1 // Copyright (C) The Lightning Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
30 "git.arvados.org/arvados.git/sdk/go/arvados"
31 "github.com/arvados/lightning/hgvs"
32 "github.com/james-bowman/nlp"
33 "github.com/kshedden/gonpy"
34 "github.com/sirupsen/logrus"
35 log "github.com/sirupsen/logrus"
36 "golang.org/x/crypto/blake2b"
37 "gonum.org/v1/gonum/mat"
40 const annotationMaxTileSpan = 100
42 type sliceNumpy struct {
47 pvalueMinFrequency float64
55 trainingSet []int // samples index => training set index, or -1 if not in training set
57 pvalue func(onehot []bool) float64
61 func (cmd *sliceNumpy) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
62 err := cmd.run(prog, args, stdin, stdout, stderr)
64 fmt.Fprintf(stderr, "%s\n", err)
70 func (cmd *sliceNumpy) run(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) error {
71 flags := flag.NewFlagSet("", flag.ContinueOnError)
72 flags.SetOutput(stderr)
73 pprof := flags.String("pprof", "", "serve Go profile data at http://`[addr]:port`")
74 runlocal := flags.Bool("local", false, "run on local host (default: run in an arvados container)")
75 arvadosRAM := flags.Int("arvados-ram", 750000000000, "amount of memory to request for arvados container (`bytes`)")
76 arvadosVCPUs := flags.Int("arvados-vcpus", 96, "number of VCPUs to request for arvados container")
77 projectUUID := flags.String("project", "", "project `UUID` for output data")
78 priority := flags.Int("priority", 500, "container request priority")
79 preemptible := flags.Bool("preemptible", true, "request preemptible instance")
80 inputDir := flags.String("input-dir", "./in", "input `directory`")
81 outputDir := flags.String("output-dir", "./out", "output `directory`")
82 ref := flags.String("ref", "", "reference name (if blank, choose last one that appears in input)")
83 regionsFilename := flags.String("regions", "", "only output columns/annotations that intersect regions in specified bed `file`")
84 expandRegions := flags.Int("expand-regions", 0, "expand specified regions by `N` base pairs on each side`")
85 mergeOutput := flags.Bool("merge-output", false, "merge output into one matrix.npy and one matrix.annotations.csv")
86 hgvsSingle := flags.Bool("single-hgvs-matrix", false, "also generate hgvs-based matrix")
87 hgvsChunked := flags.Bool("chunked-hgvs-matrix", false, "also generate hgvs-based matrix per chromosome")
88 onehotSingle := flags.Bool("single-onehot", false, "generate one-hot tile-based matrix")
89 onehotChunked := flags.Bool("chunked-onehot", false, "generate one-hot tile-based matrix per input chunk")
90 samplesFilename := flags.String("samples", "", "`samples.csv` file with training/validation and case/control groups (see 'lightning choose-samples')")
91 caseControlOnly := flags.Bool("case-control-only", false, "drop samples that are not in case/control groups")
92 onlyPCA := flags.Bool("pca", false, "run principal component analysis, write components to pca.npy and samples.csv")
93 flags.IntVar(&cmd.pcaComponents, "pca-components", 4, "number of PCA components to compute / use in logistic regression")
94 maxPCATiles := flags.Int("max-pca-tiles", 0, "maximum tiles to use as PCA input (filter, then drop every 2nd colum pair until below max)")
95 debugTag := flags.Int("debug-tag", -1, "log debugging details about specified tag")
96 flags.IntVar(&cmd.threads, "threads", 16, "number of memory-hungry assembly threads, and number of VCPUs to request for arvados container")
97 flags.Float64Var(&cmd.chi2PValue, "chi2-p-value", 1, "do Χ² test (or logistic regression if -samples file has PCA components) and omit columns with p-value above this threshold")
98 flags.Float64Var(&cmd.pvalueMinFrequency, "pvalue-min-frequency", 0.01, "skip p-value calculation on tile variants below this frequency in the training set")
99 flags.BoolVar(&cmd.includeVariant1, "include-variant-1", false, "include most common variant when building one-hot matrix")
100 cmd.filter.Flags(flags)
101 err := flags.Parse(args)
102 if err == flag.ErrHelp {
104 } else if err != nil {
106 } else if flags.NArg() > 0 {
107 return fmt.Errorf("errant command line arguments after parsed flags: %v", flags.Args())
112 log.Println(http.ListenAndServe(*pprof, nil))
116 if cmd.chi2PValue != 1 && *samplesFilename == "" {
117 return fmt.Errorf("cannot use provided -chi2-p-value=%f because -samples= value is empty", cmd.chi2PValue)
120 cmd.debugTag = tagID(*debugTag)
123 runner := arvadosContainerRunner{
124 Name: "lightning slice-numpy",
125 Client: arvados.NewClientFromEnv(),
126 ProjectUUID: *projectUUID,
127 RAM: int64(*arvadosRAM),
128 VCPUs: *arvadosVCPUs,
132 Preemptible: *preemptible,
134 err = runner.TranslatePaths(inputDir, regionsFilename, samplesFilename)
138 runner.Args = []string{"slice-numpy", "-local=true",
140 "-input-dir=" + *inputDir,
141 "-output-dir=/mnt/output",
142 "-threads=" + fmt.Sprintf("%d", cmd.threads),
143 "-regions=" + *regionsFilename,
144 "-expand-regions=" + fmt.Sprintf("%d", *expandRegions),
145 "-merge-output=" + fmt.Sprintf("%v", *mergeOutput),
146 "-single-hgvs-matrix=" + fmt.Sprintf("%v", *hgvsSingle),
147 "-chunked-hgvs-matrix=" + fmt.Sprintf("%v", *hgvsChunked),
148 "-single-onehot=" + fmt.Sprintf("%v", *onehotSingle),
149 "-chunked-onehot=" + fmt.Sprintf("%v", *onehotChunked),
150 "-samples=" + *samplesFilename,
151 "-case-control-only=" + fmt.Sprintf("%v", *caseControlOnly),
152 "-pca=" + fmt.Sprintf("%v", *onlyPCA),
153 "-pca-components=" + fmt.Sprintf("%d", cmd.pcaComponents),
154 "-max-pca-tiles=" + fmt.Sprintf("%d", *maxPCATiles),
155 "-chi2-p-value=" + fmt.Sprintf("%f", cmd.chi2PValue),
156 "-pvalue-min-frequency=" + fmt.Sprintf("%f", cmd.pvalueMinFrequency),
157 "-include-variant-1=" + fmt.Sprintf("%v", cmd.includeVariant1),
158 "-debug-tag=" + fmt.Sprintf("%d", cmd.debugTag),
160 runner.Args = append(runner.Args, cmd.filter.Args()...)
162 output, err = runner.Run()
166 fmt.Fprintln(stdout, output)
170 infiles, err := allFiles(*inputDir, matchGobFile)
174 if len(infiles) == 0 {
175 err = fmt.Errorf("no input files found in %s", *inputDir)
178 sort.Strings(infiles)
180 var refseq map[string][]tileLibRef
181 var reftiledata = make(map[tileLibRef][]byte, 11000000)
182 in0, err := open(infiles[0])
187 matchGenome, err := regexp.Compile(cmd.filter.MatchGenome)
189 err = fmt.Errorf("-match-genome: invalid regexp: %q", cmd.filter.MatchGenome)
193 if *samplesFilename != "" {
194 cmd.samples, err = loadSampleInfo(*samplesFilename)
198 } else if *caseControlOnly {
199 return fmt.Errorf("-case-control-only does not make sense without -samples")
204 err = DecodeLibrary(in0, strings.HasSuffix(infiles[0], ".gz"), func(ent *LibraryEntry) error {
205 if len(ent.TagSet) > 0 {
208 for _, cseq := range ent.CompactSequences {
209 if cseq.Name == *ref || *ref == "" {
210 refseq = cseq.TileSequences
213 for _, cg := range ent.CompactGenomes {
214 if matchGenome.MatchString(cg.Name) {
215 cmd.cgnames = append(cmd.cgnames, cg.Name)
218 for _, tv := range ent.TileVariants {
220 reftiledata[tileLibRef{tv.Tag, tv.Variant}] = tv.Sequence
230 err = fmt.Errorf("%s: reference sequence not found", infiles[0])
233 if len(tagset) == 0 {
234 err = fmt.Errorf("tagset not found")
238 taglib := &tagLibrary{}
239 err = taglib.setTags(tagset)
243 taglen := taglib.TagLen()
244 sort.Strings(cmd.cgnames)
246 if len(cmd.cgnames) == 0 {
247 return fmt.Errorf("fatal: 0 matching samples in library, nothing to do")
249 cmd.trainingSet = make([]int, len(cmd.cgnames))
250 if *samplesFilename == "" {
251 cmd.trainingSetSize = len(cmd.cgnames)
252 for i, name := range cmd.cgnames {
253 cmd.samples = append(cmd.samples, sampleInfo{
254 id: trimFilenameForLabel(name),
257 cmd.trainingSet[i] = i
259 } else if len(cmd.cgnames) != len(cmd.samples) {
260 return fmt.Errorf("mismatched sample list: %d samples in library, %d in %s", len(cmd.cgnames), len(cmd.samples), *samplesFilename)
262 for i, name := range cmd.cgnames {
263 if s := trimFilenameForLabel(name); s != cmd.samples[i].id {
264 return fmt.Errorf("mismatched sample list: sample %d is %q in library, %q in %s", i, s, cmd.samples[i].id, *samplesFilename)
267 if *caseControlOnly {
268 for i := 0; i < len(cmd.samples); i++ {
269 if !cmd.samples[i].isTraining && !cmd.samples[i].isValidation {
270 if i+1 < len(cmd.samples) {
271 copy(cmd.samples[i:], cmd.samples[i+1:])
272 copy(cmd.cgnames[i:], cmd.cgnames[i+1:])
274 cmd.samples = cmd.samples[:len(cmd.samples)-1]
275 cmd.cgnames = cmd.cgnames[:len(cmd.cgnames)-1]
281 cmd.trainingSetSize = 0
282 for i := range cmd.cgnames {
283 if cmd.samples[i].isTraining {
284 cmd.trainingSet[i] = cmd.trainingSetSize
285 cmd.trainingSetSize++
286 cmd.chi2Cases = append(cmd.chi2Cases, cmd.samples[i].isCase)
288 cmd.trainingSet[i] = -1
291 if cmd.pvalue == nil {
292 cmd.pvalue = func(onehot []bool) float64 {
293 return pvalue(onehot, cmd.chi2Cases)
297 if cmd.filter.MinCoverage == 1 {
298 // In the generic formula below, floating point
299 // arithmetic can effectively push the coverage
300 // threshold above 1.0, which is impossible/useless.
301 // 1.0 needs to mean exactly 100% coverage.
302 cmd.minCoverage = len(cmd.cgnames)
304 cmd.minCoverage = int(math.Ceil(cmd.filter.MinCoverage * float64(len(cmd.cgnames))))
307 if len(cmd.samples[0].pcaComponents) > 0 {
308 cmd.pvalue = glmPvalueFunc(cmd.samples, cmd.pcaComponents)
309 // Unfortunately, statsmodel/glm lib logs stuff to
310 // os.Stdout when it panics on an unsolvable
311 // problem. We recover() from the panic in glm.go, but
312 // we also need to commandeer os.Stdout to avoid
313 // producing large quantities of logs.
314 stdoutWas := os.Stdout
315 defer func() { os.Stdout = stdoutWas }()
316 os.Stdout, err = os.Open(os.DevNull)
322 // cgnamemap[name]==true for samples that we are including in
324 cgnamemap := map[string]bool{}
325 for _, name := range cmd.cgnames {
326 cgnamemap[name] = true
329 err = writeSampleInfo(cmd.samples, *outputDir)
334 log.Info("indexing reference tiles")
335 type reftileinfo struct {
336 variant tileVariantID
337 seqname string // chr1
338 pos int // distance from start of chromosome to starttag
339 tiledata []byte // acgtggcaa...
340 excluded bool // true if excluded by regions file
341 nexttag tagID // tagID of following tile (-1 for last tag of chromosome)
343 isdup := map[tagID]bool{}
344 reftile := map[tagID]*reftileinfo{}
345 for seqname, cseq := range refseq {
347 lastreftag := tagID(-1)
348 for _, libref := range cseq {
349 if cmd.filter.MaxTag >= 0 && libref.Tag > tagID(cmd.filter.MaxTag) {
352 tiledata := reftiledata[libref]
353 if len(tiledata) == 0 {
354 err = fmt.Errorf("missing tiledata for tag %d variant %d in %s in ref", libref.Tag, libref.Variant, seqname)
357 foundthistag := false
358 taglib.FindAll(tiledata[:len(tiledata)-1], func(tagid tagID, offset, _ int) {
359 if !foundthistag && tagid == libref.Tag {
363 if dupref, ok := reftile[tagid]; ok {
364 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)
365 delete(reftile, tagid)
367 log.Printf("found tag %d at offset %d inside tile variant %+v on %s @ %d", tagid, offset, libref, seqname, pos+offset+1)
371 if isdup[libref.Tag] {
372 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique", libref, seqname, pos)
373 } else if reftile[libref.Tag] != nil {
374 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)
375 delete(reftile, libref.Tag)
376 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique", libref, seqname, pos)
377 isdup[libref.Tag] = true
379 reftile[libref.Tag] = &reftileinfo{
381 variant: libref.Variant,
387 reftile[lastreftag].nexttag = libref.Tag
389 lastreftag = libref.Tag
391 pos += len(tiledata) - taglen
393 log.Printf("... %s done, len %d", seqname, pos+taglen)
397 if *regionsFilename != "" {
398 log.Printf("loading regions from %s", *regionsFilename)
399 mask, err = makeMask(*regionsFilename, *expandRegions)
403 log.Printf("before applying mask, len(reftile) == %d", len(reftile))
404 log.Printf("deleting reftile entries for regions outside %d intervals", mask.Len())
405 for _, rt := range reftile {
406 if !mask.Check(strings.TrimPrefix(rt.seqname, "chr"), rt.pos, rt.pos+len(rt.tiledata)) {
410 log.Printf("after applying mask, len(reftile) == %d", len(reftile))
413 type hgvsColSet map[hgvs.Variant][2][]int8
414 encodeHGVS := throttle{Max: len(refseq)}
415 encodeHGVSTodo := map[string]chan hgvsColSet{}
416 tmpHGVSCols := map[string]*os.File{}
418 for seqname := range refseq {
420 f, err = os.Create(*outputDir + "/tmp." + seqname + ".gob")
424 defer os.Remove(f.Name())
425 bufw := bufio.NewWriterSize(f, 1<<24)
426 enc := gob.NewEncoder(bufw)
427 tmpHGVSCols[seqname] = f
428 todo := make(chan hgvsColSet, 128)
429 encodeHGVSTodo[seqname] = todo
430 encodeHGVS.Go(func() error {
431 for colset := range todo {
432 err := enc.Encode(colset)
434 encodeHGVS.Report(err)
445 var toMerge [][]int16
446 if *mergeOutput || *hgvsSingle {
447 toMerge = make([][]int16, len(infiles))
449 var onehotIndirect [][2][]uint32 // [chunkIndex][axis][index]
450 var onehotChunkSize []uint32
451 var onehotXrefs [][]onehotXref
452 if *onehotSingle || *onlyPCA {
453 onehotIndirect = make([][2][]uint32, len(infiles))
454 onehotChunkSize = make([]uint32, len(infiles))
455 onehotXrefs = make([][]onehotXref, len(infiles))
457 chunkStartTag := make([]tagID, len(infiles))
459 throttleMem := throttle{Max: cmd.threads} // TODO: estimate using mem and data size
460 throttleNumpyMem := throttle{Max: cmd.threads/2 + 1}
461 log.Info("generating annotations and numpy matrix for each slice")
462 var errSkip = errors.New("skip infile")
464 for infileIdx, infile := range infiles {
465 infileIdx, infile := infileIdx, infile
466 throttleMem.Go(func() error {
467 seq := make(map[tagID][]TileVariant, 50000)
468 cgs := make(map[string]CompactGenome, len(cmd.cgnames))
469 f, err := open(infile)
474 log.Infof("%04d: reading %s", infileIdx, infile)
475 err = DecodeLibrary(f, strings.HasSuffix(infile, ".gz"), func(ent *LibraryEntry) error {
476 for _, tv := range ent.TileVariants {
481 // corresponding ref tile, if
482 // mask is in play (we can't
483 // determine coordinates for
485 if mask != nil && reftile[tv.Tag] == nil {
489 // corresponding ref tile is
490 // outside target regions --
491 // unless it's a potential
493 if mask != nil && reftile[tv.Tag].excluded &&
494 (int(tv.Tag+1) >= len(tagset) ||
495 (bytes.HasSuffix(tv.Sequence, tagset[tv.Tag+1]) && reftile[tv.Tag+1] != nil && !reftile[tv.Tag+1].excluded)) {
498 if tv.Tag == cmd.debugTag {
499 log.Printf("infile %d %s tag %d variant %d hash %x", infileIdx, infile, tv.Tag, tv.Variant, tv.Blake2b[:3])
501 variants := seq[tv.Tag]
502 if len(variants) == 0 {
503 variants = make([]TileVariant, 100)
505 for len(variants) <= int(tv.Variant) {
506 variants = append(variants, TileVariant{})
508 variants[int(tv.Variant)] = tv
509 seq[tv.Tag] = variants
511 for _, cg := range ent.CompactGenomes {
512 if cmd.filter.MaxTag >= 0 && cg.StartTag > tagID(cmd.filter.MaxTag) {
515 if !cgnamemap[cg.Name] {
518 // pad to full slice size
519 // to avoid out-of-bounds
521 if sliceSize := 2 * int(cg.EndTag-cg.StartTag); len(cg.Variants) < sliceSize {
522 cg.Variants = append(cg.Variants, make([]tileVariantID, sliceSize-len(cg.Variants))...)
530 } else if err != nil {
531 return fmt.Errorf("%04d: DecodeLibrary(%s): err", infileIdx, infile)
533 tagstart := cgs[cmd.cgnames[0]].StartTag
534 tagend := cgs[cmd.cgnames[0]].EndTag
535 chunkStartTag[infileIdx] = tagstart
539 log.Infof("%04d: renumber/dedup variants for tags %d-%d", infileIdx, tagstart, tagend)
540 variantRemap := make([][]tileVariantID, tagend-tagstart)
541 throttleCPU := throttle{Max: runtime.GOMAXPROCS(0)}
542 for tag, variants := range seq {
543 tag, variants := tag, variants
544 throttleCPU.Go(func() error {
546 count := make(map[[blake2b.Size256]byte]int, len(variants))
550 count[blake2b.Sum256(rt.tiledata)] = 0
553 for cgname, cg := range cgs {
554 idx := int(tag-tagstart) * 2
555 for allele := 0; allele < 2; allele++ {
556 v := cg.Variants[idx+allele]
557 if v > 0 && len(variants[v].Sequence) > 0 {
558 count[variants[v].Blake2b]++
561 if v > 0 && tag == cmd.debugTag {
562 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])
566 if alleleCoverage < cmd.minCoverage*2 {
567 idx := int(tag-tagstart) * 2
568 for _, cg := range cgs {
570 cg.Variants[idx+1] = 0
572 if tag == cmd.debugTag {
573 log.Printf("tag %d alleleCoverage %d < min %d, sample data wiped", tag, alleleCoverage, cmd.minCoverage*2)
578 // hash[i] will be the hash of
579 // the variant(s) that should
580 // be at rank i (0-based).
581 hash := make([][blake2b.Size256]byte, 0, len(count))
582 for b := range count {
583 hash = append(hash, b)
585 sort.Slice(hash, func(i, j int) bool {
586 bi, bj := &hash[i], &hash[j]
587 if ci, cj := count[*bi], count[*bj]; ci != cj {
590 return bytes.Compare((*bi)[:], (*bj)[:]) < 0
593 // rank[b] will be the 1-based
594 // new variant number for
595 // variants whose hash is b.
596 rank := make(map[[blake2b.Size256]byte]tileVariantID, len(hash))
597 for i, h := range hash {
598 rank[h] = tileVariantID(i + 1)
600 if tag == cmd.debugTag {
601 for h, r := range rank {
602 log.Printf("tag %d rank(%x) = %v", tag, h[:3], r)
605 // remap[v] will be the new
606 // variant number for original
608 remap := make([]tileVariantID, len(variants))
609 for i, tv := range variants {
610 remap[i] = rank[tv.Blake2b]
612 if tag == cmd.debugTag {
613 for in, out := range remap {
615 log.Printf("tag %d remap %d => %d", tag, in, out)
619 variantRemap[tag-tagstart] = remap
621 refrank := rank[blake2b.Sum256(rt.tiledata)]
622 if tag == cmd.debugTag {
623 log.Printf("tag %d reftile variant %d => %d", tag, rt.variant, refrank)
632 var onehotChunk [][]int8
633 var onehotXref []onehotXref
635 var annotationsFilename string
637 annotationsFilename = "/dev/null"
639 annotationsFilename = fmt.Sprintf("%s/matrix.%04d.annotations.csv", *outputDir, infileIdx)
640 log.Infof("%04d: writing %s", infileIdx, annotationsFilename)
642 annof, err := os.Create(annotationsFilename)
646 annow := bufio.NewWriterSize(annof, 1<<20)
648 for tag := tagstart; tag < tagend; tag++ {
650 if rt == nil && mask != nil {
651 // With no ref tile, we don't
652 // have coordinates to say
653 // this is in the desired
654 // regions -- so it's not.
655 // TODO: handle ref spanning
659 if rt != nil && rt.excluded {
660 // TODO: don't skip yet --
661 // first check for spanning
662 // tile variants that
663 // intersect non-excluded ref
667 if cmd.filter.MaxTag >= 0 && tag > tagID(cmd.filter.MaxTag) {
670 remap := variantRemap[tag-tagstart]
672 // was not assigned above,
673 // because minCoverage
677 maxv := tileVariantID(0)
678 for _, v := range remap {
683 if *onehotChunked || *onehotSingle || *onlyPCA {
684 onehot, xrefs := cmd.tv2homhet(cgs, maxv, remap, tag, tagstart, seq)
685 if tag == cmd.debugTag {
686 log.WithFields(logrus.Fields{
689 }).Info("tv2homhet()")
691 onehotChunk = append(onehotChunk, onehot...)
692 onehotXref = append(onehotXref, xrefs...)
699 // Reference does not use any
700 // variant of this tile
702 // TODO: diff against the
703 // relevant portion of the
704 // ref's spanning tile
708 fmt.Fprintf(annow, "%d,%d,%d,=,%s,%d,,,\n", tag, outcol, rt.variant, rt.seqname, rt.pos)
710 reftilestr := strings.ToUpper(string(rt.tiledata))
712 done := make([]bool, maxv+1)
713 variantDiffs := make([][]hgvs.Variant, maxv+1)
714 for v, tv := range variants {
716 if v == 0 || v == rt.variant || done[v] {
721 if len(tv.Sequence) < taglen {
724 // if reftilestr doesn't end
725 // in the same tag as tv,
726 // extend reftilestr with
727 // following ref tiles until
728 // it does (up to an arbitrary
729 // sanity-check limit)
730 reftilestr := reftilestr
731 endtagstr := strings.ToUpper(string(tv.Sequence[len(tv.Sequence)-taglen:]))
732 for i, rt := 0, rt; i < annotationMaxTileSpan && !strings.HasSuffix(reftilestr, endtagstr) && rt.nexttag >= 0; i++ {
733 rt = reftile[rt.nexttag]
737 reftilestr += strings.ToUpper(string(rt.tiledata[taglen:]))
739 if mask != nil && !mask.Check(strings.TrimPrefix(rt.seqname, "chr"), rt.pos, rt.pos+len(reftilestr)) {
742 if !strings.HasSuffix(reftilestr, endtagstr) {
743 fmt.Fprintf(annow, "%d,%d,%d,,%s,%d,,,\n", tag, outcol, v, rt.seqname, rt.pos)
746 if lendiff := len(reftilestr) - len(tv.Sequence); lendiff < -1000 || lendiff > 1000 {
747 fmt.Fprintf(annow, "%d,%d,%d,,%s,%d,,,\n", tag, outcol, v, rt.seqname, rt.pos)
750 diffs, _ := hgvs.Diff(reftilestr, strings.ToUpper(string(tv.Sequence)), 0)
751 for i := range diffs {
752 diffs[i].Position += rt.pos
754 for _, diff := range diffs {
755 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)
758 variantDiffs[v] = diffs
762 // We can now determine, for each HGVS
763 // variant (diff) in this reftile
764 // region, whether a given genome
765 // phase/allele (1) has the variant, (0) has
766 // =ref or a different variant in that
767 // position, or (-1) is lacking
768 // coverage / couldn't be diffed.
769 hgvsCol := hgvsColSet{}
770 for _, diffs := range variantDiffs {
771 for _, diff := range diffs {
772 if _, ok := hgvsCol[diff]; ok {
775 hgvsCol[diff] = [2][]int8{
776 make([]int8, len(cmd.cgnames)),
777 make([]int8, len(cmd.cgnames)),
781 for row, name := range cmd.cgnames {
782 variants := cgs[name].Variants[(tag-tagstart)*2:]
783 for ph := 0; ph < 2; ph++ {
785 if int(v) >= len(remap) {
791 // hgvsCol[*][ph][row] is already 0
792 } else if len(variantDiffs[v]) == 0 {
793 // lacking coverage / couldn't be diffed
794 for _, col := range hgvsCol {
798 for _, diff := range variantDiffs[v] {
799 hgvsCol[diff][ph][row] = 1
804 for diff, colpair := range hgvsCol {
805 allele2homhet(colpair)
806 if !cmd.filterHGVScolpair(colpair) {
807 delete(hgvsCol, diff)
810 if len(hgvsCol) > 0 {
811 encodeHGVSTodo[rt.seqname] <- hgvsCol
826 // transpose onehotChunk[col][row] to numpy[row*ncols+col]
827 rows := len(cmd.cgnames)
828 cols := len(onehotChunk)
829 log.Infof("%04d: preparing onehot numpy (rows=%d, cols=%d, mem=%d)", infileIdx, rows, cols, rows*cols)
830 throttleNumpyMem.Acquire()
831 out := onehotcols2int8(onehotChunk)
832 fnm := fmt.Sprintf("%s/onehot.%04d.npy", *outputDir, infileIdx)
833 err = writeNumpyInt8(fnm, out, rows, cols)
837 fnm = fmt.Sprintf("%s/onehot-columns.%04d.npy", *outputDir, infileIdx)
838 err = writeNumpyInt32(fnm, onehotXref2int32(onehotXref), 4, len(onehotXref))
843 throttleNumpyMem.Release()
845 if *onehotSingle || *onlyPCA {
846 onehotIndirect[infileIdx] = onehotChunk2Indirect(onehotChunk)
847 onehotChunkSize[infileIdx] = uint32(len(onehotChunk))
848 onehotXrefs[infileIdx] = onehotXref
849 n := len(onehotIndirect[infileIdx][0])
850 log.Infof("%04d: keeping onehot coordinates in memory (n=%d, mem=%d)", infileIdx, n, n*8*2)
852 if !(*onehotSingle || *onehotChunked || *onlyPCA) || *mergeOutput || *hgvsSingle {
853 log.Infof("%04d: preparing numpy (rows=%d, cols=%d)", infileIdx, len(cmd.cgnames), 2*outcol)
854 throttleNumpyMem.Acquire()
855 rows := len(cmd.cgnames)
857 out := make([]int16, rows*cols)
858 for row, name := range cmd.cgnames {
860 for col, v := range cgs[name].Variants {
861 tag := tagstart + tagID(col/2)
862 if cmd.filter.MaxTag >= 0 && tag > tagID(cmd.filter.MaxTag) {
865 if rt := reftile[tag]; rt == nil || rt.excluded {
869 out[outidx] = 0 // tag not found / spanning tile
870 } else if variants, ok := seq[tag]; ok && int(v) < len(variants) && len(variants[v].Sequence) > 0 {
871 out[outidx] = int16(variantRemap[tag-tagstart][v])
873 out[outidx] = -1 // low quality tile variant
875 if tag == cmd.debugTag {
876 log.Printf("tag %d row %d col %d outidx %d v %d out %d", tag, row, col, outidx, v, out[outidx])
884 throttleNumpyMem.Release()
885 if *mergeOutput || *hgvsSingle {
886 log.Infof("%04d: matrix fragment %d rows x %d cols", infileIdx, rows, cols)
887 toMerge[infileIdx] = out
889 if !*mergeOutput && !*onehotChunked && !*onehotSingle {
890 fnm := fmt.Sprintf("%s/matrix.%04d.npy", *outputDir, infileIdx)
891 err = writeNumpyInt16(fnm, out, rows, cols)
898 log.Infof("%s: done (%d/%d)", infile, int(atomic.AddInt64(&done, 1)), len(infiles))
902 if err = throttleMem.Wait(); err != nil {
907 log.Info("flushing hgvsCols temp files")
908 for seqname := range refseq {
909 close(encodeHGVSTodo[seqname])
911 err = encodeHGVS.Wait()
915 for seqname := range refseq {
916 log.Infof("%s: reading hgvsCols from temp file", seqname)
917 f := tmpHGVSCols[seqname]
918 _, err = f.Seek(0, io.SeekStart)
922 var hgvsCols hgvsColSet
923 dec := gob.NewDecoder(bufio.NewReaderSize(f, 1<<24))
925 err = dec.Decode(&hgvsCols)
930 log.Infof("%s: sorting %d hgvs variants", seqname, len(hgvsCols))
931 variants := make([]hgvs.Variant, 0, len(hgvsCols))
932 for v := range hgvsCols {
933 variants = append(variants, v)
935 sort.Slice(variants, func(i, j int) bool {
936 vi, vj := &variants[i], &variants[j]
937 if vi.Position != vj.Position {
938 return vi.Position < vj.Position
939 } else if vi.Ref != vj.Ref {
940 return vi.Ref < vj.Ref
942 return vi.New < vj.New
945 rows := len(cmd.cgnames)
946 cols := len(variants) * 2
947 log.Infof("%s: building hgvs matrix (rows=%d, cols=%d, mem=%d)", seqname, rows, cols, rows*cols)
948 out := make([]int8, rows*cols)
949 for varIdx, variant := range variants {
950 hgvsCols := hgvsCols[variant]
951 for row := range cmd.cgnames {
952 for ph := 0; ph < 2; ph++ {
953 out[row*cols+varIdx+ph] = hgvsCols[ph][row]
957 err = writeNumpyInt8(fmt.Sprintf("%s/hgvs.%s.npy", *outputDir, seqname), out, rows, cols)
963 fnm := fmt.Sprintf("%s/hgvs.%s.annotations.csv", *outputDir, seqname)
964 log.Infof("%s: writing hgvs column labels to %s", seqname, fnm)
965 var hgvsLabels bytes.Buffer
966 for varIdx, variant := range variants {
967 fmt.Fprintf(&hgvsLabels, "%d,%s:g.%s\n", varIdx, seqname, variant.String())
969 err = ioutil.WriteFile(fnm, hgvsLabels.Bytes(), 0666)
976 if *mergeOutput || *hgvsSingle {
977 var annow *bufio.Writer
980 annoFilename := fmt.Sprintf("%s/matrix.annotations.csv", *outputDir)
981 annof, err = os.Create(annoFilename)
985 annow = bufio.NewWriterSize(annof, 1<<20)
988 rows := len(cmd.cgnames)
990 for _, chunk := range toMerge {
991 cols += len(chunk) / rows
993 log.Infof("merging output matrix (rows=%d, cols=%d, mem=%d) and annotations", rows, cols, rows*cols*2)
996 out = make([]int16, rows*cols)
998 hgvsCols := map[string][2][]int16{} // hgvs -> [[g0,g1,g2,...], [g0,g1,g2,...]] (slice of genomes for each phase)
1000 for outIdx, chunk := range toMerge {
1001 chunkcols := len(chunk) / rows
1003 for row := 0; row < rows; row++ {
1004 copy(out[row*cols+startcol:], chunk[row*chunkcols:(row+1)*chunkcols])
1007 toMerge[outIdx] = nil
1009 annotationsFilename := fmt.Sprintf("%s/matrix.%04d.annotations.csv", *outputDir, outIdx)
1010 log.Infof("reading %s", annotationsFilename)
1011 buf, err := os.ReadFile(annotationsFilename)
1016 err = os.Remove(annotationsFilename)
1021 for _, line := range bytes.Split(buf, []byte{'\n'}) {
1025 fields := bytes.SplitN(line, []byte{','}, 9)
1026 tag, _ := strconv.Atoi(string(fields[0]))
1027 incol, _ := strconv.Atoi(string(fields[1]))
1028 tileVariant, _ := strconv.Atoi(string(fields[2]))
1029 hgvsID := string(fields[3])
1030 seqname := string(fields[4])
1031 pos, _ := strconv.Atoi(string(fields[5]))
1034 // Null entry for un-diffable
1039 // Null entry for ref tile
1042 if mask != nil && !mask.Check(strings.TrimPrefix(seqname, "chr"), pos, pos+len(refseq)) {
1043 // The tile intersects one of
1044 // the selected regions, but
1045 // this particular HGVS
1046 // variant does not.
1049 hgvsColPair := hgvsCols[hgvsID]
1050 if hgvsColPair[0] == nil {
1051 // values in new columns start
1052 // out as -1 ("no data yet")
1053 // or 0 ("=ref") here, may
1054 // change to 1 ("hgvs variant
1055 // present") below, either on
1056 // this line or a future line.
1057 hgvsColPair = [2][]int16{make([]int16, len(cmd.cgnames)), make([]int16, len(cmd.cgnames))}
1058 rt, ok := reftile[tagID(tag)]
1060 err = fmt.Errorf("bug: seeing annotations for tag %d, but it has no reftile entry", tag)
1063 for ph := 0; ph < 2; ph++ {
1064 for row := 0; row < rows; row++ {
1065 v := chunk[row*chunkcols+incol*2+ph]
1066 if tileVariantID(v) == rt.variant {
1067 hgvsColPair[ph][row] = 0
1069 hgvsColPair[ph][row] = -1
1073 hgvsCols[hgvsID] = hgvsColPair
1075 hgvsref := hgvs.Variant{
1077 Ref: string(refseq),
1078 New: string(refseq),
1080 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])
1084 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])
1086 for ph := 0; ph < 2; ph++ {
1087 for row := 0; row < rows; row++ {
1088 v := chunk[row*chunkcols+incol*2+ph]
1089 if int(v) == tileVariant {
1090 hgvsColPair[ph][row] = 1
1096 startcol += chunkcols
1107 err = writeNumpyInt16(fmt.Sprintf("%s/matrix.npy", *outputDir), out, rows, cols)
1115 cols = len(hgvsCols) * 2
1116 log.Printf("building hgvs-based matrix: %d rows x %d cols", rows, cols)
1117 out = make([]int16, rows*cols)
1118 hgvsIDs := make([]string, 0, cols/2)
1119 for hgvsID := range hgvsCols {
1120 hgvsIDs = append(hgvsIDs, hgvsID)
1122 sort.Strings(hgvsIDs)
1123 var hgvsLabels bytes.Buffer
1124 for idx, hgvsID := range hgvsIDs {
1125 fmt.Fprintf(&hgvsLabels, "%d,%s\n", idx, hgvsID)
1126 for ph := 0; ph < 2; ph++ {
1127 hgvscol := hgvsCols[hgvsID][ph]
1128 for row, val := range hgvscol {
1129 out[row*cols+idx*2+ph] = val
1133 err = writeNumpyInt16(fmt.Sprintf("%s/hgvs.npy", *outputDir), out, rows, cols)
1138 fnm := fmt.Sprintf("%s/hgvs.annotations.csv", *outputDir)
1139 log.Printf("writing hgvs labels: %s", fnm)
1140 err = ioutil.WriteFile(fnm, hgvsLabels.Bytes(), 0777)
1146 if *onehotSingle || *onlyPCA {
1148 for _, part := range onehotIndirect {
1149 nzCount += len(part[0])
1151 onehot := make([]uint32, nzCount*2) // [r,r,r,...,c,c,c,...]
1152 var xrefs []onehotXref
1153 chunkOffset := uint32(0)
1155 for i, part := range onehotIndirect {
1156 for i := range part[1] {
1157 part[1][i] += chunkOffset
1159 copy(onehot[outcol:], part[0])
1160 copy(onehot[outcol+nzCount:], part[1])
1161 xrefs = append(xrefs, onehotXrefs[i]...)
1163 outcol += len(part[0])
1164 chunkOffset += onehotChunkSize[i]
1168 onehotXrefs[i] = nil
1169 debug.FreeOSMemory()
1172 fnm := fmt.Sprintf("%s/onehot.npy", *outputDir)
1173 err = writeNumpyUint32(fnm, onehot, 2, nzCount)
1177 fnm = fmt.Sprintf("%s/onehot-columns.npy", *outputDir)
1178 err = writeNumpyInt32(fnm, onehotXref2int32(xrefs), 5, len(xrefs))
1182 fnm = fmt.Sprintf("%s/stats.json", *outputDir)
1183 j, err := json.Marshal(map[string]interface{}{
1184 "pvalueCallCount": cmd.pvalueCallCount,
1189 err = os.WriteFile(fnm, j, 0777)
1196 for _, c := range onehot[nzCount:] {
1202 return fmt.Errorf("cannot do PCA: one-hot matrix is empty")
1204 log.Printf("have %d one-hot cols", cols)
1206 for *maxPCATiles > 0 && cols > *maxPCATiles*2 {
1207 cols = (cols + 1) / 2
1211 // we work with pairs of columns
1214 log.Printf("creating full matrix (%d rows) and training matrix (%d rows) with %d cols, stride %d", len(cmd.cgnames), cmd.trainingSetSize, cols, stride)
1215 mtxFull := mat.NewDense(len(cmd.cgnames), cols, nil)
1216 mtxTrain := mat.NewDense(cmd.trainingSetSize, cols, nil)
1217 for i, c := range onehot[nzCount:] {
1218 if int(c/2)%stride == 0 {
1219 outcol := int(c/2)/stride*2 + int(c)%2
1220 mtxFull.Set(int(onehot[i]), outcol, 1)
1221 if trainRow := cmd.trainingSet[int(onehot[i])]; trainRow >= 0 {
1222 mtxTrain.Set(trainRow, outcol, 1)
1226 log.Print("fitting")
1227 transformer := nlp.NewPCA(cmd.pcaComponents)
1228 transformer.Fit(mtxTrain.T())
1229 log.Printf("transforming")
1230 pca, err := transformer.Transform(mtxFull.T())
1235 outrows, outcols := pca.Dims()
1236 log.Printf("copying result to numpy output array: %d rows, %d cols", outrows, outcols)
1237 out := make([]float64, outrows*outcols)
1238 for i := 0; i < outrows; i++ {
1239 for j := 0; j < outcols; j++ {
1240 out[i*outcols+j] = pca.At(i, j)
1243 fnm := fmt.Sprintf("%s/pca.npy", *outputDir)
1244 log.Printf("writing numpy: %s", fnm)
1245 output, err := os.OpenFile(fnm, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, 0777)
1249 npw, err := gonpy.NewWriter(nopCloser{output})
1251 return fmt.Errorf("gonpy.NewWriter: %w", err)
1253 npw.Shape = []int{outrows, outcols}
1254 err = npw.WriteFloat64(out)
1256 return fmt.Errorf("WriteFloat64: %w", err)
1258 err = output.Close()
1264 log.Print("copying pca components to sampleInfo")
1265 for i := range cmd.samples {
1266 cmd.samples[i].pcaComponents = make([]float64, outcols)
1267 for c := 0; c < outcols; c++ {
1268 cmd.samples[i].pcaComponents[i] = pca.At(i, c)
1273 err = writeSampleInfo(cmd.samples, *outputDir)
1279 if !*mergeOutput && !*onehotChunked && !*onehotSingle && !*onlyPCA {
1280 tagoffsetFilename := *outputDir + "/chunk-tag-offset.csv"
1281 log.Infof("writing tag offsets to %s", tagoffsetFilename)
1283 f, err = os.Create(tagoffsetFilename)
1288 for idx, offset := range chunkStartTag {
1289 _, err = fmt.Fprintf(f, "%q,%d\n", fmt.Sprintf("matrix.%04d.npy", idx), offset)
1291 err = fmt.Errorf("write %s: %w", tagoffsetFilename, err)
1297 err = fmt.Errorf("close %s: %w", tagoffsetFilename, err)
1305 type sampleInfo struct {
1311 pcaComponents []float64
1314 // Read samples.csv file with case/control and training/validation
1316 func loadSampleInfo(samplesFilename string) ([]sampleInfo, error) {
1318 f, err := open(samplesFilename)
1322 buf, err := io.ReadAll(f)
1328 for _, csv := range bytes.Split(buf, []byte{'\n'}) {
1333 split := strings.Split(string(csv), ",")
1335 return nil, fmt.Errorf("%d fields < 4 in %s line %d: %q", len(split), samplesFilename, lineNum, csv)
1337 if split[0] == "Index" && split[1] == "SampleID" && split[2] == "CaseControl" && split[3] == "TrainingValidation" {
1340 idx, err := strconv.Atoi(split[0])
1343 return nil, fmt.Errorf("header does not look right: %q", csv)
1345 return nil, fmt.Errorf("%s line %d: index: %s", samplesFilename, lineNum, err)
1348 return nil, fmt.Errorf("%s line %d: index %d out of order", samplesFilename, lineNum, idx)
1350 var pcaComponents []float64
1352 for _, s := range split[4:] {
1353 f, err := strconv.ParseFloat(s, 64)
1355 return nil, fmt.Errorf("%s line %d: cannot parse float %q: %s", samplesFilename, lineNum, s, err)
1357 pcaComponents = append(pcaComponents, f)
1360 si = append(si, sampleInfo{
1362 isCase: split[2] == "1",
1363 isControl: split[2] == "0",
1364 isTraining: split[3] == "1",
1365 isValidation: split[3] == "0" && len(split[2]) > 0, // fix errant 0s in input
1366 pcaComponents: pcaComponents,
1372 func writeSampleInfo(samples []sampleInfo, outputDir string) error {
1373 fnm := outputDir + "/samples.csv"
1374 log.Infof("writing sample metadata to %s", fnm)
1375 f, err := os.Create(fnm)
1381 if len(samples) > 0 {
1382 for i := range samples[0].pcaComponents {
1383 pcaLabels += fmt.Sprintf(",PCA%d", i)
1386 _, err = fmt.Fprintf(f, "Index,SampleID,CaseControl,TrainingValidation%s\n", pcaLabels)
1390 for i, si := range samples {
1394 } else if si.isControl {
1399 } else if si.isValidation {
1403 for _, pcaval := range si.pcaComponents {
1404 pcavals += fmt.Sprintf(",%f", pcaval)
1406 _, err = fmt.Fprintf(f, "%d,%s,%s,%s%s\n", i, si.id, cc, tv, pcavals)
1408 return fmt.Errorf("write %s: %w", fnm, err)
1413 return fmt.Errorf("close %s: %w", fnm, err)
1419 func (cmd *sliceNumpy) filterHGVScolpair(colpair [2][]int8) bool {
1420 if cmd.chi2PValue >= 1 {
1423 col0 := make([]bool, 0, len(cmd.chi2Cases))
1424 col1 := make([]bool, 0, len(cmd.chi2Cases))
1425 cases := make([]bool, 0, len(cmd.chi2Cases))
1426 for i, c := range cmd.chi2Cases {
1427 if colpair[0][i] < 0 {
1430 col0 = append(col0, colpair[0][i] != 0)
1431 col1 = append(col1, colpair[1][i] != 0)
1432 cases = append(cases, c)
1434 return len(cases) >= cmd.minCoverage &&
1435 (pvalue(col0, cases) <= cmd.chi2PValue || pvalue(col1, cases) <= cmd.chi2PValue)
1438 func writeNumpyUint32(fnm string, out []uint32, rows, cols int) error {
1439 output, err := os.Create(fnm)
1443 defer output.Close()
1444 bufw := bufio.NewWriterSize(output, 1<<26)
1445 npw, err := gonpy.NewWriter(nopCloser{bufw})
1449 log.WithFields(log.Fields{
1453 "bytes": rows * cols * 4,
1454 }).Infof("writing numpy: %s", fnm)
1455 npw.Shape = []int{rows, cols}
1456 npw.WriteUint32(out)
1461 return output.Close()
1464 func writeNumpyInt32(fnm string, out []int32, rows, cols int) error {
1465 output, err := os.Create(fnm)
1469 defer output.Close()
1470 bufw := bufio.NewWriterSize(output, 1<<26)
1471 npw, err := gonpy.NewWriter(nopCloser{bufw})
1475 log.WithFields(log.Fields{
1479 "bytes": rows * cols * 4,
1480 }).Infof("writing numpy: %s", fnm)
1481 npw.Shape = []int{rows, cols}
1487 return output.Close()
1490 func writeNumpyInt16(fnm string, out []int16, rows, cols int) error {
1491 output, err := os.Create(fnm)
1495 defer output.Close()
1496 bufw := bufio.NewWriterSize(output, 1<<26)
1497 npw, err := gonpy.NewWriter(nopCloser{bufw})
1501 log.WithFields(log.Fields{
1505 "bytes": rows * cols * 2,
1506 }).Infof("writing numpy: %s", fnm)
1507 npw.Shape = []int{rows, cols}
1513 return output.Close()
1516 func writeNumpyInt8(fnm string, out []int8, rows, cols int) error {
1517 output, err := os.Create(fnm)
1521 defer output.Close()
1522 bufw := bufio.NewWriterSize(output, 1<<26)
1523 npw, err := gonpy.NewWriter(nopCloser{bufw})
1527 log.WithFields(log.Fields{
1531 "bytes": rows * cols,
1532 }).Infof("writing numpy: %s", fnm)
1533 npw.Shape = []int{rows, cols}
1539 return output.Close()
1542 func allele2homhet(colpair [2][]int8) {
1543 a, b := colpair[0], colpair[1]
1544 for i, av := range a {
1546 if av < 0 || bv < 0 {
1549 } else if av > 0 && bv > 0 {
1552 } else if av > 0 || bv > 0 {
1556 // ref (or a different variant in same position)
1557 // (this is a no-op) a[i], b[i] = 0, 0
1562 type onehotXref struct {
1564 variant tileVariantID
1569 const onehotXrefSize = unsafe.Sizeof(onehotXref{})
1571 // Build onehot matrix (m[tileVariantIndex][genome] == 0 or 1) for all
1572 // variants of a single tile/tag#.
1574 // Return nil if no tile variant passes Χ² filter.
1575 func (cmd *sliceNumpy) tv2homhet(cgs map[string]CompactGenome, maxv tileVariantID, remap []tileVariantID, tag, chunkstarttag tagID, seq map[tagID][]TileVariant) ([][]int8, []onehotXref) {
1576 if tag == cmd.debugTag {
1577 tv := make([]tileVariantID, len(cmd.cgnames)*2)
1578 for i, name := range cmd.cgnames {
1579 copy(tv[i*2:(i+1)*2], cgs[name].Variants[(tag-chunkstarttag)*2:])
1581 log.WithFields(logrus.Fields{
1582 "cgs[i].Variants[tag*2+j]": tv,
1586 "chunkstarttag": chunkstarttag,
1587 }).Info("tv2homhet()")
1589 if maxv < 1 || (maxv < 2 && !cmd.includeVariant1) {
1590 // everyone has the most common variant (of the variants we don't drop)
1593 tagoffset := tag - chunkstarttag
1595 for _, cg := range cgs {
1597 for _, v := range cg.Variants[tagoffset*2 : tagoffset*2+2] {
1598 if v > 0 && int(v) < len(seq[tag]) && len(seq[tag][v].Sequence) > 0 {
1606 if coverage < cmd.minCoverage {
1609 // "observed" array for p-value calculation (training set
1611 obs := make([][]bool, (maxv+1)*2) // 2 slices (hom + het) for each variant#
1612 // one-hot output (all samples)
1613 outcols := make([][]int8, (maxv+1)*2)
1614 for i := range obs {
1615 obs[i] = make([]bool, cmd.trainingSetSize)
1616 outcols[i] = make([]int8, len(cmd.cgnames))
1618 for cgid, name := range cmd.cgnames {
1619 tsid := cmd.trainingSet[cgid]
1620 cgvars := cgs[name].Variants[tagoffset*2:]
1621 tv0, tv1 := remap[cgvars[0]], remap[cgvars[1]]
1622 for v := tileVariantID(1); v <= maxv; v++ {
1623 if tv0 == v && tv1 == v {
1625 obs[v*2][tsid] = true
1627 outcols[v*2][cgid] = 1
1628 } else if tv0 == v || tv1 == v {
1630 obs[v*2+1][tsid] = true
1632 outcols[v*2+1][cgid] = 1
1637 var xref []onehotXref
1638 for col := 2; col < len(obs); col++ {
1639 // col 0,1 correspond to tile variant 0, i.e.,
1640 // no-call; col 2,3 correspond to the most common
1641 // variant; so we (normally) start at col 4.
1642 if col < 4 && !cmd.includeVariant1 {
1645 if col&1 == 0 && cmd.pvalueMinFrequency < 1 && homhet2maf(obs[col:col+2]) < cmd.pvalueMinFrequency {
1646 // Skip both columns (hom and het) if allele
1647 // frequency is below threshold
1651 atomic.AddInt64(&cmd.pvalueCallCount, 1)
1652 p := cmd.pvalue(obs[col])
1653 if cmd.chi2PValue < 1 && !(p < cmd.chi2PValue) {
1656 onehot = append(onehot, outcols[col])
1657 xref = append(xref, onehotXref{
1659 variant: tileVariantID(col >> 1),
1667 func homhet2maf(onehot [][]bool) float64 {
1668 if len(onehot[0]) == 0 {
1672 for i := range onehot[0] {
1676 } else if onehot[1][i] {
1681 return float64(n) / float64(len(onehot[0])*2)
1684 // convert a []onehotXref with length N to a numpy-style []int32
1685 // matrix with N columns, one row per field of onehotXref struct.
1687 // Hom/het row contains hom=0, het=1.
1689 // P-value row contains 1000000x actual p-value.
1690 func onehotXref2int32(xrefs []onehotXref) []int32 {
1692 xdata := make([]int32, 5*xcols)
1693 for i, xref := range xrefs {
1694 xdata[i] = int32(xref.tag)
1695 xdata[xcols+i] = int32(xref.variant)
1697 xdata[xcols*2+i] = 1
1699 xdata[xcols*3+i] = int32(xref.pvalue * 1000000)
1700 xdata[xcols*4+i] = int32(-math.Log10(xref.pvalue) * 1000000)
1705 // transpose onehot data from in[col][row] to numpy-style
1706 // out[row*cols+col].
1707 func onehotcols2int8(in [][]int8) []int8 {
1713 out := make([]int8, rows*cols)
1714 for row := 0; row < rows; row++ {
1715 outrow := out[row*cols:]
1716 for col, incol := range in {
1717 outrow[col] = incol[row]
1723 // Return [2][]uint32{rowIndices, colIndices} indicating which
1724 // elements of matrixT[c][r] have non-zero values.
1725 func onehotChunk2Indirect(matrixT [][]int8) [2][]uint32 {
1727 for c, col := range matrixT {
1728 for r, val := range col {
1730 nz[0] = append(nz[0], uint32(r))
1731 nz[1] = append(nz[1], uint32(c))