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 | 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 6x 6x 6x 8x 8x 7x 7x 8x 6x 6x 5x 5x 7x 7x 7x 7x 1x 1x 1x 6x 7x 7x 7x 7x 2x 3x 3x 2x 2x 3x 1x 1x 1x 1x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 2x 2x 38x 38x 38x 38x 2x 2x 2x 2x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 2x 4x 4x 4x 4x 4x 4x 4x 4x 4x 4x 4x 4x 4x 4x 4x 4x 4x 4x 4x 4x 6x 6x 6x 6x 6x 6x 6x 6x 6x 1x 1x 6x 4x 9x 9x 9x 9x 9x 9x 9x 9x 9x 9x 4x 4x 4x 1x 1x 1x 1x 4x 4x 1x 1x 4x 4x 6x 6x 6x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 1x 1x 1x 2x 2x 34x 34x 34x 43x 43x 43x 34x 34x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 17x 15x 15x 15x 15x 19x 19x 19x 19x 14x 14x 14x 14x 280x 280x 241x 241x 241x 241x 280x 280x 14x 14x 241x 241x 241x 14x 14x 14x 14x 14x 5x 5x 5x 5x 7x 7x 420x 420x 420x 420x 7x 7x 5x 19x 5x 19x 19x 5x 5x 5x 7x 9x 9x 9x 9x 9x 9x 9x 9x 9x 9x 9x 9x 9x 660x 660x 660x 660x 51600x 51600x 51600x 51600x 51600x 51600x 51600x 51180x 51180x 51180x 51600x 660x 9x 7x 19x 5x 5x 5x 5x 5x 26x 26x 26x 26x 26x 26x 26x 26x 26x 21x 21x 6x 19x 9x 9x 7x 9x 21x 21x 9x 9x 9x 9x 9x 7x 9x 9x 3x 3x 9x 9x 9x 2x 21x 21x 21x 21x 21x 21x 21x 21x 21x 21x 21x 2x 2x 2x 2x 2x 2x 21x 21x 21x 19x 21x 21x 2x 2x 2x 2x 19x 19x 19x 19x 19x 19x 21x 661x 661x 661x 661x 661x 661x 661x 661x 661x 661x 661x 661x 661x 21x 21x 14x 241x 241x 20x 20x 20x 1x 14x 19x 21x 19x 19x 21x 21x 21x 19x 19x 21x 19x 19x 19x 19x 19x 19x 21x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 919x 919x 919x 919x 919x 919x 919x 919x 919x 919x 919x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 404x 404x 404x 404x 404x 404x 404x 404x 404x 404x 404x 404x 17x 17x 17x 17x 17x 17x 17x 17x 17x 21x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 160x 160x 160x 160x 2x 2x 2x 2x 200x 200x 200x 200x 200x 200x 200x 200x 2x 2x 2x 2x 2x 2x 2x 2x 1x 1x 1x 2x 2x 2x 2x 2x 19x 19x 21x 659x 659x 659x 19x 19x 19x 19x 15x 12x 3x 4x 21x 2x 2x 2x 2x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 21x 21x 21x 21x 21x 21x 21x 21x 21x 21x 21x 17x 17x 17x 17x 17x 604x 604x 604x 17x 17x 17x 17x 17x 14x 17x 3x 3x 14x 15x 241x 241x 241x 241x 241x 240x 240x 240x 241x 241x 14x 14x 17x 17x 17x 17x 17x 619x 619x 34532x 34532x 34532x 34532x 34532x 34532x 34532x 34532x 34532x 619x 17x 17x 14x 14x 14x 14x 14x 14x 14x 17x 17x 17x 17x 619x 34532x 34532x 34532x 34532x 34532x 619x 17x 17x 42x 42x 42x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 14x 3x 17x 17x 17x 17x 17x 17x 17x 17x 17x 3x 14x 14x 14x 14x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 2x 2x 2x 2x 2x 40x 40x 380x 380x 380x 380x 380x 380x 380x 380x 380x 380x 40x 2x 2x 19x 19x 19x 661x 661x 1547x 1522x 1522x 1547x 659x 659x 659x 1547x 661x 661x 19x 19x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 40x 380x 380x 380x 380x 380x 380x 40x 2x 2x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 619x 34532x 34532x 8699x 8699x 34532x 34532x 619x 17x 17x 9079x 9079x 9079x 9079x | /**
* The runners behind `tasks/structure.ts` (§25.6).
*
* The map is read one square per sequence, at the map's bin size, through
* the same road every other map task takes (`squareFor`); the model comes
* in as the feature set the structure row is, with its `model` of places.
* Everything numeric is `structure/versusMap.ts`, held still by its tests.
*/
import { structureFromBeads, type ModelBead } from '../../annotations/g3d'
import type { FeatureSet, StructureModel } from '../../annotations/types'
import { DEFAULT_COUNT_COLUMN, DEFAULT_WEIGHT_COLUMN } from '../../metacooler/matrix'
import { parseRegionString, splitRegionList } from '../../metacooler/range'
import { logBinCurve, logLogSlope } from '../expected'
import {
boolParam,
chosenInput,
contigIdFor,
missingInput,
numParam,
strParam,
type JobContext,
type TaskRunner,
} from '../job'
import type {
CurveOutput,
CurveSeries,
IntervalsOutput,
MatrixOutput,
ProfileOutput,
TableOutput,
TaskOutput,
} from '../types'
import {
dispersionAt,
fitDispersion,
likelihoodWalk,
mdsWalk,
momentsBySeparation,
randomStart,
storeWishes,
type Dispersion,
type MdsFit,
type NegBinomFit,
type SeparationMoments,
} from '../../structure/negbinom'
import { numpyRandom } from '../random'
import { openFitThreads } from '../likelihoodThreads'
import {
classicalMdsSweeps,
compactInPlace,
distancesOverContacts,
fillDistancesInPlace,
normaliseCoordinates,
smacofStep,
spearman,
stressOf,
type Coordinates,
type DenseMatrix,
type WeightMode,
} from '../../structure/reconstruct'
import {
agreement,
beadsOf,
binBeads,
centroid,
contactsBySeparation,
distanceBySeparation,
distanceFrom,
distanceMatrix,
inverseDistances,
localCompaction,
paintProfile,
} from '../../structure/versusMap'
import { blockCovers, spanOfContig, squareFor, usableBlock, type Span } from './blocks'
import { formatBp } from './labels'
import { regionLabel } from '../../genome/assembly'
/**
* The bar's share of pastis' MDS start, in the units the likelihood's steps are.
*
* The MDS stops on pastis' tolerance rather than on a count, so it has no
* denominator to walk; it usually stops within a hundred steps, so the first
* hundred units are its and anything after them waits at the last.
*/
const MDS_SHARE = 100
/** What a run says about why the optimiser stopped. */
const STOP_WORDS: Record<NegBinomFit['stop'], string> = {
gradient: 'the gradient: a maximum of the likelihood',
value: 'pastis’ tolerance: a step bought less than a double can hold',
steps: 'the step budget — raise it for a closer fit',
search: 'a line search that found nowhere better to go',
}
const NO_STRUCTURE = {
unchosen: 'No structure was chosen. Load a g3d or bed3d and pick it under Structure.',
gone: 'The structure that was chosen is no longer in the list.',
}
/** The sequences the model and the map share, with the map's span of each. */
function sharedSequences(job: JobContext, set: FeatureSet): Array<{ name: string; cid: number; span: Span }> {
const out: Array<{ name: string; cid: number; span: Span }> = []
for (const name of set.contigs) {
const cid = job.mc.chromnames.indexOf(name)
if (cid < 0) continue
const [from, to] = job.mc.contigExtent(cid)
if (to > from) out.push({ name, cid, span: { from, to } })
}
return out
}
/** The haplotypes to run over: each of a phased file, or the one of a haploid. */
function haplotypesOf(model: StructureModel): Array<{ index: number; label: string }> {
if (model.haplotypes.length <= 1) return [{ index: -1, label: '' }]
return model.haplotypes.map((label, index) => ({ index, label }))
}
function structureOf(job: JobContext): { set: FeatureSet; model: StructureModel; label: string } | null {
const found = chosenInput(job.inputs.features, job.params, 'structure')
const missing = missingInput(found, NO_STRUCTURE)
if (missing !== null || found.value === undefined) {
job.note(missing ?? NO_STRUCTURE.unchosen)
return null
}
const set = found.value
if (set.model === undefined) {
job.note('The chosen row is not a structure: it has no coordinates to read.')
return null
}
const id = typeof job.params.structure === 'string' ? job.params.structure : ''
return { set, model: set.model, label: job.inputs.labels?.[id] ?? 'structure' }
}
export const runStructureMap: TaskRunner = async (job) => {
const held = structureOf(job)
if (held === null) return []
const { set, model, label } = held
const sequences = sharedSequences(job, set)
if (sequences.length === 0) {
job.note('The model and the map share no sequence name. Check the names in the Tracks card.')
return []
}
const balance = boolParam(job.params, 'balance', false)
const value = job.framing.value === '' ? DEFAULT_COUNT_COLUMN : job.framing.value
const alpha = numParam(job.params, 'alpha', 3)
// The first of what the box names: the box is seeded with every sequence
// on screen (§25.6), and the picture is of one.
const wanted = (splitRegionList(strParam(job.params, 'region', ''))[0] ?? '').split(':')[0].trim()
const pictured = sequences.find((sequence) => sequence.name === wanted) ?? sequences[0]
const haplotypes = haplotypesOf(model)
const binsize = job.mc.binsize
const outputs: TaskOutput[] = []
const rows: TableOutput['rows'] = []
let done = 0
for (const sequence of sequences) {
job.throwIfCancelled()
job.progress(done++, sequences.length, sequence.name)
const bins = sequence.span.to - sequence.span.from
const square = await squareFor(job, sequence.span, {
block: usableBlock(job, [sequence.span], balance),
balance,
value,
})
for (const haplotype of haplotypes) {
const beads = beadsOf(set, sequence.name, haplotype.index)
if (beads.length === 0) continue
const points = binBeads(set, model, beads, binsize, bins)
const distances = distanceMatrix(points)
const fit = agreement(distances, square.values, bins)
rows.push({
sequence: sequence.name,
haplotype: haplotype.label,
bins: Array.from(points.present).filter((p) => p === 1).length,
pairs: fit.pairs,
rho: fit.rho,
})
if (sequence !== pictured) continue
const suffix = haplotype.label === '' ? '' : ` · ${haplotype.label}`
const ticks = Array.from({ length: bins }, (_, i) => formatBp(i * binsize))
outputs.push(
matrixOf(`structure-distance-${haplotype.index}`, `Model distances · ${sequence.name}${suffix}`, bins, distances, ticks, 'viridis'),
matrixOf(
`structure-predicted-${haplotype.index}`,
`Predicted contacts 1/d^${alpha} · ${sequence.name}${suffix}`,
bins,
inverseDistances(distances, alpha),
ticks,
'fall',
),
)
outputs.push(...curvesOf(sequence.name, suffix, haplotype.index, distances, square.values, bins, binsize))
}
}
job.progress(sequences.length, sequences.length)
const table: TableOutput = {
kind: 'table',
id: 'structure-agreement',
label: `Model versus map · ${label}`,
columns: [
{ key: 'sequence', label: 'Sequence', kind: 'text' },
...(haplotypes.length > 1 ? [{ key: 'haplotype', label: 'Haplotype', kind: 'text' as const }] : []),
{ key: 'bins', label: 'Bins modelled', kind: 'number', digits: 0 },
{ key: 'pairs', label: 'Pairs compared', kind: 'number', digits: 0 },
{
key: 'rho',
label: 'Spearman ρ (1/d vs contacts)',
kind: 'number',
digits: 3,
hint: 'Rank correlation between the inverse model distance and the observed contacts, over pairs at least two bins apart. 1 is a model that orders every pair as the map does.',
},
],
rows,
}
outputs.push(table)
return outputs
}
function matrixOf(
id: string,
label: string,
n: number,
values: Float32Array,
ticks: string[],
colormap: string,
): MatrixOutput {
return {
kind: 'matrix',
id,
label,
nrows: n,
ncols: n,
values,
xLabel: 'Position',
yLabel: 'Position',
xTicks: ticks,
yTicks: ticks,
colormap,
}
}
/** Distance by separation and the map's own P(s) over the same bins, log-binned, with the slope. */
function curvesOf(
name: string,
suffix: string,
haplotype: number,
distances: Float32Array,
contacts: Float32Array,
bins: number,
binsize: number,
): CurveOutput[] {
const spatial = distanceBySeparation(distances, bins)
const observed = contactsBySeparation(contacts, bins)
const toSeries = (values: Float64Array, seriesName: string): CurveSeries | null => {
const curve = { maxOffset: bins - 1, sum: new Float64Array(bins), count: new Float64Array(bins), mean: values }
for (let s = 1; s < bins; s++) {
if (Number.isNaN(values[s])) continue
curve.sum[s] = values[s]
curve.count[s] = 1
}
const binned = logBinCurve(curve, binsize, 1.25, 1)
if (binned.x.length < 2) return null
return { name: seriesName, x: binned.x, y: binned.y, n: bins }
}
const out: CurveOutput[] = []
const spatialSeries = toSeries(spatial, `${name}${suffix}`)
if (spatialSeries !== null) {
const slope = logLogSlope(spatialSeries.x, spatialSeries.y)
out.push({
kind: 'curve',
id: `structure-dvs-${haplotype}`,
label: `Model distance by separation · ${name}${suffix}`,
series: [spatialSeries],
xLabel: 'Genomic separation (bp)',
yLabel: 'Mean distance (model units)',
xScale: 'log',
yScale: 'log',
})
if (slope.x.length > 0) {
out.push({
kind: 'curve',
id: `structure-dvs-slope-${haplotype}`,
label: `Log-log slope of the model’s distance · ${name}${suffix}`,
series: [{ name: `${name}${suffix}`, x: slope.x, y: slope.y }],
xLabel: 'Genomic separation (bp)',
yLabel: 'd log d / d log s',
xScale: 'log',
yScale: 'linear',
})
}
}
const observedSeries = toSeries(observed, `${name} · map`)
if (observedSeries !== null) {
out.push({
kind: 'curve',
id: `structure-ps-${haplotype}`,
label: `The map’s P(s) over the same bins · ${name}`,
series: [observedSeries],
xLabel: 'Genomic separation (bp)',
yLabel: 'Mean contact per bin pair',
xScale: 'log',
yScale: 'log',
})
}
return out
}
export const runSpatialTracks: TaskRunner = async (job) => {
const held = structureOf(job)
if (held === null) return []
const { set, model, label } = held
const sequences = sharedSequences(job, set)
if (sequences.length === 0) {
job.note('The model and the map share no sequence name. Check the names in the Tracks card.')
return []
}
const binsize = job.mc.binsize
const nBins = job.mc.nBins
const window = numParam(job.params, 'window', 5)
const locusText = strParam(job.params, 'locus', '').trim()
const haplotypes = haplotypesOf(model)
const outputs: TaskOutput[] = []
for (const haplotype of haplotypes) {
const suffix = haplotype.label === '' ? '' : ` · ${haplotype.label}`
const all = sequences.flatMap((sequence) => beadsOf(set, sequence.name, haplotype.index))
const centre = centroid(model, all)
const radial = new Float32Array(nBins).fill(Number.NaN)
const compaction = new Float32Array(nBins).fill(Number.NaN)
const toLocus = new Float32Array(nBins).fill(Number.NaN)
const anchor = locusText === '' ? null : anchorPoint(set, model, locusText, haplotype.index)
for (const sequence of sequences) {
job.throwIfCancelled()
const beads = beadsOf(set, sequence.name, haplotype.index)
if (beads.length === 0) continue
const bins = sequence.span.to - sequence.span.from
if (centre !== null) {
paintProfile(set, beads, distanceFrom(model, beads, centre), binsize, sequence.span.from, bins, radial)
}
paintProfile(set, beads, localCompaction(model, beads, window), binsize, sequence.span.from, bins, compaction)
if (anchor !== null) {
paintProfile(set, beads, distanceFrom(model, beads, anchor), binsize, sequence.span.from, bins, toLocus)
}
}
const profile = (id: string, name: string, values: Float32Array, colour: string): ProfileOutput => ({
kind: 'profile',
id: `${id}-${haplotype.index}`,
label: `${name}${suffix} · ${label}`,
values,
extent: [0, nBins],
binsize,
unit: 'model units',
colour,
group: 'structure',
})
if (centre !== null) outputs.push(profile('structure-radial', 'Radial position', radial, '#1f5fa9'))
outputs.push(profile('structure-compaction', `Local compaction (±${Math.round(window)} beads)`, compaction, '#0f8a5f'))
if (anchor !== null) {
// *The centre of*, because the anchor is a centroid: a 400 kb window named
// bare read as though the distance were to a point.
outputs.push(
profile('structure-distance', `Distance to the centre of ${locusName(locusText)}`, toLocus, '#c0392b'),
)
}
}
if (locusText !== '' && !outputs.some((output) => output.id.startsWith('structure-distance'))) {
job.note(`No bead of the model lies at ${locusName(locusText)}, so the distance track was not made.`)
}
return outputs
}
/**
* A locus as a reader's own coordinates, for a name that will be read.
*
* What a runner is handed is cooler's grammar — `IV:1000000-1400000`, which
* `runnerParams` made of what was typed or picked — and naming a structure row
* and three lanes with it put the one string on the Structures card that had no
* separators directly under a framing line that did, a base away from it. So
* the name is §26.1's display form, `IV:1,000,001-1,400,000`, through the
* helper every locus box uses; a box that takes several names each part, in
* the order given; a bare sequence or genome name has no coordinates and comes
* back as itself.
*/
function locusName(machine: string): string {
return splitRegionList(machine).map(regionLabel).join(', ')
}
/**
* The model's point at a locus — 0-based and half-open, as a runner is handed
* one (`runnerParams`) — the centroid of every bead the locus covers, which is
* what the tracks below call its centre.
*/
function anchorPoint(
set: FeatureSet,
model: StructureModel,
text: string,
haplotype: number,
): [number, number, number] | null {
const [name, rest] = text.split(':')
const contig = name?.trim() ?? ''
const beads = beadsOf(set, contig, haplotype)
if (beads.length === 0) return null
const range = (rest ?? '').replace(/,/g, '').split(/[-–]/)
const start = Number(range[0])
const end = range.length > 1 ? Number(range[1]) : start
if (!Number.isFinite(start)) return centroid(model, beads)
const from = Math.min(start, end)
const to = Math.max(start, end)
const hit = beads.filter((i) => set.end[i] > from && set.start[i] < to)
return centroid(model, hit.length > 0 ? hit : beads.filter((i) => set.start[i] <= from && set.end[i] > from))
}
// ---------------------------------------------------------------------------
// A model from the map
// ---------------------------------------------------------------------------
/**
* The contacts a model is solved over: one cell per bead, each one bin, with
* the square of contacts between every two.
*
* `binFrom`/`binTo` name each cell's bins, since the cells need not be one
* contiguous run — several sequences named leave the ones between them out,
* and the whole file is every sequence's run one after the other. A cell is
* always exactly one bin of the selected resolution, whichever way the
* contacts arrive; the pair is kept rather than an index because the bins are
* not contiguous.
*/
interface Cells {
n: number
binFrom: Int32Array
binTo: Int32Array
values: Float32Array
}
/** The sequences of the file as bin runs, for placing a cell on one. */
function contigRuns(job: JobContext): Array<{ cid: number; from: number; to: number }> {
const out: Array<{ cid: number; from: number; to: number }> = []
for (let cid = 0; cid < job.mc.nContigs; cid++) {
const [from, to] = job.mc.contigExtent(cid)
if (to > from) out.push({ cid, from, to })
}
return out
}
/**
* Where the contacts come from: the view's block as drawn, or the file read
* and pooled.
*
* The block is what the reader is looking at, and it is taken as it is —
* pooled cells included, which is the one place `binsPerCell > 1` is
* honoured rather than refused — whenever it can honestly answer: it is
* square, balanced the way the run asks, and covers every span asked for.
* Otherwise the file is **read span by span**, one rectangle per pair of
* spans (so two sequences far apart in the file cost their two squares and
* the rectangle between them, never the sequences in between), and pooled
* `k` bins to a bead as it is read — which is how a whole genome at 1 kb
* becomes a model at 5 kb without a 12,000² block ever existing. The mean
* of the counted bins, so a balanced matrix pools as a balanced one.
*/
async function contactCells(
job: JobContext,
spans: Span[] | null,
balance: boolean,
value: string,
/** The `n²` matrices the method will hold, for the note if the first will not fit. */
matrices: number,
): Promise<{ cells: Cells; how: string } | { refusal: string }> {
const block = job.inputs.block
// **Never a pooled block**, which is `usableBlock`'s rule and every other
// task's (`run/blocks.ts`): a block drawn four bins to the cell is coarser
// than the file, and `binsPerCell` exists precisely so a task can decline to
// pretend otherwise. This one used to accept it, and so was the only task in
// the app whose answer came at the resolution of the *window* — a number out
// of a raster cap — rather than at the resolution the reader selected. A
// block at one bin to the cell is the file, so borrowing that one is free
// and changes nothing.
const square =
block !== undefined &&
block.binsPerCell === 1 &&
block.rowExtent[0] === block.colExtent[0] &&
block.rowExtent[1] === block.colExtent[1] &&
block.nrows === block.ncols &&
job.framing.balance === balance
const wholeFile = spans === null
const wanted: Span[] = wholeFile ? contigRuns(job).map((run) => ({ from: run.from, to: run.to })) : spans
if (wanted.length === 0) return { refusal: 'The file has no sequence to model.' }
// The block, where it covers what was asked for. The whole file is the
// block only where the block *is* the whole file.
if (square && block !== undefined && wanted.every((span) => blockCovers(block, span))) {
const rows: number[] = []
const binFrom: number[] = []
const binTo: number[] = []
for (let r = 0; r < block.nrows; r++) {
const from = block.rowExtent[0] + r
if (!wanted.some((span) => from < span.to && from + 1 > span.from)) continue
rows.push(r)
binFrom.push(from)
binTo.push(from + 1)
}
const n = rows.length
const values = allocate(n)
if (values === null) return { refusal: tooBig(n, matrices) }
for (let a = 0; a < n; a++) {
const at = rows[a] * block.ncols
for (let b = 0; b < n; b++) values[a * n + b] = block.values[at + rows[b]]
}
return {
cells: { n, binFrom: Int32Array.from(binFrom), binTo: Int32Array.from(binTo), values },
how: 'the block as drawn, one bin to a bead',
}
}
// The file, read and pooled.
const binFrom: number[] = []
const binTo: number[] = []
const cellsOfSpan: number[][] = []
for (const span of wanted) {
const own: number[] = []
for (let from = span.from; from < span.to; from++) {
own.push(binFrom.length)
binFrom.push(from)
binTo.push(from + 1)
}
cellsOfSpan.push(own)
}
const n = binFrom.length
if (n === 0) return { refusal: 'The framing holds no bins to model.' }
const values = allocate(n)
const counts = values === null ? null : allocate(n)
if (values === null || counts === null) return { refusal: tooBig(n, matrices) }
const total = (wanted.length * (wanted.length + 1)) / 2
let done = 0
for (let a = 0; a < wanted.length; a++) {
for (let b = a; b < wanted.length; b++) {
await job.breathe()
job.progress(done++, total + 2, `reading ${done} of ${total} blocks`)
const rowSpan = wanted[a]
const colSpan = wanted[b]
const dense = await job.mc.matrix({
rows: [rowSpan.from, rowSpan.to],
cols: [colSpan.from, colSpan.to],
balance,
value,
})
const rowCells = cellsOfSpan[a]
const colCells = cellsOfSpan[b]
for (let r = 0; r < dense.nrows; r++) {
const rc = rowCells[r]
if (rc === undefined) continue
const rowAt = r * dense.ncols
for (let c = 0; c < dense.ncols; c++) {
const v = dense.values[rowAt + c]
if (!Number.isFinite(v)) continue
const cc = colCells[c]
if (cc === undefined) continue
values[rc * n + cc] += v
counts[rc * n + cc] += 1
if (rc !== cc) {
values[cc * n + rc] += v
counts[cc * n + rc] += 1
}
}
}
}
}
for (let at = 0; at < values.length; at++) values[at] = counts[at] > 0 ? values[at] / counts[at] : Number.NaN
return {
cells: { n, binFrom: Int32Array.from(binFrom), binTo: Int32Array.from(binTo), values },
how: 'read from the file, one bin to a bead',
}
}
/**
* An `n × n` matrix of singles, or `null` where the browser will not give one.
*
* The replacement for the bead ceiling. A ceiling was a number this file
* guessed about somebody else's machine; asking and being refused is the same
* machine answering for itself. A `RangeError` — the length is past what a
* typed array may be — and an out-of-memory both arrive here as a throw.
*/
function allocate(n: number): Float32Array | null {
try {
// On memory other threads can read, where this one may share it — the
// command line; a page is not served with the headers that allow it — so
// that pastis-NB's helpers read the counts rather than copies of them
// (`analysis/likelihoodThreads.ts`). A refusal falls back to a buffer of
// this thread's own, which is the same matrix to everything but a helper.
const isolated = (globalThis as { crossOriginIsolated?: boolean }).crossOriginIsolated
if (typeof SharedArrayBuffer === 'function' && isolated !== false) {
try {
return new Float32Array(new SharedArrayBuffer(4 * n * n))
} catch {
// An ordinary buffer, then.
}
}
return new Float32Array(n * n)
} catch {
return null
}
}
/**
* The note for a browser that would not give the run its matrices: `matrices`
* of them, `n²` singles each — two under ShRec3D, one under pastis-NB.
*/
function tooBig(n: number, matrices: number): string {
const gb = (matrices * 4 * n * n) / 1e9
return (
`${n.toLocaleString()} beads needs about ${gb.toFixed(1)} GB ${matrices === 1 ? 'for its matrix' : 'of matrices'}, and this browser ` +
'would not give it. Select a coarser resolution, or name fewer sequences.'
)
}
/**
* The spans a typed region names — one per comma-separated part, a sequence
* or a locus — or `null` for everything; the missing name where the file
* lacks one.
*/
function regionSpans(job: JobContext, text: string): Span[] | null | { missing: string } {
if (text === '') return null
const out: Span[] = []
for (const trimmed of splitRegionList(text)) {
const span = regionSpan(job, trimmed)
if (span === undefined) return { missing: parseRegionString(trimmed)[0] }
if (span !== null) out.push(span)
}
return out.length === 0 ? null : out
}
/** The span one typed part names, or `null` for everything; `undefined` for a name the file lacks. */
function regionSpan(job: JobContext, text: string): Span | null | undefined {
if (text === '') return null
const [name, start, end] = parseRegionString(text)
const cid = contigIdFor(job.mc, name)
if (cid < 0) return undefined
const whole = spanOfContig(job.mc, cid)
if (whole === null) return undefined
if (start === null) return whole
const binsize = job.mc.binsize
const from = Math.max(whole.from, whole.from + Math.floor(start / binsize))
const to = end === null ? whole.to : Math.min(whole.to, whole.from + Math.ceil(end / binsize))
return to > from ? { from, to } : undefined
}
export const runReconstruct: TaskRunner = async (job) => {
const balance = boolParam(job.params, 'balance', false)
const value = job.framing.value === '' ? DEFAULT_COUNT_COLUMN : job.framing.value
const alpha = Math.max(0.5, numParam(job.params, 'alpha', 3))
const iterations = Math.max(0, Math.round(numParam(job.params, 'iterations', 150)))
const weightText = strParam(job.params, 'weights', 'contacts')
const weightMode: WeightMode = weightText === 'sqrt' || weightText === 'equal' ? weightText : 'contacts'
// The runner settles the method itself rather than trusting the control, for
// the reason every other one does: a job can reach the worker without passing
// through a form, and a stored session may name a method a later version
// dropped. Anything that is not ShRec3D is the default, which is pastis-NB.
const nb = strParam(job.params, 'method', 'nb') !== 'shrec3d'
// pastis' `random_state`: where its MDS places the beads before it starts.
const seed = Math.max(0, Math.round(numParam(job.params, 'seed', 0)))
const regionText = strParam(job.params, 'region', '').trim()
const spans = regionSpans(job, regionText)
if (spans !== null && !Array.isArray(spans)) {
job.note(
`This file has no sequence called ${spans.missing}, under that spelling or ` +
`any this viewer knows of. It has ` +
`${job.mc.chromnames.slice(0, 6).join(', ')}${job.mc.chromnames.length > 6 ? ', …' : ''}.`,
)
return []
}
// **pastis-NB reads the raw counts**, whatever the view is balanced with, and
// folds the weights into the mean instead (`biasForCells`). A balanced value
// is a count times two weights: not an integer, scaled by an arbitrary
// constant, and — the part that matters — not something `variance > mean` can
// be asked of, so a balanced map is never found overdispersed and the
// negative binomial quietly becomes a Poisson. ShRec3D is unaffected: a
// distance from `c^(−1/α)` is a ratio, and balancing is what makes those
// ratios comparable between bins.
const readBalanced = nb ? false : balance
const source = await contactCells(job, spans, readBalanced, value, nb ? 1 : 2)
if ('refusal' in source) {
job.note(source.refusal)
return []
}
const { cells, how } = source
if (job.framing.normalise !== null) {
job.note(
`The view is drawn ${job.framing.normalise}-normalised and the model is solved on the counts ` +
'underneath: no read can ask for that normalisation.',
)
}
if (nb && balance) {
job.note(
'The view is drawn balanced and pastis-NB is solved on the raw counts underneath, with the ' +
'file’s weights folded into each pair’s expected count instead (μ = bᵢ·bⱼ·d^(−α), ' +
'bᵢ = 1/wᵢ scaled to a mean of 1, as pastis scales them). A balanced value is a count times two weights, and overdispersion is a ' +
'statement about a count: handed one, the negative binomial would find none and become a ' +
'Poisson. The coverage is still corrected — it is corrected in the mean rather than in ' +
'the data (§25.6).',
)
}
// Each cell to its sequence: the one holding its first bin. A pooled cell
// that straddles two sequences is filed under the first and cut at its end,
// which is the honest reading of a cell the block itself drew across them.
const runs = contigRuns(job)
const binsize = job.mc.binsize
const group = new Int32Array(cells.n).fill(-1)
const index = new Int32Array(cells.n)
const beads: ModelBead[] = []
const perRun = new Map<number, number>()
for (let i = 0; i < cells.n; i++) {
const binFrom = cells.binFrom[i]
const run = runs.find((r) => r.from <= binFrom && binFrom < r.to)
if (run === undefined) continue
const name = job.mc.chromnames[run.cid]
const length = job.mc.chromsizes.get(name) ?? (run.to - run.from) * binsize
const start = (binFrom - run.from) * binsize
const end = Math.min(length, (Math.min(run.to, cells.binTo[i]) - run.from) * binsize)
group[i] = run.cid
const ordinal = perRun.get(run.cid) ?? 0
index[i] = ordinal
perRun.set(run.cid, ordinal + 1)
beads.push({ contig: name, start, end: Math.max(start + 1, end), x: 0, y: 0, z: 0 })
}
// The coverage bias, per cell, from the weights the file was balanced with.
// `null` on a file that has none, where every bead is equally well read as
// far as anything here can tell.
const bias = nb ? biasForCells(job, cells) : null
if (bias !== null) {
// A bin the balancer **rejected** — a `NaN` weight, from low mappability or
// no coverage — is one the balanced view of this map does not draw at all,
// because `applyWeights` makes its every cell `NaN`. Reading raw counts
// would put it back, and a model would gain beads that the same map
// balanced does not have. So it is blanked here and falls out through the
// same door a gap does, counted in the same note.
for (let i = 0; i < cells.n; i++) {
if (Number.isFinite(bias[i])) continue
for (let j = 0; j < cells.n; j++) {
cells.values[i * cells.n + j] = Number.NaN
cells.values[j * cells.n + i] = Number.NaN
}
}
}
// Beads the walk placed, in cell order — `beads[b]` is cell `cellOf[b]`.
const cellOf = Int32Array.from({ length: cells.n }, (_, i) => i).filter((i) => group[i] >= 0)
if (cellOf.length < 4) {
job.note('Fewer than four cells of the block lie on a sequence of this file; nothing to model.')
return []
}
const m = cellOf.length
// Which bead a cell became, for reading the walk's beads back by cell.
const beadOfCell = new Int32Array(cells.n).fill(-1)
for (let b = 0; b < cellOf.length; b++) beadOfCell[cellOf[b]] = b
job.progress(0, 1, nb ? 'the counts' : 'contacts to distances')
await job.breathe()
// The cells that are on a sequence **and** have a counted pair, in one list,
// and every matrix built once at that size.
//
// It used to take two subsets in turn — the counts over the placed cells,
// the distances over those, then the constrained ones out of *both* — which
// is six dense matrices for a job that needs two. Each is `n²`, so at ten
// thousand beads each is four hundred megabytes, and the difference is the
// difference between a model and a tab that dies making one.
const pick = constrainedIn(cells.values, cells.n, cellOf)
const dropped = m - pick.length
if (pick.length < 4) {
job.note('Fewer than four beads have any counted pair; nothing to model.')
return []
}
const n = pick.length
const keptGroup = Int32Array.from(pick, (i) => group[i])
const keptIndex = Int32Array.from(pick, (i) => index[i])
let coords: Coordinates
let methodRows: Array<Record<string, string>>
let fit: { rho: number; pairs: number }
if (nb) {
// **One matrix**: the counts, compacted onto the kept beads over the buffer
// they were read into. The likelihood reads them above the diagonal and
// pastis' MDS start reads its wishes below it, so neither needs a matrix of
// its own and no distance is ever filled in.
const counts = countsInPlace(cells.values, cells.n, pick)
// pastis scales the biases to a mean of one, which is what lets `β` be one:
// a distance of 1 is then where a pair of average coverage expects one
// contact, and the model's size is fitted rather than chosen.
const keptBias = bias === null ? null : meanOne(Float64Array.from(pick, (i) => bias[i]))
const moments = momentsBySeparation(counts, n, keptGroup, keptIndex, keptBias)
const dispersion = fitDispersion(moments)
const evidence = zerosOf(counts, n)
storeWishes(counts, n, keptBias, alpha)
// The helpers are started now, and lent to the MDS start and then to the
// likelihood; each evaluation of either is shared among them and this
// thread, and awaited: the walk asks, and is answered when every band is
// in. The same bits on any number of threads, so the thread count is said
// in the bar and nowhere a result is written.
const threads = openFitThreads(n, counts)
try {
// pastis' start: beads uniform in a cube, numpy's draws from the reader's
// seed, then its MDS to its own loose tolerance. Between every evaluation,
// not inside one: the cancel lands here (`job.breathe`).
const total = MDS_SHARE + iterations
const mds = threads.mds(counts)
const mdsThreads = mds.threads === 1 ? '' : ` · ${mds.threads} threads`
const mdsSteps = mdsWalk(n, randomStart(n, numpyRandom(seed)))
let stress = Number.POSITIVE_INFINITY
let mdsAsk = mdsSteps.next()
while (mdsAsk.done !== true) {
await job.breathe()
const { point, gradient, iteration, current } = mdsAsk.value
const value = await mds.evaluate(point, gradient)
stress = Math.min(stress, current, value)
job.progress(
Math.min(MDS_SHARE - 1, iteration),
total,
`MDS start · step ${iteration}${mdsThreads} · stress ${stress.toPrecision(3)}`,
)
mdsAsk = mdsSteps.next(value)
}
const start: MdsFit = mdsAsk.value
// The likelihood, on top of the model the MDS just made and at its scale:
// the wishes and the mean are in the same units, so nothing is rescaled
// between the two.
let likelihood: NegBinomFit | null = null
coords = start.coords
if (iterations > 0) {
const negBinom = threads.likelihood({ counts, n, bias: keptBias, dispersion, alpha })
const nbThreads = negBinom.threads === 1 ? '' : ` · ${negBinom.threads} threads`
const walk = likelihoodWalk(n, start.coords, { iterations })
let best = Number.POSITIVE_INFINITY
let ask = walk.next()
while (ask.done !== true) {
// Between every *evaluation*, not every iteration: a line search can
// be several passes over the pairs.
await job.breathe()
const { point, gradient, iteration, current } = ask.value
const value = await negBinom.evaluate(point, gradient)
best = Math.min(best, current, value)
const at = Math.min(iterations, iteration)
job.progress(
MDS_SHARE + at,
total,
`pastis-NB · step ${at} of at most ${iterations}${nbThreads} · −log L ${best.toExponential(3)}`,
)
ask = walk.next(value)
}
likelihood = ask.value
coords = likelihood.coords
}
job.progress(total, total)
fit = fitOfCounts(counts, n, coords, keptBias)
methodRows = nbRows(evidence, alpha, dispersion, moments, keptBias !== null, seed, start, likelihood, iterations)
} finally {
threads.close()
}
} else {
// ShRec3D wants a *weight* per pair, and a pair the map counted nothing for
// has no weight because it has no distance. The weights are taken while the
// counts are still counts, and the counts are then turned into the
// distances where they already lie — compacted onto the kept beads and put
// through the law, both over the buffer they are read from. At twelve
// thousand beads a matrix is half a gigabyte, so a third one is the
// difference between a model and a tab that dies making one.
const pairs = weightsOf(cells.values, cells.n, pick, weightMode)
if (pairs === null) return refuse(job, n)
compactInPlace(cells.values, cells.n, pick)
const kept = cells.values.subarray(0, n * n)
distancesOverContacts(kept, n, alpha)
// The whole of the work, as one scale the bar can walk: the sweeps of the
// classical start, then the refinement.
const sweeps = 80
const total = sweeps + iterations + 1
job.progress(0, total, 'classical scaling')
await job.breathe()
// Filled over itself: which pairs the map counted is what the `NaN`s said,
// and the weights say it too — zero for a pair it never saw — so the copy
// that kept them was a second matrix holding a fact already held.
fillDistancesInPlace(kept, n, keptGroup, keptIndex)
const scaling = classicalMdsSweeps(kept, n, sweeps)
let sweep = scaling.next()
while (sweep.done !== true) {
// Between every sweep, for the same two reasons the refinement breathes
// between steps: the bar moves, and a cancel has somewhere to land.
await job.breathe()
job.progress(sweep.value.sweep, total, `classical scaling · sweep ${sweep.value.sweep} of ${sweep.value.sweeps}`)
sweep = scaling.next()
}
coords = sweep.value
let stress = stressOf(kept, n, coords, pairs)
let steps = 0
for (; steps < iterations; steps++) {
// Between every step, not inside: the cancel lands here (`job.breathe`).
await job.breathe()
const before = stress
smacofStep(kept, n, coords, pairs, before)
const after = stressOf(kept, n, coords, pairs)
stress = after
job.progress(sweeps + steps + 1, total, `refining · step ${steps + 1} of ${iterations} · stress ${after.toPrecision(3)}`)
if (before - after < 1e-6 * before) {
steps++
break
}
}
job.progress(total, total)
fit = fitOf(kept, n, coords, pairs)
methodRows = [
{ what: 'Distance law', value: `d = c^(−1/${alpha})` },
{
what: 'Pairs weighted by',
value:
weightMode === 'contacts'
? 'the contact count'
: weightMode === 'sqrt'
? 'the square root of the contact count'
: 'nothing: every counted pair alike',
},
{ what: 'Refinement steps', value: `${steps.toLocaleString()}` },
// Three significant digits rather than an exponent: `3.994e-2` beside
// `0.881` reads as a number on a different scale, and a stress is read
// as the fraction it is. `toPrecision` keeps the exponent only for the
// magnitudes where a fixed point would be all zeros.
{ what: 'Weighted stress', value: stress.toPrecision(3) },
]
}
normaliseCoordinates(coords)
const placed: ModelBead[] = []
for (let b = 0; b < n; b++) {
const bead = beads[beadOfCell[pick[b]]]
placed.push({ ...bead, x: coords.x[b], y: coords.y[b], z: coords.z[b] })
}
const set = structureFromBeads(placed, binsize)
const sequences = new Set(placed.map((bead) => bead.contig))
const subject =
spans === null
? sequences.size === 1
? [...sequences][0]
: `${sequences.size} sequences`
: locusName(regionText)
if (dropped > 0) {
job.note(
`${dropped.toLocaleString()} of ${m.toLocaleString()} cells had no counted pair at all — ` +
'filtered bins, gaps — and are not in the model.',
)
}
const model: IntervalsOutput = {
kind: 'intervals',
id: 'reconstructed-model',
label: `3D model · ${subject}`,
payload: { kind: 'features', data: set },
entryKind: 'structure',
noun: { one: 'bead', many: 'beads' },
}
const table: TableOutput = {
kind: 'table',
id: 'reconstructed-fit',
label: `Model from the map · ${subject}`,
columns: [
{ key: 'what', label: 'Quantity', kind: 'text' },
{ key: 'value', label: 'Value', kind: 'text' },
],
rows: [
{ what: 'Method', value: nb ? 'pastis-NB · negative binomial likelihood' : 'ShRec3D · MDS with SMACOF' },
{ what: 'Contacts', value: how },
{ what: 'Beads', value: `${n.toLocaleString()} of ${cells.n.toLocaleString()} cells, ${formatBp(binsize)} each` },
{ what: 'Sequences', value: `${sequences.size.toLocaleString()}` },
{ what: 'Pairs counted', value: `${fit.pairs.toLocaleString()} of ${((n * (n - 1)) / 2).toLocaleString()}` },
...methodRows,
{ what: 'Spearman ρ (1/d vs contacts)', value: Number.isNaN(fit.rho) ? '—' : fit.rho.toFixed(3) },
],
}
return [model, table]
}
/**
* Each cell's coverage bias, `1/w`, from the weights the file was balanced
* with — or `null` where it carries none.
*
* This is where the balancing goes under pastis-NB, and it has to go *somewhere*
* because the counts themselves must stay raw ({@link NegBinomModel.bias}): the
* mean of a pair becomes `b_i·b_j·d^(−α)`, which is pastis' own handling of
* ICE biases and leaves the counts as counts. The runner scales them to a mean
* of one, as pastis' pipeline does, once the kept beads are known.
*
* A pooled cell is several bins, so its bias is the **mean of its bins'**, over
* the bins the balancer kept. A cell it kept none of gets `NaN`, which the
* caller reads as *this bead is not in the model* — the same answer the balanced
* road gives by turning the bin's every cell into a `NaN`.
*
* One column read, bounded to the bins in play, and no failure a run should die
* of: a file with no `/bins/weight` is an unbalanced file, and an unbalanced
* file's beads are as well read as each other as far as anything here can tell.
*/
function biasForCells(job: JobContext, cells: Cells): Float64Array | null {
if (cells.n === 0) return null
let lo = cells.binFrom[0]
let hi = cells.binTo[0]
for (let i = 1; i < cells.n; i++) {
if (cells.binFrom[i] < lo) lo = cells.binFrom[i]
if (cells.binTo[i] > hi) hi = cells.binTo[i]
}
if (!(hi > lo)) return null
let weights: Float64Array
try {
const column = job.mc.binColumn(DEFAULT_WEIGHT_COLUMN, lo, hi)
if (typeof column[0] === 'string') return null
weights = column as Float64Array
} catch {
return null
}
const out = new Float64Array(cells.n)
for (let i = 0; i < cells.n; i++) {
let sum = 0
let seen = 0
for (let bin = cells.binFrom[i]; bin < cells.binTo[i]; bin++) {
const w = weights[bin - lo]
if (!Number.isFinite(w) || !(w > 0)) continue
sum += 1 / w
seen++
}
out[i] = seen > 0 ? sum / seen : Number.NaN
}
return out
}
/**
* The counts over the kept beads, **with a zero and a gap still distinct**,
* compacted over the buffer they were read into.
*
* {@link weightsOf}'s question asked differently: whether a cell is finite,
* where that one asks whether it is positive. The difference is the whole of
* what pastis-NB reads a map for. A cell holding `0` is a pair the map looked
* at and counted nothing on — under a likelihood that is *evidence*, and the
* pair pushes its beads apart — while a `NaN` is a pair a filtered bin or a gap
* made unobservable, which says nothing and is skipped. ShRec3D cannot tell
* them apart because neither is a distance; that is why it has to fill.
*
* In place, so the method holds one `n²` matrix and not two: the read was
* already a private copy (`contactCells`), and nothing else wants it.
*/
function countsInPlace(values: Float32Array, n: number, keep: Int32Array): Float32Array {
compactInPlace(values, n, keep)
const m = keep.length
const out = values.subarray(0, m * m)
for (let a = 0; a < m; a++) {
out[a * m + a] = 0
for (let b = a + 1; b < m; b++) {
const p = out[a * m + b]
const q = out[b * m + a]
const okP = Number.isFinite(p)
const okQ = Number.isFinite(q)
const c = !okP && !okQ ? Number.NaN : okP && okQ ? (p + q) / 2 : okP ? p : q
// A balanced cell is a count times two positive weights and cannot be
// negative; a rounding at the bottom of one can be, and a negative count
// is not a thing the likelihood has a value for.
const kept = Number.isNaN(c) ? c : c > 0 ? c : 0
out[a * m + b] = kept
out[b * m + a] = kept
}
}
return out
}
/** The biases over their mean, as pastis' pipeline scales them (`bias /= bias.mean()`). */
function meanOne(bias: Float64Array): Float64Array {
let sum = 0
for (const b of bias) sum += b
const mean = sum / bias.length
if (mean > 0) for (let i = 0; i < bias.length; i++) bias[i] /= mean
return bias
}
/** The pairs the map could see, and how many of them it counted nothing on. */
function zerosOf(counts: DenseMatrix, n: number): { observed: number; zeros: number } {
let observed = 0
let zeros = 0
for (let i = 0; i < n; i++) {
for (let j = i + 1; j < n; j++) {
const c = counts[i * n + j]
if (!Number.isFinite(c)) continue
observed++
if (c === 0) zeros++
}
}
return { observed, zeros }
}
/** A number as a reader should see it in a sentence: three figures, no exponent where none is needed. */
function figure(value: number): string {
return value >= 100 ? value.toFixed(0) : value >= 10 ? value.toFixed(1) : value.toPrecision(3)
}
/**
* What a pastis-NB run has to say for itself, as rows of the fit table.
*
* Three of these are here because they are the only way a reader can tell
* whether the method did anything a Poisson would not have. **Zeros used as
* evidence** is the count of pairs ShRec3D would have had to invent a distance
* for. **Overdispersion** is the one curve pastis-NB adds over pastis-PM2,
* written out with the `r` it gives near the diagonal and far from it; where it
* reports *none* the two methods have converged and the reader should read the
* model as a Poisson's. **Stopped on** is whether the fit finished or ran out
* of steps, which on a big model is the difference between a maximum and a
* waypoint.
*/
function nbRows(
evidence: { observed: number; zeros: number },
alpha: number,
dispersion: Dispersion | null,
moments: SeparationMoments,
balanced: boolean,
seed: number,
start: MdsFit,
fit: NegBinomFit | null,
iterations: number,
): Array<Record<string, string>> {
const last = moments.mean.length - 1
const rows: Array<Record<string, string>> = [
{
what: 'Mean of a pair',
value: balanced ? `μ = bᵢ · bⱼ · d^(−${alpha})` : `μ = d^(−${alpha})`,
},
{
what: 'Counts',
value: balanced
? 'raw, with the file’s balancing weights folded into the mean as bᵢ = 1/wᵢ, scaled to a mean of 1'
: 'raw: this file carries no balancing weights',
},
{
what: 'Zeros used as evidence',
value: `${evidence.zeros.toLocaleString()} of the ${evidence.observed.toLocaleString()} pairs the map could see`,
},
{
what: 'Overdispersion (r)',
value:
dispersion === null
? 'none: these counts are Poisson, so this is pastis-PM2'
: `r = ${figure(Math.exp(dispersion.c0))} · λ^${dispersion.c1.toFixed(3)}, λ the mean, fitted over ` +
`${moments.mean.length.toLocaleString()} separations: ` +
`${figure(dispersionAt(dispersion, moments.mean[0]))} near the diagonal, ` +
`${figure(dispersionAt(dispersion, moments.mean[last]))} at ${moments.separation[last].toLocaleString()} bins ` +
'— a smaller r is more overdispersed',
},
{
what: 'MDS start',
value:
`pastis’ MDS from random positions, seed ${seed}: ${start.iterations.toLocaleString()} steps, ` +
`${start.evaluations.toLocaleString()} evaluations`,
},
]
if (fit === null) return rows
rows.push(
{
what: 'Likelihood steps',
value: `${fit.iterations.toLocaleString()} of at most ${iterations.toLocaleString()}, over ${fit.evaluations.toLocaleString()} evaluations`,
},
{ what: 'Stopped on', value: STOP_WORDS[fit.stop] },
// Less `Σ log c!`, which no structure can change and pastis leaves out too,
// so this is the very number its objective returns and can be negative.
{ what: '−log L (less Σ log c!, as pastis gives it)', value: fit.value.toExponential(4) },
)
return rows
}
/**
* The pairs' weights over the kept beads, or `null` where the matrix will not fit.
*
* Read out of the counts at their *own* indices, because this is what has to
* happen before the counts are compacted and turned into distances over
* themselves.
*/
function weightsOf(values: Float32Array, n: number, keep: Int32Array, mode: WeightMode): Float32Array | null {
const m = keep.length
const out = allocate(m)
if (out === null) return null
for (let a = 0; a < m; a++) {
const row = keep[a] * n
for (let b = a + 1; b < m; b++) {
const p = values[row + keep[b]]
const q = values[keep[b] * n + keep[a]]
const okP = Number.isFinite(p) && p > 0
const okQ = Number.isFinite(q) && q > 0
if (!okP && !okQ) continue
const count = okP && okQ ? (p + q) / 2 : okP ? p : q
const w = mode === 'contacts' ? count : mode === 'sqrt' ? Math.sqrt(count) : 1
out[a * m + b] = w
out[b * m + a] = w
}
}
return out
}
/** The note a run leaves when the browser will not give it the matrices it needs. */
function refuse(job: JobContext, n: number): TaskOutput[] {
job.note(tooBig(n, 2))
return []
}
/**
* The cells of `within` that have a counted pair with another of them.
*
* Asked of the counts rather than of the distances, which is what lets the
* distance matrix be built once at its final size: the power law turns a
* positive count into a distance and everything else into `NaN`, so "has a
* counted pair" is a question about the counts and needs no matrix to answer.
*/
function constrainedIn(values: Float32Array, n: number, within: Int32Array): Int32Array {
const out: number[] = []
for (const i of within) {
let any = false
for (const j of within) {
if (j === i) continue
const a = values[i * n + j]
const b = values[j * n + i]
if ((Number.isFinite(a) && a > 0) || (Number.isFinite(b) && b > 0)) {
any = true
break
}
}
if (any) out.push(i)
}
return Int32Array.from(out)
}
/**
* How the model orders the counted pairs against the map: Spearman's ρ between
* `1/d` in the model and the contact each pair had, over at most 200,000 pairs
* taken at an even stride so a big block does not sort millions. ShRec3D's
* reading, over the distances its counts became.
*/
function fitOf(
target: DenseMatrix,
n: number,
coords: Coordinates,
/** Which pairs the map counted: non-zero where it did. The filled matrix no longer says. */
weights: DenseMatrix,
): { rho: number; pairs: number } {
let pairs = 0
for (let i = 0; i < n; i++) for (let j = i + 1; j < n; j++) if (weights[i * n + j] > 0) pairs++
const stride = Math.max(1, Math.ceil(pairs / 200000))
const a: number[] = []
const b: number[] = []
let seen = 0
for (let i = 0; i < n; i++) {
for (let j = i + 1; j < n; j++) {
const delta = target[i * n + j]
if (!(weights[i * n + j] > 0)) continue
if (seen++ % stride !== 0) continue
a.push(inverseDistance(coords, i, j))
// The count is the distance undone; its rank is the count's rank, so
// `−delta` ranks as the count does and costs no `pow`.
b.push(-delta)
}
}
return { rho: spearman(a, b), pairs }
}
/**
* The same ρ under pastis-NB, read off the counts above the diagonal.
*
* **Against `raw/(bᵢbⱼ)`**, which is what makes the number comparable across
* the two methods. pastis-NB read raw counts and put the balancing weights into
* the mean, so the quantity its model is fitted against is `raw/(bᵢbⱼ)` —
* precisely the *balanced* value, since balancing is `raw·wᵢwⱼ` and `b = 1/w`,
* up to the one constant the biases were scaled by, which no rank sees.
* Ranking against the raw counts instead would score the model on coverage it
* was right to have divided out, and would report a worse fit the more uneven
* the file is. Given the bias, this is the same number *Model versus map*
* computes for a loaded model on a balanced map.
*/
function fitOfCounts(
counts: DenseMatrix,
n: number,
coords: Coordinates,
bias: Float64Array | null,
): { rho: number; pairs: number } {
let pairs = 0
for (let i = 0; i < n; i++) for (let j = i + 1; j < n; j++) if (counts[i * n + j] > 0) pairs++
const stride = Math.max(1, Math.ceil(pairs / 200000))
const a: number[] = []
const b: number[] = []
let seen = 0
for (let i = 0; i < n; i++) {
for (let j = i + 1; j < n; j++) {
const c = counts[i * n + j]
if (!(c > 0)) continue
if (seen++ % stride !== 0) continue
a.push(inverseDistance(coords, i, j))
b.push(bias === null ? c : c / (bias[i] * bias[j]))
}
}
return { rho: spearman(a, b), pairs }
}
function inverseDistance(coords: Coordinates, i: number, j: number): number {
const d = Math.hypot(coords.x[i] - coords.x[j], coords.y[i] - coords.y[j], coords.z[i] - coords.z[j])
return d > 0 ? 1 / d : Number.POSITIVE_INFINITY
}
|