Press n or j to go to the next uncovered block, b, p or k for the previous block.
| 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 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 | 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 62x 62x 62x 62x 62x 62x 62x 62x 62x 8x 54x 62x 62x 62x 62x 21x 41x 2x 39x 3x 36x 62x 1x 1x 1x 1x 1x 1x 62x 1x 1x 1x 1x 1x 62x 3x 3x 3x 3x 3x 62x 62x 62x 60x 60x 60x 60x 62x 62x 62x 62x 62x 62x 62x 62x 62x 62x 62x 62x 60x 1347x 1347x 1325x 1325x 1325x 1325x 1325x 1325x 1325x 1325x 1325x 1347x 1347x 1347x 1347x 1347x 1347x 1347x 1347x 1347x 11x 11x 11x 11x 11x 11x 11x 11x 11x 11x 11x 5x 60x 62x 62x 62x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 62x 62x 62x 62x 150x 150x 150x 150x 1336x 1336x 1336x 1336x 1336x 1336x 1336x 11x 11x 11x 1325x 1336x 41x 41x 44x 44x 44x 44x 44x 9x 9x 9x 44x 41x 1325x 1325x 1325x 1325x 1325x 1325x 1325x 1336x 150x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 3x 3x 2x 2x 2x 3x 1x 1x 60x 8x 8x 8x 60x 36x 3x 33x 4x 29x 24x 60x 8x 8x 52x 60x 8x 8x 2x 6x 52x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 8x 8x 8x 8x 8x 8x 8x 52x 52x 52x 52x 52x 52x 60x 60x 60x 60x 60x 24x 24x 24x 24x 24x 24x 24x 24x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 60x 1325x 1325x 1325x 60x 60x 60x 60x 60x 60x 60x 60x 60x 16x 16x 16x 16x 60x 1x 1x 1x 1x 60x 60x 41x 41x 88x 88x 88x 88x 88x 88x 88x 88x 88x 88x 88x 440x 440x 440x 88x 60x 60x 60x 60x 60x 60x 14x 14x 15x 15x 15x 15x 15x 15x 15x 15x 60x 14x 60x 60x 60x 60x 21x 21x 21x 21x 1048x 1048x 1048x 1048x 1048x 1048x 1048x 1048x 1048x 1048x 1048x 1048x 1048x 1048x 1048x 1048x 21x 21x 24x 24x 24x 24x 24x 24x 7x 7x 7x 1x 6x 24x 24x 24x 422x 422x 422x 422x 17x 17x 41x 422x 422x 2501x 2501x 2501x 36x 36x 36x 36x 36x 406x 406x 406x 406x 406x 406x 406x 406x 406x 406x 406x 406x 406x 406x 406x 36x 36x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 2518x 5036x 5036x 5036x 60x 60x 60x 60x 60x 60x 60x 60x 1358x 1358x 1358x 1358x 1358x 1358x 1358x 1358x 32x 32x 8x 8x 8x 8x 8x 8x 8x 8x 8x 8x 32x 32x 32x 32x 32x 32x 32x 32x 32x 32x 24x 24x 1326x 1326x 1326x 1326x 1326x 1326x 1326x 1358x 1304x 1304x 11x 11x 11x 11x 11x 11x 11x 11x 11x 1x 1x 1x 6x 6x 6x 6x 6x 30x 30x 8x 30x 6x 6x 2x 2x 10x 2x 2x 1336x 1336x 1336x 60x 60x 60x 60x 1336x 1336x 1276x 1276x 1276x 1276x 1276x 1276x 1276x 1186x 1186x 1186x 1186x 1276x 150x 150x 60x 60x 150x 150x 150x 150x 150x 150x 150x 150x 150x 150x 150x 150x 150x 150x 1336x 1336x 1336x 1336x 1336x 1336x 1336x 1336x 1336x 1336x 1336x 1336x 1336x 1336x 1336x 1336x 1336x 1336x 36496x 36496x 6907256x 6907256x 57141x 57141x 57141x 6850115x 6906515x 6228134x 6228134x 6228134x 621981x 6907256x 6907256x 36496x 1336x 1336x 24x 24x 24x 24x 24x 24x 24x 24x 8x 8x 8x 8x 8x 8x 8x 508x 508x 508x 508x 508x 508x 494x 508x 8x 8x 8x 14x 14x 14x 14x 14x 14x 14x 14x 14x 82x 82x 82x 14x 14x 14x 14x 8x 8x 5x 5x 5x 5x 5x 16x 16x 16x 16x 16x 16x 16x 16x 16x 16x 16x 16x 16x 5x 5x 5x 3x 3x 3x 3x 9x 9x 9x 9x 9x 9x 9x 9x 9x 9x 9x 9x 9x 9x 9x 9x 3x 3x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 1x 1x 1x 1x 1x 1x 2x 4x 4x 2x 2x 2x 2x 2x 2x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 4x 2x 2x 2x 11x 11x 11x 6x | /**
* Aggregate peak analysis, run.
*
* Three things happen here that the kernel deliberately does not know about: the
* anchors are turned into windows, the windows are *read*, and the reads are
* coalesced.
*
* **Coalescing is an optimisation of the reads and not of the resolution**, and
* the distinction is the whole reason the task exists in this shape (§18.6).
* Windows that sit near each other on both axes are read as one block and sliced
* apart afterwards — the same cells, at the same bin size, in fewer requests.
* What is never done is widening the *query* to a bounding box, which would come
* back pooled and hand every window a coarser bin than the file has.
*
* Loop anchors cluster, which is what makes this worth doing: on a real yeast
* micro-C loop set the clusters collapse several thousand windows into a few
* hundred reads. On a local file it changes nothing measurable, because a read
* there is a `slice()` — which is why the estimate prices the un-coalesced upper
* bound and lets the run come in under it, rather than promising a saving that
* depends on how the anchors happen to be distributed.
*
* **The domain pileup is the third window shape and the one with a body.** A set
* of TADs — intervals between two called boundaries — piled up in the scaled
* mode gives the picture everybody draws next to a loop APA: the domain as the
* square in the middle, its two boundaries as the cross through it, and the
* neighbouring domains in the corners. Three things make it that picture rather
* than a plausible smear, and each is written where it happens: the window is
* `[start − fL, end + fL]` on **both axes** so the body is centred
* ({@link planFor}); the body's two ends are pinned to their **own grid lines**
* for every domain whatever its length ({@link cut}); and a distance-matched
* control is shifted by a **multiple of the window** rather than by a length in
* base pairs, because a control 100 kb from a megabase-wide window is mostly the
* window.
*
* **A window is cut as coolpuppy cuts one**, because coolpuppy is what these
* pileups are held to (§18.10) and the two parted company wherever a window
* reached the diagonal. Nothing is taken from below it: the map read here is
* drawn whole, both triangles, so a loop window crossing the diagonal used to
* pile up the cells below it — contacts it already held above, mirrored — and
* coolpuppy's, cut from the upper triangle, left them out, which on a real loop
* set was ten per cent in the cells nearest the diagonal. The first
* {@link minDiagOf} diagonals are left out as well, at coolpuppy's default of
* two. And observed / expected divides **each window, cell by cell**, by its
* sequence's own expected at that cell's own diagonal, then averages — see
* {@link cut} for what it did before, which was not that.
*/
import type { FeatureSet, PairSet } from '../../annotations/types'
import { DETAIL_SEP } from '../../annotations/types'
import { formatPosition } from '../../genome/assembly'
import { contextOf, type NameContext } from '../../genome/naming'
import { DEFAULT_COUNT_COLUMN } from '../../metacooler/matrix'
import { DEFAULT_DIVERGING_COLORMAP } from '../../render/colormaps'
import {
PileupAccumulator,
domainScore,
evenWindow,
p2ll,
ratioOf,
resampleWindow,
resampleWindowTo,
scaledEdges,
scaledLayout,
shiftedControls,
zoomWindow,
type AnchorPair,
type ScaledLayout,
} from '../apa'
import {
boolParam,
chosenInput,
contigIdFor,
idParam,
listParam,
missingInput,
numParam,
strParam,
type JobContext,
type TaskRunner,
} from '../job'
import { loopSpectrum } from '../loopSpectrum'
import { makeRandom } from '../random'
import {
anchorAt,
flankFractionOf,
isChromosight,
isDomainPileup,
isScaled,
minDiagOf,
minDistOf,
onDiagonal,
poolsAnchors,
stretchesEvenly,
usesPairSet,
allPairs,
consecutivePairs,
distanceBins,
MAX_GENERATED_PAIRS,
windowBins,
} from '../tasks/apa'
import type { MarginProfile, MatrixOutput, TableOutput, TaskOutput } from '../types'
import { chromosightPileup } from './apaChromosight'
import { diagonalMeans, usableBlock, type Span } from './blocks'
import { offsetTicks } from './labels'
/** A rectangle of the matrix, in absolute bins. */
interface Rect {
rowFrom: number
rowTo: number
colFrom: number
colTo: number
}
/**
* Where a scaled window's body sits, in absolute bins.
*
* Carried on the plan rather than recomputed at cut time because the *control*
* windows have a body too — the same length as the anchor they were drawn from,
* shifted with it — and a control resampled on a different layout from the real
* window it is the denominator of would divide two different pictures.
*/
interface Body {
from: number
to: number
/** Flank each side, in bins. Fractional, so every body lands on one grid line. */
pad: number
}
/** One window to pile up, and what it came from. */
interface Plan {
rect: Rect
/** For the per-anchor table; absent for a control. */
label?: string
separation: number
control: boolean
group: number
/** The sequence the window is on, in absolute bins: whose expected it takes. */
bounds: Span
/** The scaled on-diagonal layout this window is resampled on; absent otherwise. */
body?: Body
/**
* A domain window stretched evenly onto the grid by interpolation, gaps and
* all ({@link stretchesEvenly}), rather than pinned or resampled by area.
*/
zoomed?: boolean
}
/**
* The largest block a coalesced read is allowed to become, per side.
*
* Coalescing trades bytes for requests, and past a point the trade is bad: two
* windows 400 bins apart share one 500-bin block, which is 250,000 cells read to
* deliver 5,000. The cap keeps the merged block within a small multiple of what
* the windows actually need, so the saving is in requests and not in waste.
*/
const MERGE_CAP_BINS = 512
export const runApa: TaskRunner = async (job) => {
// Which **input** the run reads, which since the third anchor mode is no
// longer the same question as which side of the diagonal it draws on: *every
// pair* reads a feature set and its windows are off the diagonal.
const usingPairs = usesPairSet(job.params)
// By the **parameter key** — `pairs` or `features`, whichever mode is in
// force — and never by the entry id in `params`; see `chosenInput`.
const heldPairs = chosenInput(job.inputs.pairs, job.params, 'pairs')
const heldFeatures = chosenInput(job.inputs.features, job.params, 'features')
const pairs = usingPairs ? heldPairs.value : undefined
const features = usingPairs ? undefined : heldFeatures.value
if (pairs === undefined && features === undefined) {
// Named rather than returned empty. A task that finishes in a millisecond
// with no figure and no sentence is indistinguishable from one that is
// broken, and this is the commonest way to reach that state: the drawer
// opens on a task whose input has not been chosen yet.
job.note(
missingInput(usingPairs ? heldPairs : heldFeatures, {
unchosen: usingPairs
? 'No loop set was chosen. Load a BEDPE — a loop table, a set of anchor pairs — and pick it under Pairs.'
: 'No feature set was chosen. Load a BED or a GTF and pick it under Features.',
gone: `The ${usingPairs ? 'loop' : 'feature'} set chosen is no longer loaded. Re-open the file, or pick another.`,
}) ?? '',
)
return []
}
const balance = boolParam(job.params, 'balance', false)
const value = job.framing.value === '' ? DEFAULT_COUNT_COLUMN : job.framing.value
const side = isScaled(job.params)
? Math.max(11, Math.round(numParam(job.params, 'grid', 51)))
: windowBins(job.params) * 2 + 1
const norm = strParam(job.params, 'norm', 'matched')
const madePairs = allPairs(job.params)
const neighbours = consecutivePairs(job.params)
const anchors = usingPairs
? pairAnchors(job, pairs as PairSet)
: madePairs
? generatedPairs(job, features as FeatureSet)
: neighbours
? consecutiveAnchors(job, features as FeatureSet)
: featureAnchors(job, features as FeatureSet)
if (anchors.length === 0) {
job.note(
madePairs
? 'No two features of this set are on one sequence within the separation range, so there ' +
'is no pair to pile up. Widen the range, or check that the set and the file spell ' +
'their sequences alike.'
: neighbours
? 'No two features of this set are on one sequence this file has, so no feature has a ' +
'next one to pair with. Check that the set and the file spell their sequences alike.'
: 'No anchor of this set lands inside a sequence this file has, so there is nothing to pile up.',
)
return []
}
if (madePairs) {
job.note(
`${anchors.length.toLocaleString()} pairs made from ` +
`${(features as FeatureSet).length.toLocaleString()} features.`,
)
}
if (neighbours) {
job.note(
`${anchors.length.toLocaleString()} pairs, each feature with the next along its sequence, ` +
`made from ${(features as FeatureSet).length.toLocaleString()} features.`,
)
}
// The chromosight pileup is a different computation over the same anchors —
// the map is detrended before anything is cut out of it, so it reads whole
// sequences rather than windows — and it answers with its own three outputs.
// Branching here rather than at the top is deliberate: the anchor resolution
// above, including §5.9's naming, is shared and must stay shared.
if (isChromosight(job.params)) {
return chromosightPileup(
job,
anchors.map((anchor) => ({
row: anchor.pair.row,
col: anchor.pair.col,
bounds: anchor.bounds,
label: anchor.label,
})),
)
}
// **The one floor a loop set has**: a loop nearer the diagonal than the
// nearest loop (`minDist`, five bins unless typed) is left out. coolpuppy's
// `mindist='auto'` is further out — twice the half-width and two bins, where a
// window stops reaching the diagonals it would have gaps cut in — and is what
// the validation types. It is a floor on *where* a loop is, not a sample of the
// set, and it is said: how many, how near, and how to have them all.
const piled = (usingPairs || neighbours) && !isScaled(job.params) ? nearEnough(job, anchors) : anchors
if (piled.length === 0) return []
// Every anchor of every group. There used to be a cap here, and it was not a
// budget so much as a silent change of subject: a pileup of the first five
// thousand anchors is a pileup of whichever ones sorted first, under a caption
// saying how many were left out — which nobody compares across two runs.
// Two ways of dividing the same anchors, and never both at once: N separation
// bins crossed with K rank groups is N×K panels nobody reads, and neither
// division means anything once it has been taken inside the other. Separation
// wins where it is asked for, and the form hides the ranking controls off the
// diagonal in any case.
const bins = distanceBins(job.params)
const kept =
bins > 1
? bySeparation(piled, bins)
: groupAnchors(job, piled, onDiagonal(job.params) ? (features as FeatureSet) : null)
const groupNames = bins > 1 ? separationNames(kept, (span) => job.toBp(span)) : null
// The grid's own division into flank, body, flank — one layout for the whole
// run, which is what puts every domain's two boundaries on the same two grid
// lines however long the domain is. Unused off the diagonal and in the fixed
// mode, where a window has no body.
const layout = scaledLayout(side, flankFractionOf(job.params))
const scaledWindows = isScaled(job.params)
const scaledDomains = isDomainPileup(job.params)
const plans: Plan[] = []
const rng = makeRandom(Math.round(numParam(job.params, 'seed', 1)))
const controlsPer = norm === 'matched' ? Math.max(1, Math.round(numParam(job.params, 'controls', 5))) : 0
// Bins in, bins out: `shiftedControls` has always wanted a bin offset, and the
// form asks for one now, so the `toBins` that used to sit here is gone rather
// than moved (§18.2).
const minShift = Math.max(0, Math.round(numParam(job.params, 'minShift', 50)))
const maxShift = Math.max(minShift, Math.round(numParam(job.params, 'maxShift', 500)))
const minShiftScale = Math.max(0, numParam(job.params, 'minShiftScale', 1))
const maxShiftScale = Math.max(minShiftScale, numParam(job.params, 'maxShiftScale', 3))
for (let g = 0; g < kept.length; g++) {
for (const anchor of kept[g]) {
const plan = planFor(job, anchor, side, layout)
if (plan === null) continue
plans.push({
...plan,
label: anchor.label,
separation: anchor.pair.separation,
control: false,
group: g,
bounds: anchor.bounds,
})
if (controlsPer === 0) continue
// **The shift is a multiple of the window in the scaled mode and a count
// of bins in the fixed one**, and the difference is not a convenience. A
// domain pileup's window is as wide as the domain plus its flanks — often
// a thousand bins — so a control drawn fifty bins away is ninety per cent
// the same cells as the thing it is the denominator of, and the ratio it
// produces is 1 everywhere with the domain washed out of it. A fixed
// window is a couple of dozen bins, and there a count of bins is exactly
// the right way to ask: it is the same window at every resolution.
const width = scaledWindows ? Math.max(1, widthOf(plan.rect)) : 0
const from = scaledWindows ? Math.round(width * minShiftScale) : minShift
const to = scaledWindows ? Math.max(from, Math.round(width * maxShiftScale)) : maxShift
for (const control of shiftedControls(anchor.pair, anchor.bounds, {
perPair: controlsPer,
minShift: from,
maxShift: to,
rng,
})) {
const controlPlan = planFor(job, shiftAnchor(anchor, control), side, layout)
if (controlPlan !== null) {
plans.push({
...controlPlan,
separation: control.separation,
control: true,
group: g,
bounds: anchor.bounds,
})
}
}
}
}
if (plans.length === 0) {
job.note('Every window ran off the end of its sequence. Use a narrower window.')
return []
}
// **Observed/expected on a scaled grid is for a domain pileup**, which is
// coolpuppy's and is held to it ({@link stretchesEvenly}); for scaled pairs
// the estimate refuses it before Run rather than here, there being no pileup
// of coolpuppy's like that window. This is the stored session that names it
// anyway: said out loud and the raw pileup shown, which is the honest half of
// what was asked for.
const oeOnAScaledGrid = norm === 'oe' && isScaled(job.params) && !onDiagonal(job.params)
if (oeOnAScaledGrid) {
job.note(
'Observed / expected on a scaled grid is for a feature set on the diagonal, and these are ' +
'pairs. The raw pileup is shown — use the distance-matched control, which is drawn at ' +
'the same separations by construction.',
)
}
const real = kept.map(() => new PileupAccumulator(side, side))
const controls = kept.map(() => new PileupAccumulator(side, side))
// The tracks whose aggregate is drawn along the pileup's two edges, and one
// running mean per group, per track, per axis. Empty where none was chosen,
// which is every run this task has ever done until now.
const tracks = marginTracks(job)
const alongRows = kept.map(() => tracks.map(() => new PileupAccumulator(1, side)))
const alongCols = kept.map(() => tracks.map(() => new PileupAccumulator(1, side)))
const pooled = poolsAnchors(job.params)
const perAnchor: Array<{ label: string; separation: number; score: number }> = []
// How every window is cut, the same for the real ones and their controls,
// since a control masked differently from the window it divides would put a
// pattern in the ratio that is in neither.
const shape: CutOptions = {
minDiag: minDiagOf(job.params),
mirrored: onDiagonal(job.params),
}
const block = usableBlock(job, plans.map(spanOf), balance)
const expected =
norm === 'oe' && !oeOnAScaledGrid ? await expectedBySequence(job, plans, { block, balance, value }) : null
const batches = coalesce(plans)
let done = 0
for (const batch of batches) {
// The pileup is arithmetic over windows already in hand, so nothing below
// returns to the event loop on its own and a cancel would sit in the queue
// until the last batch had been summed. See `analysis/job.ts`.
await job.breathe()
job.progress(done, plans.length, `${done.toLocaleString()} of ${plans.length.toLocaleString()} windows`)
const dense = await readRect(job, batch.rect, { block, balance, value })
for (const plan of batch.plans) {
done++
// Every window that could be cut is summed, however little of it carries
// a value. There used to be a minimum here, and what it dropped was the
// windows in the sparsest regions — which are exactly the ones a weak
// pileup is weak because of, so removing them made every pileup look
// better than the data. `PileupAccumulator` counts per cell, so a
// half-empty window contributes what it has and nothing more.
const window = cut(dense, batch.rect, plan, side, layout, {
...shape,
expected: expected?.get(plan.bounds.from),
})
if (window === null) continue
if (plan.control) {
controls[plan.group].add(window)
continue
}
real[plan.group].add(window)
// **Only the real windows**, and the same ones: a marginal averaged over
// the controls as well would be a profile of the shifted anchors, which
// is the denominator of the picture rather than the picture. Cut from the
// plan's own rectangle and resampled by the plan's own edges, so a cell of
// the profile is the cell of the pileup above it by construction.
if (tracks.length > 0) {
const edges = edgesFor(plan, layout)
for (let t = 0; t < tracks.length; t++) {
const down = marginOf(tracks[t].values, plan.rect.rowFrom, plan.rect.rowTo, side, edges)
const across = marginOf(tracks[t].values, plan.rect.colFrom, plan.rect.colTo, side, edges)
alongRows[plan.group][t].add(down)
alongCols[plan.group][t].add(across)
// Pooled, every window contributes **both** of its anchors to **both**
// edges, so the two profiles come out one mean over twice as many
// anchors rather than two means over half as many each. Done here, a
// window at a time, rather than by averaging the two finished profiles
// at the end: the two accumulators count per cell, and a cell one of
// them never saw would be weighted as though it had. See
// {@link poolsAnchors}.
if (pooled) {
alongRows[plan.group][t].add(across)
alongCols[plan.group][t].add(down)
}
}
}
if (plan.label !== undefined) {
perAnchor.push({
label: plan.label,
separation: plan.separation * job.mc.binsize,
score: scaledDomains ? domainScore(window, layout) : p2ll(window, side, side),
})
}
}
}
job.progress(plans.length, plans.length, 'done')
job.throwIfCancelled()
const outputs: TaskOutput[] = []
for (let g = 0; g < kept.length; g++) {
const observed = real[g].finish()
if (observed.n === 0) continue
const name =
groupNames !== null
? `APA · ${groupNames[g]}`
: kept.length === 1
? 'APA'
: `APA · group ${g + 1}`
let values = observed.values
let centre: number | undefined
let label = `${name} · raw`
if (norm === 'matched') {
const control = controls[g].finish()
if (control.n > 0) {
values = ratioOf(observed.values, control.values)
centre = 1
label = `${name} · over distance-matched control`
} else {
job.note('No control window could be placed, so the raw pileup is shown.')
}
} else if (expected !== null) {
// Already divided, window by window, as each was cut — so the mean is
// the mean of the ratios, which is coolpuppy's `ooe`, and not a ratio of
// means, which is what dividing the finished pileup would give.
centre = 1
label = `${name} · observed / expected`
}
// **What the middle of the picture is**, which stopped being "the centre" the
// moment a window could be cut about a feature's start. An axis labelled
// *bins from the centre* over a pileup at four thousand transcription start
// sites is a caption for a figure nobody drew.
const from = onDiagonal(job.params)
? anchorAt(job.params) === 'start'
? 'the feature’s start'
: anchorAt(job.params) === 'end'
? 'the feature’s end'
: 'the feature’s centre'
: 'the centre'
const axis = isScaled(job.params)
? scaledDomains
? `Position along the feature (0 → 1), flanked by ${flankFractionOf(job.params)}×`
: 'Scaled position'
: `Distance from ${from}`
// **In base pairs, not in bins.** A fixed window is a window of the genome,
// and what a reader asks of a pileup is how far from the anchor the thing
// they are looking at lies; a count of bins answers that only for somebody
// holding the resolution in their head, and differently at every
// resolution. A *scaled* window has no such distance — its axis is the
// feature, 0 to 1 — and keeps its own labels ({@link scaledTickLabels}).
const ticks = isScaled(job.params)
? scaledDomains
? stretchesEvenly(job.params)
? evenTickLabels(flankFractionOf(job.params))
: scaledTickLabels(layout, flankFractionOf(job.params))
: []
: offsetTicks(side, job.mc.binsize)
outputs.push({
kind: 'matrix',
id: `apa-${g}`,
label,
nrows: side,
ncols: side,
values,
xLabel: axis,
yLabel: axis,
xTicks: ticks,
yTicks: ticks,
colormap: centre === undefined ? 'default' : DEFAULT_DIVERGING_COLORMAP,
...(centre === undefined ? {} : { centre }),
// The aggregate signal above and beside the aggregate map, over the same
// windows: the figure this task could not draw, and the one most papers
// that use an APA actually print.
...(() => {
const margins = marginsOf(tracks, alongRows[g], alongCols[g])
return margins === undefined ? {} : { margins }
})(),
stats: [
{ label: 'windows', value: observed.n, digits: 0, hint: 'Anchors that contributed.' },
scaledDomains
? {
label: 'in / out',
value: domainScore(values, layout),
digits: 2,
hint:
'The domain’s own corner against the same block across its boundary. ' +
'Above 1 means contact stops at the edge.',
}
: {
label: 'P2LL',
value: p2ll(values, side, side),
digits: 2,
hint: 'Centre against the lower-left corner. Above 1 is enrichment.',
},
],
} satisfies MatrixOutput)
}
if (perAnchor.length > 0) {
// **The loop spectrum**, before the table and only off the diagonal, where a
// separation exists to plot against. Every anchor already carries its own
// P2LL and its own separation — the table below prints both — and the one
// thing the pileup above cannot show is how the first varies with the
// second. A loop set whose 30 kb pairs are loops and whose 2 Mb pairs are
// noise averages into a single plausible dot, and the two numbers that say
// so are already in hand (`analysis/loopSpectrum.ts`).
if (!onDiagonal(job.params)) {
const spectrum = loopSpectrum(perAnchor, {
id: 'apa-loop-spectrum',
label: 'Loop spectrum',
noun: allPairs(job.params) ? 'pairs' : 'loops',
yLabel: 'P2LL — centre against the lower-left corner',
})
if (spectrum !== null) outputs.push(spectrum)
}
perAnchor.sort((a, b) => (Number.isNaN(b.score) ? -1 : b.score) - (Number.isNaN(a.score) ? -1 : a.score))
outputs.push({
kind: 'table',
id: 'apa-anchors',
label: 'Per anchor',
columns: [
{ key: 'anchor', label: 'Anchor', kind: 'text' },
{ key: 'separation', label: 'Separation (bp)', kind: 'number', digits: 0 },
{ key: 'score', label: scaledDomains ? 'In / out' : 'P2LL', kind: 'number', digits: 3 },
],
rows: perAnchor.map((row) => ({
anchor: row.label,
separation: row.separation,
score: row.score,
})),
} satisfies TableOutput)
}
return outputs
}
// ---------------------------------------------------------------------------
// Track aggregates along the edges
// ---------------------------------------------------------------------------
/** One chosen track: what it is called, and one value per bin of this map. */
interface MarginTrack {
name: string
values: Float32Array
}
/**
* The tracks whose aggregate is drawn along the pileup's edges.
*
* **By the parameter key plus the entry id**, which is how a `multiple` entry
* parameter's inputs are filed (`AnalysisRunner`'s `chosenEntries`) — and a
* track arrives as a *vector over the map's bins* rather than as a file handle,
* which is what lets a task in this venue read one at all (§18.6): the bigWig
* lives in the annotations worker and is megabytes, and what a marginal profile
* needs is one number per bin, which the lane renderer already computes.
*
* A row dropped between the form being filled in and Run being pressed is simply
* absent, and is left out rather than drawn as a flat line.
*/
function marginTracks(job: JobContext): MarginTrack[] {
const chosen = listParam(job.params, 'tracks')
const out: MarginTrack[] = []
for (const id of chosen) {
const values = job.inputs.vectors?.[`tracks:${id}`]
if (values === undefined) continue
out.push({ name: job.inputs.labels?.[id] ?? id, values })
}
if (out.length < chosen.length) {
// Said out loud, because the pileup comes back looking perfectly finished
// with one fewer line above it than the form promised — and a reader who
// dropped a track between filling the form in and pressing Run has no other
// way to find out which of the two things happened.
job.note(
`${(chosen.length - out.length).toLocaleString()} of ${chosen.length.toLocaleString()} ` +
'tracks could not be read and are not drawn beside the pileup. Re-open the file, or ' +
'pick another under Track aggregates.',
)
}
return out
}
/**
* The grid edges one window's profile is resampled onto, or `undefined`.
*
* The plan's own, so a marginal cell is the cell of the pileup above it: a
* domain window is flank, body, flank with the body pinned to its own two grid
* lines ({@link cut}), and a profile stretched evenly over the same window would
* put the boundary a column away from where the picture draws it.
*/
function edgesFor(plan: Plan, layout: ScaledLayout): Float64Array | undefined {
if (plan.body === undefined) return undefined
return scaledEdges(
layout,
plan.body.from - plan.rect.rowFrom,
plan.body.to - plan.rect.rowFrom,
plan.body.pad,
)
}
/**
* One window's stretch of a track, resampled onto the pileup's grid.
*
* Bins outside the vector read as `NaN` rather than as zero — a window running
* off the end of the last sequence has no coverage there, and zero is a
* measurement. `PileupAccumulator` skips them per cell, so a window that is half
* off the end contributes the half it has.
*
* The resampling is `analysis/apa.ts`'s, over a source one row tall: area
* weighted and gap preserving, which is the same arithmetic the matrix window
* goes through and therefore the same answer about where a feature's edge lands.
*/
function marginOf(
values: Float32Array,
from: number,
to: number,
side: number,
edges: Float64Array | undefined,
): Float64Array {
const span = to - from
if (span <= 0) return new Float64Array(side).fill(Number.NaN)
const source = new Float64Array(span)
for (let i = 0; i < span; i++) {
const at = from + i
source[i] = at >= 0 && at < values.length ? values[at] : Number.NaN
}
if (edges === undefined) return resampleWindow(source, 1, span, 1, side)
return resampleWindowTo(source, 1, span, Float64Array.from([0, 1]), edges)
}
/**
* The finished profiles for one group, or `undefined` where there are none.
*
* A profile with no window behind it is left out rather than drawn empty: a flat
* line along the top of a pileup is a claim about the signal, and *no anchor of
* this group could be profiled* is not that claim.
*/
function marginsOf(
tracks: readonly MarginTrack[],
rows: readonly PileupAccumulator[],
cols: readonly PileupAccumulator[],
): { top?: MarginProfile[]; right?: MarginProfile[] } | undefined {
if (tracks.length === 0) return undefined
const top: MarginProfile[] = []
const right: MarginProfile[] = []
for (let t = 0; t < tracks.length; t++) {
// The **columns** run along the top and the **rows** down the right side,
// which is what puts each profile beside the axis it is a profile of.
const alongTop = cols[t].finish()
const alongSide = rows[t].finish()
if (alongTop.n > 0) top.push({ name: tracks[t].name, values: Float64Array.from(alongTop.values) })
if (alongSide.n > 0) {
right.push({ name: tracks[t].name, values: Float64Array.from(alongSide.values) })
}
}
if (top.length === 0 && right.length === 0) return undefined
return {
...(top.length === 0 ? {} : { top }),
...(right.length === 0 ? {} : { right }),
}
}
// ---------------------------------------------------------------------------
// Anchors
// ---------------------------------------------------------------------------
interface Anchor {
pair: AnchorPair
bounds: Span
label: string
rank: number
/** Bins the feature spans, for the scaled mode. */
span: number
/**
* For a loaded pair, how far apart its two anchors' centres are, in base
* pairs: what coolpuppy measures its `mindist` on, and so what the nearest
* loop is held to here. Absent for anything else.
*/
centreDistance?: number
/**
* The rows and columns a *scaled* window's body covers, half-open, in bins.
*
* The feature's own bins on both axes for an on-diagonal anchor — which is
* what makes the domain land in the middle of the picture rather than in the
* corner of it — and the single anchor bin on each axis for a pair, whose
* scaled window is a square about the loop and has no body to resample.
*/
rows: Span
cols: Span
/**
* The feature's own start and end on its sequence, in base pairs, for an
* on-diagonal anchor: what a window stretched evenly is scaled about
* ({@link evenWindow}), which works in base pairs as bioframe does.
*/
extent?: { start: number; end: number }
}
function pairAnchors(job: JobContext, set: PairSet): Anchor[] {
const out: Anchor[] = []
// The set's own list is what says how *it* spells sequences, and a `.bedpe`
// loaded against a `chr1` lead is walked here against whichever panel of the
// deck this job is running on (§18.2). Computed once per set, not per row.
const from = contextOf(set.contigs)
for (let i = 0; i < set.length; i++) {
const a = binOf(job, set.contigs[set.contig1[i]], set.start1[i], set.end1[i], from)
const b = binOf(job, set.contigs[set.contig2[i]], set.start2[i], set.end2[i], from)
// Both anchors on one sequence, because a pileup off the diagonal between
// two *different* sequences is a trans pileup and a different question — one
// no loop caller emits and no APA convention describes.
if (a === null || b === null || a.contig !== b.contig) continue
const row = Math.min(a.bin, b.bin)
const col = Math.max(a.bin, b.bin)
out.push({
pair: { row, col, separation: col - row },
bounds: a.bounds,
// An anchor the file gave no name is named by its first base as a reader
// reads one (§26.1), here and in the two places below. It was the stored
// start, 0-based, in the *Per anchor* table of a drawer whose every other
// coordinate is 1-based.
label: set.name[i] === '' ? formatPosition(set.contigs[set.contig1[i]], set.start1[i]) : set.name[i],
rank: set.score[i],
span: col - row,
centreDistance: Math.abs((set.start2[i] + set.end2[i]) / 2 - (set.start1[i] + set.end1[i]) / 2),
rows: { from: row, to: row + 1 },
cols: { from: col, to: col + 1 },
})
}
return out
}
/**
* The loops no nearer the diagonal than the nearest loop, and a note for the rest.
*
* coolpuppy's `mindist`, measured as coolpuppy measures it — between the two
* anchors' centres, in base pairs, a loop exactly at the floor kept — so that a
* BEDPE whose anchors are not bin-aligned is cut at the same loop in both. The
* floor is asked in bins, as every matrix parameter is (§18.2), and is base
* pairs here at this file's resolution.
*/
function nearEnough(job: JobContext, anchors: Anchor[]): Anchor[] {
const nearest = minDistOf(job.params)
const floor = nearest * job.mc.binsize
const out = anchors.filter((anchor) => (anchor.centreDistance ?? Number.POSITIVE_INFINITY) >= floor)
const left = anchors.length - out.length
if (left === 0) return out
const how = `nearer the diagonal than ${nearest.toLocaleString()} bins (${shortBp(floor)})`
job.note(
out.length === 0
? `Every loop of this set is ${how} so there is nothing to pile up. Set the nearest loop lower, ` +
'or to 0 for every one.'
: `${left.toLocaleString()} of ${anchors.length.toLocaleString()} loops are ${how} and are left ` +
'out. Set the nearest loop to 0 to pile up every one.',
)
return out
}
/**
* The base a window is cut about: the feature's start, its centre or its end.
*
* **Strand decides which end is which**, exactly as `analysis/metaplot.ts`'s
* reference point does, because *start* means the 5′ end and a minus-strand gene
* runs the other way. A set with no strand column reads as plus throughout,
* which is what an unstranded peak set means by its own coordinates.
*
* The end takes the feature's **last base** rather than its exclusive end, which
* is a bin off on any feature whose end lands on a bin boundary — and on a gene
* set that is most of them, since annotation ends are round numbers.
*
* The centre is {@link midpointBase}: coolpuppy's bin for it, never the next one.
*/
function anchorBase(set: FeatureSet, i: number, at: 'start' | 'centre' | 'end'): number {
const start = set.start[i]
const end = set.end[i]
if (at === 'centre') return midpointBase(start, end)
const last = Math.max(start, end - 1)
const reversed = set.strand[i] < 0
if (at === 'start') return reversed ? last : start
return reversed ? start : last
}
/**
* The base an interval's midpoint is binned by: the midpoint **rounded down**.
*
* coolpuppy bins a feature by `floor((start + end) / 2 / binsize)`, the exact
* midpoint, and rounding it down to a base first lands in the same bin whatever
* the bin size. Rounding it to the nearest base did not: a feature an odd
* number of bases long whose midpoint sits half a base before a bin edge went
* into the next bin — one CTCF peak of 2,106 on GM12878's chr19 at 2 kb, enough
* to move a pileup's cells by a per cent from coolpuppy's.
*/
function midpointBase(start: number, end: number): number {
return Math.floor((start + end) / 2)
}
function featureAnchors(job: JobContext, set: FeatureSet): Anchor[] {
const out: Anchor[] = []
const from = contextOf(set.contigs)
const reference = anchorAt(job.params)
for (let i = 0; i < set.length; i++) {
const at = binOf(job, set.contigs[set.contig[i]], set.start[i], set.end[i], from, {
point: anchorBase(set, i, reference),
})
if (at === null) continue
out.push({
pair: { row: at.bin, col: at.bin, separation: 0 },
bounds: at.bounds,
label: set.name[i] === '' ? formatPosition(set.contigs[set.contig[i]], set.start[i]) : set.name[i],
rank: set.score[i],
// The feature's **own bins**, not `round(length / binsize)`. The two differ
// by one wherever an interval straddles a bin edge, and the difference is
// not academic here: the body is what the scaled window is centred on and
// resampled against, so a span taken from the length and a centre taken
// from the midpoint disagree about where the feature is by up to a bin.
span: at.body.to - at.body.from,
rows: at.body,
cols: at.body,
extent: { start: set.start[i], end: set.end[i] },
})
}
return out
}
/**
* The bin an interval's centre falls in, and the sequence's bin bounds.
*
* `named` is the naming context of the *set* the interval came out of, so a BED
* of `chr1` loci resolves against a file that writes `1` (§5.9). The sequence
* comes back under **this file's** spelling rather than the set's, which is what
* makes the two anchors of a pair comparable: two names that resolve to one
* sequence are one sequence, and comparing the strings would call them cis and
* trans by turns.
*/
function binOf(
job: JobContext,
contig: string,
start: number,
end: number,
named?: NameContext,
options: { point?: number } = {},
): { bin: number; body: Span; bounds: Span; contig: string } | null {
const cid = contigIdFor(job.mc, contig, named)
if (cid < 0) return null
const [from, to] = job.mc.contigExtent(cid)
// The reader's own reference point where there is one, and the midpoint
// otherwise — which is what a loop set's anchor has and what this task did
// before the choice existed. The *body* below is the feature either way: the
// scaled mode resamples the whole interval, and where it is cut about is a
// question only the fixed mode asks.
const bin = job.mc.coords.binIndex(cid, options.point ?? midpointBase(start, end))
if (bin < from || bin >= to) return null
// The bins the interval itself covers, clamped to the sequence and never
// empty: a feature narrower than one bin is one bin, since that is the finest
// thing this file can say about where it is.
const first = clamp(job.mc.coords.binIndex(cid, Math.max(0, Math.min(start, end))), from, to - 1)
const last = clamp(job.mc.coords.binIndex(cid, Math.max(0, Math.max(start, end) - 1)), first, to - 1)
return {
bin,
body: { from: first, to: last + 1 },
bounds: { from, to },
contig: job.mc.chromnames[cid],
}
}
function clamp(value: number, lo: number, hi: number): number {
return Math.min(hi, Math.max(lo, value))
}
/** The anchor groups a run draws, after ranking and subsetting. */
function groupAnchors(job: JobContext, anchors: Anchor[], set: FeatureSet | null): Anchor[][] {
const rankBy = idParam(job.params, 'rankBy')
const subset = strParam(job.params, 'subset', 'all')
if (rankBy === null || subset === 'all' || set === null) return [anchors]
const ranked = rankBy === 'score' ? anchors : withDetailRank(anchors, set, rankBy)
const sorted = [...ranked].sort((a, b) => {
if (Number.isNaN(a.rank)) return Number.isNaN(b.rank) ? 0 : 1
if (Number.isNaN(b.rank)) return -1
return b.rank - a.rank
})
if (subset === 'top') {
return [sorted.slice(0, Math.max(1, Math.round(numParam(job.params, 'count', 500))))]
}
const groups = Math.max(2, Math.round(numParam(job.params, 'groups', 4)))
const size = Math.ceil(sorted.length / groups)
const out: Anchor[][] = []
for (let g = 0; g < groups; g++) {
const slice = sorted.slice(g * size, (g + 1) * size)
if (slice.length > 0) out.push(slice)
}
return out
}
/** The ranking column, dug out of the `detail` string a parser joined. */
function withDetailRank(anchors: Anchor[], set: FeatureSet, column: string): Anchor[] {
return anchors.map((anchor, i) => {
for (const part of (set.detail[i] ?? '').split(DETAIL_SEP)) {
const at = part.search(/[=:]/)
if (at < 0) continue
if (part.slice(0, at).trim() !== column) continue
const value = Number.parseFloat(part.slice(at + 1).trim())
return { ...anchor, rank: Number.isFinite(value) ? value : Number.NaN }
}
return { ...anchor, rank: Number.NaN }
})
}
// ---------------------------------------------------------------------------
// Windows and reads
// ---------------------------------------------------------------------------
/**
* The window one anchor gets: the rectangle to read, and where its body sits.
*
* Three shapes, and the middle one is the whole of this change.
*
* **Fixed** is the centre bin plus or minus a half-width, on both axes.
*
* **Scaled, on the diagonal** — a *domain* pileup — is the feature's own bins
* with a flank each side, on both axes: `[start − fL, end + fL]²`, so the domain
* is the square in the middle of the picture and its two boundaries are the
* cross through it. It used to be `[centre − fL, centre + L + fL]`, which is the
* same width and half a domain to the right of where it belongs: the body sat
* against the top-left corner with the whole flank on the far side, and a set of
* real TADs piled up into a picture with the domain in the wrong place and no
* way to tell from the figure.
*
* **Scaled, off the diagonal** is a square about each anchor whose side is a
* fraction of the separation — the rectangle between two anchors, which is what
* the reader asked for when they said the window should scale with the loop.
*
* The pad is deliberately **fractional**: the read is whole bins, but the grid
* lines the body is resampled onto are `start` and `end` exactly, so a 17-bin
* domain and a 431-bin one put their boundaries on the same two columns instead
* of within half a bin of them.
*/
function planFor(
job: JobContext,
anchor: Anchor,
side: number,
layout: ScaledLayout,
): { rect: Rect; body?: Body; zoomed?: boolean } | null {
const { from, to } = anchor.bounds
if (isScaled(job.params)) {
// Nothing to scale *by*: a pair whose two anchors land in the same bin has
// no separation for the window to be a fraction of, and a window of one bin
// stretched over the grid is a flat square averaged in as though it were
// evidence. Dropped, as it was before there was a body.
if (anchor.span <= 0) return null
// **A domain over the expected is scaled about its middle and stretched
// evenly**, as coolpuppy cuts it ({@link stretchesEvenly}): the window is the
// feature's base pairs times `1 + 2f`, snapped out to whole bins, and a
// window that runs off its sequence is dropped rather than clipped.
if (stretchesEvenly(job.params) && anchor.extent !== undefined) {
const bins = evenWindow(anchor.extent.start, anchor.extent.end, flankFractionOf(job.params), job.mc.binsize)
const rect = {
rowFrom: from + bins.from,
rowTo: from + bins.to,
colFrom: from + bins.from,
colTo: from + bins.to,
}
if (rect.rowFrom < from || rect.rowTo > to || rect.rowTo <= rect.rowFrom) return null
return { rect, zoomed: true }
}
// No flank cells on the grid means no flank read: a reader who asked for
// none gets the body, not a band of bins that is read and then dropped.
const pad = layout.flank === 0 ? 0 : anchor.span * flankFractionOf(job.params)
const rect = {
rowFrom: Math.floor(anchor.rows.from - pad),
rowTo: Math.ceil(anchor.rows.to + pad),
colFrom: Math.floor(anchor.cols.from - pad),
colTo: Math.ceil(anchor.cols.to + pad),
}
if (rect.rowFrom < from || rect.colTo > to) return null
if (rect.rowTo <= rect.rowFrom || rect.colTo <= rect.colFrom) return null
// Only the on-diagonal window has a body to resample against, and it is the
// *mode* that says so rather than the separation: a pair's two anchors can
// fall in one bin, and reading that as "on the diagonal" would resample a
// loop window against a body one bin wide.
if (!isDomainPileup(job.params)) return { rect }
return { rect, body: { from: anchor.rows.from, to: anchor.rows.to, pad } }
}
const half = (side - 1) / 2
const rect = {
rowFrom: anchor.pair.row - half,
rowTo: anchor.pair.row + half + 1,
colFrom: anchor.pair.col - half,
colTo: anchor.pair.col + half + 1,
}
// Clipped windows are dropped rather than padded. A pileup is an average over
// identically shaped things, and half a window averaged into the middle of the
// picture is signal in the wrong cells.
if (rect.rowFrom < from || rect.colTo > to) return null
return { rect }
}
/** One anchor moved to where a control landed, body and all. */
function shiftAnchor(anchor: Anchor, control: AnchorPair): Anchor {
const delta = control.row - anchor.pair.row
return {
...anchor,
pair: control,
rows: { from: anchor.rows.from + delta, to: anchor.rows.to + delta },
cols: { from: anchor.cols.from + delta, to: anchor.cols.to + delta },
}
}
/** The widest side of a rectangle, in bins. */
function widthOf(rect: Rect): number {
return Math.max(rect.rowTo - rect.rowFrom, rect.colTo - rect.colFrom)
}
/**
* The axis labels for a scaled on-diagonal grid, in units of the feature.
*
* Five of them, evenly spaced, because that is what a matrix output's ticks are
* — one label per fraction `i / 4` of the axis.
*
* **The axis is piecewise, not linear, and reading it as linear is wrong by a
* cell.** The grid gives the flank a whole number of cells and the body what is
* left, and those two counts are not in the ratio the flank fraction is — a
* 41-cell grid at half a body each side is 10, 21, 10 — so a cell of flank and a
* cell of body are different lengths of feature. The map is therefore three
* straight lines meeting at the two boundaries, which is exactly how the window
* was resampled ({@link scaledEdges}); the ends come out at `−f` and `1 + f`
* because that is where they are.
*/
function scaledTickLabels(layout: ScaledLayout, flankFraction: number): string[] {
if (layout.flank === 0) return ['0', '0.25', '0.5', '0.75', '1']
const bodyStart = layout.flank / layout.side
const bodyEnd = (layout.flank + layout.body) / layout.side
const at = (fraction: number): number => {
if (fraction <= bodyStart) return -flankFraction * (1 - fraction / bodyStart)
if (fraction >= bodyEnd) return 1 + (flankFraction * (fraction - bodyEnd)) / (1 - bodyEnd)
return (fraction - bodyStart) / (bodyEnd - bodyStart)
}
// `Number(...)` rather than the fixed string, so `1.00` reads as `1` and a
// rounded `-0.00` reads as `0` rather than as a minus sign nobody can explain.
return [0, 0.25, 0.5, 0.75, 1].map((fraction) => String(Number(at(fraction).toFixed(2))))
}
/**
* The same five labels for a window stretched evenly ({@link stretchesEvenly}),
* where the axis is one straight line from `−f` to `1 + f`: at the default half
* a feature each side the feature's two ends are the second and fourth labels,
* `0` and `1`, whatever the grid.
*/
function evenTickLabels(flankFraction: number): string[] {
return [0, 0.25, 0.5, 0.75, 1].map((fraction) =>
String(Number((fraction * (1 + 2 * flankFraction) - flankFraction).toFixed(2))),
)
}
function spanOf(plan: Plan): Span {
return { from: Math.min(plan.rect.rowFrom, plan.rect.colFrom), to: Math.max(plan.rect.rowTo, plan.rect.colTo) }
}
/** Windows near each other on both axes, gathered into one read. */
function coalesce(plans: Plan[]): Array<{ rect: Rect; plans: Plan[] }> {
const sorted = [...plans].sort((a, b) => a.rect.rowFrom - b.rect.rowFrom || a.rect.colFrom - b.rect.colFrom)
const out: Array<{ rect: Rect; plans: Plan[] }> = []
for (const plan of sorted) {
const last = out[out.length - 1]
if (last !== undefined) {
const merged = {
rowFrom: Math.min(last.rect.rowFrom, plan.rect.rowFrom),
rowTo: Math.max(last.rect.rowTo, plan.rect.rowTo),
colFrom: Math.min(last.rect.colFrom, plan.rect.colFrom),
colTo: Math.max(last.rect.colTo, plan.rect.colTo),
}
if (merged.rowTo - merged.rowFrom <= MERGE_CAP_BINS && merged.colTo - merged.colFrom <= MERGE_CAP_BINS) {
last.rect = merged
last.plans.push(plan)
continue
}
}
out.push({ rect: { ...plan.rect }, plans: [plan] })
}
return out
}
interface Dense {
values: Float32Array
nrows: number
ncols: number
}
async function readRect(
job: JobContext,
rect: Rect,
options: { block: ReturnType<typeof usableBlock>; balance: boolean; value: string },
): Promise<Dense> {
if (options.block !== null) {
const block = options.block
const nrows = rect.rowTo - rect.rowFrom
const ncols = rect.colTo - rect.colFrom
const values = new Float32Array(nrows * ncols)
for (let r = 0; r < nrows; r++) {
const at = (rect.rowFrom - block.rowExtent[0] + r) * block.ncols + (rect.colFrom - block.colExtent[0])
values.set(block.values.subarray(at, at + ncols), r * ncols)
}
return { values, nrows, ncols }
}
const dense = await job.mc.matrix({
rows: [rect.rowFrom, rect.rowTo],
cols: [rect.colFrom, rect.colTo],
balance: options.balance,
value: options.value,
})
return { values: dense.values, nrows: dense.nrows, ncols: dense.ncols }
}
/** How a window is cut out of the map before it is piled up. */
interface CutOptions {
/** Diagonals, the main one first, whose cells are left out; see {@link minDiagOf}. */
minDiag: number
/**
* Whether the window sits on the diagonal and is its own mirror.
*
* There the cells below the diagonal are the cells above it, transposed, and
* the window is kept whole with only the near diagonals left out on both
* sides — which is coolpuppy's local pileup exactly, since coolpuppy cuts the
* upper triangle and then folds the finished picture onto its transpose. Off
* the diagonal nothing is taken from below it.
*/
mirrored: boolean
/**
* The expected of the window's sequence, by diagonal, where every window is
* divided by it; `undefined` otherwise.
*/
expected?: Float64Array
}
/**
* One window out of a coalesced read, resampled to the pileup's shape.
*
* Two resamplings, and which one runs is decided by whether the plan has a
* **body**. Without one the window is stretched evenly onto the grid, which is
* right for a fixed window (it is already the right size) and for a scaled pair
* (a square about the loop, with nothing inside it that has to land anywhere in
* particular). With one — a domain — the grid is flank, body, flank and the
* body's two ends are pinned to their own grid lines, so every domain in the set
* contributes its boundary to the same column.
*
* Before either, the window is cut as coolpuppy cuts one ({@link CutOptions}),
* and divided by its expected where the run asked for observed / expected:
* **each cell by the expected at its own diagonal**, `|col − row|` in the map,
* on its own sequence. That replaced a division of the *finished* pileup by an
* expected looked up at each cell's distance from the pileup's centre, which is
* a cell's diagonal only for a window on the diagonal. For a loop forty bins out
* it divided the loop's own pixel by the expected of the main diagonal, and the
* expected came from the windows' own cells, which were not read that close to
* the diagonal at all — so the centre of a loop pileup came out at 0.2 on a
* mouse loop set coolpuppy puts at 2.9.
*/
function cut(
dense: Dense,
read: Rect,
plan: Plan,
side: number,
layout: ScaledLayout,
options: CutOptions,
): Float64Array | null {
const want = plan.rect
const rows = want.rowTo - want.rowFrom
const cols = want.colTo - want.colFrom
if (rows <= 0 || cols <= 0) return null
const rowSkip = want.rowFrom - read.rowFrom
const colSkip = want.colFrom - read.colFrom
if (rowSkip < 0 || colSkip < 0 || rowSkip + rows > dense.nrows || colSkip + cols > dense.ncols) {
return null
}
const cutOut = new Float64Array(rows * cols)
const { minDiag, mirrored, expected } = options
for (let r = 0; r < rows; r++) {
const at = (rowSkip + r) * dense.ncols + colSkip
for (let c = 0; c < cols; c++) {
const diagonal = want.colFrom + c - (want.rowFrom + r)
if ((mirrored ? Math.abs(diagonal) : diagonal) < minDiag) {
cutOut[r * cols + c] = Number.NaN
continue
}
const observed = dense.values[at + c]
if (expected === undefined) {
cutOut[r * cols + c] = observed
continue
}
// A diagonal with nothing on it gives no ratio rather than an infinity,
// for the reason `ratioOf` gives.
const denominator = expected[Math.abs(diagonal)]
cutOut[r * cols + c] = denominator > 0 ? observed / denominator : Number.NaN
}
}
if (plan.zoomed === true) return zoomWindow(cutOut, rows, side)
if (plan.body === undefined) return resampleWindow(cutOut, rows, cols, side, side)
// In cells of this window rather than absolute bins, which is what the edges
// are measured in.
const edges = scaledEdges(
layout,
plan.body.from - want.rowFrom,
plan.body.to - want.rowFrom,
plan.body.pad,
)
return resampleWindowTo(cutOut, rows, cols, edges, edges)
}
// ---------------------------------------------------------------------------
// Observed over expected
// ---------------------------------------------------------------------------
/**
* Each sequence's expected, by diagonal, as deep as its farthest window reaches.
*
* Keyed by the sequence's first bin, which is what a plan carries. **The whole
* sequence, and no more of it than the windows need**: cooltools' expected is a
* property of the sequence, so a window is divided by what that sequence does
* at the window's diagonals everywhere along it — not by what the windows
* themselves hold there, which over a loop set is the mean of the places the set
* picked and enriched by construction. The depth is the one economy, and it is
* an exact one: a diagonal no window reaches is a diagonal nothing is divided by.
*/
async function expectedBySequence(
job: JobContext,
plans: readonly Plan[],
options: { block: ReturnType<typeof usableBlock>; balance: boolean; value: string },
): Promise<Map<number, Float64Array>> {
const deepest = new Map<number, { bounds: Span; hi: number }>()
for (const plan of plans) {
// The farthest a cell of this window is from the diagonal, either side of it.
const reach = Math.max(
Math.abs(plan.rect.colTo - 1 - plan.rect.rowFrom),
Math.abs(plan.rect.rowTo - 1 - plan.rect.colFrom),
)
const held = deepest.get(plan.bounds.from)
if (held === undefined) deepest.set(plan.bounds.from, { bounds: plan.bounds, hi: reach })
else if (reach > held.hi) held.hi = reach
}
const out = new Map<number, Float64Array>()
let done = 0
for (const [from, { bounds, hi }] of deepest) {
job.progress(done, deepest.size, `the expected, sequence ${done + 1} of ${deepest.size}`)
out.set(
from,
await diagonalMeans(job, bounds, {
hi,
block: options.block,
balance: options.balance,
value: options.value,
onTile: async () => {
await job.breathe()
job.throwIfCancelled()
},
}),
)
done++
}
return out
}
// ---------------------------------------------------------------------------
// Pairs made rather than loaded
// ---------------------------------------------------------------------------
/** One feature where the file has it: its bin, and the point its pairs are measured between. */
interface Placed {
contig: string
centre: number
/**
* The point unrounded — half way along the feature where the window is cut
* about its centre — which is where coolpuppy measures a pair's separation
* from, so the nearest loop keeps the same pairs as it does.
*/
exact: number
bin: number
bounds: Span
body: Span
label: string
}
/**
* A feature set's features on this file, in order along each sequence.
*
* Resolved once per feature rather than once per pair: `contigIdFor` walks a
* naming context and a set of forty thousand features inside a 2 Mb window
* produces millions of pairs, so resolving inside the inner loop is the one
* decision that would make a run minutes rather than seconds. Sorted by this
* file's own sequence name and then by position, because the set's order is the
* *set's* and two of its names may resolve to one sequence here (§5.9): sorting
* on what was resolved is what keeps a walk along a sequence correct in exactly
* that case.
*/
function placedFeatures(job: JobContext, set: FeatureSet): Placed[] {
const from = contextOf(set.contigs)
const placed: Placed[] = []
const reference = anchorAt(job.params)
for (let i = 0; i < set.length; i++) {
// The **same point** the window is cut about is the point the separation is
// measured between, and it has to be: a range of 20 kb to 2 Mb between gene
// centres and a pileup at gene starts would be two different sets of pairs
// wearing one caption.
const point = anchorBase(set, i, reference)
const at = binOf(job, set.contigs[set.contig[i]], set.start[i], set.end[i], from, { point })
if (at === null) continue
placed.push({
contig: at.contig,
centre: point,
exact: reference === 'centre' ? (set.start[i] + set.end[i]) / 2 : point,
bin: at.bin,
bounds: at.bounds,
body: at.body,
label: set.name[i] === '' ? formatPosition(set.contigs[set.contig[i]], set.start[i]) : set.name[i],
})
}
placed.sort((a, b) => (a.contig === b.contig ? a.centre - b.centre : a.contig < b.contig ? -1 : 1))
return placed
}
/**
* Each feature of a set with the next one along its sequence.
*
* One pair a neighbour, the two in the set's order along the sequence, and no
* sampling: a set of n features on one sequence makes n − 1 pairs. A pair is
* kept whatever its separation and held to the nearest loop afterwards, as a
* loaded BEDPE of the same pairs is (`nearEnough`), measured between the two
* features' own centres as coolpuppy measures it — so a pileup of these pairs
* and coolpuppy's of a BEDPE listing them are of the same pairs.
*/
function consecutiveAnchors(job: JobContext, set: FeatureSet): Anchor[] {
const placed = placedFeatures(job, set)
const out: Anchor[] = []
for (let i = 0; i + 1 < placed.length; i++) {
const a = placed[i]
const b = placed[i + 1]
if (b.contig !== a.contig) continue
const row = Math.min(a.bin, b.bin)
const col = Math.max(a.bin, b.bin)
out.push({
pair: { row, col, separation: col - row },
bounds: a.bounds,
label: `${a.label} × ${b.label}`,
rank: Number.NaN,
span: col - row,
centreDistance: Math.abs(b.exact - a.exact),
rows: { from: row, to: row + 1 },
cols: { from: col, to: col + 1 },
})
}
return out
}
/**
* Every pair of a feature set on one sequence, inside the separation range.
*
* The sliding window is the whole of the implementation and the whole of what
* makes it affordable: the set is sorted by `(contig, start)` already (§8.4), so
* for each feature the partners are a contiguous run of the ones after it, and
* the run ends at the first one past `maxSep`. That makes the cost proportional
* to the number of pairs **produced** rather than to `n²` — which matters,
* because the two differ by four orders of magnitude on a genome-wide peak set
* with a two-hundred-bin reach.
*
* **The range is in bins and the filter is on the pair's own bin offset**, which
* is the thing the pileup is a picture of: `pair.separation` is `col − row`, the
* separation bands are bin offsets, and the diagonal a pair is dropped for
* sitting on is one bin wide. The base-pair reading this replaces compared the
* distance between two feature centres, so two features 30 kb apart were inside
* a 20 kb floor on a 50 kb map and in the same bin — a pair on the diagonal,
* which the last guard in the loop had to catch separately. The offsets are
* monotone in the centres, so the `break` that makes the walk affordable still
* holds.
*
* Past {@link MAX_GENERATED_PAIRS} the pairs are **sampled rather than
* truncated**, and the distinction is the one §18.8 draws about the cap this
* task used to have on its anchors: taking the first two hundred thousand is
* taking whichever sequences sort first, and a pileup of chromosome I under a
* caption saying *200,000 pairs* is a change of subject. Reservoir sampling with
* the run's own seed gives a uniform sample of the whole set, which is a smaller
* version of the same figure rather than a different one.
*/
function generatedPairs(job: JobContext, set: FeatureSet): Anchor[] {
const minSep = Math.max(0, Math.round(numParam(job.params, 'minSep', 2)))
const maxSep = Math.max(minSep + 1, Math.round(numParam(job.params, 'maxSep', 200)))
job.note(
`Pairs ${minSep.toLocaleString()} to ${maxSep.toLocaleString()} bins apart, which is ` +
`${shortBp(minSep * job.mc.binsize)} to ${shortBp(maxSep * job.mc.binsize)} at this ` +
`file’s ${job.mc.binsize.toLocaleString()} bp resolution.`,
)
const rng = makeRandom(Math.round(numParam(job.params, 'seed', 1)))
const placed = placedFeatures(job, set)
const out: Anchor[] = []
let seen = 0
const offer = (anchor: Anchor): void => {
seen++
if (out.length < MAX_GENERATED_PAIRS) {
out.push(anchor)
return
}
// Reservoir: the `k`-th candidate replaces a held one with probability
// `cap / k`, which leaves every candidate equally likely to survive.
const at = Math.floor(rng() * seen)
if (at < MAX_GENERATED_PAIRS) out[at] = anchor
}
for (let i = 0; i < placed.length; i++) {
const a = placed[i]
for (let j = i + 1; j < placed.length; j++) {
const b = placed[j]
if (b.contig !== a.contig) break
const row = Math.min(a.bin, b.bin)
const col = Math.max(a.bin, b.bin)
const separation = col - row
if (separation > maxSep) break
if (separation < minSep) continue
// A pair in one bin is not an off-diagonal pileup — it is the diagonal,
// which is the largest thing in the matrix and would dominate every cell
// it contributed to. Only reachable at a floor of zero now that the floor
// is in bins; kept, because zero is a number a reader can type.
if (row === col) continue
offer({
pair: { row, col, separation: col - row },
bounds: a.bounds,
label: `${a.label} × ${b.label}`,
rank: Number.NaN,
span: col - row,
// One bin each, like a loaded pair's anchors: what is being asked about
// is the contact between two places, and the scaled mode's rectangle
// between them is built from these two spans by the same code path.
rows: { from: row, to: row + 1 },
cols: { from: col, to: col + 1 },
})
}
}
if (seen > out.length) {
job.note(
`${seen.toLocaleString()} pairs are inside the separation range and ` +
`${out.length.toLocaleString()} were sampled, uniformly and with this run’s seed. ` +
'Narrow the range for all of them.',
)
}
return out
}
/**
* The anchors split into `bins` log-spaced separation bands.
*
* **Log-spaced, on the quantiles of nothing** — the edges come from the observed
* minimum and maximum separation rather than from equal counts — because the
* quantity being divided is a separation, and contact falls off with it as a
* power law. Equal-count bins over a set whose separations pile up at one end
* produce a first band spanning a decade and a last one spanning a factor of
* two, and the pileups are then not comparable with each other, which is the
* entire reason for splitting them.
*
* Empty bands are dropped rather than drawn empty, so a set with a gap in its
* separations produces fewer panels than asked for rather than a blank one.
*/
function bySeparation(anchors: Anchor[], bins: number): Anchor[][] {
let lo = Number.POSITIVE_INFINITY
let hi = 0
for (const anchor of anchors) {
const separation = anchor.pair.separation
if (separation <= 0) continue
if (separation < lo) lo = separation
if (separation > hi) hi = separation
}
if (!Number.isFinite(lo) || hi <= lo) return [anchors]
const logLo = Math.log(lo)
const step = (Math.log(hi) - logLo) / bins
const out: Anchor[][] = Array.from({ length: bins }, () => [])
for (const anchor of anchors) {
const separation = anchor.pair.separation
// A pair whose two anchors land in one bin has no separation to band by, and
// it goes in the nearest band rather than being dropped: §18.8's rule is
// that every anchor is piled up, and a run that quietly discarded the
// shortest-range pairs would be weakest exactly where the signal is
// strongest.
const at =
separation <= 0
? 0
: Math.min(bins - 1, Math.floor((Math.log(separation) - logLo) / step))
out[at].push(anchor)
}
return out.filter((band) => band.length > 0)
}
/** What each band is called, in base pairs, from the separations it holds. */
function separationNames(bands: Anchor[][], toBp: (bins: number) => number): string[] {
return bands.map((band) => {
let lo = Number.POSITIVE_INFINITY
let hi = 0
for (const anchor of band) {
const separation = anchor.pair.separation
if (separation < lo) lo = separation
if (separation > hi) hi = separation
}
// The band's **own** extremes rather than the edges it was cut at, so the
// label says what is in it rather than what could have been.
return `${shortBp(toBp(Math.max(0, lo)))}–${shortBp(toBp(hi))}`
})
}
/** `1.2 Mb`, `340 kb`, `900 bp` — the band label, short enough to be a title. */
function shortBp(bp: number): string {
if (!Number.isFinite(bp)) return '?'
if (bp >= 1e6) return `${(bp / 1e6).toFixed(bp >= 1e7 ? 0 : 1)} Mb`
if (bp >= 1e3) return `${(bp / 1e3).toFixed(bp >= 1e4 ? 0 : 1)} kb`
return `${Math.round(bp)} bp`
}
|