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 | 2x 2x 2x 2x 2x 2x 2x 18x 18x 18x 18x 18x 18x 18x 18x 18x 18x 18x 18x 2x 19x 19x 1x 1x 1x 1x 1x 18x 18x 19x 19x 19x 19x 19x 19x 19x 10x 8x 8x 16x 3x 3x 1x 1x 1x 1x 1x 3x 18x 18x 18x 18x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 2x 5x 5x 8x 8x 8x 8x 8x 8x 8x 8x 4x 4x 8x 8x 8x 8x 4x 40x 40x 4x 8x 8x 8x 5x 5x 5x 5x 5x 5x 5x 5x 5x 5x 5x 5x 5x 48x 48x 48x 48x 48x 5x 5x 4x 4x 4x 4x 64x 24x 24x 24x 24x 24x 4x 1x 1x 1x 3x 3x 4x 48x 48x 48x 48x 3x 3x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 14x 8x 8x 8x 8x 8x 8x 8x 8x 8x 8x 8x 8x 8x 8x 8x 514x 514x 514x 8x 8x 8x 506x 506x 514x 514x 506x 514x 8x 8x 8x 8x 8x 8x 8x 8x 8x | /**
* The 4C task, run.
*
* Two paths, and the difference between them is the whole design of the task
* (`analysis/tasks/fourc.ts`) — on a cooler. **A metagenome always takes the
* second**, whatever the parameters say, because the contacts a MAG makes with
* the rest of the community are the reading such a file exists for and the cis
* half alone is one assembly pretending it was sequenced on its own.
*
* The cis path shares `bandSums` with
* `render/viewpoint.ts` — the *same function* the lane on the map uses, imported
* rather than reimplemented, so the exported bigWig and the strip above the
* picture cannot disagree about what a band sum is. The trans path cannot: there
* is no dense block over a whole metagenome to take a band out of, so it walks
* the pixels.
*
* **The walk stops early, and that is the point.** Pixels are ordered by
* `bin1`, and every pixel this profile needs has `bin1 < anchorEnd` — either
* because `bin1` is inside the anchor, or because `bin1` is before it and `bin2`
* lands in it. Past that point nothing can contribute, so `onChunk` returns
* `false` and the read ends there: on average half the file rather than all of
* it, which is exactly what §18.6 prices.
*
* **Except under the coverage correction**, which is a statement about every bin
* and therefore about every pixel. A marginal accumulated from half the file is
* half a marginal, and a profile divided by one would be corrected on the bins
* before the anchor and uncorrected after it — one quantity by the axis, two by
* the arithmetic, with nothing on the picture to say which bin is which. That is
* the failure {@link wholeProfile} already exists to prevent for the decay, and
* the answer is the same: the pass runs to the end, and the estimate says so.
*/
import { DEFAULT_COUNT_COLUMN } from '../../metacooler/matrix'
import { parseMetacoolerRange } from '../../metacooler/range'
import { regionLabel } from '../../genome/assembly'
import { ExpectedAccumulator, expectedAt } from '../expected'
import { boolParam, strParam, type JobContext, type TaskRunner } from '../job'
import type { ProfileOutput, TaskOutput } from '../types'
import { bandSums, ViewpointAccumulator } from '../viewpoint'
import type { BlockInput } from '../worker/protocol'
import { squareFor, usableBlock, validBins, type Span } from './blocks'
/** The anchor, resolved to absolute bins, and the scope "cis" means for it. */
interface Anchor {
band: Span
/** The sequence — or the whole MAG — a cis-only run stays inside. */
scope: Span
label: string
}
/**
* The anchor, and what *cis* means around it.
*
* A locus names a contig, so cis is that contig. A **genome** may also be named,
* and then cis is the whole MAG rather than whichever of its contigs the first
* bin happens to fall in — on a metagenome the second reading would silently
* answer a question about one contig of an organism when the reader asked about
* the organism, which is the distinction §3.2 exists to make.
*/
function resolveAnchor(job: JobContext, text: string): Anchor {
const mc = job.mc
const parsed = parseMetacoolerRange(mc, text)
const band: Span = {
from: parsed.extent[0],
to: Math.max(parsed.extent[0] + 1, parsed.extent[1]),
}
if (parsed.kind === 'genome') {
const [from, to] = mc.genomeExtent(parsed.label)
return { band, scope: { from, to }, label: parsed.label }
}
const cid = mc.coords.contigOfBin(band.from)
const [from, to] = mc.contigExtent(cid)
return { band, scope: { from, to }, label: mc.chromnames[cid] }
}
export const runFourC: TaskRunner = async (job) => {
const text = strParam(job.params, 'anchor', '').trim()
if (text === '') {
job.note(
'No anchor was chosen. Type a locus, or a whole sequence by name, or press “pick on map” ' +
'and drag along either axis.',
)
return []
}
const balance = boolParam(job.params, 'balance', false)
// **A metagenome is read whole, and raw.** Neither key is offered on such a
// file (`tasks/fourc.ts`) and neither is trusted here: a stored session, a
// form filled in on a cooler and a default all leave `cis` sitting at `true`,
// and honouring it would give back the profile a reader has just told us is
// the wrong one — one MAG against itself, with the rest of the community, the
// thing the map is *of*, left as gaps. `normalise` goes the same way: the
// decay of one assembly's contigs is not the expectation for contact between
// assemblies, so the ratio would carry two meanings in one track.
const whole = job.mc.hasGenomes
const cis = !whole && boolParam(job.params, 'cis', true)
// One control, four answers, read as two independent divisions — which is what
// they are. A metagenome is settled to neither, for the reason above.
const normalise = whole ? 'none' : strParam(job.params, 'normalise', 'none')
const overExpected = normalise === 'ps' || normalise === 'ps+coverage'
const overCoverage = normalise === 'coverage' || normalise === 'ps+coverage'
const value = job.framing.value === '' ? DEFAULT_COUNT_COLUMN : job.framing.value
const anchor = resolveAnchor(job, text)
const values = cis
? await cisProfile(job, anchor, balance, value, overExpected, overCoverage)
: await wholeProfile(job, anchor, balance, overExpected, overCoverage)
// Last, over whatever the profile came to be — counts or ratios, one half or
// both — so the gap is the same gap whichever divisions ran (§18.8).
const excluded = boolParam(job.params, 'excludeAnchor', false)
if (excluded) {
excludeAnchor(values, anchor.band)
if (values.every((each) => Number.isNaN(each))) {
job.note(
`The anchor covers everything this profile reads, so leaving it out leaves nothing` +
`${cis ? ' — turn off “cis only”, or name a narrower anchor.' : '.'}`,
)
}
}
const profile: ProfileOutput = {
kind: 'profile',
id: `fourc-${anchor.band.from}-${anchor.band.to}`,
// The suffix marks a *choice*, so a metagenome — where every 4C is over the
// whole file — does not carry it. There it would name the only thing the
// task does, on every track, which is a label that distinguishes nothing.
// The anchor as the reader reads it, `IV:1,000,001-1,010,000` (§26.1), and
// not as this runner is handed it: what reaches a runner is cooler's
// grammar (`runnerParams`), and a lane's name is read beside every other
// locus on the screen.
label:
`4C · ${regionLabel(text)}${cis || whole ? '' : ' (with trans)'}` +
`${overExpected ? ' · O/E' : ''}${overCoverage ? ' · per coverage' : ''}` +
`${excluded ? ' · anchor excluded' : ''}`,
values,
extent: [0, job.mc.nBins],
binsize: job.mc.binsize,
// A ratio has no unit, and calling one "contacts" would put a number
// centred on 1 under an axis labelled as a count. The coverage divisor is
// centred on one, so it does **not** cost the track its unit: a corrected
// count is still a count, at the depth the map has on average.
unit: overExpected ? '' : 'contacts',
}
const outputs: TaskOutput[] = [profile]
return outputs
}
/**
* The anchor's own bins, and one bin either side of them, as gaps (§18.8).
*
* Gaps and not zeros: nothing was measured away, and a zero is a claim that
* the anchor touches itself not at all. The bins are the profile's — absolute,
* as the anchor's band is — and a band at either end of the file stops there.
*/
export function excludeAnchor(values: Float32Array, band: Span): void {
values.fill(Number.NaN, Math.max(0, band.from - 1), Math.min(values.length, band.to + 1))
}
/**
* The whole file, with both halves of the profile against their own expectation.
*
* Without normalisation this is the pixel walk and nothing else. With it, the
* two halves have two different expectations and **both** have to be applied, or
* the track holds raw counts on the anchor's own sequence and ratios everywhere
* else — one quantity by the axis, two quantities by the arithmetic, and no
* mark on the picture to say which bin is which. That is what this used to do:
* `overTransMean` skipped the cis half on the stated grounds that the decay had
* already been divided out of it, which was true on the cis-only path and never
* on this one, where nothing had read a square at all.
*
* So the square gets read. It is one on-diagonal block for one sequence against
* a pass that is already touching half the pixels in the file, and it is only
* read when a reader has asked for the ratio.
*/
async function wholeProfile(
job: JobContext,
anchor: Anchor,
balance: boolean,
overExpected: boolean,
overCoverage: boolean,
): Promise<Float32Array> {
// Before the walk rather than after it, so the progress line runs forwards:
// `transProfile` reports its way through the file and this would otherwise
// send it back to the start once the reader had watched it finish.
// `DEFAULT_COUNT_COLUMN` and not `value`: {@link transProfile} walks the
// pixels' `count` whatever the map is drawn from, and says so. A cis half read
// from another column would put the two quantities back in one track by a
// second route — the one this function exists to close.
const scored = overExpected
? await cisProfile(job, anchor, balance, DEFAULT_COUNT_COLUMN, true)
: null
const marginals = overCoverage ? new Float64Array(job.mc.nBins) : null
const read = await transProfile(job, anchor, balance, marginals)
let out = scored === null ? read : overTransMean(job, read, anchor)
if (scored !== null) {
// The anchor's own sequence, over its own decay; everything else over the
// trans mean {@link overTransMean} just applied.
for (let bin = anchor.scope.from; bin < Math.min(anchor.scope.to, out.length); bin++) {
out[bin] = scored[bin]
}
}
// Last, and over the whole vector: the divisor is centred on one, so it
// corrects a raw profile and a ratio alike without moving either off its own
// scale. See {@link relativeCoverage}.
if (marginals !== null) out = overCoverageOf(job, out, marginals)
return out
}
/**
* A profile divided by the decay, cell by cell.
*
* The curve is accumulated from **the square this task has already read**, which
* is the anchor's own sequence — not the P(s) task's genome-wide one. A small
* chromosome divided by the average behaviour of the large ones comes out with a
* slope that is a property of the set rather than of this locus, and the slope
* is exactly what a reader turns this on to remove.
*
* A bin whose expectation is unknown becomes a gap rather than keeping its
* numerator: half of a ratio is not a ratio, and a track holding both raw counts
* and O/E values would be read as one quantity.
*/
function overDecay(sums: Float64Array, square: BlockInput, band: [number, number]): Float64Array {
// `addBlock` rather than a loop written here: it files each cell under the
// separation its *absolute* bins imply and counts a mirrored square exactly
// once per stored copy, which are the two things a second implementation of
// this gets wrong. It is also the one that cannot read a rectangular block
// down the wrong stride.
const accumulator = new ExpectedAccumulator(Math.max(0, square.nrows - 1))
accumulator.addBlock({
nrows: square.nrows,
ncols: square.ncols,
values: square.values,
rowExtent: [square.rowExtent[0], square.rowExtent[1]],
colExtent: [square.colExtent[0], square.colExtent[1]],
})
const curve = accumulator.finish()
// The anchor's own middle is where every separation is measured from. A band
// is usually a bin or two wide, so its centre is the honest single point.
const centre = (band[0] + band[1] - 1) / 2
const out = new Float64Array(sums.length)
for (let i = 0; i < sums.length; i++) {
const expected = expectedAt(curve, Math.abs(i - centre))
out[i] =
Number.isFinite(sums[i]) && Number.isFinite(expected) && expected !== 0
? sums[i] / expected
: Number.NaN
}
return out
}
/**
* The trans half of a profile, over its own mean.
*
* Two chromosomes have no separation between them, so there is no curve to read
* and the expectation is the mean of what this anchor reaches off its own
* sequence — the same rule the compartment chain applies to a trans block. The
* cis half is left exactly as it came in, because {@link wholeProfile} overwrites
* it with the decay-normalised one; both halves then come out as ratios against
* their own expectation and can be read on one axis.
*/
function overTransMean(job: JobContext, values: Float32Array, anchor: Anchor): Float32Array {
let total = 0
let seen = 0
for (let bin = 0; bin < values.length; bin++) {
if (bin >= anchor.scope.from && bin < anchor.scope.to) continue
const value = values[bin]
if (!Number.isFinite(value)) continue
total += value
seen++
}
if (seen === 0) {
job.note('Nothing was reached outside this sequence, so there is no trans expectation to divide by.')
return values
}
const mean = total / seen
const out = new Float32Array(values.length)
for (let bin = 0; bin < values.length; bin++) {
// Only what is outside the anchor's own sequence: the cis half is replaced
// wholesale by {@link wholeProfile} with its decay-normalised counterpart.
const inside = bin >= anchor.scope.from && bin < anchor.scope.to
const value = values[bin]
out[bin] = inside || !Number.isFinite(value) ? value : mean === 0 ? Number.NaN : value / mean
}
return out
}
/**
* The anchor's own sequence, from one square.
*
* `bandSums` takes the anchor as **block-local** indices on the axis the band is
* taken over, which is what the lane hands it too; converting once here is what
* keeps that function free of any idea where the block came from.
*/
async function cisProfile(
job: JobContext,
anchor: Anchor,
balance: boolean,
value: string,
overExpected = false,
overCoverage = false,
): Promise<Float32Array> {
job.progress(0, 1, `reading ${anchor.label}`)
// The view's block only where it holds the column being asked for.
// `usableBlock` vouches for the resolution and the balancing and not for the
// value column — it is whatever the map happens to be drawn from — so a caller
// that wants `count` off a map drawn from `expected` has to read.
const framed = job.framing.value === '' ? DEFAULT_COUNT_COLUMN : job.framing.value
const block = value === framed ? usableBlock(job, [anchor.scope], balance) : null
const square = await squareFor(job, anchor.scope, { block, balance, value })
job.throwIfCancelled()
const local: [number, number] = [
Math.max(0, anchor.band.from - anchor.scope.from),
Math.min(anchor.scope.to - anchor.scope.from, anchor.band.to - anchor.scope.from),
]
const sums = bandSums(
{
nrows: square.nrows,
ncols: square.ncols,
values: square.values,
rowExtent: [square.rowExtent[0], square.rowExtent[1]],
colExtent: [square.colExtent[0], square.colExtent[1]],
},
local,
'cols',
)
const scored = overExpected ? overDecay(sums, square, local) : sums
// Over the whole assembly, with everything outside this sequence left as a
// gap. Not zero: a cis-only run has said nothing at all about the other
// sequences, and a track claiming zero contact there would be a claim it did
// not make.
const out = new Float32Array(job.mc.nBins).fill(Number.NaN)
for (let i = 0; i < scored.length; i++) out[anchor.scope.from + i] = scored[i]
if (!overCoverage) return out
// **The coverage of this sequence, out of the square already in hand.** A cis
// profile is a statement about one sequence, so the depth to divide by is the
// depth *within* it — the column sums of the same square the profile came
// from. Reaching for a genome-wide marginal here would cost a pass over the
// whole file to correct a reading that never left one chromosome, and would
// divide by a number that includes contacts this profile did not look at.
const marginals = new Float64Array(job.mc.nBins)
for (let r = 0; r < square.nrows; r++) {
const base = r * square.ncols
for (let c = 0; c < square.ncols; c++) {
const cell = square.values[base + c]
if (Number.isFinite(cell)) marginals[anchor.scope.from + c] += cell
}
}
return overCoverageOf(job, out, marginals)
}
/**
* A profile divided by how much each bin is seen at all.
*
* The divisor is the bin's own total contact **relative to the mean** — see
* {@link relativeCoverage} — so this is a bias correction and not a change of
* units: a raw profile stays in contacts, an O/E ratio stays centred on one, and
* the axis label goes on meaning what it said.
*
* A bin with no coverage becomes a gap rather than keeping its numerator, which
* is the rule {@link overDecay} states for the same reason: half of a ratio is
* not a ratio, and it would be indistinguishable on the axis from the corrected
* values around it.
*/
function overCoverageOf(
job: JobContext,
values: Float32Array,
marginals: Float64Array,
): Float32Array {
const relative = relativeCoverage(marginals)
if (relative === null) {
job.note(
'No bin carried any coverage, so there was nothing to correct for and the profile is as ' +
'it was counted.',
)
return values
}
const out = new Float32Array(values.length).fill(Number.NaN)
let dropped = 0
for (let bin = 0; bin < values.length; bin++) {
const value = values[bin]
if (!Number.isFinite(value)) continue
const divisor = relative[bin]
if (!(divisor > 0)) {
dropped++
continue
}
out[bin] = value / divisor
}
if (dropped > 0) {
job.note(
`${dropped.toLocaleString()} ${dropped === 1 ? 'bin has' : 'bins have'} a value but no ` +
'coverage to divide it by, so they are gaps rather than uncorrected numbers.',
)
}
return out
}
/**
* Per-bin coverage as a multiplier centred on one, or `null` where there is none.
*
* The mean is taken over the bins that **have** coverage rather than over every
* bin in the file, and that is the whole subtlety of this function. A cis run
* fills the marginal for one sequence and leaves the rest of the assembly at
* zero; averaging over all of them would put the mean at a fraction of the true
* one and multiply the whole profile by the reciprocal — a correction that
* changes the scale of the track by however much of the file the anchor's
* sequence happens to be, which is not a property of anything being measured.
*/
function relativeCoverage(marginals: Float64Array): Float64Array | null {
let total = 0
let seen = 0
for (let bin = 0; bin < marginals.length; bin++) {
const value = marginals[bin]
if (!Number.isFinite(value) || value <= 0) continue
total += value
seen++
}
if (seen === 0) return null
const mean = total / seen
const out = new Float64Array(marginals.length)
for (let bin = 0; bin < marginals.length; bin++) {
const value = marginals[bin]
out[bin] = Number.isFinite(value) && value > 0 ? value / mean : 0
}
return out
}
/**
* Every contact the anchor makes, wherever it lands.
*
* One accumulator over every bin in the file, fed from a partial pass. A pixel
* contributes when either of its ends is inside the anchor, and the *other* end
* is the bin that receives it — which for a pixel with both ends inside the
* anchor means it contributes to both, exactly as the dense band sum would count
* it twice on a mirrored square.
*/
async function transProfile(
job: JobContext,
anchor: Anchor,
balance: boolean,
/**
* Total contact per bin, filled in as the walk goes — or `null`, which is what
* makes the walk stop early.
*
* The two are one decision and it is written as one parameter for that reason:
* a marginal is a statement about every pixel in the file, so asking for one
* *is* asking for the whole pass, and a caller that could ask for the marginal
* and the early stop separately is a caller that can ask for half a marginal.
*/
marginals: Float64Array | null = null,
): Promise<Float32Array> {
const mc = job.mc
const weights = validBins(mc, balance)
if (job.framing.value !== '' && job.framing.value !== DEFAULT_COUNT_COLUMN) {
job.note(
`The map is drawn from the ${job.framing.value} column, but a whole-file pass reads ` +
`${DEFAULT_COUNT_COLUMN}. This profile is of the counts.`,
)
}
const accumulator = new ViewpointAccumulator(mc.nBins, anchor.band)
let stopped = false
await mc.scanPixels({
onChunk: (chunk) => {
job.throwIfCancelled()
for (let i = 0; i < chunk.length; i++) {
const bin1 = chunk.bin1[i]
const reachable = accumulator.reachable(bin1)
if (!reachable && marginals === null) {
stopped = true
return false
}
const bin2 = chunk.bin2[i]
const raw = chunk.count[i]
const value = weights === null ? raw : raw * weights[bin1] * weights[bin2]
// Both ends, and a self-contact once: the same identity `cisMarginal`
// relies on, which is what makes a marginal off an upper-triangular
// sparse table equal to one off a mirrored dense block.
if (marginals !== null && Number.isFinite(value)) {
marginals[bin1] += value
if (bin2 !== bin1) marginals[bin2] += value
}
if (reachable) accumulator.add(bin1, bin2, value)
}
return true
},
onProgress: (done, total) => {
if (!stopped) job.progress(done, total, 'reading pixels')
},
})
job.throwIfCancelled()
job.progress(1, 1, 'done')
return accumulator.finish(weights === null ? null : (bin) => Number.isFinite(weights[bin]))
}
|