1 // Copyright (C) The Lightning Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
28 "git.arvados.org/arvados.git/sdk/go/arvados"
29 "github.com/arvados/lightning/hgvs"
30 "github.com/kshedden/gonpy"
31 log "github.com/sirupsen/logrus"
32 "golang.org/x/crypto/blake2b"
35 type sliceNumpy struct {
38 chi2CaseControlColumn string
39 chi2CaseControlFile string
46 func (cmd *sliceNumpy) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
50 fmt.Fprintf(stderr, "%s\n", err)
53 flags := flag.NewFlagSet("", flag.ContinueOnError)
54 flags.SetOutput(stderr)
55 pprof := flags.String("pprof", "", "serve Go profile data at http://`[addr]:port`")
56 runlocal := flags.Bool("local", false, "run on local host (default: run in an arvados container)")
57 projectUUID := flags.String("project", "", "project `UUID` for output data")
58 priority := flags.Int("priority", 500, "container request priority")
59 inputDir := flags.String("input-dir", "./in", "input `directory`")
60 outputDir := flags.String("output-dir", "./out", "output `directory`")
61 ref := flags.String("ref", "", "reference name (if blank, choose last one that appears in input)")
62 regionsFilename := flags.String("regions", "", "only output columns/annotations that intersect regions in specified bed `file`")
63 expandRegions := flags.Int("expand-regions", 0, "expand specified regions by `N` base pairs on each side`")
64 mergeOutput := flags.Bool("merge-output", false, "merge output into one matrix.npy and one matrix.annotations.csv")
65 hgvsSingle := flags.Bool("single-hgvs-matrix", false, "also generate hgvs-based matrix")
66 hgvsChunked := flags.Bool("chunked-hgvs-matrix", false, "also generate hgvs-based matrix per chromosome")
67 onehotSingle := flags.Bool("single-onehot", false, "generate one-hot tile-based matrix")
68 onehotChunked := flags.Bool("chunked-onehot", false, "generate one-hot tile-based matrix per input chunk")
69 flags.IntVar(&cmd.threads, "threads", 16, "number of memory-hungry assembly threads")
70 flags.StringVar(&cmd.chi2CaseControlFile, "chi2-case-control-file", "", "tsv file or directory indicating cases and controls for Χ² test (if directory, all .tsv files will be read)")
71 flags.StringVar(&cmd.chi2CaseControlColumn, "chi2-case-control-column", "", "name of case/control column in case-control files for Χ² test (value must be 0 for control, 1 for case)")
72 flags.Float64Var(&cmd.chi2PValue, "chi2-p-value", 1, "do Χ² test and omit columns with p-value above this threshold")
73 cmd.filter.Flags(flags)
74 err = flags.Parse(args)
75 if err == flag.ErrHelp {
78 } else if err != nil {
84 log.Println(http.ListenAndServe(*pprof, nil))
88 if cmd.chi2PValue != 1 && (cmd.chi2CaseControlFile == "" || cmd.chi2CaseControlColumn == "") {
89 log.Errorf("cannot use provided -chi2-p-value=%f because -chi2-case-control-file= or -chi2-case-control-column= value is empty", cmd.chi2PValue)
94 runner := arvadosContainerRunner{
95 Name: "lightning slice-numpy",
96 Client: arvados.NewClientFromEnv(),
97 ProjectUUID: *projectUUID,
104 err = runner.TranslatePaths(inputDir, regionsFilename, &cmd.chi2CaseControlFile)
108 runner.Args = []string{"slice-numpy", "-local=true",
110 "-input-dir=" + *inputDir,
111 "-output-dir=/mnt/output",
112 "-threads=" + fmt.Sprintf("%d", cmd.threads),
113 "-regions=" + *regionsFilename,
114 "-expand-regions=" + fmt.Sprintf("%d", *expandRegions),
115 "-merge-output=" + fmt.Sprintf("%v", *mergeOutput),
116 "-single-hgvs-matrix=" + fmt.Sprintf("%v", *hgvsSingle),
117 "-chunked-hgvs-matrix=" + fmt.Sprintf("%v", *hgvsChunked),
118 "-single-onehot=" + fmt.Sprintf("%v", *onehotSingle),
119 "-chunked-onehot=" + fmt.Sprintf("%v", *onehotChunked),
120 "-chi2-case-control-file=" + cmd.chi2CaseControlFile,
121 "-chi2-case-control-column=" + cmd.chi2CaseControlColumn,
122 "-chi2-p-value=" + fmt.Sprintf("%f", cmd.chi2PValue),
124 runner.Args = append(runner.Args, cmd.filter.Args()...)
126 output, err = runner.Run()
130 fmt.Fprintln(stdout, output)
134 infiles, err := allFiles(*inputDir, matchGobFile)
138 if len(infiles) == 0 {
139 err = fmt.Errorf("no input files found in %s", *inputDir)
142 sort.Strings(infiles)
144 var refseq map[string][]tileLibRef
145 var reftiledata = make(map[tileLibRef][]byte, 11000000)
146 in0, err := open(infiles[0])
151 matchGenome, err := regexp.Compile(cmd.filter.MatchGenome)
153 err = fmt.Errorf("-match-genome: invalid regexp: %q", cmd.filter.MatchGenome)
159 DecodeLibrary(in0, strings.HasSuffix(infiles[0], ".gz"), func(ent *LibraryEntry) error {
160 if len(ent.TagSet) > 0 {
163 for _, cseq := range ent.CompactSequences {
164 if cseq.Name == *ref || *ref == "" {
165 refseq = cseq.TileSequences
168 for _, cg := range ent.CompactGenomes {
169 if matchGenome.MatchString(cg.Name) {
170 cmd.cgnames = append(cmd.cgnames, cg.Name)
173 for _, tv := range ent.TileVariants {
175 reftiledata[tileLibRef{tv.Tag, tv.Variant}] = tv.Sequence
185 err = fmt.Errorf("%s: reference sequence not found", infiles[0])
188 if len(tagset) == 0 {
189 err = fmt.Errorf("tagset not found")
193 taglib := &tagLibrary{}
194 err = taglib.setTags(tagset)
198 taglen := taglib.TagLen()
200 if len(cmd.cgnames) == 0 {
201 err = fmt.Errorf("no genomes found matching regexp %q", cmd.filter.MatchGenome)
204 sort.Strings(cmd.cgnames)
205 err = cmd.useCaseControlFiles()
209 cmd.minCoverage = int(math.Ceil(cmd.filter.MinCoverage * float64(len(cmd.cgnames))))
212 labelsFilename := *outputDir + "/samples.csv"
213 log.Infof("writing labels to %s", labelsFilename)
215 f, err = os.Create(labelsFilename)
220 for i, name := range cmd.cgnames {
222 if cmd.chi2Cases != nil && cmd.chi2Cases[i] {
225 _, err = fmt.Fprintf(f, "%d,%q,%d\n", i, trimFilenameForLabel(name), cc)
227 err = fmt.Errorf("write %s: %w", labelsFilename, err)
233 err = fmt.Errorf("close %s: %w", labelsFilename, err)
238 log.Info("indexing reference tiles")
239 type reftileinfo struct {
240 variant tileVariantID
241 seqname string // chr1
242 pos int // distance from start of chromosome to starttag
243 tiledata []byte // acgtggcaa...
245 isdup := map[tagID]bool{}
246 reftile := map[tagID]*reftileinfo{}
247 for seqname, cseq := range refseq {
249 for _, libref := range cseq {
250 if libref.Tag > tagID(cmd.filter.MaxTag) {
253 tiledata := reftiledata[libref]
254 if len(tiledata) == 0 {
255 err = fmt.Errorf("missing tiledata for tag %d variant %d in %s in ref", libref.Tag, libref.Variant, seqname)
258 foundthistag := false
259 taglib.FindAll(tiledata[:len(tiledata)-1], func(tagid tagID, offset, _ int) {
260 if !foundthistag && tagid == libref.Tag {
264 if dupref, ok := reftile[tagid]; ok {
265 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)
266 delete(reftile, tagid)
268 log.Printf("found tag %d at offset %d inside tile variant %+v on %s @ %d", tagid, offset, libref, seqname, pos+offset+1)
272 if isdup[libref.Tag] {
273 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique", libref, seqname, pos)
274 } else if reftile[libref.Tag] != nil {
275 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)
276 delete(reftile, libref.Tag)
277 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique", libref, seqname, pos)
278 isdup[libref.Tag] = true
280 reftile[libref.Tag] = &reftileinfo{
282 variant: libref.Variant,
287 pos += len(tiledata) - taglen
289 log.Printf("... %s done, len %d", seqname, pos+taglen)
293 if *regionsFilename != "" {
294 log.Printf("loading regions from %s", *regionsFilename)
295 mask, err = makeMask(*regionsFilename, *expandRegions)
299 log.Printf("before applying mask, len(reftile) == %d", len(reftile))
300 log.Printf("deleting reftile entries for regions outside %d intervals", mask.Len())
301 for tag, rt := range reftile {
302 if !mask.Check(strings.TrimPrefix(rt.seqname, "chr"), rt.pos, rt.pos+len(rt.tiledata)) {
306 log.Printf("after applying mask, len(reftile) == %d", len(reftile))
309 type hgvsColSet map[hgvs.Variant][2][]int8
310 encodeHGVS := throttle{Max: len(refseq)}
311 encodeHGVSTodo := map[string]chan hgvsColSet{}
312 tmpHGVSCols := map[string]*os.File{}
314 for seqname := range refseq {
316 f, err = os.Create(*outputDir + "/tmp." + seqname + ".gob")
320 defer os.Remove(f.Name())
321 bufw := bufio.NewWriterSize(f, 1<<24)
322 enc := gob.NewEncoder(bufw)
323 tmpHGVSCols[seqname] = f
324 todo := make(chan hgvsColSet, 128)
325 encodeHGVSTodo[seqname] = todo
326 encodeHGVS.Go(func() error {
327 for colset := range todo {
328 err := enc.Encode(colset)
330 encodeHGVS.Report(err)
341 var toMerge [][]int16
342 if *mergeOutput || *hgvsSingle {
343 toMerge = make([][]int16, len(infiles))
345 var onehotIndirect [][2][]uint32 // [chunkIndex][axis][index]
346 var onehotXrefs [][]onehotXref
348 onehotIndirect = make([][2][]uint32, len(infiles))
349 onehotXrefs = make([][]onehotXref, len(infiles))
352 throttleMem := throttle{Max: cmd.threads} // TODO: estimate using mem and data size
353 throttleNumpyMem := throttle{Max: cmd.threads/2 + 1}
354 log.Info("generating annotations and numpy matrix for each slice")
356 for infileIdx, infile := range infiles {
357 infileIdx, infile := infileIdx, infile
358 throttleMem.Go(func() error {
359 seq := make(map[tagID][]TileVariant, 50000)
360 cgs := make(map[string]CompactGenome, len(cmd.cgnames))
361 f, err := open(infile)
366 log.Infof("%04d: reading %s", infileIdx, infile)
367 err = DecodeLibrary(f, strings.HasSuffix(infile, ".gz"), func(ent *LibraryEntry) error {
368 for _, tv := range ent.TileVariants {
372 if mask != nil && reftile[tv.Tag] == nil {
378 variants := seq[tv.Tag]
379 if len(variants) == 0 {
380 variants = make([]TileVariant, 100)
382 for len(variants) <= int(tv.Variant) {
383 variants = append(variants, TileVariant{})
385 variants[int(tv.Variant)] = tv
386 seq[tv.Tag] = variants
388 for _, cg := range ent.CompactGenomes {
389 if !matchGenome.MatchString(cg.Name) {
392 // pad to full slice size
393 // to avoid out-of-bounds
395 if sliceSize := 2 * int(cg.EndTag-cg.StartTag); len(cg.Variants) < sliceSize {
396 cg.Variants = append(cg.Variants, make([]tileVariantID, sliceSize-len(cg.Variants))...)
405 tagstart := cgs[cmd.cgnames[0]].StartTag
406 tagend := cgs[cmd.cgnames[0]].EndTag
410 log.Infof("%04d: renumber/dedup variants for tags %d-%d", infileIdx, tagstart, tagend)
411 variantRemap := make([][]tileVariantID, tagend-tagstart)
412 throttleCPU := throttle{Max: runtime.GOMAXPROCS(0)}
413 for tag, variants := range seq {
414 tag, variants := tag, variants
415 throttleCPU.Acquire()
417 defer throttleCPU.Release()
418 count := make(map[[blake2b.Size256]byte]int, len(variants))
422 count[blake2b.Sum256(rt.tiledata)] = 0
425 for _, cg := range cgs {
426 idx := int(tag-tagstart) * 2
427 for allele := 0; allele < 2; allele++ {
428 v := cg.Variants[idx+allele]
429 if v > 0 && len(variants[v].Sequence) > 0 {
430 count[variants[v].Blake2b]++
434 // hash[i] will be the hash of
435 // the variant(s) that should
436 // be at rank i (0-based).
437 hash := make([][blake2b.Size256]byte, 0, len(count))
438 for b := range count {
439 hash = append(hash, b)
441 sort.Slice(hash, func(i, j int) bool {
442 bi, bj := &hash[i], &hash[j]
443 if ci, cj := count[*bi], count[*bj]; ci != cj {
446 return bytes.Compare((*bi)[:], (*bj)[:]) < 0
449 // rank[b] will be the 1-based
450 // new variant number for
451 // variants whose hash is b.
452 rank := make(map[[blake2b.Size256]byte]tileVariantID, len(hash))
453 for i, h := range hash {
454 rank[h] = tileVariantID(i + 1)
456 // remap[v] will be the new
457 // variant number for original
459 remap := make([]tileVariantID, len(variants))
460 for i, tv := range variants {
461 remap[i] = rank[tv.Blake2b]
463 variantRemap[tag-tagstart] = remap
465 rt.variant = rank[blake2b.Sum256(rt.tiledata)]
471 var onehotChunk [][]int8
472 var onehotXref []onehotXref
474 annotationsFilename := fmt.Sprintf("%s/matrix.%04d.annotations.csv", *outputDir, infileIdx)
475 log.Infof("%04d: writing %s", infileIdx, annotationsFilename)
476 annof, err := os.Create(annotationsFilename)
480 annow := bufio.NewWriterSize(annof, 1<<20)
482 for tag := tagstart; tag < tagend; tag++ {
484 if rt == nil && mask != nil {
485 // Excluded by specified regions
488 if tag > tagID(cmd.filter.MaxTag) {
491 remap := variantRemap[tag-tagstart]
492 maxv := tileVariantID(0)
493 for _, v := range remap {
498 if *onehotChunked || *onehotSingle {
499 onehot, xrefs := cmd.tv2homhet(cgs, maxv, remap, tag, tagstart)
500 onehotChunk = append(onehotChunk, onehot...)
501 onehotXref = append(onehotXref, xrefs...)
504 // Reference does not use any
505 // variant of this tile
509 fmt.Fprintf(annow, "%d,%d,%d,=,%s,%d,,,\n", tag, outcol, rt.variant, rt.seqname, rt.pos)
511 reftilestr := strings.ToUpper(string(rt.tiledata))
513 done := make([]bool, maxv+1)
514 variantDiffs := make([][]hgvs.Variant, maxv+1)
515 for v, tv := range variants {
517 if v == rt.variant || done[v] {
522 if len(tv.Sequence) < taglen || !bytes.HasSuffix(rt.tiledata, tv.Sequence[len(tv.Sequence)-taglen:]) {
523 fmt.Fprintf(annow, "%d,%d,%d,,%s,%d,,,\n", tag, outcol, v, rt.seqname, rt.pos)
526 if lendiff := len(rt.tiledata) - len(tv.Sequence); lendiff < -1000 || lendiff > 1000 {
527 fmt.Fprintf(annow, "%d,%d,%d,,%s,%d,,,\n", tag, outcol, v, rt.seqname, rt.pos)
530 diffs, _ := hgvs.Diff(reftilestr, strings.ToUpper(string(tv.Sequence)), 0)
531 for i := range diffs {
532 diffs[i].Position += rt.pos
534 for _, diff := range diffs {
535 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)
538 variantDiffs[v] = diffs
542 // We can now determine, for each HGVS
543 // variant (diff) in this reftile
544 // region, whether a given genome
545 // phase/allele (1) has the variant, (0) has
546 // =ref or a different variant in that
547 // position, or (-1) is lacking
548 // coverage / couldn't be diffed.
549 hgvsCol := hgvsColSet{}
550 for _, diffs := range variantDiffs {
551 for _, diff := range diffs {
552 if _, ok := hgvsCol[diff]; ok {
555 hgvsCol[diff] = [2][]int8{
556 make([]int8, len(cmd.cgnames)),
557 make([]int8, len(cmd.cgnames)),
561 for row, name := range cmd.cgnames {
562 variants := cgs[name].Variants[(tag-tagstart)*2:]
563 for ph := 0; ph < 2; ph++ {
565 if int(v) >= len(remap) {
571 // hgvsCol[*][ph][row] is already 0
572 } else if len(variantDiffs[v]) == 0 {
573 // lacking coverage / couldn't be diffed
574 for _, col := range hgvsCol {
578 for _, diff := range variantDiffs[v] {
579 hgvsCol[diff][ph][row] = 1
584 for diff, colpair := range hgvsCol {
585 allele2homhet(colpair)
586 if !cmd.filterHGVScolpair(colpair) {
587 delete(hgvsCol, diff)
590 if len(hgvsCol) > 0 {
591 encodeHGVSTodo[rt.seqname] <- hgvsCol
606 // transpose onehotChunk[col][row] to numpy[row*ncols+col]
607 rows := len(cmd.cgnames)
608 cols := len(onehotChunk)
609 log.Infof("%04d: preparing onehot numpy (rows=%d, cols=%d, mem=%d)", infileIdx, len(cmd.cgnames), len(onehotChunk), len(cmd.cgnames)*len(onehotChunk))
610 throttleNumpyMem.Acquire()
611 out := onehotcols2int8(onehotChunk)
612 fnm := fmt.Sprintf("%s/onehot.%04d.npy", *outputDir, infileIdx)
613 err = writeNumpyInt8(fnm, out, rows, cols)
617 fnm = fmt.Sprintf("%s/onehot-columns.%04d.npy", *outputDir, infileIdx)
618 err = writeNumpyInt32(fnm, onehotXref2int32(onehotXref), 4, len(onehotXref))
623 throttleNumpyMem.Release()
626 onehotIndirect[infileIdx] = onehotChunk2Indirect(onehotChunk)
627 onehotXrefs[infileIdx] = onehotXref
628 n := len(onehotIndirect[infileIdx][0])
629 log.Infof("%04d: keeping onehot coordinates in memory (n=%d, mem=%d)", infileIdx, n, n*8)
631 if !(*onehotSingle || *onehotChunked) || *mergeOutput || *hgvsSingle {
632 log.Infof("%04d: preparing numpy", infileIdx)
633 throttleNumpyMem.Acquire()
634 rows := len(cmd.cgnames)
636 out := make([]int16, rows*cols)
637 for row, name := range cmd.cgnames {
638 out := out[row*cols:]
640 for col, v := range cgs[name].Variants {
641 tag := tagstart + tagID(col/2)
642 if mask != nil && reftile[tag] == nil {
645 if variants, ok := seq[tag]; ok && len(variants) > int(v) && len(variants[v].Sequence) > 0 {
646 out[outcol] = int16(variantRemap[tag-tagstart][v])
656 throttleNumpyMem.Release()
657 if *mergeOutput || *hgvsSingle {
658 log.Infof("%04d: matrix fragment %d rows x %d cols", infileIdx, rows, cols)
659 toMerge[infileIdx] = out
661 if !*mergeOutput && !*onehotChunked && !*onehotSingle {
662 fnm := fmt.Sprintf("%s/matrix.%04d.npy", *outputDir, infileIdx)
663 err = writeNumpyInt16(fnm, out, rows, cols)
670 log.Infof("%s: done (%d/%d)", infile, int(atomic.AddInt64(&done, 1)), len(infiles))
674 if err = throttleMem.Wait(); err != nil {
679 log.Info("flushing hgvsCols temp files")
680 for seqname := range refseq {
681 close(encodeHGVSTodo[seqname])
683 err = encodeHGVS.Wait()
687 for seqname := range refseq {
688 log.Infof("%s: reading hgvsCols from temp file", seqname)
689 f := tmpHGVSCols[seqname]
690 _, err = f.Seek(0, io.SeekStart)
694 var hgvsCols hgvsColSet
695 dec := gob.NewDecoder(bufio.NewReaderSize(f, 1<<24))
697 err = dec.Decode(&hgvsCols)
702 log.Infof("%s: sorting %d hgvs variants", seqname, len(hgvsCols))
703 variants := make([]hgvs.Variant, 0, len(hgvsCols))
704 for v := range hgvsCols {
705 variants = append(variants, v)
707 sort.Slice(variants, func(i, j int) bool {
708 vi, vj := &variants[i], &variants[j]
709 if vi.Position != vj.Position {
710 return vi.Position < vj.Position
711 } else if vi.Ref != vj.Ref {
712 return vi.Ref < vj.Ref
714 return vi.New < vj.New
717 rows := len(cmd.cgnames)
718 cols := len(variants) * 2
719 log.Infof("%s: building hgvs matrix (rows=%d, cols=%d, mem=%d)", seqname, rows, cols, rows*cols)
720 out := make([]int8, rows*cols)
721 for varIdx, variant := range variants {
722 hgvsCols := hgvsCols[variant]
723 for row := range cmd.cgnames {
724 for ph := 0; ph < 2; ph++ {
725 out[row*cols+varIdx+ph] = hgvsCols[ph][row]
729 err = writeNumpyInt8(fmt.Sprintf("%s/hgvs.%s.npy", *outputDir, seqname), out, rows, cols)
735 fnm := fmt.Sprintf("%s/hgvs.%s.annotations.csv", *outputDir, seqname)
736 log.Infof("%s: writing hgvs column labels to %s", seqname, fnm)
737 var hgvsLabels bytes.Buffer
738 for varIdx, variant := range variants {
739 fmt.Fprintf(&hgvsLabels, "%d,%s:g.%s\n", varIdx, seqname, variant.String())
741 err = ioutil.WriteFile(fnm, hgvsLabels.Bytes(), 0666)
748 if *mergeOutput || *hgvsSingle {
749 var annow *bufio.Writer
752 annoFilename := fmt.Sprintf("%s/matrix.annotations.csv", *outputDir)
753 annof, err = os.Create(annoFilename)
757 annow = bufio.NewWriterSize(annof, 1<<20)
760 rows := len(cmd.cgnames)
762 for _, chunk := range toMerge {
763 cols += len(chunk) / rows
765 log.Infof("merging output matrix (rows=%d, cols=%d, mem=%d) and annotations", rows, cols, rows*cols*2)
768 out = make([]int16, rows*cols)
770 hgvsCols := map[string][2][]int16{} // hgvs -> [[g0,g1,g2,...], [g0,g1,g2,...]] (slice of genomes for each phase)
772 for outIdx, chunk := range toMerge {
773 chunkcols := len(chunk) / rows
775 for row := 0; row < rows; row++ {
776 copy(out[row*cols+startcol:], chunk[row*chunkcols:(row+1)*chunkcols])
779 toMerge[outIdx] = nil
781 annotationsFilename := fmt.Sprintf("%s/matrix.%04d.annotations.csv", *outputDir, outIdx)
782 log.Infof("reading %s", annotationsFilename)
783 buf, err := os.ReadFile(annotationsFilename)
788 err = os.Remove(annotationsFilename)
793 for _, line := range bytes.Split(buf, []byte{'\n'}) {
797 fields := bytes.SplitN(line, []byte{','}, 9)
798 tag, _ := strconv.Atoi(string(fields[0]))
799 incol, _ := strconv.Atoi(string(fields[1]))
800 tileVariant, _ := strconv.Atoi(string(fields[2]))
801 hgvsID := string(fields[3])
802 seqname := string(fields[4])
803 pos, _ := strconv.Atoi(string(fields[5]))
806 // Null entry for un-diffable
811 // Null entry for ref tile
814 if mask != nil && !mask.Check(strings.TrimPrefix(seqname, "chr"), pos, pos+len(refseq)) {
815 // The tile intersects one of
816 // the selected regions, but
817 // this particular HGVS
821 hgvsColPair := hgvsCols[hgvsID]
822 if hgvsColPair[0] == nil {
823 // values in new columns start
824 // out as -1 ("no data yet")
825 // or 0 ("=ref") here, may
826 // change to 1 ("hgvs variant
827 // present") below, either on
828 // this line or a future line.
829 hgvsColPair = [2][]int16{make([]int16, len(cmd.cgnames)), make([]int16, len(cmd.cgnames))}
830 rt, ok := reftile[tagID(tag)]
832 err = fmt.Errorf("bug: seeing annotations for tag %d, but it has no reftile entry", tag)
835 for ph := 0; ph < 2; ph++ {
836 for row := 0; row < rows; row++ {
837 v := chunk[row*chunkcols+incol*2+ph]
838 if tileVariantID(v) == rt.variant {
839 hgvsColPair[ph][row] = 0
841 hgvsColPair[ph][row] = -1
845 hgvsCols[hgvsID] = hgvsColPair
847 hgvsref := hgvs.Variant{
852 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])
856 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])
858 for ph := 0; ph < 2; ph++ {
859 for row := 0; row < rows; row++ {
860 v := chunk[row*chunkcols+incol*2+ph]
861 if int(v) == tileVariant {
862 hgvsColPair[ph][row] = 1
868 startcol += chunkcols
879 err = writeNumpyInt16(fmt.Sprintf("%s/matrix.npy", *outputDir), out, rows, cols)
887 cols = len(hgvsCols) * 2
888 log.Printf("building hgvs-based matrix: %d rows x %d cols", rows, cols)
889 out = make([]int16, rows*cols)
890 hgvsIDs := make([]string, 0, cols/2)
891 for hgvsID := range hgvsCols {
892 hgvsIDs = append(hgvsIDs, hgvsID)
894 sort.Strings(hgvsIDs)
895 var hgvsLabels bytes.Buffer
896 for idx, hgvsID := range hgvsIDs {
897 fmt.Fprintf(&hgvsLabels, "%d,%s\n", idx, hgvsID)
898 for ph := 0; ph < 2; ph++ {
899 hgvscol := hgvsCols[hgvsID][ph]
900 for row, val := range hgvscol {
901 out[row*cols+idx*2+ph] = val
905 err = writeNumpyInt16(fmt.Sprintf("%s/hgvs.npy", *outputDir), out, rows, cols)
910 fnm := fmt.Sprintf("%s/hgvs.annotations.csv", *outputDir)
911 log.Printf("writing hgvs labels: %s", fnm)
912 err = ioutil.WriteFile(fnm, hgvsLabels.Bytes(), 0777)
920 for _, part := range onehotIndirect {
921 nzCount += len(part[0])
923 onehot := make([]uint32, nzCount*2) // [r,r,r,...,c,c,c,...]
924 var xrefs []onehotXref
926 for i, part := range onehotIndirect {
927 for i := range part[1] {
928 part[1][i] += uint32(outcol)
930 copy(onehot[outcol:], part[0])
931 copy(onehot[outcol+nzCount:], part[1])
932 outcol += len(part[0])
933 xrefs = append(xrefs, onehotXrefs[i]...)
940 fnm := fmt.Sprintf("%s/onehot.npy", *outputDir)
941 err = writeNumpyUint32(fnm, onehot, 2, nzCount)
945 fnm = fmt.Sprintf("%s/onehot-columns.npy", *outputDir)
946 err = writeNumpyInt32(fnm, onehotXref2int32(xrefs), 4, len(xrefs))
954 // Read case/control files, remove non-case/control entries from
955 // cmd.cgnames, and build cmd.chi2Cases.
956 func (cmd *sliceNumpy) useCaseControlFiles() error {
957 if cmd.chi2CaseControlFile == "" {
960 infiles, err := allFiles(cmd.chi2CaseControlFile, nil)
964 // index in cmd.cgnames => case(true) / control(false)
966 for _, infile := range infiles {
967 f, err := open(infile)
971 buf, err := io.ReadAll(f)
977 for _, tsv := range bytes.Split(buf, []byte{'\n'}) {
981 split := strings.Split(string(tsv), "\t")
984 for col, name := range split {
985 if name == cmd.chi2CaseControlColumn {
991 return fmt.Errorf("%s: no column named %q in header row %q", infile, cmd.chi2CaseControlColumn, tsv)
995 if len(split) <= ccCol {
1000 for i, name := range cmd.cgnames {
1001 if strings.Contains(name, pattern) {
1003 log.Warnf("pattern %q in %s matches multiple genome IDs (%qs, %q)", pattern, infile, cmd.cgnames[found], name)
1009 log.Warnf("pattern %q in %s does not match any genome IDs", pattern, infile)
1012 if split[ccCol] == "0" {
1015 if split[ccCol] == "1" {
1020 allnames := cmd.cgnames
1024 for i, name := range allnames {
1025 if cc, ok := cc[i]; ok {
1026 cmd.cgnames = append(cmd.cgnames, name)
1027 cmd.chi2Cases = append(cmd.chi2Cases, cc)
1033 log.Printf("%d cases, %d controls, %d neither (dropped)", ncases, len(cmd.cgnames)-ncases, len(allnames)-len(cmd.cgnames))
1037 func (cmd *sliceNumpy) filterHGVScolpair(colpair [2][]int8) bool {
1038 if cmd.chi2PValue >= 1 {
1041 col0 := make([]bool, 0, len(cmd.chi2Cases))
1042 col1 := make([]bool, 0, len(cmd.chi2Cases))
1043 cases := make([]bool, 0, len(cmd.chi2Cases))
1044 for i, c := range cmd.chi2Cases {
1045 if colpair[0][i] < 0 {
1048 col0 = append(col0, colpair[0][i] != 0)
1049 col1 = append(col1, colpair[1][i] != 0)
1050 cases = append(cases, c)
1052 return len(cases) >= cmd.minCoverage &&
1053 (pvalue(col0, cases) <= cmd.chi2PValue || pvalue(col1, cases) <= cmd.chi2PValue)
1056 func writeNumpyUint32(fnm string, out []uint32, rows, cols int) error {
1057 output, err := os.Create(fnm)
1061 defer output.Close()
1062 bufw := bufio.NewWriterSize(output, 1<<26)
1063 npw, err := gonpy.NewWriter(nopCloser{bufw})
1067 log.WithFields(log.Fields{
1071 "bytes": rows * cols * 4,
1072 }).Infof("writing numpy: %s", fnm)
1073 npw.Shape = []int{rows, cols}
1074 npw.WriteUint32(out)
1079 return output.Close()
1082 func writeNumpyInt32(fnm string, out []int32, rows, cols int) error {
1083 output, err := os.Create(fnm)
1087 defer output.Close()
1088 bufw := bufio.NewWriterSize(output, 1<<26)
1089 npw, err := gonpy.NewWriter(nopCloser{bufw})
1093 log.WithFields(log.Fields{
1097 "bytes": rows * cols * 4,
1098 }).Infof("writing numpy: %s", fnm)
1099 npw.Shape = []int{rows, cols}
1105 return output.Close()
1108 func writeNumpyInt16(fnm string, out []int16, rows, cols int) error {
1109 output, err := os.Create(fnm)
1113 defer output.Close()
1114 bufw := bufio.NewWriterSize(output, 1<<26)
1115 npw, err := gonpy.NewWriter(nopCloser{bufw})
1119 log.WithFields(log.Fields{
1123 "bytes": rows * cols * 2,
1124 }).Infof("writing numpy: %s", fnm)
1125 npw.Shape = []int{rows, cols}
1131 return output.Close()
1134 func writeNumpyInt8(fnm string, out []int8, rows, cols int) error {
1135 output, err := os.Create(fnm)
1139 defer output.Close()
1140 bufw := bufio.NewWriterSize(output, 1<<26)
1141 npw, err := gonpy.NewWriter(nopCloser{bufw})
1145 log.WithFields(log.Fields{
1149 "bytes": rows * cols,
1150 }).Infof("writing numpy: %s", fnm)
1151 npw.Shape = []int{rows, cols}
1157 return output.Close()
1160 func allele2homhet(colpair [2][]int8) {
1161 a, b := colpair[0], colpair[1]
1162 for i, av := range a {
1164 if av < 0 || bv < 0 {
1167 } else if av > 0 && bv > 0 {
1170 } else if av > 0 || bv > 0 {
1174 // ref (or a different variant in same position)
1175 // (this is a no-op) a[i], b[i] = 0, 0
1180 type onehotXref struct {
1182 variant tileVariantID
1187 const onehotXrefSize = unsafe.Sizeof(onehotXref{})
1189 // Build onehot matrix (m[variant*2+isHet][genome] == 0 or 1) for all
1190 // variants of a single tile/tag#.
1192 // Return nil if no tile variant passes Χ² filter.
1193 func (cmd *sliceNumpy) tv2homhet(cgs map[string]CompactGenome, maxv tileVariantID, remap []tileVariantID, tag, chunkstarttag tagID) ([][]int8, []onehotXref) {
1195 // everyone has the most common variant
1198 tagoffset := tag - chunkstarttag
1200 for _, cg := range cgs {
1201 if cg.Variants[tagoffset*2] > 0 && cg.Variants[tagoffset*2+1] > 0 {
1205 if coverage < cmd.minCoverage {
1208 obs := make([][]bool, (maxv+1)*2) // 2 slices (hom + het) for each variant#
1209 for i := range obs {
1210 obs[i] = make([]bool, len(cmd.cgnames))
1212 for cgid, name := range cmd.cgnames {
1213 cgvars := cgs[name].Variants
1214 for v := tileVariantID(2); v <= maxv; v++ {
1215 if remap[cgvars[tagoffset*2]] == v && remap[cgvars[tagoffset*2+1]] == v {
1216 obs[v*2][cgid] = true
1217 } else if remap[cgvars[tagoffset*2]] == v || remap[cgvars[tagoffset*2+1]] == v {
1218 obs[v*2+1][cgid] = true
1223 var xref []onehotXref
1224 for homcol := 4; homcol < len(obs); homcol += 2 {
1225 for het := 0; het < 2; het++ {
1226 p := pvalue(obs[homcol+het], cmd.chi2Cases)
1227 if cmd.chi2PValue < 1 && !(p < cmd.chi2PValue) {
1230 onehot = append(onehot, bool2int8(obs[homcol+het]))
1231 xref = append(xref, onehotXref{
1233 variant: tileVariantID(homcol / 2),
1242 func bool2int8(in []bool) []int8 {
1243 out := make([]int8, len(in))
1244 for i, v := range in {
1252 // convert a []onehotXref with length N to a numpy-style []int32
1253 // matrix with N columns, one row per field of onehotXref struct.
1255 // Hom/het row contains hom=0, het=1.
1257 // P-value row contains 1000000x actual p-value.
1258 func onehotXref2int32(xrefs []onehotXref) []int32 {
1260 xdata := make([]int32, 4*xcols)
1261 for i, xref := range xrefs {
1262 xdata[i] = int32(xref.tag)
1263 xdata[xcols+i] = int32(xref.variant)
1265 xdata[xcols*2+i] = 1
1267 xdata[xcols*3+i] = int32(xref.pvalue * 1000000)
1272 // transpose onehot data from in[col][row] to numpy-style
1273 // out[row*cols+col].
1274 func onehotcols2int8(in [][]int8) []int8 {
1280 out := make([]int8, rows*cols)
1281 for row := 0; row < rows; row++ {
1282 outrow := out[row*cols:]
1283 for col, incol := range in {
1284 outrow[col] = incol[row]
1290 // Return [2][]uint32{rowIndices, colIndices} indicating which
1291 // elements of matrixT[c][r] have non-zero values.
1292 func onehotChunk2Indirect(matrixT [][]int8) [2][]uint32 {
1294 for c, col := range matrixT {
1295 for r, val := range col {
1297 nz[0] = append(nz[0], uint32(r))
1298 nz[1] = append(nz[1], uint32(c))