_dimension: int = 256
output_dimension: int = 256
attention_heads: int = 8
inducing_points: int = 32
feedforward_dimension: int = 512
attention_blocks: int = 2
position_bins: int = 64
dropout: float = 0.1
@classmethod
def from_dict(cls, values: dict[str, object]) -> "ModelConfig":
return cls(**{key: values[key] for key in asdict(cls()) if key in values})
class GeneProjection(nn.Module):
def __init__(self, config: ModelConfig) -> None:
super().__init__()
self.projection = nn.Linear(config.esm_dimension, config.hidden_dimension)
self.normalization = nn.LayerNorm(config.hidden_dimension)
self.position = nn.Embedding(config.position_bins, config.hidden_dimension)
self.dropout = nn.Dropout(config.dropout)
self.position_bins = config.position_bins
def forward(self, embeddings: torch.Tensor, positions: torch.Tensor) -> torch.Tensor:
bins = (positions * (self.position_bins - 1)).long().clamp(0, self.position_bins - 1)
projected = F.gelu(self.normalization(self.projection(embeddings)))
return self.dropout(projected + self.position(bins))
class PfamGeneProjection(nn.Module):
"""Project ESM genes together with a BGC-level Pfam inventory embedding."""
def __init__(self, config: ModelConfig) -> None:
super().__init__()
if config.pfam_vocab_size < 3:
raise ValueError("Pfam vocabulary must contain padding, unknown, and one domain token")
self.pfam_embedding = nn.Embedding(
config.pfam_vocab_size, config.pfam_dimension, padding_idx=0
)
self.projection = nn.Linear(
config.esm_dimension + config.pfam_dimension, config.hidden_dimension
)
self.normalization = nn.LayerNorm(config.hidden_dimension)
self.position = nn.Embedding(config.position_bins, config.hidden_dimension)
self.dropout = nn.Dropout(config.dropout)
self.position_bins = config.position_bins
def forward(
self,
embeddings: torch.Tensor,
positions: torch.Tensor,
pfam_tokens: torch.Tensor,
) -> torch.Tensor:
bins = (positions * (self.position_bins - 1)).long().clamp(0, self.position_bins - 1)
token_mask = pfam_tokens.ne(0).unsqueeze(-1)
token_values = self.pfam_embedding(pfam_tokens).masked_fill(~token_mask, 0.0)
counts = token_mask.sum(dim=1).clamp_min(1)
pfam_summary = token_values.sum(dim=1) / counts
pfam_summary = pfam_summary.unsqueeze(1).expand(-1, embeddings.shape[1], -1)
merged = torch.cat((embeddings, pfam_summary), dim=-1)
projected = F.gelu(self.normalization(self.projection(merged)))
return self.dropout(projected + self.position(bins))
class InducedSelfAttentionBlock(nn.Module):
def __init__(self, config: ModelConfig) -> None:
super().__init__()
dimension = config.hidden_dimension
self.inducing = nn.Parameter(torch.empty(config.inducing_points, dimension))
nn.init.xavier_uniform_(self.inducing)
self.inducing_norm = nn.LayerNorm(dimension)
self.input_norm = nn.LayerNorm(dimension)
self.summary_attention = nn.MultiheadAttention(
dimension, config.attention_heads, config.dropout, batch_first=True
)
self.output_attention = nn.MultiheadAttention(
dimension, config.attention_heads, config.dropout, batch_first=True
)
self.output_norm = nn.LayerNorm(dimension)
self.feedforward = nn.Sequential(
nn.Linear(dimension, config.feedforward_dimension),
nn.GELU(),
nn.Dropout(config.dropout),
nn.Linear(config.feedforward_dimension, dimension),
nn.Dropout(config.dropout),
)
def forward(self, values: torch.Tensor, padding_mask: torch.Tensor | None) -> torch.Tensor:
batch_size = values.shape[0]
inducing = self.inducing.unsqueeze(0).expand(batch_size, -1, -1)
normalized_values = self.input_norm(values)
summary, _ = self.summary_attention(
inducing, normalized_values, normalized_values, key_padding_mask=padding_mask
)
normalized_summary = self.inducing_norm(summary)
update, _ = self.output_attention(
normalized_values, normalized_summary, normalized_summary
)
output = values + update
return output + self.feedforward(self.output_norm(output))
class PoolingByAttention(nn.Module):
def __init__(self, config: ModelConfig) -> None:
super().__init__()
self.seed = nn.Parameter(torch.empty(1, config.hidden_dimension))
nn.init.xavier_uniform_(self.seed)
self.normalization = nn.LayerNorm(config.hidden_dimension)
self.attention = nn.MultiheadAttention(
config.hidden_dimension, config.attention_heads, config.dropout, batch_first=True
)
def forward(self, values: torch.Tensor, padding_mask: torch.Tensor | None) -> torch.Tensor:
seed = self.seed.unsqueeze(0).expand(values.shape[0], -1, -1)
output, _ = self.attention(
seed,
self.normalization(values),
self.normalization(values),
key_padding_mask=padding_mask,
)
return output.squeeze(1)
class LeakageFreeBGCSetNet(nn.Module):
"""Map a variable-length BGC gene set to a normalized embedding."""
input_names = ("gene_embeddings", "relative_positions", "padding_mask")
def __init__(self, config: ModelConfig) -> None:
super().__init__()
self.config = config
self.gene_projection = GeneProjection(config)
self.blocks = nn.ModuleList(
InducedSelfAttentionBlock(config) for _ in range(config.attention_blocks)
)
self.pooling = PoolingByAttention(config)
self.output = nn.Linear(config.hidden_dimension, config.output_dimension)
def encode_genes(
self,
gene_embeddings: torch.Tensor,
relative_positions: torch.Tensor,
padding_mask: torch.Tensor | None = None,
) -> torch.Tensor:
values = self.gene_projection(gene_embeddings, relative_positions)
for block in self.blocks:
values = block(values, padding_mask)
return values
def forward(
self,
gene_embeddings: torch.Tensor,
relative_positions: torch.Tensor,
padding_mask: torch.Tensor | None = None,
) -> torch.Tensor:
genes = self.encode_genes(gene_embeddings, relative_positions, padding_mask)
pooled = self.pooling(genes, padding_mask)
return F.normalize(self.output(pooled), p=2, dim=-1)
class MaskedGenePredictionHead(nn.Module):
def __init__(self, config: ModelConfig) -> None:
super().__init__()
self.layers = nn.Sequential(
nn.Linear(config.hidden_dimension, config.feedforward_dimension),
nn.GELU(),
nn.Linear(config.feedforward_dimension, config.esm_dimension),
)
def forward(self, contextual_embeddings: torch.Tensor) -> torch.Tensor:
return self.layers(contextual_embeddings)
class GatedDeepSets(nn.Module):
"""Permutation-invariant learned pooling without gene-gene interactions."""
input_names = ("gene_embeddings", "relative_positions", "padding_mask")
def __init__(self, config: ModelConfig) -> None:
super().__init__()
self.config = config
self.gene_projection = GeneProjection(config)
self.gate = nn.Sequential(
nn.Linear(config.hidden_dimension, config.hidden_dimension // 2),
nn.GELU(),
nn.Linear(config.hidden_dimension // 2, 1),
)
self.output = nn.Linear(config.hidden_dimension, config.output_dimension)
def encode_genes(self, gene_embeddings, relative_positions, padding_mask=None):
return self.gene_projection(gene_embeddings, relative_positions)
def forward(self, gene_embeddings, relative_positions, padding_mask=None):
genes = self.encode_genes(gene_embeddings, relative_positions, padding_mask)
logits = self.gate(genes).squeeze(-1)
if padding_mask is not None:
logits = logits.masked_fill(padding_mask, -torch.finfo(logits.dtype).max)
weights = torch.softmax(logits, dim=-1)
if padding_mask is not None:
weights = weights.masked_fill(padding_mask, 0.0)
pooled = (genes * weights.unsqueeze(-1)).sum(dim=1)
return F.normalize(self.output(pooled), p=2, dim=-1)
class RotarySelfAttentionBlock(nn.Module):
"""Bidirectional self-attention with rotary position phases."""
def __init__(self, config: ModelConfig) -> None:
super().__init__()
d = config.hidden_dimension
self.heads = config.attention_heads
self.head_dim = d // self.heads
if self.head_dim % 2:
raise ValueError("Rotary attention requires an even per-head dimension")
self.norm = nn.LayerNorm(d)
self.qkv = nn.Linear(d, 3 * d)
self.output = nn.Linear(d, d)
self.dropout = nn.Dropout(config.dropout)
self.ffn_norm = nn.LayerNorm(d)
self.ffn = nn.Sequential(
nn.Linear(d, config.feedforward_dimension), nn.GELU(),
nn.Dropout(config.dropout), nn.Linear(config.feedforward_dimension, d),
nn.Dropout(config.dropout),
)
def _rotate(self, values: torch.Tensor, positions: torch.Tensor) -> torch.Tensor:
half = self.head_dim // 2
inv = 1.0 / (10000.0 ** (
torch.arange(half, device=values.device, dtype=values.dtype) / half
))
angles = positions[:, None, :, None] * 128.0 * inv[None, None, None, :]
cos, sin = angles.cos(), angles.sin()
first, second = values[..., :half], values[..., half:]
return torch.cat((first * cos - second * sin, first * sin + second * cos), dim=-1)
def forward(self, values: torch.Tensor, positions: torch.Tensor, padding_mask=None) -> torch.Tensor:
normalized = self.norm(values)
batch, length, dimension = normalized.shape
qkv = self.qkv(normalized).view(batch, length, 3, self.heads, self.head_dim)
q, k, v = qkv.unbind(dim=2)
q, k = q.transpose(1, 2), k.transpose(1, 2)
v = v.transpose(1, 2)
q, k = self._rotate(q, positions), self._rotate(k, positions)
scores = torch.matmul(q, k.transpose(-2, -1)) / (self.head_dim ** 0.5)
if padding_mask is not None:
scores = scores.masked_fill(
padding_mask[:, None, None, :], -torch.finfo(scores.dtype).max
)
attention = self.dropout(torch.softmax(scores, dim=-1))
contextual = torch.matmul(attention, v).transpose(1, 2).reshape(batch, length, dimension)
output = values + self.output(contextual)
output = output + self.ffn(self.ffn_norm(output))
if padding_mask is not None:
output = output.masked_fill(padding_mask.unsqueeze(-1), 0.0)
return output
class RoPETransformer(nn.Module):
"""BGC-MAP-inspired bidirectional RoPE encoder with attentive pooling."""
input_names = ("gene_embeddings", "relative_positions", "padding_mask")
def __init__(self, config: ModelConfig) -> None:
super().__init__()
self.config = config
self.gene_projection = GeneProjection(config)
self.blocks = nn.ModuleList(
RotarySelfAttentionBlock(config) for _ in range(config.attention_blocks)
)
self.pooling = PoolingByAttention(config)
self.output = nn.Linear(config.hidden_dimension, config.output_dimension)
def encode_genes(self, gene_embeddings, relative_positions, padding_mask=None):
values = self.gene_projection(gene_embeddings, relative_positions)
for block in self.blocks:
values = block(values, relative_positions, padding_mask)
return values
def forward(self, gene_embeddings, relative_positions, padding_mask=None):
genes = self.encode_genes(gene_embeddings, relative_positions, padding_mask)
return F.normalize(self.output(self.pooling(genes, padding_mask)), p=2, dim=-1)
class CrossAttentionPool(nn.Module):
"""Learned context queries cross-attend to a self-attended BGC sequence."""
input_names = ("gene_embeddings", "relative_positions", "padding_mask")
def __init__(self, config: ModelConfig) -> None:
super().__init__()
self.config = config
d = config.hidden_dimension
self.gene_projection = GeneProjection(config)
layer = nn.TransformerEncoderLayer(
d, config.attention_heads, config.feedforward_dimension, config.dropout,
batch_first=True, norm_first=True, activation="gelu",
)
self.encoder = nn.TransformerEncoder(layer, config.attention_blocks)
self.queries = nn.Parameter(torch.empty(2, d))
nn.init.xavier_uniform_(self.queries)
self.cross = nn.MultiheadAttention(d, config.attention_heads, config.dropout, batch_first=True)
self.output = nn.Linear(d, config.output_dimension)
def encode_genes(self, gene_embeddings, relative_positions, padding_mask=None):
values = self.gene_projection(gene_embeddings, relative_positions)
return self.encoder(values, src_key_padding_mask=padding_mask)
def forward(self, gene_embeddings, relative_positions, padding_mask=None):
genes = self.encode_genes(gene_embeddings, relative_positions, padding_mask)
queries = self.queries.unsqueeze(0).expand(genes.shape[0], -1, -1)
attended, _ = self.cross(queries, genes, genes, key_padding_mask=padding_mask)
return F.normalize(self.output(attended.mean(dim=1)), p=2, dim=-1)
class LocalGlobalEncoder(nn.Module):
"""PST-inspired local adjacency message passing followed by global pooling."""
input_names = ("gene_embeddings", "relative_positions", "padding_mask")
def __init__(self, config: ModelConfig) -> None:
super().__init__()
self.config = config
d = config.hidden_dimension
self.gene_projection = GeneProjection(config)
self.local = nn.Conv1d(d, d, kernel_size=3, padding=1, groups=1)
self.local_gate = nn.Sequential(nn.Linear(d, d), nn.Sigmoid())
self.global_block = InducedSelfAttentionBlock(config)
self.pooling = PoolingByAttention(config)
self.output = nn.Linear(d, config.output_dimension)
def encode_genes(self, gene_embeddings, relative_positions, padding_mask=None):
values = self.gene_projection(gene_embeddings, relative_positions)
local = self.local(values.transpose(1, 2)).transpose(1, 2)
values = values + local * self.local_gate(values)
return self.global_block(values, padding_mask)
def forward(self, gene_embeddings, relative_positions, padding_mask=None):
genes = self.encode_genes(gene_embeddings, relative_positions, padding_mask)
return F.normalize(self.output(self.pooling(genes, padding_mask)), p=2, dim=-1)
class DilatedCNN(nn.Module):
"""BiGCARP-inspired residual dilated convolutional encoder."""
input_names = ("gene_embeddings", "relative_positions", "padding_mask")
def __init__(self, config: ModelConfig) -> None:
super().__init__()
self.config = config
d = config.hidden_dimension
self.gene_projection = GeneProjection(config)
self.blocks = nn.ModuleList(
nn.Sequential(
nn.Conv1d(d, d, 3, padding=dilation, dilation=dilation),
nn.GELU(), nn.Dropout(config.dropout), nn.Conv1d(d, d, 1),
) for dilation in (1, 2, 4, 8, 16)
)
self.pooling = PoolingByAttention(config)
self.output = nn.Linear(d, config.output_dimension)
def encode_genes(self, gene_embeddings, relative_positions, padding_mask=None):
values = self.gene_projection(gene_embeddings, relative_positions)
for block in self.blocks:
updated = block(values.transpose(1, 2)).transpose(1, 2)
values = values + updated
if padding_mask is not None:
values = values.masked_fill(padding_mask.unsqueeze(-1), 0.0)
return values
def forward(self, gene_embeddings, relative_positions, padding_mask=None):
genes = self.encode_genes(gene_embeddings, relative_positions, padding_mask)
return F.normalize(self.output(self.pooling(genes, padding_mask)), p=2, dim=-1)
class HierarchicalLocalGlobal(nn.Module):
"""Multi-scale local convolutions followed by a small global Transformer."""
input_names = ("gene_embeddings", "relative_positions", "padding_mask")
def __init__(self, config: ModelConfig) -> None:
super().__init__()
self.config = config
d = config.hidden_dimension
self.gene_projection = GeneProjection(config)
self.local3 = nn.Conv1d(d, d // 2, 3, padding=1)
self.local5 = nn.Conv1d(d, d // 2, 5, padding=2)
self.merge = nn.Linear(d, d)
layer = nn.TransformerEncoderLayer(
d, config.attention_heads, config.feedforward_dimension, config.dropout,
batch_first=True, norm_first=True, activation="gelu",
)
self.global_encoder = nn.TransformerEncoder(
layer, max(1, config.attention_blocks // 2)
)
self.pooling = PoolingByAttention(config)
self.output = nn.Linear(d, config.output_dimension)
def encode_genes(self, gene_embeddings, relative_positions, padding_mask=None):
values = self.gene_projection(gene_embeddings, relative_positions)
transposed = values.transpose(1, 2)
local = torch.cat((self.local3(transposed), self.local5(transposed)), dim=1).transpose(1, 2)
values = values + self.merge(local)
return self.global_encoder(values, src_key_padding_mask=padding_mask)
def forward(self, gene_embeddings, relative_positions, padding_mask=None):
genes = self.encode_genes(gene_embeddings, relative_positions, padding_mask)
return F.normalize(self.output(self.pooling(genes, padding_mask)), p=2, dim=-1)
class PfamAugmentedSetNet(nn.Module):
"""SetNet whose gene projection is conditioned on the BGC Pfam inventory."""
input_names = ("gene_embeddings", "relative_positions", "padding_mask", "pfam_tokens")
uses_pfam = True
def __init__(self, config: ModelConfig) -> None:
super().__init__()
self.config = config
self.gene_projection = PfamGeneProjection(config)
self.blocks = nn.ModuleList(
InducedSelfAttentionBlock(config) for _ in range(config.attention_blocks)
)
self.pooling = PoolingByAttention(config)
self.output = nn.Linear(config.hidden_dimension, config.output_dimension)
def encode_genes(
self,
gene_embeddings: torch.Tensor,
relative_positions: torch.Tensor,
padding_mask: torch.Tensor | None = None,
pfam_tokens: torch.Tensor | None = None,
) -> torch.Tensor:
if pfam_tokens is None:
raise ValueError("Pfam-augmented SetNet requires pfam_tokens")
values = self.gene_projection(gene_embeddings, relative_positions, pfam_tokens)
for block in self.blocks:
values = block(values, padding_mask)
return values
def forward(
self,
gene_embeddings: torch.Tensor,
relative_positions: torch.Tensor,
padding_mask: torch.Tensor | None = None,
pfam_tokens: torch.Tensor | None = None,
) -> torch.Tensor:
genes = self.encode_genes(
gene_embeddings, relative_positions, padding_mask, pfam_tokens
)
pooled = self.pooling(genes, padding_mask)
return F.normalize(self.output(pooled), p=2, dim=-1)
class WeightedPfamJaccard(nn.Module):
"""Learn nonnegative Pfam importance weights for differentiable set Jaccard."""
input_names = ("pfam_tokens",)
uses_pfam = True
def __init__(self, config: ModelConfig) -> None:
super().__init__()
if config.pfam_vocab_size < 3:
raise ValueError("Pfam vocabulary must contain padding, unknown, and one domain token")
initial = float(torch.log(torch.expm1(torch.tensor(1.0))))
self.raw_weights = nn.Parameter(
torch.full((config.pfam_vocab_size,), initial, dtype=torch.float32)
)
with torch.no_grad():
self.raw_weights[0] = -20.0
def domain_weights(self) -> torch.Tensor:
positive = F.softplus(self.raw_weights)
return torch.cat((positive[:1] * 0.0, positive[1:]))
def pairwise_jaccard(self, pfam_tokens: torch.Tensor) -> torch.Tensor:
if pfam_tokens.ndim != 2:
raise ValueError("Pfam tokens must have shape [batch, domains]")
batch_size = pfam_tokens.shape[0]
vocabulary = self.raw_weights.shape[0]
presence = torch.zeros(
batch_size, vocabulary, device=pfam_tokens.device, dtype=torch.float32
)
presence.scatter_(1, pfam_tokens.clamp_min(0), 1.0)
presence[:, 0] = 0.0
weighted = presence * self.domain_weights().to(pfam_tokens.device)
totals = weighted.sum(dim=1)
intersection = weighted @ presence.T
union = totals[:, None] + totals[None, :] - intersection
return intersection / union.clamp_min(1e-8)
def forward(self, pfam_tokens: torch.Tensor) -> torch.Tensor:
return self.domain_weights()
ARCHITECTURES = {
"setnet": LeakageFreeBGCSetNet,
"weighted_pfam_jaccard": WeightedPfamJaccard,
"pfam_setnet": PfamAugmentedSetNet,
"gated_deepsets": GatedDeepSets,
"rope_transformer": RoPETransformer,
"cross_attention": CrossAttentionPool,
"local_global": LocalGlobalEncoder,
"dilated_cnn": DilatedCNN,
"hierarchical": HierarchicalLocalGlobal,
}
def build_model(config: ModelConfig) -> nn.Module:
try:
model_class = ARCHITECTURES[config.architecture]
except KeyError as exc:
raise ValueError(f"Unknown architecture: {config.architecture}") from exc
return model_class(config)
|