1 // Copyright (C) The Lightning Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
29 "git.arvados.org/arvados.git/sdk/go/arvados"
30 "github.com/arvados/lightning/hgvs"
31 "github.com/james-bowman/nlp"
32 "github.com/kshedden/gonpy"
33 "github.com/sirupsen/logrus"
34 log "github.com/sirupsen/logrus"
35 "golang.org/x/crypto/blake2b"
36 "gonum.org/v1/gonum/mat"
39 const annotationMaxTileSpan = 100
41 type sliceNumpy struct {
52 trainingSet []int // samples index => training set index, or -1 if not in training set
56 func (cmd *sliceNumpy) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
57 err := cmd.run(prog, args, stdin, stdout, stderr)
59 fmt.Fprintf(stderr, "%s\n", err)
65 func (cmd *sliceNumpy) run(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) error {
66 flags := flag.NewFlagSet("", flag.ContinueOnError)
67 flags.SetOutput(stderr)
68 pprof := flags.String("pprof", "", "serve Go profile data at http://`[addr]:port`")
69 runlocal := flags.Bool("local", false, "run on local host (default: run in an arvados container)")
70 arvadosRAM := flags.Int("arvados-ram", 750000000000, "amount of memory to request for arvados container (`bytes`)")
71 arvadosVCPUs := flags.Int("arvados-vcpus", 96, "number of VCPUs to request for arvados container")
72 projectUUID := flags.String("project", "", "project `UUID` for output data")
73 priority := flags.Int("priority", 500, "container request priority")
74 inputDir := flags.String("input-dir", "./in", "input `directory`")
75 outputDir := flags.String("output-dir", "./out", "output `directory`")
76 ref := flags.String("ref", "", "reference name (if blank, choose last one that appears in input)")
77 regionsFilename := flags.String("regions", "", "only output columns/annotations that intersect regions in specified bed `file`")
78 expandRegions := flags.Int("expand-regions", 0, "expand specified regions by `N` base pairs on each side`")
79 mergeOutput := flags.Bool("merge-output", false, "merge output into one matrix.npy and one matrix.annotations.csv")
80 hgvsSingle := flags.Bool("single-hgvs-matrix", false, "also generate hgvs-based matrix")
81 hgvsChunked := flags.Bool("chunked-hgvs-matrix", false, "also generate hgvs-based matrix per chromosome")
82 onehotSingle := flags.Bool("single-onehot", false, "generate one-hot tile-based matrix")
83 onehotChunked := flags.Bool("chunked-onehot", false, "generate one-hot tile-based matrix per input chunk")
84 samplesFilename := flags.String("samples", "", "`samples.csv` file with training/validation and case/control groups (see 'lightning choose-samples')")
85 onlyPCA := flags.Bool("pca", false, "generate pca matrix")
86 pcaComponents := flags.Int("pca-components", 4, "number of PCA components")
87 maxPCATiles := flags.Int("max-pca-tiles", 0, "maximum tiles to use as PCA input (filter, then drop every 2nd colum pair until below max)")
88 debugTag := flags.Int("debug-tag", -1, "log debugging details about specified tag")
89 flags.IntVar(&cmd.threads, "threads", 16, "number of memory-hungry assembly threads, and number of VCPUs to request for arvados container")
90 flags.Float64Var(&cmd.chi2PValue, "chi2-p-value", 1, "do Χ² test and omit columns with p-value above this threshold")
91 flags.BoolVar(&cmd.includeVariant1, "include-variant-1", false, "include most common variant when building one-hot matrix")
92 cmd.filter.Flags(flags)
93 err := flags.Parse(args)
94 if err == flag.ErrHelp {
96 } else if err != nil {
102 log.Println(http.ListenAndServe(*pprof, nil))
106 if cmd.chi2PValue != 1 && *samplesFilename == "" {
107 return fmt.Errorf("cannot use provided -chi2-p-value=%f because -samples= value is empty", cmd.chi2PValue)
110 cmd.debugTag = tagID(*debugTag)
113 runner := arvadosContainerRunner{
114 Name: "lightning slice-numpy",
115 Client: arvados.NewClientFromEnv(),
116 ProjectUUID: *projectUUID,
117 RAM: int64(*arvadosRAM),
118 VCPUs: *arvadosVCPUs,
123 err = runner.TranslatePaths(inputDir, regionsFilename, samplesFilename)
127 runner.Args = []string{"slice-numpy", "-local=true",
129 "-input-dir=" + *inputDir,
130 "-output-dir=/mnt/output",
131 "-threads=" + fmt.Sprintf("%d", cmd.threads),
132 "-regions=" + *regionsFilename,
133 "-expand-regions=" + fmt.Sprintf("%d", *expandRegions),
134 "-merge-output=" + fmt.Sprintf("%v", *mergeOutput),
135 "-single-hgvs-matrix=" + fmt.Sprintf("%v", *hgvsSingle),
136 "-chunked-hgvs-matrix=" + fmt.Sprintf("%v", *hgvsChunked),
137 "-single-onehot=" + fmt.Sprintf("%v", *onehotSingle),
138 "-chunked-onehot=" + fmt.Sprintf("%v", *onehotChunked),
139 "-samples=" + *samplesFilename,
140 "-pca=" + fmt.Sprintf("%v", *onlyPCA),
141 "-pca-components=" + fmt.Sprintf("%d", *pcaComponents),
142 "-max-pca-tiles=" + fmt.Sprintf("%d", *maxPCATiles),
143 "-chi2-p-value=" + fmt.Sprintf("%f", cmd.chi2PValue),
144 "-include-variant-1=" + fmt.Sprintf("%v", cmd.includeVariant1),
145 "-debug-tag=" + fmt.Sprintf("%d", cmd.debugTag),
147 runner.Args = append(runner.Args, cmd.filter.Args()...)
149 output, err = runner.Run()
153 fmt.Fprintln(stdout, output)
157 infiles, err := allFiles(*inputDir, matchGobFile)
161 if len(infiles) == 0 {
162 err = fmt.Errorf("no input files found in %s", *inputDir)
165 sort.Strings(infiles)
167 var refseq map[string][]tileLibRef
168 var reftiledata = make(map[tileLibRef][]byte, 11000000)
169 in0, err := open(infiles[0])
174 matchGenome, err := regexp.Compile(cmd.filter.MatchGenome)
176 err = fmt.Errorf("-match-genome: invalid regexp: %q", cmd.filter.MatchGenome)
180 if *samplesFilename != "" {
181 cmd.samples, err = cmd.loadSampleInfo(*samplesFilename)
189 err = DecodeLibrary(in0, strings.HasSuffix(infiles[0], ".gz"), func(ent *LibraryEntry) error {
190 if len(ent.TagSet) > 0 {
193 for _, cseq := range ent.CompactSequences {
194 if cseq.Name == *ref || *ref == "" {
195 refseq = cseq.TileSequences
198 for _, cg := range ent.CompactGenomes {
199 if matchGenome.MatchString(cg.Name) {
200 cmd.cgnames = append(cmd.cgnames, cg.Name)
203 for _, tv := range ent.TileVariants {
205 reftiledata[tileLibRef{tv.Tag, tv.Variant}] = tv.Sequence
215 err = fmt.Errorf("%s: reference sequence not found", infiles[0])
218 if len(tagset) == 0 {
219 err = fmt.Errorf("tagset not found")
223 taglib := &tagLibrary{}
224 err = taglib.setTags(tagset)
228 taglen := taglib.TagLen()
229 sort.Strings(cmd.cgnames)
231 if len(cmd.cgnames) == 0 {
232 return fmt.Errorf("fatal: 0 matching samples in library, nothing to do")
234 cmd.trainingSet = make([]int, len(cmd.cgnames))
235 if *samplesFilename == "" {
236 cmd.trainingSetSize = len(cmd.cgnames)
237 for i, name := range cmd.cgnames {
238 cmd.samples = append(cmd.samples, sampleInfo{
239 id: trimFilenameForLabel(name),
242 cmd.trainingSet[i] = i
244 } else if len(cmd.cgnames) != len(cmd.samples) {
245 return fmt.Errorf("mismatched sample list: %d samples in library, %d in %s", len(cmd.cgnames), len(cmd.samples), *samplesFilename)
247 cmd.trainingSetSize = 0
248 for i, name := range cmd.cgnames {
249 if s := trimFilenameForLabel(name); s != cmd.samples[i].id {
250 return fmt.Errorf("mismatched sample list: sample %d is %q in library, %q in %s", i, s, cmd.samples[i].id, *samplesFilename)
252 if cmd.samples[i].isTraining {
253 cmd.trainingSet[i] = cmd.trainingSetSize
254 cmd.trainingSetSize++
256 cmd.trainingSet[i] = -1
260 if cmd.filter.MinCoverage == 1 {
261 // In the generic formula below, floating point
262 // arithmetic can effectively push the coverage
263 // threshold above 1.0, which is impossible/useless.
264 // 1.0 needs to mean exactly 100% coverage.
265 cmd.minCoverage = len(cmd.cgnames)
267 cmd.minCoverage = int(math.Ceil(cmd.filter.MinCoverage * float64(len(cmd.cgnames))))
270 log.Info("indexing reference tiles")
271 type reftileinfo struct {
272 variant tileVariantID
273 seqname string // chr1
274 pos int // distance from start of chromosome to starttag
275 tiledata []byte // acgtggcaa...
276 excluded bool // true if excluded by regions file
277 nexttag tagID // tagID of following tile (-1 for last tag of chromosome)
279 isdup := map[tagID]bool{}
280 reftile := map[tagID]*reftileinfo{}
281 for seqname, cseq := range refseq {
283 lastreftag := tagID(-1)
284 for _, libref := range cseq {
285 if cmd.filter.MaxTag >= 0 && libref.Tag > tagID(cmd.filter.MaxTag) {
288 tiledata := reftiledata[libref]
289 if len(tiledata) == 0 {
290 err = fmt.Errorf("missing tiledata for tag %d variant %d in %s in ref", libref.Tag, libref.Variant, seqname)
293 foundthistag := false
294 taglib.FindAll(tiledata[:len(tiledata)-1], func(tagid tagID, offset, _ int) {
295 if !foundthistag && tagid == libref.Tag {
299 if dupref, ok := reftile[tagid]; ok {
300 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique, also found inside %+v from %s @ %d", tileLibRef{Tag: tagid, Variant: dupref.variant}, dupref.seqname, dupref.pos, libref, seqname, pos+offset+1)
301 delete(reftile, tagid)
303 log.Printf("found tag %d at offset %d inside tile variant %+v on %s @ %d", tagid, offset, libref, seqname, pos+offset+1)
307 if isdup[libref.Tag] {
308 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique", libref, seqname, pos)
309 } else if reftile[libref.Tag] != nil {
310 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique", tileLibRef{Tag: libref.Tag, Variant: reftile[libref.Tag].variant}, reftile[libref.Tag].seqname, reftile[libref.Tag].pos)
311 delete(reftile, libref.Tag)
312 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique", libref, seqname, pos)
313 isdup[libref.Tag] = true
315 reftile[libref.Tag] = &reftileinfo{
317 variant: libref.Variant,
323 reftile[lastreftag].nexttag = libref.Tag
325 lastreftag = libref.Tag
327 pos += len(tiledata) - taglen
329 log.Printf("... %s done, len %d", seqname, pos+taglen)
333 if *regionsFilename != "" {
334 log.Printf("loading regions from %s", *regionsFilename)
335 mask, err = makeMask(*regionsFilename, *expandRegions)
339 log.Printf("before applying mask, len(reftile) == %d", len(reftile))
340 log.Printf("deleting reftile entries for regions outside %d intervals", mask.Len())
341 for _, rt := range reftile {
342 if !mask.Check(strings.TrimPrefix(rt.seqname, "chr"), rt.pos, rt.pos+len(rt.tiledata)) {
346 log.Printf("after applying mask, len(reftile) == %d", len(reftile))
349 type hgvsColSet map[hgvs.Variant][2][]int8
350 encodeHGVS := throttle{Max: len(refseq)}
351 encodeHGVSTodo := map[string]chan hgvsColSet{}
352 tmpHGVSCols := map[string]*os.File{}
354 for seqname := range refseq {
356 f, err = os.Create(*outputDir + "/tmp." + seqname + ".gob")
360 defer os.Remove(f.Name())
361 bufw := bufio.NewWriterSize(f, 1<<24)
362 enc := gob.NewEncoder(bufw)
363 tmpHGVSCols[seqname] = f
364 todo := make(chan hgvsColSet, 128)
365 encodeHGVSTodo[seqname] = todo
366 encodeHGVS.Go(func() error {
367 for colset := range todo {
368 err := enc.Encode(colset)
370 encodeHGVS.Report(err)
381 var toMerge [][]int16
382 if *mergeOutput || *hgvsSingle {
383 toMerge = make([][]int16, len(infiles))
385 var onehotIndirect [][2][]uint32 // [chunkIndex][axis][index]
386 var onehotChunkSize []uint32
387 var onehotXrefs [][]onehotXref
388 if *onehotSingle || *onlyPCA {
389 onehotIndirect = make([][2][]uint32, len(infiles))
390 onehotChunkSize = make([]uint32, len(infiles))
391 onehotXrefs = make([][]onehotXref, len(infiles))
393 chunkStartTag := make([]tagID, len(infiles))
395 throttleMem := throttle{Max: cmd.threads} // TODO: estimate using mem and data size
396 throttleNumpyMem := throttle{Max: cmd.threads/2 + 1}
397 log.Info("generating annotations and numpy matrix for each slice")
398 var errSkip = errors.New("skip infile")
400 for infileIdx, infile := range infiles {
401 infileIdx, infile := infileIdx, infile
402 throttleMem.Go(func() error {
403 seq := make(map[tagID][]TileVariant, 50000)
404 cgs := make(map[string]CompactGenome, len(cmd.cgnames))
405 f, err := open(infile)
410 log.Infof("%04d: reading %s", infileIdx, infile)
411 err = DecodeLibrary(f, strings.HasSuffix(infile, ".gz"), func(ent *LibraryEntry) error {
412 for _, tv := range ent.TileVariants {
417 // corresponding ref tile, if
418 // mask is in play (we can't
419 // determine coordinates for
421 if mask != nil && reftile[tv.Tag] == nil {
425 // corresponding ref tile is
426 // outside target regions --
427 // unless it's a potential
429 if mask != nil && reftile[tv.Tag].excluded &&
430 (int(tv.Tag+1) >= len(tagset) ||
431 (bytes.HasSuffix(tv.Sequence, tagset[tv.Tag+1]) && reftile[tv.Tag+1] != nil && !reftile[tv.Tag+1].excluded)) {
434 if tv.Tag == cmd.debugTag {
435 log.Printf("infile %d %s tag %d variant %d hash %x", infileIdx, infile, tv.Tag, tv.Variant, tv.Blake2b[:3])
437 variants := seq[tv.Tag]
438 if len(variants) == 0 {
439 variants = make([]TileVariant, 100)
441 for len(variants) <= int(tv.Variant) {
442 variants = append(variants, TileVariant{})
444 variants[int(tv.Variant)] = tv
445 seq[tv.Tag] = variants
447 for _, cg := range ent.CompactGenomes {
448 if cmd.filter.MaxTag >= 0 && cg.StartTag > tagID(cmd.filter.MaxTag) {
451 if !matchGenome.MatchString(cg.Name) {
454 // pad to full slice size
455 // to avoid out-of-bounds
457 if sliceSize := 2 * int(cg.EndTag-cg.StartTag); len(cg.Variants) < sliceSize {
458 cg.Variants = append(cg.Variants, make([]tileVariantID, sliceSize-len(cg.Variants))...)
466 } else if err != nil {
467 return fmt.Errorf("%04d: DecodeLibrary(%s): err", infileIdx, infile)
469 tagstart := cgs[cmd.cgnames[0]].StartTag
470 tagend := cgs[cmd.cgnames[0]].EndTag
471 chunkStartTag[infileIdx] = tagstart
475 log.Infof("%04d: renumber/dedup variants for tags %d-%d", infileIdx, tagstart, tagend)
476 variantRemap := make([][]tileVariantID, tagend-tagstart)
477 throttleCPU := throttle{Max: runtime.GOMAXPROCS(0)}
478 for tag, variants := range seq {
479 tag, variants := tag, variants
480 throttleCPU.Go(func() error {
482 count := make(map[[blake2b.Size256]byte]int, len(variants))
486 count[blake2b.Sum256(rt.tiledata)] = 0
489 for cgname, cg := range cgs {
490 idx := int(tag-tagstart) * 2
491 for allele := 0; allele < 2; allele++ {
492 v := cg.Variants[idx+allele]
493 if v > 0 && len(variants[v].Sequence) > 0 {
494 count[variants[v].Blake2b]++
497 if v > 0 && tag == cmd.debugTag {
498 log.Printf("tag %d cg %s allele %d tv %d hash %x count is now %d", tag, cgname, allele, v, variants[v].Blake2b[:3], count[variants[v].Blake2b])
502 if alleleCoverage < cmd.minCoverage*2 {
503 idx := int(tag-tagstart) * 2
504 for _, cg := range cgs {
506 cg.Variants[idx+1] = 0
508 if tag == cmd.debugTag {
509 log.Printf("tag %d alleleCoverage %d < min %d, sample data wiped", tag, alleleCoverage, cmd.minCoverage*2)
514 // hash[i] will be the hash of
515 // the variant(s) that should
516 // be at rank i (0-based).
517 hash := make([][blake2b.Size256]byte, 0, len(count))
518 for b := range count {
519 hash = append(hash, b)
521 sort.Slice(hash, func(i, j int) bool {
522 bi, bj := &hash[i], &hash[j]
523 if ci, cj := count[*bi], count[*bj]; ci != cj {
526 return bytes.Compare((*bi)[:], (*bj)[:]) < 0
529 // rank[b] will be the 1-based
530 // new variant number for
531 // variants whose hash is b.
532 rank := make(map[[blake2b.Size256]byte]tileVariantID, len(hash))
533 for i, h := range hash {
534 rank[h] = tileVariantID(i + 1)
536 if tag == cmd.debugTag {
537 for h, r := range rank {
538 log.Printf("tag %d rank(%x) = %v", tag, h[:3], r)
541 // remap[v] will be the new
542 // variant number for original
544 remap := make([]tileVariantID, len(variants))
545 for i, tv := range variants {
546 remap[i] = rank[tv.Blake2b]
548 if tag == cmd.debugTag {
549 for in, out := range remap {
551 log.Printf("tag %d remap %d => %d", tag, in, out)
555 variantRemap[tag-tagstart] = remap
557 refrank := rank[blake2b.Sum256(rt.tiledata)]
558 if tag == cmd.debugTag {
559 log.Printf("tag %d reftile variant %d => %d", tag, rt.variant, refrank)
568 var onehotChunk [][]int8
569 var onehotXref []onehotXref
571 var annotationsFilename string
573 annotationsFilename = "/dev/null"
575 annotationsFilename = fmt.Sprintf("%s/matrix.%04d.annotations.csv", *outputDir, infileIdx)
576 log.Infof("%04d: writing %s", infileIdx, annotationsFilename)
578 annof, err := os.Create(annotationsFilename)
582 annow := bufio.NewWriterSize(annof, 1<<20)
584 for tag := tagstart; tag < tagend; tag++ {
586 if rt == nil && mask != nil {
587 // With no ref tile, we don't
588 // have coordinates to say
589 // this is in the desired
590 // regions -- so it's not.
591 // TODO: handle ref spanning
595 if rt != nil && rt.excluded {
596 // TODO: don't skip yet --
597 // first check for spanning
598 // tile variants that
599 // intersect non-excluded ref
603 if cmd.filter.MaxTag >= 0 && tag > tagID(cmd.filter.MaxTag) {
606 remap := variantRemap[tag-tagstart]
607 maxv := tileVariantID(0)
608 for _, v := range remap {
613 if *onehotChunked || *onehotSingle || *onlyPCA {
614 onehot, xrefs := cmd.tv2homhet(cgs, maxv, remap, tag, tagstart, seq)
615 if tag == cmd.debugTag {
616 log.WithFields(logrus.Fields{
619 }).Info("tv2homhet()")
621 onehotChunk = append(onehotChunk, onehot...)
622 onehotXref = append(onehotXref, xrefs...)
629 // Reference does not use any
630 // variant of this tile
632 // TODO: diff against the
633 // relevant portion of the
634 // ref's spanning tile
638 fmt.Fprintf(annow, "%d,%d,%d,=,%s,%d,,,\n", tag, outcol, rt.variant, rt.seqname, rt.pos)
640 reftilestr := strings.ToUpper(string(rt.tiledata))
642 done := make([]bool, maxv+1)
643 variantDiffs := make([][]hgvs.Variant, maxv+1)
644 for v, tv := range variants {
646 if v == 0 || v == rt.variant || done[v] {
651 if len(tv.Sequence) < taglen {
654 // if reftilestr doesn't end
655 // in the same tag as tv,
656 // extend reftilestr with
657 // following ref tiles until
658 // it does (up to an arbitrary
659 // sanity-check limit)
660 reftilestr := reftilestr
661 endtagstr := strings.ToUpper(string(tv.Sequence[len(tv.Sequence)-taglen:]))
662 for i, rt := 0, rt; i < annotationMaxTileSpan && !strings.HasSuffix(reftilestr, endtagstr) && rt.nexttag >= 0; i++ {
663 rt = reftile[rt.nexttag]
667 reftilestr += strings.ToUpper(string(rt.tiledata[taglen:]))
669 if mask != nil && !mask.Check(strings.TrimPrefix(rt.seqname, "chr"), rt.pos, rt.pos+len(reftilestr)) {
672 if !strings.HasSuffix(reftilestr, endtagstr) {
673 fmt.Fprintf(annow, "%d,%d,%d,,%s,%d,,,\n", tag, outcol, v, rt.seqname, rt.pos)
676 if lendiff := len(reftilestr) - len(tv.Sequence); lendiff < -1000 || lendiff > 1000 {
677 fmt.Fprintf(annow, "%d,%d,%d,,%s,%d,,,\n", tag, outcol, v, rt.seqname, rt.pos)
680 diffs, _ := hgvs.Diff(reftilestr, strings.ToUpper(string(tv.Sequence)), 0)
681 for i := range diffs {
682 diffs[i].Position += rt.pos
684 for _, diff := range diffs {
685 fmt.Fprintf(annow, "%d,%d,%d,%s:g.%s,%s,%d,%s,%s,%s\n", tag, outcol, v, rt.seqname, diff.String(), rt.seqname, diff.Position, diff.Ref, diff.New, diff.Left)
688 variantDiffs[v] = diffs
692 // We can now determine, for each HGVS
693 // variant (diff) in this reftile
694 // region, whether a given genome
695 // phase/allele (1) has the variant, (0) has
696 // =ref or a different variant in that
697 // position, or (-1) is lacking
698 // coverage / couldn't be diffed.
699 hgvsCol := hgvsColSet{}
700 for _, diffs := range variantDiffs {
701 for _, diff := range diffs {
702 if _, ok := hgvsCol[diff]; ok {
705 hgvsCol[diff] = [2][]int8{
706 make([]int8, len(cmd.cgnames)),
707 make([]int8, len(cmd.cgnames)),
711 for row, name := range cmd.cgnames {
712 variants := cgs[name].Variants[(tag-tagstart)*2:]
713 for ph := 0; ph < 2; ph++ {
715 if int(v) >= len(remap) {
721 // hgvsCol[*][ph][row] is already 0
722 } else if len(variantDiffs[v]) == 0 {
723 // lacking coverage / couldn't be diffed
724 for _, col := range hgvsCol {
728 for _, diff := range variantDiffs[v] {
729 hgvsCol[diff][ph][row] = 1
734 for diff, colpair := range hgvsCol {
735 allele2homhet(colpair)
736 if !cmd.filterHGVScolpair(colpair) {
737 delete(hgvsCol, diff)
740 if len(hgvsCol) > 0 {
741 encodeHGVSTodo[rt.seqname] <- hgvsCol
756 // transpose onehotChunk[col][row] to numpy[row*ncols+col]
757 rows := len(cmd.cgnames)
758 cols := len(onehotChunk)
759 log.Infof("%04d: preparing onehot numpy (rows=%d, cols=%d, mem=%d)", infileIdx, rows, cols, rows*cols)
760 throttleNumpyMem.Acquire()
761 out := onehotcols2int8(onehotChunk)
762 fnm := fmt.Sprintf("%s/onehot.%04d.npy", *outputDir, infileIdx)
763 err = writeNumpyInt8(fnm, out, rows, cols)
767 fnm = fmt.Sprintf("%s/onehot-columns.%04d.npy", *outputDir, infileIdx)
768 err = writeNumpyInt32(fnm, onehotXref2int32(onehotXref), 4, len(onehotXref))
773 throttleNumpyMem.Release()
775 if *onehotSingle || *onlyPCA {
776 onehotIndirect[infileIdx] = onehotChunk2Indirect(onehotChunk)
777 onehotChunkSize[infileIdx] = uint32(len(onehotChunk))
778 onehotXrefs[infileIdx] = onehotXref
779 n := len(onehotIndirect[infileIdx][0])
780 log.Infof("%04d: keeping onehot coordinates in memory (n=%d, mem=%d)", infileIdx, n, n*8*2)
782 if !(*onehotSingle || *onehotChunked || *onlyPCA) || *mergeOutput || *hgvsSingle {
783 log.Infof("%04d: preparing numpy (rows=%d, cols=%d)", infileIdx, len(cmd.cgnames), 2*outcol)
784 throttleNumpyMem.Acquire()
785 rows := len(cmd.cgnames)
787 out := make([]int16, rows*cols)
788 for row, name := range cmd.cgnames {
790 for col, v := range cgs[name].Variants {
791 tag := tagstart + tagID(col/2)
792 if cmd.filter.MaxTag >= 0 && tag > tagID(cmd.filter.MaxTag) {
795 if rt := reftile[tag]; rt == nil || rt.excluded {
799 out[outidx] = 0 // tag not found / spanning tile
800 } else if variants, ok := seq[tag]; ok && int(v) < len(variants) && len(variants[v].Sequence) > 0 {
801 out[outidx] = int16(variantRemap[tag-tagstart][v])
803 out[outidx] = -1 // low quality tile variant
805 if tag == cmd.debugTag {
806 log.Printf("tag %d row %d col %d outidx %d v %d out %d", tag, row, col, outidx, v, out[outidx])
814 throttleNumpyMem.Release()
815 if *mergeOutput || *hgvsSingle {
816 log.Infof("%04d: matrix fragment %d rows x %d cols", infileIdx, rows, cols)
817 toMerge[infileIdx] = out
819 if !*mergeOutput && !*onehotChunked && !*onehotSingle {
820 fnm := fmt.Sprintf("%s/matrix.%04d.npy", *outputDir, infileIdx)
821 err = writeNumpyInt16(fnm, out, rows, cols)
828 log.Infof("%s: done (%d/%d)", infile, int(atomic.AddInt64(&done, 1)), len(infiles))
832 if err = throttleMem.Wait(); err != nil {
837 log.Info("flushing hgvsCols temp files")
838 for seqname := range refseq {
839 close(encodeHGVSTodo[seqname])
841 err = encodeHGVS.Wait()
845 for seqname := range refseq {
846 log.Infof("%s: reading hgvsCols from temp file", seqname)
847 f := tmpHGVSCols[seqname]
848 _, err = f.Seek(0, io.SeekStart)
852 var hgvsCols hgvsColSet
853 dec := gob.NewDecoder(bufio.NewReaderSize(f, 1<<24))
855 err = dec.Decode(&hgvsCols)
860 log.Infof("%s: sorting %d hgvs variants", seqname, len(hgvsCols))
861 variants := make([]hgvs.Variant, 0, len(hgvsCols))
862 for v := range hgvsCols {
863 variants = append(variants, v)
865 sort.Slice(variants, func(i, j int) bool {
866 vi, vj := &variants[i], &variants[j]
867 if vi.Position != vj.Position {
868 return vi.Position < vj.Position
869 } else if vi.Ref != vj.Ref {
870 return vi.Ref < vj.Ref
872 return vi.New < vj.New
875 rows := len(cmd.cgnames)
876 cols := len(variants) * 2
877 log.Infof("%s: building hgvs matrix (rows=%d, cols=%d, mem=%d)", seqname, rows, cols, rows*cols)
878 out := make([]int8, rows*cols)
879 for varIdx, variant := range variants {
880 hgvsCols := hgvsCols[variant]
881 for row := range cmd.cgnames {
882 for ph := 0; ph < 2; ph++ {
883 out[row*cols+varIdx+ph] = hgvsCols[ph][row]
887 err = writeNumpyInt8(fmt.Sprintf("%s/hgvs.%s.npy", *outputDir, seqname), out, rows, cols)
893 fnm := fmt.Sprintf("%s/hgvs.%s.annotations.csv", *outputDir, seqname)
894 log.Infof("%s: writing hgvs column labels to %s", seqname, fnm)
895 var hgvsLabels bytes.Buffer
896 for varIdx, variant := range variants {
897 fmt.Fprintf(&hgvsLabels, "%d,%s:g.%s\n", varIdx, seqname, variant.String())
899 err = ioutil.WriteFile(fnm, hgvsLabels.Bytes(), 0666)
906 if *mergeOutput || *hgvsSingle {
907 var annow *bufio.Writer
910 annoFilename := fmt.Sprintf("%s/matrix.annotations.csv", *outputDir)
911 annof, err = os.Create(annoFilename)
915 annow = bufio.NewWriterSize(annof, 1<<20)
918 rows := len(cmd.cgnames)
920 for _, chunk := range toMerge {
921 cols += len(chunk) / rows
923 log.Infof("merging output matrix (rows=%d, cols=%d, mem=%d) and annotations", rows, cols, rows*cols*2)
926 out = make([]int16, rows*cols)
928 hgvsCols := map[string][2][]int16{} // hgvs -> [[g0,g1,g2,...], [g0,g1,g2,...]] (slice of genomes for each phase)
930 for outIdx, chunk := range toMerge {
931 chunkcols := len(chunk) / rows
933 for row := 0; row < rows; row++ {
934 copy(out[row*cols+startcol:], chunk[row*chunkcols:(row+1)*chunkcols])
937 toMerge[outIdx] = nil
939 annotationsFilename := fmt.Sprintf("%s/matrix.%04d.annotations.csv", *outputDir, outIdx)
940 log.Infof("reading %s", annotationsFilename)
941 buf, err := os.ReadFile(annotationsFilename)
946 err = os.Remove(annotationsFilename)
951 for _, line := range bytes.Split(buf, []byte{'\n'}) {
955 fields := bytes.SplitN(line, []byte{','}, 9)
956 tag, _ := strconv.Atoi(string(fields[0]))
957 incol, _ := strconv.Atoi(string(fields[1]))
958 tileVariant, _ := strconv.Atoi(string(fields[2]))
959 hgvsID := string(fields[3])
960 seqname := string(fields[4])
961 pos, _ := strconv.Atoi(string(fields[5]))
964 // Null entry for un-diffable
969 // Null entry for ref tile
972 if mask != nil && !mask.Check(strings.TrimPrefix(seqname, "chr"), pos, pos+len(refseq)) {
973 // The tile intersects one of
974 // the selected regions, but
975 // this particular HGVS
979 hgvsColPair := hgvsCols[hgvsID]
980 if hgvsColPair[0] == nil {
981 // values in new columns start
982 // out as -1 ("no data yet")
983 // or 0 ("=ref") here, may
984 // change to 1 ("hgvs variant
985 // present") below, either on
986 // this line or a future line.
987 hgvsColPair = [2][]int16{make([]int16, len(cmd.cgnames)), make([]int16, len(cmd.cgnames))}
988 rt, ok := reftile[tagID(tag)]
990 err = fmt.Errorf("bug: seeing annotations for tag %d, but it has no reftile entry", tag)
993 for ph := 0; ph < 2; ph++ {
994 for row := 0; row < rows; row++ {
995 v := chunk[row*chunkcols+incol*2+ph]
996 if tileVariantID(v) == rt.variant {
997 hgvsColPair[ph][row] = 0
999 hgvsColPair[ph][row] = -1
1003 hgvsCols[hgvsID] = hgvsColPair
1005 hgvsref := hgvs.Variant{
1007 Ref: string(refseq),
1008 New: string(refseq),
1010 fmt.Fprintf(annow, "%d,%d,%d,%s:g.%s,%s,%d,%s,%s,%s\n", tag, incol+startcol/2, rt.variant, seqname, hgvsref.String(), seqname, pos, refseq, refseq, fields[8])
1014 fmt.Fprintf(annow, "%d,%d,%d,%s,%s,%d,%s,%s,%s\n", tag, incol+startcol/2, tileVariant, hgvsID, seqname, pos, refseq, fields[7], fields[8])
1016 for ph := 0; ph < 2; ph++ {
1017 for row := 0; row < rows; row++ {
1018 v := chunk[row*chunkcols+incol*2+ph]
1019 if int(v) == tileVariant {
1020 hgvsColPair[ph][row] = 1
1026 startcol += chunkcols
1037 err = writeNumpyInt16(fmt.Sprintf("%s/matrix.npy", *outputDir), out, rows, cols)
1045 cols = len(hgvsCols) * 2
1046 log.Printf("building hgvs-based matrix: %d rows x %d cols", rows, cols)
1047 out = make([]int16, rows*cols)
1048 hgvsIDs := make([]string, 0, cols/2)
1049 for hgvsID := range hgvsCols {
1050 hgvsIDs = append(hgvsIDs, hgvsID)
1052 sort.Strings(hgvsIDs)
1053 var hgvsLabels bytes.Buffer
1054 for idx, hgvsID := range hgvsIDs {
1055 fmt.Fprintf(&hgvsLabels, "%d,%s\n", idx, hgvsID)
1056 for ph := 0; ph < 2; ph++ {
1057 hgvscol := hgvsCols[hgvsID][ph]
1058 for row, val := range hgvscol {
1059 out[row*cols+idx*2+ph] = val
1063 err = writeNumpyInt16(fmt.Sprintf("%s/hgvs.npy", *outputDir), out, rows, cols)
1068 fnm := fmt.Sprintf("%s/hgvs.annotations.csv", *outputDir)
1069 log.Printf("writing hgvs labels: %s", fnm)
1070 err = ioutil.WriteFile(fnm, hgvsLabels.Bytes(), 0777)
1076 if *onehotSingle || *onlyPCA {
1078 for _, part := range onehotIndirect {
1079 nzCount += len(part[0])
1081 onehot := make([]uint32, nzCount*2) // [r,r,r,...,c,c,c,...]
1082 var xrefs []onehotXref
1083 chunkOffset := uint32(0)
1085 for i, part := range onehotIndirect {
1086 for i := range part[1] {
1087 part[1][i] += chunkOffset
1089 copy(onehot[outcol:], part[0])
1090 copy(onehot[outcol+nzCount:], part[1])
1091 xrefs = append(xrefs, onehotXrefs[i]...)
1093 outcol += len(part[0])
1094 chunkOffset += onehotChunkSize[i]
1098 onehotXrefs[i] = nil
1099 debug.FreeOSMemory()
1102 fnm := fmt.Sprintf("%s/onehot.npy", *outputDir)
1103 err = writeNumpyUint32(fnm, onehot, 2, nzCount)
1107 fnm = fmt.Sprintf("%s/onehot-columns.npy", *outputDir)
1108 err = writeNumpyInt32(fnm, onehotXref2int32(xrefs), 5, len(xrefs))
1115 for _, c := range onehot[nzCount:] {
1121 return fmt.Errorf("cannot do PCA: one-hot matrix is empty")
1123 log.Printf("have %d one-hot cols", cols)
1125 for *maxPCATiles > 0 && cols > *maxPCATiles*2 {
1126 cols = (cols + 1) / 2
1129 log.Printf("creating full matrix (%d rows) and training matrix (%d rows) with %d cols, stride %d", len(cmd.cgnames), cmd.trainingSetSize, cols, stride)
1130 mtxFull := mat.NewDense(len(cmd.cgnames), cols, nil)
1131 mtxTrain := mat.NewDense(cmd.trainingSetSize, cols, nil)
1132 for i, c := range onehot[nzCount:] {
1133 if int(c/2)%stride == 0 {
1134 outcol := int(c/2)/stride*2 + int(c)%2
1135 mtxFull.Set(int(onehot[i]), outcol, 1)
1136 if trainRow := cmd.trainingSet[int(onehot[i])]; trainRow >= 0 {
1137 mtxTrain.Set(trainRow, outcol, 1)
1141 log.Print("fitting")
1142 transformer := nlp.NewPCA(*pcaComponents)
1143 transformer.Fit(mtxTrain.T())
1144 log.Printf("transforming")
1145 pca, err := transformer.Transform(mtxFull.T())
1150 outrows, outcols := pca.Dims()
1151 log.Printf("copying result to numpy output array: %d rows, %d cols", outrows, outcols)
1152 out := make([]float64, outrows*outcols)
1153 for i := 0; i < outrows; i++ {
1154 for j := 0; j < outcols; j++ {
1155 out[i*outcols+j] = pca.At(i, j)
1158 fnm := fmt.Sprintf("%s/pca.npy", *outputDir)
1159 log.Printf("writing numpy: %s", fnm)
1160 output, err := os.OpenFile(fnm, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, 0777)
1164 npw, err := gonpy.NewWriter(nopCloser{output})
1166 return fmt.Errorf("gonpy.NewWriter: %w", err)
1168 npw.Shape = []int{outrows, outcols}
1169 err = npw.WriteFloat64(out)
1171 return fmt.Errorf("WriteFloat64: %w", err)
1173 err = output.Close()
1179 samplesOutFilename := *outputDir + "/samples.csv"
1180 log.Infof("writing sample metadata to %s", samplesOutFilename)
1182 f, err = os.Create(samplesOutFilename)
1187 for i, si := range cmd.samples {
1191 } else if si.isControl {
1200 for c := 0; c < outcols; c++ {
1201 pcavals += fmt.Sprintf(",%f", pca.At(i, c))
1203 _, err = fmt.Fprintf(f, "%d,%s,%s,%s%s\n", i, si.id, cc, tv, pcavals)
1205 err = fmt.Errorf("write %s: %w", samplesOutFilename, err)
1211 err = fmt.Errorf("close %s: %w", samplesOutFilename, err)
1217 if !*mergeOutput && !*onehotChunked && !*onehotSingle && !*onlyPCA {
1218 tagoffsetFilename := *outputDir + "/chunk-tag-offset.csv"
1219 log.Infof("writing tag offsets to %s", tagoffsetFilename)
1221 f, err = os.Create(tagoffsetFilename)
1226 for idx, offset := range chunkStartTag {
1227 _, err = fmt.Fprintf(f, "%q,%d\n", fmt.Sprintf("matrix.%04d.npy", idx), offset)
1229 err = fmt.Errorf("write %s: %w", tagoffsetFilename, err)
1235 err = fmt.Errorf("close %s: %w", tagoffsetFilename, err)
1243 type sampleInfo struct {
1249 pcaComponents []float64
1252 // Read samples.csv file with case/control and training/validation
1254 func (cmd *sliceNumpy) loadSampleInfo(samplesFilename string) ([]sampleInfo, error) {
1256 f, err := open(samplesFilename)
1260 buf, err := io.ReadAll(f)
1266 for _, csv := range bytes.Split(buf, []byte{'\n'}) {
1268 split := strings.Split(string(csv), ",")
1269 if len(split) != 4 {
1270 return nil, fmt.Errorf("%d fields != 4 in %s line %d: %q", len(split), samplesFilename, lineNum, csv)
1272 if split[0] == "Index" && split[1] == "SampleID" && split[2] == "CaseControl" && split[3] == "TrainingValidation" {
1275 idx, err := strconv.Atoi(split[0])
1278 return nil, fmt.Errorf("header does not look right: %q", csv)
1280 return nil, fmt.Errorf("%s line %d: index: %s", samplesFilename, lineNum, err)
1283 return nil, fmt.Errorf("%s line %d: index %d out of order", samplesFilename, lineNum, idx)
1285 si = append(si, sampleInfo{
1287 isCase: split[2] == "1",
1288 isControl: split[2] == "0",
1289 isTraining: split[3] == "1",
1290 isValidation: split[3] == "0",
1296 func (cmd *sliceNumpy) filterHGVScolpair(colpair [2][]int8) bool {
1297 if cmd.chi2PValue >= 1 {
1300 col0 := make([]bool, 0, len(cmd.chi2Cases))
1301 col1 := make([]bool, 0, len(cmd.chi2Cases))
1302 cases := make([]bool, 0, len(cmd.chi2Cases))
1303 for i, c := range cmd.chi2Cases {
1304 if colpair[0][i] < 0 {
1307 col0 = append(col0, colpair[0][i] != 0)
1308 col1 = append(col1, colpair[1][i] != 0)
1309 cases = append(cases, c)
1311 return len(cases) >= cmd.minCoverage &&
1312 (pvalue(col0, cases) <= cmd.chi2PValue || pvalue(col1, cases) <= cmd.chi2PValue)
1315 func writeNumpyUint32(fnm string, out []uint32, rows, cols int) error {
1316 output, err := os.Create(fnm)
1320 defer output.Close()
1321 bufw := bufio.NewWriterSize(output, 1<<26)
1322 npw, err := gonpy.NewWriter(nopCloser{bufw})
1326 log.WithFields(log.Fields{
1330 "bytes": rows * cols * 4,
1331 }).Infof("writing numpy: %s", fnm)
1332 npw.Shape = []int{rows, cols}
1333 npw.WriteUint32(out)
1338 return output.Close()
1341 func writeNumpyInt32(fnm string, out []int32, rows, cols int) error {
1342 output, err := os.Create(fnm)
1346 defer output.Close()
1347 bufw := bufio.NewWriterSize(output, 1<<26)
1348 npw, err := gonpy.NewWriter(nopCloser{bufw})
1352 log.WithFields(log.Fields{
1356 "bytes": rows * cols * 4,
1357 }).Infof("writing numpy: %s", fnm)
1358 npw.Shape = []int{rows, cols}
1364 return output.Close()
1367 func writeNumpyInt16(fnm string, out []int16, rows, cols int) error {
1368 output, err := os.Create(fnm)
1372 defer output.Close()
1373 bufw := bufio.NewWriterSize(output, 1<<26)
1374 npw, err := gonpy.NewWriter(nopCloser{bufw})
1378 log.WithFields(log.Fields{
1382 "bytes": rows * cols * 2,
1383 }).Infof("writing numpy: %s", fnm)
1384 npw.Shape = []int{rows, cols}
1390 return output.Close()
1393 func writeNumpyInt8(fnm string, out []int8, rows, cols int) error {
1394 output, err := os.Create(fnm)
1398 defer output.Close()
1399 bufw := bufio.NewWriterSize(output, 1<<26)
1400 npw, err := gonpy.NewWriter(nopCloser{bufw})
1404 log.WithFields(log.Fields{
1408 "bytes": rows * cols,
1409 }).Infof("writing numpy: %s", fnm)
1410 npw.Shape = []int{rows, cols}
1416 return output.Close()
1419 func allele2homhet(colpair [2][]int8) {
1420 a, b := colpair[0], colpair[1]
1421 for i, av := range a {
1423 if av < 0 || bv < 0 {
1426 } else if av > 0 && bv > 0 {
1429 } else if av > 0 || bv > 0 {
1433 // ref (or a different variant in same position)
1434 // (this is a no-op) a[i], b[i] = 0, 0
1439 type onehotXref struct {
1441 variant tileVariantID
1446 const onehotXrefSize = unsafe.Sizeof(onehotXref{})
1448 // Build onehot matrix (m[tileVariantIndex][genome] == 0 or 1) for all
1449 // variants of a single tile/tag#.
1451 // Return nil if no tile variant passes Χ² filter.
1452 func (cmd *sliceNumpy) tv2homhet(cgs map[string]CompactGenome, maxv tileVariantID, remap []tileVariantID, tag, chunkstarttag tagID, seq map[tagID][]TileVariant) ([][]int8, []onehotXref) {
1453 if tag == cmd.debugTag {
1454 tv := make([]tileVariantID, len(cmd.cgnames)*2)
1455 for i, name := range cmd.cgnames {
1456 copy(tv[i*2:(i+1)*2], cgs[name].Variants[(tag-chunkstarttag)*2:])
1458 log.WithFields(logrus.Fields{
1459 "cgs[i].Variants[tag*2+j]": tv,
1463 "chunkstarttag": chunkstarttag,
1464 }).Info("tv2homhet()")
1466 if maxv < 1 || (maxv < 2 && !cmd.includeVariant1) {
1467 // everyone has the most common variant (of the variants we don't drop)
1470 tagoffset := tag - chunkstarttag
1472 for _, cg := range cgs {
1474 for _, v := range cg.Variants[tagoffset*2 : tagoffset*2+2] {
1475 if v > 0 && int(v) < len(seq[tag]) && len(seq[tag][v].Sequence) > 0 {
1483 if coverage < cmd.minCoverage {
1486 obs := make([][]bool, (maxv+1)*2) // 2 slices (hom + het) for each variant#
1487 for i := range obs {
1488 obs[i] = make([]bool, cmd.trainingSetSize)
1490 for cgid, name := range cmd.cgnames {
1491 tsid := cmd.trainingSet[cgid]
1495 cgvars := cgs[name].Variants[tagoffset*2:]
1496 tv0, tv1 := remap[cgvars[0]], remap[cgvars[1]]
1497 for v := tileVariantID(1); v <= maxv; v++ {
1498 if tv0 == v && tv1 == v {
1499 obs[v*2][tsid] = true
1500 } else if tv0 == v || tv1 == v {
1501 obs[v*2+1][tsid] = true
1506 var xref []onehotXref
1507 for col := 2; col < len(obs); col++ {
1508 // col 0,1 correspond to tile variant 0, i.e.,
1509 // no-call; col 2,3 correspond to the most common
1510 // variant; so we (normally) start at col 4.
1511 if col < 4 && !cmd.includeVariant1 {
1514 p := pvalue(obs[col], cmd.chi2Cases)
1515 if cmd.chi2PValue < 1 && !(p < cmd.chi2PValue) {
1518 onehot = append(onehot, bool2int8(obs[col]))
1519 xref = append(xref, onehotXref{
1521 variant: tileVariantID(col >> 1),
1529 func bool2int8(in []bool) []int8 {
1530 out := make([]int8, len(in))
1531 for i, v := range in {
1539 // convert a []onehotXref with length N to a numpy-style []int32
1540 // matrix with N columns, one row per field of onehotXref struct.
1542 // Hom/het row contains hom=0, het=1.
1544 // P-value row contains 1000000x actual p-value.
1545 func onehotXref2int32(xrefs []onehotXref) []int32 {
1547 xdata := make([]int32, 5*xcols)
1548 for i, xref := range xrefs {
1549 xdata[i] = int32(xref.tag)
1550 xdata[xcols+i] = int32(xref.variant)
1552 xdata[xcols*2+i] = 1
1554 xdata[xcols*3+i] = int32(xref.pvalue * 1000000)
1555 xdata[xcols*4+i] = int32(-math.Log10(xref.pvalue) * 1000000)
1560 // transpose onehot data from in[col][row] to numpy-style
1561 // out[row*cols+col].
1562 func onehotcols2int8(in [][]int8) []int8 {
1568 out := make([]int8, rows*cols)
1569 for row := 0; row < rows; row++ {
1570 outrow := out[row*cols:]
1571 for col, incol := range in {
1572 outrow[col] = incol[row]
1578 // Return [2][]uint32{rowIndices, colIndices} indicating which
1579 // elements of matrixT[c][r] have non-zero values.
1580 func onehotChunk2Indirect(matrixT [][]int8) [2][]uint32 {
1582 for c, col := range matrixT {
1583 for r, val := range col {
1585 nz[0] = append(nz[0], uint32(r))
1586 nz[1] = append(nz[1], uint32(c))