File size: 69,247 Bytes
a328241 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 |
import logging
import torch
import torch.nn as nn
import torch.nn.functional as F
import json
import time
import pytorch_lightning as pl
from torch.optim import AdamW
from torchmetrics import MeanSquaredError, PearsonCorrCoef, SpearmanCorrCoef, R2Score
logger = logging.getLogger(__name__)
# ===================== VERSION STRING FOR CLUSTER VERIFICATION =====================
ARCH_VERSION = "2024-12-24-stability-fix"
print(f"[ARCH] architectures.py loaded: {ARCH_VERSION}")
# ====================================================================================
class Interp1d(torch.autograd.Function):
@staticmethod
def forward(ctx, x, y, xnew):
is_flat = {}
vals = {'x': x, 'y': y, 'xnew': xnew}
for name, arr in vals.items():
is_flat[name] = (arr.dim() == 1)
if is_flat[name]:
vals[name] = arr.unsqueeze(0)
x_2d, y_2d, xnew_2d = vals['x'], vals['y'], vals['xnew']
B, Nx = x_2d.shape
# SAFETY: Handle edge case where sequence length is < 5
if Nx < 5:
# Return constant interpolation (repeat/average the values)
ynew_2d = y_2d.mean(dim=1, keepdim=True).expand(-1, xnew_2d.shape[1])
ctx.save_for_backward(x_2d, y_2d, xnew_2d,
torch.zeros_like(xnew_2d, dtype=torch.long),
torch.zeros_like(xnew_2d))
ctx.Nx_was_small = True
if is_flat['x'] and is_flat['xnew']:
ynew_2d = ynew_2d.squeeze(0)
return ynew_2d
ctx.Nx_was_small = False
idx = torch.searchsorted(x_2d, xnew_2d, right=False) - 1
idx = idx.clamp(min=0, max=Nx-2)
xL = torch.gather(x_2d, 1, idx)
xR = torch.gather(x_2d, 1, idx+1)
yL = torch.gather(y_2d, 1, idx)
yR = torch.gather(y_2d, 1, idx+1)
denom = (xR - xL)
denom[denom == 0] = 1e-12
t = (xnew_2d - xL)/denom
ynew_2d = yL + (yR - yL)*t
ctx.save_for_backward(x_2d, y_2d, xnew_2d, idx, t)
if is_flat['x'] and is_flat['xnew']:
ynew_2d = ynew_2d.squeeze(0)
return ynew_2d
@staticmethod
def backward(ctx, grad_out):
x_2d, y_2d, xnew_2d, idx, t = ctx.saved_tensors
grad_x = grad_y = grad_xnew = None
# Handle edge case from forward
if getattr(ctx, 'Nx_was_small', False):
if ctx.needs_input_grad[1]:
grad_y = grad_out.sum(dim=-1, keepdim=True).expand_as(y_2d)
return grad_x, grad_y, grad_xnew
if ctx.needs_input_grad[1]:
grad_y_tmp = torch.zeros_like(y_2d)
idxp1 = (idx + 1).clamp(max=y_2d.shape[1] - 1) # SAFETY: clamp idxp1
# Calculate gradients
grad_yL = (1.0 - t) * grad_out
grad_yR = t * grad_out
# Ensure consistent dtype between source and destination tensors
grad_yL = grad_yL.to(dtype=grad_y_tmp.dtype)
grad_yR = grad_yR.to(dtype=grad_y_tmp.dtype)
grad_y_tmp.scatter_add_(1, idx, grad_yL)
grad_y_tmp.scatter_add_(1, idxp1, grad_yR)
grad_y = grad_y_tmp
return grad_x, grad_y, grad_xnew
def interp1d(x, y, xnew):
return Interp1d.apply(x, y, xnew)
class SWE_Pooling(nn.Module):
"""
Sliced-Wasserstein Embedding (SWE) Pooling.
Maps token embeddings [B, L, d_in] => [B, num_slices].
"""
def __init__(self, d_in, num_slices, num_ref_points, freeze_swe=False):
super().__init__()
self.num_slices = num_slices
self.num_ref_points = num_ref_points
ref = torch.linspace(-1,1,num_ref_points).unsqueeze(1).repeat(1,num_slices)
self.reference = nn.Parameter(ref, requires_grad=not freeze_swe)
self.theta = nn.utils.weight_norm(nn.Linear(d_in, num_slices, bias=False), dim=0)
self.theta.weight_g.data = torch.ones_like(self.theta.weight_g.data)
self.theta.weight_g.requires_grad=False
nn.init.normal_(self.theta.weight_v)
self.weight = nn.Linear(num_ref_points,1,bias=False)
if freeze_swe:
self.theta.weight_v.requires_grad=False
self.reference.requires_grad=False
def forward(self, X, mask=None):
B, N, D = X.shape
device = X.device
X_slices = self.theta(X) # => [B,N,num_slices]
X_slices_sorted, _ = torch.sort(X_slices, dim=1)
x_coord = torch.linspace(0,1,N,device=device).unsqueeze(0).repeat(B*self.num_slices,1)
X_flat = X_slices_sorted.permute(0,2,1).reshape(B*self.num_slices, N)
xnew = torch.linspace(0,1,self.num_ref_points,device=device).unsqueeze(0).repeat(B*self.num_slices,1)
y_intp = interp1d(x_coord, X_flat, xnew)
X_slices_sorted_interp = y_intp.view(B,self.num_slices,self.num_ref_points).permute(0,2,1)
r_expanded = self.reference.expand_as(X_slices_sorted_interp)
embeddings = (r_expanded - X_slices_sorted_interp).permute(0,2,1) # => [B,num_slices,num_ref_points]
weighted = self.weight(embeddings).sum(dim=-1) # => [B, num_slices]
return weighted
#############################################################################
# Enhanced Mutation-Aware SWE_Pooling #
#############################################################################
class MutationAwareSWEPooling(nn.Module):
"""
Enhanced Sliced-Wasserstein Embedding Pooling with explicit mutation position handling.
Maps token embeddings [B, L, d_in] => [B, num_slices].
- Preserves mutation position information through weighted aggregation
- Uses mutation positions to guide the pooling process
"""
def __init__(self, d_in, num_slices, num_ref_points, freeze_swe=False):
super().__init__()
self.num_slices = num_slices
self.num_ref_points = num_ref_points
self.d_esm = 1152 # FIXED: Hardcode to 1152 to avoid channel indexing bugs with context window
# Standard SWE components
ref = torch.linspace(-1, 1, num_ref_points).unsqueeze(1).repeat(1, num_slices)
self.reference = nn.Parameter(ref, requires_grad=not freeze_swe)
# For ESM features (without mutation channel)
self.theta = nn.utils.weight_norm(nn.Linear(self.d_esm, num_slices, bias=False), dim=0)
self.theta.weight_g.data = torch.ones_like(self.theta.weight_g.data)
self.theta.weight_g.requires_grad = False
nn.init.normal_(self.theta.weight_v)
# Mutation-aware components
self.mutation_importance = nn.Sequential(
nn.Linear(1, 32),
nn.ReLU(),
nn.Linear(32, num_slices),
nn.Sigmoid()
)
# Position-specific weighting for each slice
self.pos_weighting = nn.Linear(1, num_slices, bias=False)
# FIXED: Direct projection of mutation channel (the 1153rd dim)
# Previously, this channel was only used as a multiplier, meaning if ESM
# features had no diff, the result was 0. Now we project it directly.
self.mut_projection = nn.Linear(1, num_slices, bias=False)
# Final weighting
self.weight = nn.Linear(num_ref_points, 1, bias=False)
if freeze_swe:
self.theta.weight_v.requires_grad = False
self.reference.requires_grad = False
def forward(self, X, mask=None):
"""
X: [B, L, d_in] where d_in = d_esm + 1 (mutation channel)
mask: [B, L] boolean mask
"""
B, N, D = X.shape
device = X.device
# Check if using context window (additional channel)
use_context = (D > self.d_esm + 1)
if use_context:
# Split ESM features and channels
X_esm = X[:, :, :-2] # [B, L, d_esm]
X_mut = X[:, :, -2:-1] # [B, L, 1] - mutation indicator
else:
# Split ESM features and mutation channel
X_esm = X[:, :, :-1] # [B, L, d_esm]
X_mut = X[:, :, -1:] # [B, L, 1] - mutation indicator
# Regular SWE on ESM features
X_slices = self.theta(X_esm) # => [B, L, num_slices]
# Compute mutation importance weights
mut_weights = self.mutation_importance(X_mut) # [B, L, num_slices]
# Create position encodings (0 to 1 for each sequence)
pos_tensor = torch.linspace(0, 1, N, device=device).view(1, N, 1).expand(B, N, 1)
pos_weights = self.pos_weighting(pos_tensor) # [B, L, num_slices]
# Apply mutation-aware weighting to slices
# Use both mutation indicator and position information
# BUGFIX: We ALSO add the projected mutation signal directly.
# This ensures the model 'sees' the 1.0 signal even if ESM features are identical.
X_slices = X_slices * (1.0 + mut_weights * pos_weights) + self.mut_projection(X_mut)
# Sort slices as in standard SWE
X_slices_sorted, _ = torch.sort(X_slices, dim=1)
# Continue with standard SWE interpolation
x_coord = torch.linspace(0, 1, N, device=device).unsqueeze(0).repeat(B*self.num_slices, 1)
X_flat = X_slices_sorted.permute(0, 2, 1).reshape(B*self.num_slices, N)
xnew = torch.linspace(0, 1, self.num_ref_points, device=device).unsqueeze(0).repeat(B*self.num_slices, 1)
y_intp = interp1d(x_coord, X_flat, xnew)
X_slices_sorted_interp = y_intp.view(B, self.num_slices, self.num_ref_points).permute(0, 2, 1)
r_expanded = self.reference.expand_as(X_slices_sorted_interp)
embeddings = (r_expanded - X_slices_sorted_interp).permute(0, 2, 1) # => [B, num_slices, num_ref_points]
weighted = self.weight(embeddings).sum(dim=-1) # => [B, num_slices]
return weighted
#############################################################################
# Mutation-Specific Cross-Attention with Gating #
#############################################################################
class MutationSpecificAttention(nn.Module):
"""
Enhanced cross-attention that explicitly handles mutation positions with gating.
- Keeps ESM embeddings (1152-dim) and mutation channel separate
- Uses specific mutation positions to guide attention
- Preserves position-specific information throughout the network
- Adds gating mechanism to control information flow
- Includes memory-efficient computation for long sequences
"""
def __init__(self, d_model=1152, num_heads=4, dropout=0.1):
super().__init__()
assert d_model % num_heads == 0, "d_model must be divisible by num_heads"
self.d_model = d_model
self.num_heads = num_heads
self.head_dim = d_model // num_heads
# Core attention for ESM embeddings only (1152-dim)
self.query = nn.Linear(d_model, d_model)
self.key = nn.Linear(d_model, d_model)
self.value = nn.Linear(d_model, d_model)
# Absolute position encoding
self.pos_encoder = nn.Sequential(
nn.Linear(1, 32),
nn.ReLU(),
nn.Linear(32, d_model)
)
# Mutation-position specific attention
self.mut_encoder = nn.Sequential(
nn.Linear(2, 64), # Input: [mut_binary, position_normalized]
nn.ReLU(),
nn.Linear(64, num_heads)
)
# Gating mechanism to control information flow
self.gate = nn.Sequential(
nn.Linear(d_model*2, d_model),
nn.Sigmoid()
)
self.dropout = nn.Dropout(dropout)
self.out_proj = nn.Linear(d_model, d_model)
def split_heads(self, x):
"""Split the last dimension into (heads, head_dim)"""
batch_size, seq_len, _ = x.shape
x = x.view(batch_size, seq_len, self.num_heads, self.head_dim)
return x.permute(0, 2, 1, 3) # [batch, heads, seq_len, head_dim]
def merge_heads(self, x):
"""Merge the (heads, head_dim) into d_model"""
batch_size, _, seq_len, _ = x.shape
x = x.permute(0, 2, 1, 3) # [batch, seq_len, heads, head_dim]
return x.reshape(batch_size, seq_len, self.d_model)
def forward(self, q_esm, k_esm, v_esm, q_mut, k_mut, mask=None):
"""
Inputs:
q_esm, k_esm, v_esm: ESM embeddings [B, L, 1152]
q_mut, k_mut: Mutation information [B, L, 1]
mask: Optional attention mask [B, L] or [B, 1, L]
"""
batch_size = q_esm.shape[0]
q_len, k_len = q_esm.shape[1], k_esm.shape[1]
# Create position tensors (0-1 range for each sequence)
q_pos = torch.linspace(0, 1, q_len, device=q_esm.device).view(1, -1, 1).expand(batch_size, q_len, 1)
k_pos = torch.linspace(0, 1, k_len, device=k_esm.device).view(1, -1, 1).expand(batch_size, k_len, 1)
# Position encoding
q_pos_enc = self.pos_encoder(q_pos)
k_pos_enc = self.pos_encoder(k_pos)
# Add position encodings to ESM features
q_esm_pos = q_esm + q_pos_enc
k_esm_pos = k_esm + k_pos_enc
# Process core ESM embeddings with position information
q = self.split_heads(self.query(q_esm_pos)) # [B, h, q_len, d_k]
k = self.split_heads(self.key(k_esm_pos)) # [B, h, k_len, d_k]
v = self.split_heads(self.value(v_esm)) # [B, h, v_len, d_v]
# Concatenate mutation indicator with position
q_mut_pos = torch.cat([q_mut, q_pos], dim=-1) # [B, q_len, 2]
k_mut_pos = torch.cat([k_mut, k_pos], dim=-1) # [B, k_len, 2]
# Encode position-aware mutation information
q_mut_enc = self.mut_encoder(q_mut_pos) # [B, q_len, num_heads]
k_mut_enc = self.mut_encoder(k_mut_pos) # [B, k_len, num_heads]
# Standard scaled dot-product attention
d_k = q.size(-1)
scores = torch.matmul(q, k.transpose(-2, -1)) / (d_k ** 0.5) # [B, h, q_len, k_len]
# Create mutation-position attention bias
# This explicitly boosts attention between positions based on mutation status
mut_attn_bias = torch.matmul(
q_mut_enc.permute(0, 2, 1).unsqueeze(3), # [B, h, q_len, 1]
k_mut_enc.permute(0, 2, 1).unsqueeze(2) # [B, h, 1, k_len]
) # [B, h, q_len, k_len]
# Apply mutation bias to attention scores
# This makes mutations and their surrounding context attend more to each other
scores = scores + mut_attn_bias
# Apply mask if provided
if mask is not None:
# Fix mask dimension to match scores
# mask shape should be [B, L] or [B, 1, L]
if mask.dim() == 2: # [B, L]
# For keys mask [B, k_len] -> [B, 1, 1, k_len]
mask = mask.unsqueeze(1).unsqueeze(2)
elif mask.dim() == 3 and mask.size(1) == 1: # [B, 1, L]
# For keys mask [B, 1, k_len] -> [B, 1, 1, k_len]
mask = mask.unsqueeze(2)
# Expand mask to match scores dimensions
# [B, 1, 1, k_len] -> [B, h, q_len, k_len]
mask = mask.expand(-1, scores.size(1), scores.size(2), -1)
# FIXED: Use -1e4 instead of -1e9 to avoid half-precision overflow
scores = scores.masked_fill(mask == 0, -1e4)
# Apply softmax and dropout
attention_weights = F.softmax(scores, dim=-1)
attention_weights = self.dropout(attention_weights)
# Apply attention to values
context = torch.matmul(attention_weights, v) # [B, h, q_len, d_v]
context = self.merge_heads(context) # [B, q_len, d_model]
attn_output = self.out_proj(context)
# Apply gating mechanism (new addition)
# Concatenate the original query with the attention output to determine the gate
gate_input = torch.cat([q_esm, attn_output], dim=-1)
gate_value = self.gate(gate_input)
# Memory optimization for long sequences
# Processing the gating operation in chunks to prevent OOM errors
if q_len > 1000: # Only use chunking for very long sequences
chunk_size = 500
output_chunks = []
for i in range(0, q_len, chunk_size):
end_idx = min(i + chunk_size, q_len)
# Process chunks
chunk_gate = gate_value[:, i:end_idx, :]
chunk_attn = attn_output[:, i:end_idx, :]
chunk_q = q_esm[:, i:end_idx, :]
# Apply gating equation to this chunk
chunk_output = chunk_gate * chunk_attn + (1 - chunk_gate) * chunk_q
output_chunks.append(chunk_output)
# Combine chunks
output = torch.cat(output_chunks, dim=1)
else:
# Original operation for shorter sequences
output = gate_value * attn_output + (1 - gate_value) * q_esm
return output
class MutationSpecificCrossAttentionBlock(nn.Module):
"""
Cross-attention block with explicit mutation position handling.
Each block processes ESM embeddings and mutation channels separately,
with special emphasis on mutation positions.
"""
def __init__(self, d_model=1152, num_heads=4, ffn_dim=2048, dropout=0.1):
super().__init__()
# Mutation-aware cross attention
self.attn_c12 = MutationSpecificAttention(d_model, num_heads, dropout)
self.attn_c21 = MutationSpecificAttention(d_model, num_heads, dropout)
# Layer normalization for ESM embeddings
self.norm_c1 = nn.LayerNorm(d_model)
self.norm_c2 = nn.LayerNorm(d_model)
# FFN for ESM embeddings
self.ffn_c1 = nn.Sequential(
nn.Linear(d_model, ffn_dim),
nn.ReLU(),
nn.Dropout(dropout),
nn.Linear(ffn_dim, d_model)
)
self.ffn_c2 = nn.Sequential(
nn.Linear(d_model, ffn_dim),
nn.ReLU(),
nn.Dropout(dropout),
nn.Linear(ffn_dim, d_model)
)
self.norm_ffn_c1 = nn.LayerNorm(d_model)
self.norm_ffn_c2 = nn.LayerNorm(d_model)
# Mutation importance update layer
self.mut_update = nn.Sequential(
nn.Linear(d_model + 1, 64),
nn.ReLU(),
nn.Linear(64, 1),
nn.Sigmoid()
)
def forward(self, c1_esm, c1_mut, c2_esm, c2_mut, mask1=None, mask2=None):
"""
Inputs:
c1_esm, c2_esm: ESM embeddings [B, L, 1152]
c1_mut, c2_mut: Mutation channels [B, L, 1]
mask1, mask2: Optional masks
"""
# c1->c2 cross-attention
c1_attn = self.attn_c12(c1_esm, c2_esm, c2_esm, c1_mut, c2_mut, mask2)
c1_out = self.norm_c1(c1_esm + c1_attn)
# c2->c1 cross-attention
c2_attn = self.attn_c21(c2_esm, c1_esm, c1_esm, c2_mut, c1_mut, mask1)
c2_out = self.norm_c2(c2_esm + c2_attn)
# Feed-forward
c1_ffn = self.ffn_c1(c1_out)
c1_ffn_out = self.norm_ffn_c1(c1_out + c1_ffn)
c2_ffn = self.ffn_c2(c2_out)
c2_ffn_out = self.norm_ffn_c2(c2_out + c2_ffn)
# Update mutation importance based on attention output
# This creates a feedback loop where mutation effect is refined
c1_mut_in = torch.cat([c1_ffn_out, c1_mut], dim=-1)
c2_mut_in = torch.cat([c2_ffn_out, c2_mut], dim=-1)
# Stabilized update: convex combination ensures values stay in [0, 1]
# Avoids exponential decay (old bug) and unbounded growth (additive bug)
c1_mut_updated = 0.9 * c1_mut + 0.1 * self.mut_update(c1_mut_in)
c2_mut_updated = 0.9 * c2_mut + 0.1 * self.mut_update(c2_mut_in)
return c1_ffn_out, c2_ffn_out, c1_mut_updated, c2_mut_updated
class MutationSpecificCrossAttentionStack(nn.Module):
"""
Stack of Mutation-Specific Cross-Attention blocks.
Emphasizes mutation positions throughout the network.
Now includes gradient checkpointing for memory efficiency.
"""
def __init__(self, d_model=1152, num_heads=4, ffn_dim=2048, dropout=0.1, num_layers=2):
super().__init__()
self.d_model = d_model
self.num_heads = num_heads
self.use_checkpoint = True # Enable gradient checkpointing by default
self.blocks = nn.ModuleList([
MutationSpecificCrossAttentionBlock(
d_model=d_model,
num_heads=num_heads,
ffn_dim=ffn_dim,
dropout=dropout
) for _ in range(num_layers)
])
def forward(self, c1, c2, mask1=None, mask2=None):
"""
Process protein chains with mutation-specific attention.
c1, c2: [B, L, D] where D can be 1153 (original) or 1154 (with context window)
Uses gradient checkpointing when in training mode to save memory.
"""
# Check input dimension to determine if context window is used
d_in = c1.shape[2]
use_context = (d_in > 1153)
if use_context:
# Split ESM embeddings from mutation+context channels
c1_esm, c1_channels = c1[:, :, :-2], c1[:, :, -2:] # [B, L, 1152], [B, L, 2]
c2_esm, c2_channels = c2[:, :, :-2], c2[:, :, -2:] # [B, L, 1152], [B, L, 2]
# Extract mutation channel (first channel)
c1_mut = c1_channels[:, :, :1] # [B, L, 1]
c2_mut = c2_channels[:, :, :1] # [B, L, 1]
else:
# Original behavior - just split ESM and mutation
c1_esm, c1_mut = c1[:, :, :-1], c1[:, :, -1:] # [B, L, 1152], [B, L, 1]
c2_esm, c2_mut = c2[:, :, :-1], c2[:, :, -1:] # [B, L, 1152], [B, L, 1]
# Process through attention blocks with optional checkpointing
for block in self.blocks:
# Use gradient checkpointing in training mode for memory efficiency
if self.use_checkpoint and self.training:
# Define helper function for checkpointing that handles None masks
def create_checkpoint_fn(block_fn):
def checkpoint_fn(esm1, mut1, esm2, mut2, has_mask1, has_mask2, mask1_val, mask2_val):
# Conditionally use the masks based on the has_mask flags
m1 = mask1_val if has_mask1 else None
m2 = mask2_val if has_mask2 else None
return block_fn(esm1, mut1, esm2, mut2, m1, m2)
return checkpoint_fn
# Convert None masks to flags and dummy tensors for checkpointing
has_mask1 = mask1 is not None
has_mask2 = mask2 is not None
mask1_val = mask1 if has_mask1 else torch.zeros(1, device=c1_esm.device)
mask2_val = mask2 if has_mask2 else torch.zeros(1, device=c1_esm.device)
# Apply checkpointing
c1_esm, c2_esm, c1_mut, c2_mut = torch.utils.checkpoint.checkpoint(
create_checkpoint_fn(block),
c1_esm, c1_mut, c2_esm, c2_mut,
torch.tensor(has_mask1, device=c1_esm.device),
torch.tensor(has_mask2, device=c1_esm.device),
mask1_val, mask2_val
)
else:
c1_esm, c2_esm, c1_mut, c2_mut = block(c1_esm, c1_mut, c2_esm, c2_mut, mask1, mask2)
# Recombine with appropriate channels
if use_context:
# Need to preserve the context channel
context_channels_c1 = c1_channels[:, :, 1:] # [B, L, 1]
context_channels_c2 = c2_channels[:, :, 1:] # [B, L, 1]
c1_out = torch.cat([c1_esm, c1_mut, context_channels_c1], dim=-1) # [B, L, 1154]
c2_out = torch.cat([c2_esm, c2_mut, context_channels_c2], dim=-1) # [B, L, 1154]
else:
# Original behavior
c1_out = torch.cat([c1_esm, c1_mut], dim=-1) # [B, L, 1153]
c2_out = torch.cat([c2_esm, c2_mut], dim=-1) # [B, L, 1153]
return c1_out, c2_out
#############################################################################
# AffinityPredictor with Improved Memory Efficiency #
#############################################################################
class AffinityPredictor(nn.Module):
"""
Enhanced AffinityPredictor with explicit mutation position handling.
embedding_method => "difference", "cosine", "cross_attention", or "cross_attention_swe".
"""
def __init__(
self,
input_dim=1153, # 1152 (ESM) + 1 (mutation)
latent_dim=1024,
num_slices=1024,
num_ref_points=128,
dropout_rate=0.2,
freeze_swe=False,
embedding_method="difference",
normalize_difference=False,
num_hidden_layers=2,
# cross-attn
num_cross_attn_layers=2,
num_attention_heads=4,
cross_ffn_dim=2048,
):
super().__init__()
self.embedding_method = embedding_method.lower()
self.normalize_difference = normalize_difference
self.input_dim = input_dim
# ESM dimension (without mutation channel)
self.esm_dim = input_dim - 1 # 1152
# Define cross-attention stack if needed
self.cross_stack = None
if "cross_attention" in self.embedding_method:
self.cross_stack = MutationSpecificCrossAttentionStack(
d_model=self.esm_dim, # 1152
num_heads=num_attention_heads,
ffn_dim=cross_ffn_dim,
dropout=dropout_rate,
num_layers=num_cross_attn_layers
)
# Enhanced Mutation-Aware SWE Pooling
self.swe_pooling = None
if self.embedding_method in ["difference", "cosine", "cross_attention_swe"]:
# For SWE, we use the full input_dim (1153)
self.swe_pooling = MutationAwareSWEPooling(
d_in=input_dim,
num_slices=num_slices,
num_ref_points=num_ref_points,
freeze_swe=freeze_swe
)
# Define aggregator MLP in-dimensions
if self.embedding_method == "cosine":
in_features = 1
else:
in_features = num_slices # difference or cross_attention_swe => [B, num_slices]
# Add projection layer for cross_attention to avoid dynamic creation
self.cross_attn_projection = None
if self.embedding_method == "cross_attention":
cross_proj_in = input_dim # Full dimension including mutation channel
cross_proj_out = in_features
self.cross_attn_projection = nn.Linear(cross_proj_in, cross_proj_out, bias=False)
# Final MLP
layers = []
current_dim = in_features
for _ in range(num_hidden_layers):
layers.append(nn.Linear(current_dim, latent_dim))
layers.append(nn.ReLU())
layers.append(nn.Dropout(dropout_rate))
current_dim = latent_dim
layers.append(nn.Linear(current_dim, 1))
self.mlp = nn.Sequential(*layers)
def forward(self, chain1, chain1_mask, chain2, chain2_mask):
"""
chain1, chain2 => [B, L, input_dim] (1153 or 1154 with context)
"""
if "cross_attention" in self.embedding_method:
# Process through mutation-specific cross-attention
c1_out, c2_out = self.cross_stack(chain1, chain2, chain1_mask, chain2_mask)
if self.embedding_method == "cross_attention_swe":
# Apply enhanced SWE pooling
rep1 = self.swe_pooling(c1_out, chain1_mask) # [B, num_slices]
rep2 = self.swe_pooling(c2_out, chain2_mask) # [B, num_slices]
# Difference aggregator
diff = rep1 - rep2
if self.normalize_difference:
diff = F.normalize(diff, p=2, dim=1)
# Final prediction
preds = self.mlp(diff).squeeze(-1)
return preds
elif self.embedding_method == "cross_attention":
# Use mutation-weighted pooling
# Extract mutation channel to guide pooling
d_in = c1_out.shape[2]
use_context = (d_in > 1153)
if use_context:
c1_mut = c1_out[:, :, -2:-1] # [B, L, 1]
c2_mut = c2_out[:, :, -2:-1] # [B, L, 1]
else:
c1_mut = c1_out[:, :, -1:] # [B, L, 1]
c2_mut = c2_out[:, :, -1:] # [B, L, 1]
# Weighted pooling - gives higher weight to mutated positions
c1_weights = F.softmax(c1_mut * 10, dim=1) # Sharpen weights
c2_weights = F.softmax(c2_mut * 10, dim=1)
c1_pool = torch.sum(c1_out * c1_weights, dim=1) # [B, 1153/1154]
c2_pool = torch.sum(c2_out * c2_weights, dim=1) # [B, 1153/1154]
# Create difference representation
diff = c1_pool - c2_pool
if self.normalize_difference:
diff = F.normalize(diff, p=2, dim=1)
# Use pre-defined projection layer instead of creating one dynamically
if self.cross_attn_projection is not None:
diff = self.cross_attn_projection(diff)
preds = self.mlp(diff).squeeze(-1)
return preds
elif self.embedding_method == "cosine":
# Enhanced SWE => [B, num_slices]
rep1 = self.swe_pooling(chain1, chain1_mask)
rep2 = self.swe_pooling(chain2, chain2_mask)
sim = F.cosine_similarity(rep1, rep2, dim=1).unsqueeze(-1)
out = self.mlp(sim).squeeze(-1)
return out
else: # "difference"
rep1 = self.swe_pooling(chain1, chain1_mask)
rep2 = self.swe_pooling(chain2, chain2_mask)
diff = rep1 - rep2
if self.normalize_difference:
diff = F.normalize(diff, p=2, dim=1)
out = self.mlp(diff).squeeze(-1)
return out
class AffinityPredictionModel(pl.LightningModule):
"""
Lightning wrapper for training. Siamese logic in training.py
"""
def __init__(self, predictor: AffinityPredictor, learning_rate=1e-4):
super().__init__()
self.predictor = predictor
self.learning_rate = learning_rate
self.loss_fn = nn.MSELoss()
self.pearson_corr = PearsonCorrCoef()
self.spearman_corr = SpearmanCorrCoef()
self.r2_score = R2Score()
self.mse_metric = MeanSquaredError()
def forward(self, chain1, chain1_mask, chain2, chain2_mask):
return self.predictor(chain1, chain1_mask, chain2, chain2_mask)
def training_step(self, batch, batch_idx):
pass
def validation_step(self, batch, batch_idx):
pass
def configure_optimizers(self):
optimizer = AdamW(self.parameters(), lr=self.learning_rate)
steps = self.trainer.estimated_stepping_batches
scheduler = torch.optim.lr_scheduler.OneCycleLR(
optimizer, max_lr=self.learning_rate, total_steps=steps
)
return {"optimizer": optimizer, "lr_scheduler": scheduler, "monitor": "val_loss"}
# Enhanced AffinityPredictor with Two-Head Architecture
# Add this to architectures.py
class DualHeadAffinityPredictor(nn.Module):
"""
Enhanced AffinityPredictor with explicit two-head architecture.
Simultaneously processes mutant and wildtype proteins to predict both ΔG and ΔΔG.
embedding_method => "difference", "cosine", "cross_attention", or "cross_attention_swe".
"""
def __init__(
self,
input_dim=1153, # 1152 (ESM) + 1 (mutation)
latent_dim=1024,
num_slices=1024,
num_ref_points=128,
dropout_rate=0.2,
freeze_swe=False,
embedding_method="difference",
normalize_difference=False,
num_hidden_layers=2,
# cross-attn
num_cross_attn_layers=2,
num_attention_heads=4,
cross_ffn_dim=2048,
use_dual_head=True, # Enable dual-head by default
ddg_signal_gain=1.0, # Initial gain for ddG signal
ddg_signal_multiplier=20.0, # FIXED: multiplier for ddG signal (vnew65.0)
):
super().__init__()
self.embedding_method = embedding_method.lower()
self.normalize_difference = normalize_difference
self.input_dim = input_dim
self.use_dual_head = use_dual_head
self._ddg_log_counter = 0
# DEBUG: Confirm this version is running
print(f"[MODEL INIT] DualHeadAffinityPredictor created: version={ARCH_VERSION}, dual_head={use_dual_head}, method={self.embedding_method}")
# ESM dimension (without mutation channel)
self.esm_dim = input_dim - 1 # 1152
# Define cross-attention stack if needed
self.cross_stack = None
if "cross_attention" in self.embedding_method:
self.cross_stack = MutationSpecificCrossAttentionStack(
d_model=self.esm_dim, # 1152
num_heads=num_attention_heads,
ffn_dim=cross_ffn_dim,
dropout=dropout_rate,
num_layers=num_cross_attn_layers
)
# Enhanced Mutation-Aware SWE Pooling
self.swe_pooling = None
if self.embedding_method in ["difference", "cosine", "cross_attention_swe"]:
# For SWE, we use the full input_dim (1153)
self.swe_pooling = MutationAwareSWEPooling(
d_in=input_dim,
num_slices=num_slices,
num_ref_points=num_ref_points,
freeze_swe=freeze_swe
)
# Define aggregator MLP in-dimensions
if self.embedding_method == "cosine":
in_features = 1
else:
in_features = num_slices # difference or cross_attention_swe => [B, num_slices]
# Add projection layer for cross_attention to avoid dynamic creation
self.cross_attn_projection = None
if self.embedding_method == "cross_attention":
cross_proj_in = input_dim # Full dimension including mutation channel
cross_proj_out = in_features
self.cross_attn_projection = nn.Linear(cross_proj_in, cross_proj_out, bias=False)
# Define dG head (main prediction head)
layers = []
current_dim = in_features
for _ in range(num_hidden_layers):
layers.append(nn.Linear(current_dim, latent_dim))
layers.append(nn.ReLU())
layers.append(nn.Dropout(dropout_rate))
current_dim = latent_dim
layers.append(nn.Linear(current_dim, 1))
self.dg_mlp = nn.Sequential(*layers)
# Define ΔΔG head for direct prediction
# FIX (vnew64.0): Shallow 2-layer MLP + residual skip connection.
# DDGACT diagnostics from vnew62-63 showed 7-layer ReLU network causes
# progressive variance collapse (std: 0.1 → 7e-05). Each ReLU zeros ~50%
# of activations, so 7 layers → 0.5^7 = 0.8% signal survival.
# Solution: (1) 2 layers only, (2) skip connection preserves raw input signal.
if self.use_dual_head:
# Shallow nonlinear pathway (2 layers)
self.ddg_hidden = nn.Sequential(
nn.Linear(in_features, latent_dim),
nn.GELU(), # GELU instead of ReLU - no zero-capping, smoother gradients
nn.Linear(latent_dim, latent_dim),
nn.GELU(),
)
# Output projection
self.ddg_out = nn.Linear(latent_dim, 1)
# Skip connection: project input directly to output dimension
self.ddg_skip = nn.Linear(in_features, 1)
# Source type embedding for conditional inference
# User-friendly types for inference:
# 0 = "mutant" - Single mutant predictions (most common)
# 1 = "wt_pairs" - Wildtype pairs with absolute binding affinity
# 2 = "antibody" - Antibody-antigen binding (CDR-focused)
self.source_type_embedding = nn.Embedding(3, 32) # 32-dim embedding
self.source_type_projection = nn.Linear(in_features + 32, in_features) # Project back to in_features
# FIXED: Restoring the historical 'learnable gain' strategy
# This allows the model to amplify the ddG signal early in Stage B
self.ddg_signal_gain = nn.Parameter(torch.tensor(float(ddg_signal_gain)))
self.ddg_signal_multiplier = float(ddg_signal_multiplier)
def _extract_features(self, chain1, chain1_mask, chain2, chain2_mask):
"""
Extract feature representation for a protein complex.
Returns a vector representation suitable for prediction.
"""
if "cross_attention" in self.embedding_method:
# Process through mutation-specific cross-attention
c1_out, c2_out = self.cross_stack(chain1, chain2, chain1_mask, chain2_mask)
if self.embedding_method == "cross_attention_swe":
# Apply enhanced SWE pooling
rep1 = self.swe_pooling(c1_out, chain1_mask) # [B, num_slices]
rep2 = self.swe_pooling(c2_out, chain2_mask) # [B, num_slices]
# Difference aggregator
diff = rep1 - rep2
if self.normalize_difference:
diff = F.normalize(diff, p=2, dim=1)
return diff
elif self.embedding_method == "cross_attention":
# Use mutation-weighted pooling
# Extract mutation channel to guide pooling
d_in = c1_out.shape[2]
use_context = (d_in > 1153)
if use_context:
c1_mut = c1_out[:, :, -2:-1] # [B, L, 1]
c2_mut = c2_out[:, :, -2:-1] # [B, L, 1]
else:
c1_mut = c1_out[:, :, -1:] # [B, L, 1]
c2_mut = c2_out[:, :, -1:] # [B, L, 1]
# Weighted pooling - gives higher weight to mutated positions
c1_weights = F.softmax(c1_mut * 10, dim=1) # Sharpen weights
c2_weights = F.softmax(c2_mut * 10, dim=1)
c1_pool = torch.sum(c1_out * c1_weights, dim=1) # [B, 1153/1154]
c2_pool = torch.sum(c2_out * c2_weights, dim=1) # [B, 1153/1154]
# Create difference representation
diff = c1_pool - c2_pool
if self.normalize_difference:
diff = F.normalize(diff, p=2, dim=1)
# Use pre-defined projection layer instead of creating one dynamically
if self.cross_attn_projection is not None:
diff = self.cross_attn_projection(diff)
return diff
elif self.embedding_method == "cosine":
# Enhanced SWE => [B, num_slices]
rep1 = self.swe_pooling(chain1, chain1_mask)
rep2 = self.swe_pooling(chain2, chain2_mask)
sim = F.cosine_similarity(rep1, rep2, dim=1).unsqueeze(-1)
return sim
else: # "difference"
rep1 = self.swe_pooling(chain1, chain1_mask)
rep2 = self.swe_pooling(chain2, chain2_mask)
diff = rep1 - rep2
if self.normalize_difference:
diff = F.normalize(diff, p=2, dim=1)
return diff
def _extract_residue_features(self, chain1, chain1_mask, chain2, chain2_mask):
"""
Extract RESIDUE-LEVEL features (before pooling) for computing differences.
Used for ddG to preserve mutation-specific information.
Returns:
c1_out, c2_out: [B, L1, D] and [B, L2, D] attended residue features
"""
if "cross_attention" in self.embedding_method:
c1_out, c2_out = self.cross_stack(chain1, chain2, chain1_mask, chain2_mask)
return c1_out, c2_out
else:
# For non-cross-attention methods, return inputs directly
return chain1, chain2
def forward(self, mut_chain1, mut_chain1_mask, mut_chain2, mut_chain2_mask,
wt_chain1=None, wt_chain1_mask=None, wt_chain2=None, wt_chain2_mask=None,
source_type_ids=None):
"""
Dual-head forward method that can handle both modes:
1. Standard mode: Just predict dG for mutant complex
2. Dual-head mode: Predict both dG and direct ddG when wildtype is provided
For ddG: Uses RESIDUE-LEVEL differences before pooling to preserve mutation info.
ddG = ddg_mlp(pool(mut_features - wt_features)) instead of
ddg_mlp(pool(mut_features) - pool(wt_features))
Args:
source_type_ids: Optional[Tensor] of shape [B], values 0/1/2 for conditioning
Returns:
If wildtype inputs are None or use_dual_head=False:
Returns mutant dG prediction only
Else:
Returns tuple of (mutant_dG, direct_ddG_prediction)
"""
# ============== OPTIMIZED: Cache residue features ==============
# Get mutant RESIDUE-LEVEL features first (used for both dG and ddG)
mut_c1_res, mut_c2_res = self._extract_residue_features(
mut_chain1, mut_chain1_mask, mut_chain2, mut_chain2_mask)
# Pool for dG prediction (reuses cached residue features)
if "cross_attention_swe" in self.embedding_method:
rep1 = self.swe_pooling(mut_c1_res, mut_chain1_mask)
rep2 = self.swe_pooling(mut_c2_res, mut_chain2_mask)
mut_features = rep1 - rep2
if self.normalize_difference:
mut_features = F.normalize(mut_features, p=2, dim=1, eps=1e-8)
else:
# Fallback pooling
mut_features = self._extract_features(mut_chain1, mut_chain1_mask, mut_chain2, mut_chain2_mask)
# ============== SOURCE TYPE CONDITIONING ==============
# Apply source type conditioning if provided
if source_type_ids is not None:
# Get source type embedding [B, 32]
src_emb = self.source_type_embedding(source_type_ids)
# Concatenate with features and project back
conditioned_features = torch.cat([mut_features, src_emb], dim=-1)
mut_features = self.source_type_projection(conditioned_features)
# Predict dG for mutant
dg_pred = self.dg_mlp(mut_features).squeeze(-1)
# If no wildtype or dual head is disabled, just return mutant dG
if not self.use_dual_head or wt_chain1 is None or wt_chain2 is None:
return dg_pred
# ============== RESIDUE-LEVEL ddG COMPUTATION ==============
# Get wildtype RESIDUE-LEVEL features (mutant already cached above)
wt_c1_res, wt_c2_res = self._extract_residue_features(
wt_chain1, wt_chain1_mask, wt_chain2, wt_chain2_mask)
# (Debug logging removed - was polluting training output)
# Compute RESIDUE-LEVEL differences BEFORE pooling
# This preserves mutation-specific changes at each position
# Handle sequence length differences by taking minimum length
L_c1 = min(mut_c1_res.shape[1], wt_c1_res.shape[1])
L_c2 = min(mut_c2_res.shape[1], wt_c2_res.shape[1])
c1_diff = mut_c1_res[:, :L_c1, :] - wt_c1_res[:, :L_c1, :] # [B, L1, D]
c2_diff = mut_c2_res[:, :L_c2, :] - wt_c2_res[:, :L_c2, :] # [B, L2, D]
# Update masks for truncated length
c1_diff_mask = mut_chain1_mask[:, :L_c1] if mut_chain1_mask is not None else None
c2_diff_mask = mut_chain2_mask[:, :L_c2] if mut_chain2_mask is not None else None
# [DIFF CHECK] Diagnostic Logging
# Measure cosine similarity at the approximate mutation site to check for embedding collapse.
if self._ddg_log_counter % 200 == 1:
with torch.no_grad():
# Extract center position
mid_idx = L_c1 // 2
# Vectors at midpoint
v_mut = mut_c1_res[:, mid_idx, :]
v_wt = wt_c1_res[:, mid_idx, :]
# Cosine similarity
cos_sim = F.cosine_similarity(v_mut, v_wt, dim=1).mean().item()
diff_norm = (v_mut - v_wt).norm(dim=1).mean().item()
logger.info(f"[DIFF CHECK] Batch {self._ddg_log_counter}: CosSim at mid={cos_sim:.5f}, DiffNorm={diff_norm:.5f}")
# NOW pool the differences
if self.swe_pooling is not None:
# =================================================================
# HYBRID POOLING: Global SWE + Local Mutation-Site-Centric (vnew37.0)
# This ensures local mutation signals are not diluted by global pooling
# =================================================================
# Concatenate chain differences along sequence dimension
combined_diff = torch.cat([c1_diff, c2_diff], dim=1) # [B, L_comb, D=1153]
if c1_diff_mask is not None and c2_diff_mask is not None:
combined_mask = torch.cat([c1_diff_mask, c2_diff_mask], dim=1)
else:
combined_mask = None
# A. Global Component: Standard SWE pooling (capture global stability context)
global_diff = self.swe_pooling(combined_diff, combined_mask) # [B, num_slices]
# B. Local Component: Mutation-Site-Centric Pooling (MSCP)
# CRITICAL FIX (v49.0): Extract indicator from RAW INPUT chains, NOT cross-attention output!
# The cross-attention stack applies 0.9 convex combination at each layer (5 layers = 0.9^5 = 59% decay).
# Using mut_chain1/mut_chain2 (raw inputs) instead of mut_c1_res/mut_c2_res (diluted outputs).
# Determine if we have context window (1154-dim) or standard (1153-dim)
d_raw = mut_chain1.shape[2]
use_context = (d_raw > 1153)
# Extract RAW indicator from INPUT chains (before cross-attention!)
if use_context:
c1_mut_raw = mut_chain1[:, :L_c1, -2:-1]
c2_mut_raw = mut_chain2[:, :L_c2, -2:-1]
else:
c1_mut_raw = mut_chain1[:, :L_c1, -1:]
c2_mut_raw = mut_chain2[:, :L_c2, -1:]
mut_indicator = torch.cat([c1_mut_raw, c2_mut_raw], dim=1) # [B, L_comb, 1]
# Stability: Clamp indicator to be non-negative and bounded.
# In case attention stack produced weird values, we force them back to [0, 2]
mut_indicator = mut_indicator.clamp(min=0.0, max=2.0)
# Weighted average focusing ONLY on the mutation sites
# use 0.001 epsilon to avoid nan for WT samples (mut_sum=0)
mut_sum = mut_indicator.sum(dim=1).clamp(min=1e-3)
# MSCP Calculation with additional stability guard
mscp_esm = (combined_diff[:, :, :1152] * mut_indicator).sum(dim=1) / mut_sum # [B, 1152]
mscp_mut = (mut_indicator * mut_indicator).sum(dim=1) / mut_sum # [B, 1] (should be ~1.0)
# Project local delta into the same slice space as global features
# using the SHARED theta projection and mut_projection from SWE
# This ensures consistent representation between global and local paths
local_diff_esm = self.swe_pooling.theta(mscp_esm)
local_diff_mut = self.swe_pooling.mut_projection(mscp_mut)
local_diff = local_diff_esm + local_diff_mut # [B, num_slices]
# C. Combine: Residual-style addition + Gain
# FIXED: Apply signal_gain AFTER normalization so it actually has effect
# Previously, L2-norm undid the scaling entirely.
diff_multiplier = getattr(self, "ddg_signal_multiplier", 20.0)
diff_features = (global_diff + local_diff) * diff_multiplier
# ===================================================================
# SIGNAL FLOW LOGGING: Track local vs global contribution
if not hasattr(self, '_ddg_log_counter'):
self._ddg_log_counter = 0
self._ddg_log_counter += 1
should_log = (self._ddg_log_counter % 200 == 1)
if should_log:
g_mag = global_diff.abs().mean().item()
l_mag = local_diff.abs().mean().item()
logger.info(f"[DDG SIGNAL] Batch {self._ddg_log_counter}: Global_mag={g_mag:.4f}, Local_mag={l_mag:.4f}")
#region agent log
try:
# Inspect what the model thinks the mutation indicator is (both tail channels)
d_raw_dbg = int(mut_chain1.shape[2])
# indicator candidate stats on chain1/chain2 for last and second-last channels
def _chan_stats(x):
return {
"min": float(x.min().item()),
"max": float(x.max().item()),
"mean": float(x.float().mean().item()),
"std": float(x.float().std().item()),
}
c1_last = _chan_stats(mut_chain1[:, :L_c1, -1])
c2_last = _chan_stats(mut_chain2[:, :L_c2, -1])
c1_last2 = _chan_stats(mut_chain1[:, :L_c1, -2]) if d_raw_dbg >= 1154 else None
c2_last2 = _chan_stats(mut_chain2[:, :L_c2, -2]) if d_raw_dbg >= 1154 else None
mut_sum_dbg = float(mut_sum.mean().item()) if "mut_sum" in locals() else None
payload = {
"sessionId": "debug-session",
"runId": "pre-fix",
"hypothesisId": "F",
"location": "architectures.py:DualHeadAffinityPredictor:mscp_indicator_debug",
"message": "Indicator channel stats (last vs second-last) to detect double-indicator / wrong channel selection",
"data": {
"ddg_log_counter": int(self._ddg_log_counter),
"d_raw": d_raw_dbg,
"use_context_flag": bool(use_context),
"c1_last": c1_last,
"c2_last": c2_last,
"c1_last2": c1_last2,
"c2_last2": c2_last2,
"mut_sum_mean": mut_sum_dbg,
},
"timestamp": int(time.time() * 1000),
}
with open("/Users/supantha/Documents/code_v2/protein/.cursor/debug.log", "a") as f:
f.write(json.dumps(payload, default=str) + "\n")
logger.info(f"[AGENTLOG MSCP] d_raw={d_raw_dbg} use_context={use_context} c1_last={c1_last} c1_last2={c1_last2} mut_sum_mean={mut_sum_dbg}")
except Exception:
pass
#endregion
if self.normalize_difference:
diff_features = F.normalize(diff_features, p=2, dim=1, eps=1e-8)
# FIXED: Apply signal_gain AFTER normalization so it actually scales the output
diff_features = diff_features * self.ddg_signal_gain
else:
# Fallback: mean pooling of differences
if c1_diff_mask is not None:
c1_diff = c1_diff * c1_diff_mask.unsqueeze(-1).float()
c1_pool = c1_diff.sum(dim=1) / c1_diff_mask.sum(dim=1, keepdim=True).clamp(min=1)
else:
c1_pool = c1_diff.mean(dim=1)
if c2_diff_mask is not None:
c2_diff = c2_diff * c2_diff_mask.unsqueeze(-1).float()
c2_pool = c2_diff.sum(dim=1) / c2_diff_mask.sum(dim=1, keepdim=True).clamp(min=1)
else:
c2_pool = c2_diff.mean(dim=1)
diff_features = c1_pool - c2_pool
if self.normalize_difference:
diff_features = F.normalize(diff_features, p=2, dim=1)
# DEBUG: Check for NaNs/Infs in diff_features
if torch.isnan(diff_features).any() or torch.isinf(diff_features).any():
print(f"[DEBUG MODEL] NaN/Inf in diff_features! Shape: {diff_features.shape}")
if c1_diff_mask is not None:
print(f" c1_mask sum: {c1_diff_mask.sum(dim=1).min().item()}")
print(f" c1_diff nan: {torch.isnan(c1_diff).any().item()}")
print(f" c2_diff nan: {torch.isnan(c2_diff).any().item()}")
# ============== SOURCE TYPE CONDITIONING FOR DDG ==============
# Apply same source conditioning to diff_features for ddG prediction
# This allows ddG head to learn different behaviors for different data sources
if source_type_ids is not None:
src_emb = self.source_type_embedding(source_type_ids)
conditioned_diff = torch.cat([diff_features, src_emb], dim=-1)
diff_features = self.source_type_projection(conditioned_diff)
# MSCP Hybrid: Skip 20x gain since MSCP provides raw, un-diluted signal
# vnew64.0: Shallow 2-layer GELU + skip connection to preserve variance
ddg_hidden_out = self.ddg_hidden(diff_features)
ddg_pred = (self.ddg_out(ddg_hidden_out) + self.ddg_skip(diff_features)).squeeze(-1)
#region agent log
# Diagnose "train variance but eval constant" which often indicates dropout-only variance or head collapse.
try:
if should_log:
# Diff feature stats across the batch
df = diff_features.detach()
df_mean = float(df.mean().item()) if df.numel() else None
df_std = float(df.std().item()) if df.numel() else None
df_abs_mean = float(df.abs().mean().item()) if df.numel() else None
# per-sample spread: average std over features (helps detect "all samples identical")
df_per_sample_std = float(df.float().std(dim=1).mean().item()) if df.dim() == 2 and df.shape[0] > 0 else None
# ddg_pred stats (across batch)
p = ddg_pred.detach()
p_mean = float(p.mean().item()) if p.numel() else None
p_std = float(p.std().item()) if p.numel() else None
# If dropout is active (training), a second forward pass should differ.
p2_std = None
p_diff_std = None
if self.training:
p2 = (self.ddg_out(self.ddg_hidden(diff_features)) + self.ddg_skip(diff_features)).squeeze(-1).detach()
p2_std = float(p2.std().item()) if p2.numel() else None
p_diff_std = float((p2 - p).std().item()) if p2.numel() else None
# Weight/bias norms (to detect collapse to near-zero weights or bias-only prediction)
lin_layers = [m for m in self.ddg_hidden.modules() if isinstance(m, nn.Linear)] if hasattr(self, "ddg_hidden") else []
w0 = lin_layers[0] if len(lin_layers) > 0 else None
wL = lin_layers[-1] if len(lin_layers) > 0 else None
w0_norm = float(w0.weight.detach().norm().item()) if w0 is not None else None
wL_norm = float(wL.weight.detach().norm().item()) if wL is not None else None
bL_norm = float(wL.bias.detach().norm().item()) if (wL is not None and wL.bias is not None) else None
payload = {
"sessionId": "debug-session",
"runId": "pre-fix",
"hypothesisId": "I",
"location": "architectures.py:DualHeadAffinityPredictor:ddg_head_eval_vs_train",
"message": "ddG head collapse vs dropout-only variance diagnostics",
"data": {
"ddg_log_counter": int(self._ddg_log_counter),
"model_training": bool(self.training),
"normalize_difference": bool(getattr(self, "normalize_difference", False)),
"ddg_signal_gain": float(self.ddg_signal_gain.detach().item()) if hasattr(self, "ddg_signal_gain") else None,
"diff_features_mean": df_mean,
"diff_features_std": df_std,
"diff_features_abs_mean": df_abs_mean,
"diff_features_per_sample_std_mean": df_per_sample_std,
"ddg_pred_mean": p_mean,
"ddg_pred_std": p_std,
"ddg_pred2_std": p2_std,
"ddg_pred_repeat_diff_std": p_diff_std,
"ddg_w0_norm": w0_norm,
"ddg_wL_norm": wL_norm,
"ddg_bL_norm": bL_norm,
},
"timestamp": int(time.time() * 1000),
}
# Log first (before file write that may fail on cluster)
logger.info(
f"[AGENTLOG DDGHEAD] train={self.training} df_std={df_std:.4f} df_ps_std={df_per_sample_std:.4f} "
f"pred_std={p_std:.4f} pred_mean={p_mean:.4f} rep_diff_std={p_diff_std if p_diff_std is not None else 'NA'} "
f"w0={w0_norm:.2f} wL={wL_norm:.2f} bL={bL_norm if bL_norm is not None else 'NA'}"
)
# Layerwise activation trace to locate where variance collapses (ReLU dead / dropout-only variance)
layer_stats = []
try:
with torch.no_grad():
x = diff_features.detach()
for li, layer in enumerate(self.ddg_hidden):
x = layer(x)
st = {
"i": int(li),
"t": layer.__class__.__name__,
"mean": float(x.mean().item()) if x.numel() else None,
"std": float(x.std().item()) if x.numel() else None,
}
if isinstance(layer, nn.ReLU):
st["zero_frac"] = float((x == 0).float().mean().item()) if x.numel() else None
layer_stats.append(st)
# Compact summary for logs (first 2 + last 2 layers)
compact = (layer_stats[:2] + (["..."] if len(layer_stats) > 4 else []) + layer_stats[-2:])
logger.info(f"[AGENTLOG DDGACT] train={self.training} layers={compact}")
except Exception:
layer_stats = []
# File write may fail on cluster - that's OK
try:
payload["data"]["ddg_layer_stats"] = layer_stats
with open("/Users/supantha/Documents/code_v2/protein/.cursor/debug.log", "a") as f:
f.write(json.dumps(payload, default=str) + "\n")
except Exception:
pass
except Exception:
pass
#endregion
# DEBUG: Check ddg_pred
if torch.isnan(ddg_pred).any() or torch.isinf(ddg_pred).any():
print(f"[DEBUG MODEL] NaN/Inf in ddg_pred!")
return dg_pred, ddg_pred
class DualHeadAffinityPredictionModel(pl.LightningModule):
"""
Lightning wrapper for dual-head training.
"""
def __init__(self, predictor: DualHeadAffinityPredictor, learning_rate=1e-4, ddg_loss_weight=1.0):
super().__init__()
self.predictor = predictor
self.learning_rate = learning_rate
self.ddg_loss_weight = ddg_loss_weight
self.loss_fn = nn.MSELoss()
self.pearson_corr = PearsonCorrCoef()
self.spearman_corr = SpearmanCorrCoef()
self.r2_score = R2Score()
self.mse_metric = MeanSquaredError()
# Save hyperparameters for checkpointing
self.save_hyperparameters(ignore=['predictor'])
def forward(self, mut_chain1, mut_chain1_mask, mut_chain2, mut_chain2_mask,
wt_chain1=None, wt_chain1_mask=None, wt_chain2=None, wt_chain2_mask=None,
source_type_ids=None):
return self.predictor(mut_chain1, mut_chain1_mask, mut_chain2, mut_chain2_mask,
wt_chain1, wt_chain1_mask, wt_chain2, wt_chain2_mask,
source_type_ids=source_type_ids)
def training_step(self, batch, batch_idx):
# Mutant data
(c1, m1, c2, m2, y_mut) = batch["mutant"]
# Wildtype data with valid mask
(cw1, w1m, cw2, w2m, y_wt) = batch["wildtype"]
has_wt = batch["has_wt"]
if self.predictor.use_dual_head and has_wt.sum() > 0:
# For samples with wildtype available, use dual-head prediction
valid_samples = has_wt.bool()
# Get predictions for valid samples
dg_pred, ddg_pred = self(
c1[valid_samples], m1[valid_samples],
c2[valid_samples], m2[valid_samples],
cw1[valid_samples], w1m[valid_samples],
cw2[valid_samples], w2m[valid_samples]
)
# Calculate losses for valid samples
dg_loss = self.loss_fn(dg_pred, y_mut[valid_samples])
# Calculate true ddG as difference between mutant and wildtype dG
true_ddg = y_mut[valid_samples] - y_wt[valid_samples]
ddg_loss = self.loss_fn(ddg_pred, true_ddg)
# Combined loss with weighting
loss = dg_loss + self.ddg_loss_weight * ddg_loss
# Process remaining samples (without wildtype) with standard prediction
if (~valid_samples).sum() > 0:
standard_dg_pred = self(
c1[~valid_samples], m1[~valid_samples],
c2[~valid_samples], m2[~valid_samples]
)
standard_loss = self.loss_fn(standard_dg_pred, y_mut[~valid_samples])
# Add to total loss, weighted by proportion of samples
n_valid = valid_samples.sum()
n_total = len(valid_samples)
loss = (n_valid / n_total) * loss + ((n_total - n_valid) / n_total) * standard_loss
else:
# Standard prediction for all samples
dg_pred = self(c1, m1, c2, m2)
loss = self.loss_fn(dg_pred, y_mut)
# Log metrics
self.log('train_loss', loss, on_step=True, on_epoch=True, prog_bar=True)
return loss
def validation_step(self, batch, batch_idx):
# Similar to training_step but with more comprehensive metrics
(c1, m1, c2, m2, y_mut) = batch["mutant"]
(cw1, w1m, cw2, w2m, y_wt) = batch["wildtype"]
has_wt = batch["has_wt"]
# Store predictions and targets for all samples
all_dg_preds = []
all_dg_targets = []
all_ddg_preds = []
all_ddg_targets = []
if self.predictor.use_dual_head and has_wt.sum() > 0:
valid_samples = has_wt.bool()
# Dual-head prediction for samples with wildtype
dg_pred, ddg_pred = self(
c1[valid_samples], m1[valid_samples],
c2[valid_samples], m2[valid_samples],
cw1[valid_samples], w1m[valid_samples],
cw2[valid_samples], w2m[valid_samples]
)
# Calculate true ddG
true_ddg = y_mut[valid_samples] - y_wt[valid_samples]
# Store predictions and targets
all_dg_preds.append(dg_pred)
all_dg_targets.append(y_mut[valid_samples])
all_ddg_preds.append(ddg_pred)
all_ddg_targets.append(true_ddg)
# Process remaining samples with standard prediction
if (~valid_samples).sum() > 0:
standard_dg_pred = self(
c1[~valid_samples], m1[~valid_samples],
c2[~valid_samples], m2[~valid_samples]
)
all_dg_preds.append(standard_dg_pred)
all_dg_targets.append(y_mut[~valid_samples])
else:
# Standard prediction for all samples
dg_pred = self(c1, m1, c2, m2)
all_dg_preds.append(dg_pred)
all_dg_targets.append(y_mut)
# For samples with wildtype, calculate implicit ddG
if has_wt.sum() > 0:
valid_samples = has_wt.bool()
wt_dg_pred = self(cw1[valid_samples], w1m[valid_samples],
cw2[valid_samples], w2m[valid_samples])
implicit_ddg_pred = dg_pred[valid_samples] - wt_dg_pred
true_ddg = y_mut[valid_samples] - y_wt[valid_samples]
all_ddg_preds.append(implicit_ddg_pred)
all_ddg_targets.append(true_ddg)
# Concatenate all predictions and targets
if all_dg_preds:
all_dg_preds = torch.cat(all_dg_preds)
all_dg_targets = torch.cat(all_dg_targets)
# Calculate dG metrics
dg_mse = self.mse_metric(all_dg_preds, all_dg_targets)
dg_pearson = self.pearson_corr(all_dg_preds, all_dg_targets)
dg_spearman = self.spearman_corr(all_dg_preds, all_dg_targets)
dg_r2 = self.r2_score(all_dg_preds, all_dg_targets)
# Log dG metrics
self.log('val_dg_mse', dg_mse, on_epoch=True, prog_bar=True)
self.log('val_dg_pearson', dg_pearson, on_epoch=True)
self.log('val_dg_spearman', dg_spearman, on_epoch=True)
self.log('val_dg_r2', dg_r2, on_epoch=True)
# Calculate ddG metrics if available
if all_ddg_preds:
all_ddg_preds = torch.cat(all_ddg_preds)
all_ddg_targets = torch.cat(all_ddg_targets)
ddg_mse = self.mse_metric(all_ddg_preds, all_ddg_targets)
ddg_pearson = self.pearson_corr(all_ddg_preds, all_ddg_targets)
ddg_spearman = self.spearman_corr(all_ddg_preds, all_ddg_targets)
ddg_r2 = self.r2_score(all_ddg_preds, all_ddg_targets)
# Log ddG metrics
self.log('val_ddg_mse', ddg_mse, on_epoch=True, prog_bar=True)
self.log('val_ddg_pearson', ddg_pearson, on_epoch=True)
self.log('val_ddg_spearman', ddg_spearman, on_epoch=True)
self.log('val_ddg_r2', ddg_r2, on_epoch=True)
# Combined validation metric for early stopping
combined_metric = dg_mse + self.ddg_loss_weight * ddg_mse
self.log('val_combined_metric', combined_metric, on_epoch=True)
return {'val_dg_mse': dg_mse if 'dg_mse' in locals() else None,
'val_ddg_mse': ddg_mse if 'ddg_mse' in locals() else None}
def test_step(self, batch, batch_idx):
# Similar to validation_step but returns more detailed metrics
return self.validation_step(batch, batch_idx)
def configure_optimizers(self):
optimizer = torch.optim.AdamW(self.parameters(), lr=self.learning_rate)
scheduler = torch.optim.lr_scheduler.OneCycleLR(
optimizer, max_lr=self.learning_rate, total_steps=self.trainer.estimated_stepping_batches
)
return {"optimizer": optimizer, "lr_scheduler": scheduler, "monitor": "val_combined_metric"} |