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
56 trainingSet []int // samples index => training set index, or -1 if not in training set
58 pvalue func(onehot []bool) float64
62 func (cmd *sliceNumpy) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
63 err := cmd.run(prog, args, stdin, stdout, stderr)
65 fmt.Fprintf(stderr, "%s\n", err)
71 func (cmd *sliceNumpy) run(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) error {
72 flags := flag.NewFlagSet("", flag.ContinueOnError)
73 flags.SetOutput(stderr)
74 pprof := flags.String("pprof", "", "serve Go profile data at http://`[addr]:port`")
75 runlocal := flags.Bool("local", false, "run on local host (default: run in an arvados container)")
76 arvadosRAM := flags.Int("arvados-ram", 750000000000, "amount of memory to request for arvados container (`bytes`)")
77 arvadosVCPUs := flags.Int("arvados-vcpus", 96, "number of VCPUs to request for arvados container")
78 projectUUID := flags.String("project", "", "project `UUID` for output data")
79 priority := flags.Int("priority", 500, "container request priority")
80 preemptible := flags.Bool("preemptible", true, "request preemptible instance")
81 inputDir := flags.String("input-dir", "./in", "input `directory`")
82 outputDir := flags.String("output-dir", "./out", "output `directory`")
83 ref := flags.String("ref", "", "reference name (if blank, choose last one that appears in input)")
84 regionsFilename := flags.String("regions", "", "only output columns/annotations that intersect regions in specified bed `file`")
85 expandRegions := flags.Int("expand-regions", 0, "expand specified regions by `N` base pairs on each side`")
86 mergeOutput := flags.Bool("merge-output", false, "merge output into one matrix.npy and one matrix.annotations.csv")
87 hgvsSingle := flags.Bool("single-hgvs-matrix", false, "also generate hgvs-based matrix")
88 hgvsChunked := flags.Bool("chunked-hgvs-matrix", false, "also generate hgvs-based matrix per chromosome")
89 onehotSingle := flags.Bool("single-onehot", false, "generate one-hot tile-based matrix")
90 onehotChunked := flags.Bool("chunked-onehot", false, "generate one-hot tile-based matrix per input chunk")
91 samplesFilename := flags.String("samples", "", "`samples.csv` file with training/validation and case/control groups (see 'lightning choose-samples')")
92 caseControlOnly := flags.Bool("case-control-only", false, "drop samples that are not in case/control groups")
93 onlyPCA := flags.Bool("pca", false, "run principal component analysis, write components to pca.npy and samples.csv")
94 flags.IntVar(&cmd.pcaComponents, "pca-components", 4, "number of PCA components to compute / use in logistic regression")
95 maxPCATiles := flags.Int("max-pca-tiles", 0, "maximum tiles to use as PCA input (filter, then drop every 2nd colum pair until below max)")
96 debugTag := flags.Int("debug-tag", -1, "log debugging details about specified tag")
97 flags.IntVar(&cmd.threads, "threads", 16, "number of memory-hungry assembly threads, and number of VCPUs to request for arvados container")
98 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")
99 flags.Float64Var(&cmd.pvalueMinFrequency, "pvalue-min-frequency", 0.01, "skip p-value calculation on tile variants below this frequency in the training set")
100 flags.Float64Var(&cmd.maxFrequency, "max-frequency", 1, "do not output variants above this frequency in the training set")
101 flags.BoolVar(&cmd.includeVariant1, "include-variant-1", false, "include most common variant when building one-hot matrix")
102 cmd.filter.Flags(flags)
103 err := flags.Parse(args)
104 if err == flag.ErrHelp {
106 } else if err != nil {
108 } else if flags.NArg() > 0 {
109 return fmt.Errorf("errant command line arguments after parsed flags: %v", flags.Args())
114 log.Println(http.ListenAndServe(*pprof, nil))
118 if cmd.chi2PValue != 1 && *samplesFilename == "" {
119 return fmt.Errorf("cannot use provided -chi2-p-value=%f because -samples= value is empty", cmd.chi2PValue)
122 cmd.debugTag = tagID(*debugTag)
125 runner := arvadosContainerRunner{
126 Name: "lightning slice-numpy",
127 Client: arvados.NewClientFromEnv(),
128 ProjectUUID: *projectUUID,
129 RAM: int64(*arvadosRAM),
130 VCPUs: *arvadosVCPUs,
134 Preemptible: *preemptible,
136 err = runner.TranslatePaths(inputDir, regionsFilename, samplesFilename)
140 runner.Args = []string{"slice-numpy", "-local=true",
142 "-input-dir=" + *inputDir,
143 "-output-dir=/mnt/output",
144 "-threads=" + fmt.Sprintf("%d", cmd.threads),
145 "-regions=" + *regionsFilename,
146 "-expand-regions=" + fmt.Sprintf("%d", *expandRegions),
147 "-merge-output=" + fmt.Sprintf("%v", *mergeOutput),
148 "-single-hgvs-matrix=" + fmt.Sprintf("%v", *hgvsSingle),
149 "-chunked-hgvs-matrix=" + fmt.Sprintf("%v", *hgvsChunked),
150 "-single-onehot=" + fmt.Sprintf("%v", *onehotSingle),
151 "-chunked-onehot=" + fmt.Sprintf("%v", *onehotChunked),
152 "-samples=" + *samplesFilename,
153 "-case-control-only=" + fmt.Sprintf("%v", *caseControlOnly),
154 "-pca=" + fmt.Sprintf("%v", *onlyPCA),
155 "-pca-components=" + fmt.Sprintf("%d", cmd.pcaComponents),
156 "-max-pca-tiles=" + fmt.Sprintf("%d", *maxPCATiles),
157 "-chi2-p-value=" + fmt.Sprintf("%f", cmd.chi2PValue),
158 "-pvalue-min-frequency=" + fmt.Sprintf("%f", cmd.pvalueMinFrequency),
159 "-max-frequency=" + fmt.Sprintf("%f", cmd.maxFrequency),
160 "-include-variant-1=" + fmt.Sprintf("%v", cmd.includeVariant1),
161 "-debug-tag=" + fmt.Sprintf("%d", cmd.debugTag),
163 runner.Args = append(runner.Args, cmd.filter.Args()...)
165 output, err = runner.Run()
169 fmt.Fprintln(stdout, output)
173 infiles, err := allFiles(*inputDir, matchGobFile)
177 if len(infiles) == 0 {
178 err = fmt.Errorf("no input files found in %s", *inputDir)
181 sort.Strings(infiles)
183 var refseq map[string][]tileLibRef
184 var reftiledata = make(map[tileLibRef][]byte, 11000000)
185 in0, err := open(infiles[0])
190 matchGenome, err := regexp.Compile(cmd.filter.MatchGenome)
192 err = fmt.Errorf("-match-genome: invalid regexp: %q", cmd.filter.MatchGenome)
196 if *samplesFilename != "" {
197 cmd.samples, err = loadSampleInfo(*samplesFilename)
201 } else if *caseControlOnly {
202 return fmt.Errorf("-case-control-only does not make sense without -samples")
207 err = DecodeLibrary(in0, strings.HasSuffix(infiles[0], ".gz"), func(ent *LibraryEntry) error {
208 if len(ent.TagSet) > 0 {
211 for _, cseq := range ent.CompactSequences {
212 if cseq.Name == *ref || *ref == "" {
213 refseq = cseq.TileSequences
216 for _, cg := range ent.CompactGenomes {
217 if matchGenome.MatchString(cg.Name) {
218 cmd.cgnames = append(cmd.cgnames, cg.Name)
221 for _, tv := range ent.TileVariants {
223 reftiledata[tileLibRef{tv.Tag, tv.Variant}] = tv.Sequence
233 err = fmt.Errorf("%s: reference sequence not found", infiles[0])
236 if len(tagset) == 0 {
237 err = fmt.Errorf("tagset not found")
241 taglib := &tagLibrary{}
242 err = taglib.setTags(tagset)
246 taglen := taglib.TagLen()
247 sort.Strings(cmd.cgnames)
249 if len(cmd.cgnames) == 0 {
250 return fmt.Errorf("fatal: 0 matching samples in library, nothing to do")
252 cmd.trainingSet = make([]int, len(cmd.cgnames))
253 if *samplesFilename == "" {
254 cmd.trainingSetSize = len(cmd.cgnames)
255 for i, name := range cmd.cgnames {
256 cmd.samples = append(cmd.samples, sampleInfo{
257 id: trimFilenameForLabel(name),
260 cmd.trainingSet[i] = i
262 } else if len(cmd.cgnames) != len(cmd.samples) {
263 return fmt.Errorf("mismatched sample list: %d samples in library, %d in %s", len(cmd.cgnames), len(cmd.samples), *samplesFilename)
265 for i, name := range cmd.cgnames {
266 if s := trimFilenameForLabel(name); s != cmd.samples[i].id {
267 return fmt.Errorf("mismatched sample list: sample %d is %q in library, %q in %s", i, s, cmd.samples[i].id, *samplesFilename)
270 if *caseControlOnly {
271 for i := 0; i < len(cmd.samples); i++ {
272 if !cmd.samples[i].isTraining && !cmd.samples[i].isValidation {
273 if i+1 < len(cmd.samples) {
274 copy(cmd.samples[i:], cmd.samples[i+1:])
275 copy(cmd.cgnames[i:], cmd.cgnames[i+1:])
277 cmd.samples = cmd.samples[:len(cmd.samples)-1]
278 cmd.cgnames = cmd.cgnames[:len(cmd.cgnames)-1]
284 cmd.trainingSetSize = 0
285 for i := range cmd.cgnames {
286 if cmd.samples[i].isTraining {
287 cmd.trainingSet[i] = cmd.trainingSetSize
288 cmd.trainingSetSize++
289 cmd.chi2Cases = append(cmd.chi2Cases, cmd.samples[i].isCase)
291 cmd.trainingSet[i] = -1
294 if cmd.pvalue == nil {
295 cmd.pvalue = func(onehot []bool) float64 {
296 return pvalue(onehot, cmd.chi2Cases)
300 if cmd.filter.MinCoverage == 1 {
301 // In the generic formula below, floating point
302 // arithmetic can effectively push the coverage
303 // threshold above 1.0, which is impossible/useless.
304 // 1.0 needs to mean exactly 100% coverage.
305 cmd.minCoverage = len(cmd.cgnames)
307 cmd.minCoverage = int(math.Ceil(cmd.filter.MinCoverage * float64(len(cmd.cgnames))))
310 if len(cmd.samples[0].pcaComponents) > 0 {
311 cmd.pvalue = glmPvalueFunc(cmd.samples, cmd.pcaComponents)
312 // Unfortunately, statsmodel/glm lib logs stuff to
313 // os.Stdout when it panics on an unsolvable
314 // problem. We recover() from the panic in glm.go, but
315 // we also need to commandeer os.Stdout to avoid
316 // producing large quantities of logs.
317 stdoutWas := os.Stdout
318 defer func() { os.Stdout = stdoutWas }()
319 os.Stdout, err = os.Open(os.DevNull)
325 // cgnamemap[name]==true for samples that we are including in
327 cgnamemap := map[string]bool{}
328 for _, name := range cmd.cgnames {
329 cgnamemap[name] = true
332 err = writeSampleInfo(cmd.samples, *outputDir)
337 log.Info("indexing reference tiles")
338 type reftileinfo struct {
339 variant tileVariantID
340 seqname string // chr1
341 pos int // distance from start of chromosome to starttag
342 tiledata []byte // acgtggcaa...
343 excluded bool // true if excluded by regions file
344 nexttag tagID // tagID of following tile (-1 for last tag of chromosome)
346 isdup := map[tagID]bool{}
347 reftile := map[tagID]*reftileinfo{}
348 for seqname, cseq := range refseq {
350 lastreftag := tagID(-1)
351 for _, libref := range cseq {
352 if cmd.filter.MaxTag >= 0 && libref.Tag > tagID(cmd.filter.MaxTag) {
355 tiledata := reftiledata[libref]
356 if len(tiledata) == 0 {
357 err = fmt.Errorf("missing tiledata for tag %d variant %d in %s in ref", libref.Tag, libref.Variant, seqname)
360 foundthistag := false
361 taglib.FindAll(tiledata[:len(tiledata)-1], func(tagid tagID, offset, _ int) {
362 if !foundthistag && tagid == libref.Tag {
366 if dupref, ok := reftile[tagid]; ok {
367 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)
368 delete(reftile, tagid)
370 log.Printf("found tag %d at offset %d inside tile variant %+v on %s @ %d", tagid, offset, libref, seqname, pos+offset+1)
374 if isdup[libref.Tag] {
375 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique", libref, seqname, pos)
376 } else if reftile[libref.Tag] != nil {
377 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)
378 delete(reftile, libref.Tag)
379 log.Printf("dropping reference tile %+v from %s @ %d, tag not unique", libref, seqname, pos)
380 isdup[libref.Tag] = true
382 reftile[libref.Tag] = &reftileinfo{
384 variant: libref.Variant,
390 reftile[lastreftag].nexttag = libref.Tag
392 lastreftag = libref.Tag
394 pos += len(tiledata) - taglen
396 log.Printf("... %s done, len %d", seqname, pos+taglen)
400 if *regionsFilename != "" {
401 log.Printf("loading regions from %s", *regionsFilename)
402 mask, err = makeMask(*regionsFilename, *expandRegions)
406 log.Printf("before applying mask, len(reftile) == %d", len(reftile))
407 log.Printf("deleting reftile entries for regions outside %d intervals", mask.Len())
408 for _, rt := range reftile {
409 if !mask.Check(strings.TrimPrefix(rt.seqname, "chr"), rt.pos, rt.pos+len(rt.tiledata)) {
413 log.Printf("after applying mask, len(reftile) == %d", len(reftile))
416 type hgvsColSet map[hgvs.Variant][2][]int8
417 encodeHGVS := throttle{Max: len(refseq)}
418 encodeHGVSTodo := map[string]chan hgvsColSet{}
419 tmpHGVSCols := map[string]*os.File{}
421 for seqname := range refseq {
423 f, err = os.Create(*outputDir + "/tmp." + seqname + ".gob")
427 defer os.Remove(f.Name())
428 bufw := bufio.NewWriterSize(f, 1<<24)
429 enc := gob.NewEncoder(bufw)
430 tmpHGVSCols[seqname] = f
431 todo := make(chan hgvsColSet, 128)
432 encodeHGVSTodo[seqname] = todo
433 encodeHGVS.Go(func() error {
434 for colset := range todo {
435 err := enc.Encode(colset)
437 encodeHGVS.Report(err)
448 var toMerge [][]int16
449 if *mergeOutput || *hgvsSingle {
450 toMerge = make([][]int16, len(infiles))
452 var onehotIndirect [][2][]uint32 // [chunkIndex][axis][index]
453 var onehotChunkSize []uint32
454 var onehotXrefs [][]onehotXref
455 if *onehotSingle || *onlyPCA {
456 onehotIndirect = make([][2][]uint32, len(infiles))
457 onehotChunkSize = make([]uint32, len(infiles))
458 onehotXrefs = make([][]onehotXref, len(infiles))
460 chunkStartTag := make([]tagID, len(infiles))
462 throttleMem := throttle{Max: cmd.threads} // TODO: estimate using mem and data size
463 throttleNumpyMem := throttle{Max: cmd.threads/2 + 1}
464 log.Info("generating annotations and numpy matrix for each slice")
465 var errSkip = errors.New("skip infile")
467 for infileIdx, infile := range infiles {
468 infileIdx, infile := infileIdx, infile
469 throttleMem.Go(func() error {
470 seq := make(map[tagID][]TileVariant, 50000)
471 cgs := make(map[string]CompactGenome, len(cmd.cgnames))
472 f, err := open(infile)
477 log.Infof("%04d: reading %s", infileIdx, infile)
478 err = DecodeLibrary(f, strings.HasSuffix(infile, ".gz"), func(ent *LibraryEntry) error {
479 for _, tv := range ent.TileVariants {
484 // corresponding ref tile, if
485 // mask is in play (we can't
486 // determine coordinates for
488 if mask != nil && reftile[tv.Tag] == nil {
492 // corresponding ref tile is
493 // outside target regions --
494 // unless it's a potential
496 if mask != nil && reftile[tv.Tag].excluded &&
497 (int(tv.Tag+1) >= len(tagset) ||
498 (bytes.HasSuffix(tv.Sequence, tagset[tv.Tag+1]) && reftile[tv.Tag+1] != nil && !reftile[tv.Tag+1].excluded)) {
501 if tv.Tag == cmd.debugTag {
502 log.Printf("infile %d %s tag %d variant %d hash %x", infileIdx, infile, tv.Tag, tv.Variant, tv.Blake2b[:3])
504 variants := seq[tv.Tag]
505 if len(variants) == 0 {
506 variants = make([]TileVariant, 100)
508 for len(variants) <= int(tv.Variant) {
509 variants = append(variants, TileVariant{})
511 variants[int(tv.Variant)] = tv
512 seq[tv.Tag] = variants
514 for _, cg := range ent.CompactGenomes {
515 if cmd.filter.MaxTag >= 0 && cg.StartTag > tagID(cmd.filter.MaxTag) {
518 if !cgnamemap[cg.Name] {
521 // pad to full slice size
522 // to avoid out-of-bounds
524 if sliceSize := 2 * int(cg.EndTag-cg.StartTag); len(cg.Variants) < sliceSize {
525 cg.Variants = append(cg.Variants, make([]tileVariantID, sliceSize-len(cg.Variants))...)
533 } else if err != nil {
534 return fmt.Errorf("%04d: DecodeLibrary(%s): err", infileIdx, infile)
536 tagstart := cgs[cmd.cgnames[0]].StartTag
537 tagend := cgs[cmd.cgnames[0]].EndTag
538 chunkStartTag[infileIdx] = tagstart
542 log.Infof("%04d: renumber/dedup variants for tags %d-%d", infileIdx, tagstart, tagend)
543 variantRemap := make([][]tileVariantID, tagend-tagstart)
544 throttleCPU := throttle{Max: runtime.GOMAXPROCS(0)}
545 for tag, variants := range seq {
546 tag, variants := tag, variants
547 throttleCPU.Go(func() error {
549 count := make(map[[blake2b.Size256]byte]int, len(variants))
553 count[blake2b.Sum256(rt.tiledata)] = 0
556 for cgname, cg := range cgs {
557 idx := int(tag-tagstart) * 2
558 for allele := 0; allele < 2; allele++ {
559 v := cg.Variants[idx+allele]
560 if v > 0 && len(variants[v].Sequence) > 0 {
561 count[variants[v].Blake2b]++
564 if v > 0 && tag == cmd.debugTag {
565 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])
569 if alleleCoverage < cmd.minCoverage*2 {
570 idx := int(tag-tagstart) * 2
571 for _, cg := range cgs {
573 cg.Variants[idx+1] = 0
575 if tag == cmd.debugTag {
576 log.Printf("tag %d alleleCoverage %d < min %d, sample data wiped", tag, alleleCoverage, cmd.minCoverage*2)
581 // hash[i] will be the hash of
582 // the variant(s) that should
583 // be at rank i (0-based).
584 hash := make([][blake2b.Size256]byte, 0, len(count))
585 for b := range count {
586 hash = append(hash, b)
588 sort.Slice(hash, func(i, j int) bool {
589 bi, bj := &hash[i], &hash[j]
590 if ci, cj := count[*bi], count[*bj]; ci != cj {
593 return bytes.Compare((*bi)[:], (*bj)[:]) < 0
596 // rank[b] will be the 1-based
597 // new variant number for
598 // variants whose hash is b.
599 rank := make(map[[blake2b.Size256]byte]tileVariantID, len(hash))
600 for i, h := range hash {
601 rank[h] = tileVariantID(i + 1)
603 if tag == cmd.debugTag {
604 for h, r := range rank {
605 log.Printf("tag %d rank(%x) = %v", tag, h[:3], r)
608 // remap[v] will be the new
609 // variant number for original
611 remap := make([]tileVariantID, len(variants))
612 for i, tv := range variants {
613 remap[i] = rank[tv.Blake2b]
615 if tag == cmd.debugTag {
616 for in, out := range remap {
618 log.Printf("tag %d remap %d => %d", tag, in, out)
622 variantRemap[tag-tagstart] = remap
624 refrank := rank[blake2b.Sum256(rt.tiledata)]
625 if tag == cmd.debugTag {
626 log.Printf("tag %d reftile variant %d => %d", tag, rt.variant, refrank)
635 var onehotChunk [][]int8
636 var onehotXref []onehotXref
638 var annotationsFilename string
640 annotationsFilename = "/dev/null"
642 annotationsFilename = fmt.Sprintf("%s/matrix.%04d.annotations.csv", *outputDir, infileIdx)
643 log.Infof("%04d: writing %s", infileIdx, annotationsFilename)
645 annof, err := os.Create(annotationsFilename)
649 annow := bufio.NewWriterSize(annof, 1<<20)
651 for tag := tagstart; tag < tagend; tag++ {
653 if rt == nil && mask != nil {
654 // With no ref tile, we don't
655 // have coordinates to say
656 // this is in the desired
657 // regions -- so it's not.
658 // TODO: handle ref spanning
662 if rt != nil && rt.excluded {
663 // TODO: don't skip yet --
664 // first check for spanning
665 // tile variants that
666 // intersect non-excluded ref
670 if cmd.filter.MaxTag >= 0 && tag > tagID(cmd.filter.MaxTag) {
673 remap := variantRemap[tag-tagstart]
675 // was not assigned above,
676 // because minCoverage
680 maxv := tileVariantID(0)
681 for _, v := range remap {
686 if *onehotChunked || *onehotSingle || *onlyPCA {
687 onehot, xrefs := cmd.tv2homhet(cgs, maxv, remap, tag, tagstart, seq)
688 if tag == cmd.debugTag {
689 log.WithFields(logrus.Fields{
692 }).Info("tv2homhet()")
694 onehotChunk = append(onehotChunk, onehot...)
695 onehotXref = append(onehotXref, xrefs...)
702 // Reference does not use any
703 // variant of this tile
705 // TODO: diff against the
706 // relevant portion of the
707 // ref's spanning tile
711 fmt.Fprintf(annow, "%d,%d,%d,=,%s,%d,,,\n", tag, outcol, rt.variant, rt.seqname, rt.pos)
713 reftilestr := strings.ToUpper(string(rt.tiledata))
715 done := make([]bool, maxv+1)
716 variantDiffs := make([][]hgvs.Variant, maxv+1)
717 for v, tv := range variants {
719 if v == 0 || v == rt.variant || done[v] {
724 if len(tv.Sequence) < taglen {
727 // if reftilestr doesn't end
728 // in the same tag as tv,
729 // extend reftilestr with
730 // following ref tiles until
731 // it does (up to an arbitrary
732 // sanity-check limit)
733 reftilestr := reftilestr
734 endtagstr := strings.ToUpper(string(tv.Sequence[len(tv.Sequence)-taglen:]))
735 for i, rt := 0, rt; i < annotationMaxTileSpan && !strings.HasSuffix(reftilestr, endtagstr) && rt.nexttag >= 0; i++ {
736 rt = reftile[rt.nexttag]
740 reftilestr += strings.ToUpper(string(rt.tiledata[taglen:]))
742 if mask != nil && !mask.Check(strings.TrimPrefix(rt.seqname, "chr"), rt.pos, rt.pos+len(reftilestr)) {
745 if !strings.HasSuffix(reftilestr, endtagstr) {
746 fmt.Fprintf(annow, "%d,%d,%d,,%s,%d,,,\n", tag, outcol, v, rt.seqname, rt.pos)
749 if lendiff := len(reftilestr) - len(tv.Sequence); lendiff < -1000 || lendiff > 1000 {
750 fmt.Fprintf(annow, "%d,%d,%d,,%s,%d,,,\n", tag, outcol, v, rt.seqname, rt.pos)
753 diffs, _ := hgvs.Diff(reftilestr, strings.ToUpper(string(tv.Sequence)), 0)
754 for i := range diffs {
755 diffs[i].Position += rt.pos
757 for _, diff := range diffs {
758 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)
761 variantDiffs[v] = diffs
765 // We can now determine, for each HGVS
766 // variant (diff) in this reftile
767 // region, whether a given genome
768 // phase/allele (1) has the variant, (0) has
769 // =ref or a different variant in that
770 // position, or (-1) is lacking
771 // coverage / couldn't be diffed.
772 hgvsCol := hgvsColSet{}
773 for _, diffs := range variantDiffs {
774 for _, diff := range diffs {
775 if _, ok := hgvsCol[diff]; ok {
778 hgvsCol[diff] = [2][]int8{
779 make([]int8, len(cmd.cgnames)),
780 make([]int8, len(cmd.cgnames)),
784 for row, name := range cmd.cgnames {
785 variants := cgs[name].Variants[(tag-tagstart)*2:]
786 for ph := 0; ph < 2; ph++ {
788 if int(v) >= len(remap) {
794 // hgvsCol[*][ph][row] is already 0
795 } else if len(variantDiffs[v]) == 0 {
796 // lacking coverage / couldn't be diffed
797 for _, col := range hgvsCol {
801 for _, diff := range variantDiffs[v] {
802 hgvsCol[diff][ph][row] = 1
807 for diff, colpair := range hgvsCol {
808 allele2homhet(colpair)
809 if !cmd.filterHGVScolpair(colpair) {
810 delete(hgvsCol, diff)
813 if len(hgvsCol) > 0 {
814 encodeHGVSTodo[rt.seqname] <- hgvsCol
829 // transpose onehotChunk[col][row] to numpy[row*ncols+col]
830 rows := len(cmd.cgnames)
831 cols := len(onehotChunk)
832 log.Infof("%04d: preparing onehot numpy (rows=%d, cols=%d, mem=%d)", infileIdx, rows, cols, rows*cols)
833 throttleNumpyMem.Acquire()
834 out := onehotcols2int8(onehotChunk)
835 fnm := fmt.Sprintf("%s/onehot.%04d.npy", *outputDir, infileIdx)
836 err = writeNumpyInt8(fnm, out, rows, cols)
840 fnm = fmt.Sprintf("%s/onehot-columns.%04d.npy", *outputDir, infileIdx)
841 err = writeNumpyInt32(fnm, onehotXref2int32(onehotXref), 4, len(onehotXref))
846 throttleNumpyMem.Release()
848 if *onehotSingle || *onlyPCA {
849 onehotIndirect[infileIdx] = onehotChunk2Indirect(onehotChunk)
850 onehotChunkSize[infileIdx] = uint32(len(onehotChunk))
851 onehotXrefs[infileIdx] = onehotXref
852 n := len(onehotIndirect[infileIdx][0])
853 log.Infof("%04d: keeping onehot coordinates in memory (n=%d, mem=%d)", infileIdx, n, n*8*2)
855 if !(*onehotSingle || *onehotChunked || *onlyPCA) || *mergeOutput || *hgvsSingle {
856 log.Infof("%04d: preparing numpy (rows=%d, cols=%d)", infileIdx, len(cmd.cgnames), 2*outcol)
857 throttleNumpyMem.Acquire()
858 rows := len(cmd.cgnames)
860 out := make([]int16, rows*cols)
861 for row, name := range cmd.cgnames {
863 for col, v := range cgs[name].Variants {
864 tag := tagstart + tagID(col/2)
865 if cmd.filter.MaxTag >= 0 && tag > tagID(cmd.filter.MaxTag) {
868 if rt := reftile[tag]; rt == nil || rt.excluded {
872 out[outidx] = 0 // tag not found / spanning tile
873 } else if variants, ok := seq[tag]; ok && int(v) < len(variants) && len(variants[v].Sequence) > 0 {
874 out[outidx] = int16(variantRemap[tag-tagstart][v])
876 out[outidx] = -1 // low quality tile variant
878 if tag == cmd.debugTag {
879 log.Printf("tag %d row %d col %d outidx %d v %d out %d", tag, row, col, outidx, v, out[outidx])
887 throttleNumpyMem.Release()
888 if *mergeOutput || *hgvsSingle {
889 log.Infof("%04d: matrix fragment %d rows x %d cols", infileIdx, rows, cols)
890 toMerge[infileIdx] = out
892 if !*mergeOutput && !*onehotChunked && !*onehotSingle {
893 fnm := fmt.Sprintf("%s/matrix.%04d.npy", *outputDir, infileIdx)
894 err = writeNumpyInt16(fnm, out, rows, cols)
901 log.Infof("%s: done (%d/%d)", infile, int(atomic.AddInt64(&done, 1)), len(infiles))
905 if err = throttleMem.Wait(); err != nil {
910 log.Info("flushing hgvsCols temp files")
911 for seqname := range refseq {
912 close(encodeHGVSTodo[seqname])
914 err = encodeHGVS.Wait()
918 for seqname := range refseq {
919 log.Infof("%s: reading hgvsCols from temp file", seqname)
920 f := tmpHGVSCols[seqname]
921 _, err = f.Seek(0, io.SeekStart)
925 var hgvsCols hgvsColSet
926 dec := gob.NewDecoder(bufio.NewReaderSize(f, 1<<24))
928 err = dec.Decode(&hgvsCols)
933 log.Infof("%s: sorting %d hgvs variants", seqname, len(hgvsCols))
934 variants := make([]hgvs.Variant, 0, len(hgvsCols))
935 for v := range hgvsCols {
936 variants = append(variants, v)
938 sort.Slice(variants, func(i, j int) bool {
939 vi, vj := &variants[i], &variants[j]
940 if vi.Position != vj.Position {
941 return vi.Position < vj.Position
942 } else if vi.Ref != vj.Ref {
943 return vi.Ref < vj.Ref
945 return vi.New < vj.New
948 rows := len(cmd.cgnames)
949 cols := len(variants) * 2
950 log.Infof("%s: building hgvs matrix (rows=%d, cols=%d, mem=%d)", seqname, rows, cols, rows*cols)
951 out := make([]int8, rows*cols)
952 for varIdx, variant := range variants {
953 hgvsCols := hgvsCols[variant]
954 for row := range cmd.cgnames {
955 for ph := 0; ph < 2; ph++ {
956 out[row*cols+varIdx+ph] = hgvsCols[ph][row]
960 err = writeNumpyInt8(fmt.Sprintf("%s/hgvs.%s.npy", *outputDir, seqname), out, rows, cols)
966 fnm := fmt.Sprintf("%s/hgvs.%s.annotations.csv", *outputDir, seqname)
967 log.Infof("%s: writing hgvs column labels to %s", seqname, fnm)
968 var hgvsLabels bytes.Buffer
969 for varIdx, variant := range variants {
970 fmt.Fprintf(&hgvsLabels, "%d,%s:g.%s\n", varIdx, seqname, variant.String())
972 err = ioutil.WriteFile(fnm, hgvsLabels.Bytes(), 0666)
979 if *mergeOutput || *hgvsSingle {
980 var annow *bufio.Writer
983 annoFilename := fmt.Sprintf("%s/matrix.annotations.csv", *outputDir)
984 annof, err = os.Create(annoFilename)
988 annow = bufio.NewWriterSize(annof, 1<<20)
991 rows := len(cmd.cgnames)
993 for _, chunk := range toMerge {
994 cols += len(chunk) / rows
996 log.Infof("merging output matrix (rows=%d, cols=%d, mem=%d) and annotations", rows, cols, rows*cols*2)
999 out = make([]int16, rows*cols)
1001 hgvsCols := map[string][2][]int16{} // hgvs -> [[g0,g1,g2,...], [g0,g1,g2,...]] (slice of genomes for each phase)
1003 for outIdx, chunk := range toMerge {
1004 chunkcols := len(chunk) / rows
1006 for row := 0; row < rows; row++ {
1007 copy(out[row*cols+startcol:], chunk[row*chunkcols:(row+1)*chunkcols])
1010 toMerge[outIdx] = nil
1012 annotationsFilename := fmt.Sprintf("%s/matrix.%04d.annotations.csv", *outputDir, outIdx)
1013 log.Infof("reading %s", annotationsFilename)
1014 buf, err := os.ReadFile(annotationsFilename)
1019 err = os.Remove(annotationsFilename)
1024 for _, line := range bytes.Split(buf, []byte{'\n'}) {
1028 fields := bytes.SplitN(line, []byte{','}, 9)
1029 tag, _ := strconv.Atoi(string(fields[0]))
1030 incol, _ := strconv.Atoi(string(fields[1]))
1031 tileVariant, _ := strconv.Atoi(string(fields[2]))
1032 hgvsID := string(fields[3])
1033 seqname := string(fields[4])
1034 pos, _ := strconv.Atoi(string(fields[5]))
1037 // Null entry for un-diffable
1042 // Null entry for ref tile
1045 if mask != nil && !mask.Check(strings.TrimPrefix(seqname, "chr"), pos, pos+len(refseq)) {
1046 // The tile intersects one of
1047 // the selected regions, but
1048 // this particular HGVS
1049 // variant does not.
1052 hgvsColPair := hgvsCols[hgvsID]
1053 if hgvsColPair[0] == nil {
1054 // values in new columns start
1055 // out as -1 ("no data yet")
1056 // or 0 ("=ref") here, may
1057 // change to 1 ("hgvs variant
1058 // present") below, either on
1059 // this line or a future line.
1060 hgvsColPair = [2][]int16{make([]int16, len(cmd.cgnames)), make([]int16, len(cmd.cgnames))}
1061 rt, ok := reftile[tagID(tag)]
1063 err = fmt.Errorf("bug: seeing annotations for tag %d, but it has no reftile entry", tag)
1066 for ph := 0; ph < 2; ph++ {
1067 for row := 0; row < rows; row++ {
1068 v := chunk[row*chunkcols+incol*2+ph]
1069 if tileVariantID(v) == rt.variant {
1070 hgvsColPair[ph][row] = 0
1072 hgvsColPair[ph][row] = -1
1076 hgvsCols[hgvsID] = hgvsColPair
1078 hgvsref := hgvs.Variant{
1080 Ref: string(refseq),
1081 New: string(refseq),
1083 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])
1087 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])
1089 for ph := 0; ph < 2; ph++ {
1090 for row := 0; row < rows; row++ {
1091 v := chunk[row*chunkcols+incol*2+ph]
1092 if int(v) == tileVariant {
1093 hgvsColPair[ph][row] = 1
1099 startcol += chunkcols
1110 err = writeNumpyInt16(fmt.Sprintf("%s/matrix.npy", *outputDir), out, rows, cols)
1118 cols = len(hgvsCols) * 2
1119 log.Printf("building hgvs-based matrix: %d rows x %d cols", rows, cols)
1120 out = make([]int16, rows*cols)
1121 hgvsIDs := make([]string, 0, cols/2)
1122 for hgvsID := range hgvsCols {
1123 hgvsIDs = append(hgvsIDs, hgvsID)
1125 sort.Strings(hgvsIDs)
1126 var hgvsLabels bytes.Buffer
1127 for idx, hgvsID := range hgvsIDs {
1128 fmt.Fprintf(&hgvsLabels, "%d,%s\n", idx, hgvsID)
1129 for ph := 0; ph < 2; ph++ {
1130 hgvscol := hgvsCols[hgvsID][ph]
1131 for row, val := range hgvscol {
1132 out[row*cols+idx*2+ph] = val
1136 err = writeNumpyInt16(fmt.Sprintf("%s/hgvs.npy", *outputDir), out, rows, cols)
1141 fnm := fmt.Sprintf("%s/hgvs.annotations.csv", *outputDir)
1142 log.Printf("writing hgvs labels: %s", fnm)
1143 err = ioutil.WriteFile(fnm, hgvsLabels.Bytes(), 0777)
1149 if *onehotSingle || *onlyPCA {
1151 for _, part := range onehotIndirect {
1152 nzCount += len(part[0])
1154 onehot := make([]uint32, nzCount*2) // [r,r,r,...,c,c,c,...]
1155 var xrefs []onehotXref
1156 chunkOffset := uint32(0)
1158 for i, part := range onehotIndirect {
1159 for i := range part[1] {
1160 part[1][i] += chunkOffset
1162 copy(onehot[outcol:], part[0])
1163 copy(onehot[outcol+nzCount:], part[1])
1164 xrefs = append(xrefs, onehotXrefs[i]...)
1166 outcol += len(part[0])
1167 chunkOffset += onehotChunkSize[i]
1171 onehotXrefs[i] = nil
1172 debug.FreeOSMemory()
1175 fnm := fmt.Sprintf("%s/onehot.npy", *outputDir)
1176 err = writeNumpyUint32(fnm, onehot, 2, nzCount)
1180 fnm = fmt.Sprintf("%s/onehot-columns.npy", *outputDir)
1181 err = writeNumpyInt32(fnm, onehotXref2int32(xrefs), 5, len(xrefs))
1185 fnm = fmt.Sprintf("%s/stats.json", *outputDir)
1186 j, err := json.Marshal(map[string]interface{}{
1187 "pvalueCallCount": cmd.pvalueCallCount,
1192 err = os.WriteFile(fnm, j, 0777)
1199 for _, c := range onehot[nzCount:] {
1205 return fmt.Errorf("cannot do PCA: one-hot matrix is empty")
1207 log.Printf("have %d one-hot cols", cols)
1209 for *maxPCATiles > 0 && cols > *maxPCATiles*2 {
1210 cols = (cols + 1) / 2
1214 // we work with pairs of columns
1217 log.Printf("creating full matrix (%d rows) and training matrix (%d rows) with %d cols, stride %d", len(cmd.cgnames), cmd.trainingSetSize, cols, stride)
1218 mtxFull := mat.NewDense(len(cmd.cgnames), cols, nil)
1219 mtxTrain := mat.NewDense(cmd.trainingSetSize, cols, nil)
1220 for i, c := range onehot[nzCount:] {
1221 if int(c/2)%stride == 0 {
1222 outcol := int(c/2)/stride*2 + int(c)%2
1223 mtxFull.Set(int(onehot[i]), outcol, 1)
1224 if trainRow := cmd.trainingSet[int(onehot[i])]; trainRow >= 0 {
1225 mtxTrain.Set(trainRow, outcol, 1)
1229 log.Print("fitting")
1230 transformer := nlp.NewPCA(cmd.pcaComponents)
1231 transformer.Fit(mtxTrain.T())
1232 log.Printf("transforming")
1233 pca, err := transformer.Transform(mtxFull.T())
1238 outrows, outcols := pca.Dims()
1239 log.Printf("copying result to numpy output array: %d rows, %d cols", outrows, outcols)
1240 out := make([]float64, outrows*outcols)
1241 for i := 0; i < outrows; i++ {
1242 for j := 0; j < outcols; j++ {
1243 out[i*outcols+j] = pca.At(i, j)
1246 fnm := fmt.Sprintf("%s/pca.npy", *outputDir)
1247 log.Printf("writing numpy: %s", fnm)
1248 output, err := os.OpenFile(fnm, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, 0777)
1252 npw, err := gonpy.NewWriter(nopCloser{output})
1254 return fmt.Errorf("gonpy.NewWriter: %w", err)
1256 npw.Shape = []int{outrows, outcols}
1257 err = npw.WriteFloat64(out)
1259 return fmt.Errorf("WriteFloat64: %w", err)
1261 err = output.Close()
1267 log.Print("copying pca components to sampleInfo")
1268 for i := range cmd.samples {
1269 cmd.samples[i].pcaComponents = make([]float64, outcols)
1270 for c := 0; c < outcols; c++ {
1271 cmd.samples[i].pcaComponents[i] = pca.At(i, c)
1276 err = writeSampleInfo(cmd.samples, *outputDir)
1282 if !*mergeOutput && !*onehotChunked && !*onehotSingle && !*onlyPCA {
1283 tagoffsetFilename := *outputDir + "/chunk-tag-offset.csv"
1284 log.Infof("writing tag offsets to %s", tagoffsetFilename)
1286 f, err = os.Create(tagoffsetFilename)
1291 for idx, offset := range chunkStartTag {
1292 _, err = fmt.Fprintf(f, "%q,%d\n", fmt.Sprintf("matrix.%04d.npy", idx), offset)
1294 err = fmt.Errorf("write %s: %w", tagoffsetFilename, err)
1300 err = fmt.Errorf("close %s: %w", tagoffsetFilename, err)
1308 type sampleInfo struct {
1314 pcaComponents []float64
1317 // Read samples.csv file with case/control and training/validation
1319 func loadSampleInfo(samplesFilename string) ([]sampleInfo, error) {
1321 f, err := open(samplesFilename)
1325 buf, err := io.ReadAll(f)
1331 for _, csv := range bytes.Split(buf, []byte{'\n'}) {
1336 split := strings.Split(string(csv), ",")
1338 return nil, fmt.Errorf("%d fields < 4 in %s line %d: %q", len(split), samplesFilename, lineNum, csv)
1340 if split[0] == "Index" && split[1] == "SampleID" && split[2] == "CaseControl" && split[3] == "TrainingValidation" {
1343 idx, err := strconv.Atoi(split[0])
1346 return nil, fmt.Errorf("header does not look right: %q", csv)
1348 return nil, fmt.Errorf("%s line %d: index: %s", samplesFilename, lineNum, err)
1351 return nil, fmt.Errorf("%s line %d: index %d out of order", samplesFilename, lineNum, idx)
1353 var pcaComponents []float64
1355 for _, s := range split[4:] {
1356 f, err := strconv.ParseFloat(s, 64)
1358 return nil, fmt.Errorf("%s line %d: cannot parse float %q: %s", samplesFilename, lineNum, s, err)
1360 pcaComponents = append(pcaComponents, f)
1363 si = append(si, sampleInfo{
1365 isCase: split[2] == "1",
1366 isControl: split[2] == "0",
1367 isTraining: split[3] == "1",
1368 isValidation: split[3] == "0" && len(split[2]) > 0, // fix errant 0s in input
1369 pcaComponents: pcaComponents,
1375 func writeSampleInfo(samples []sampleInfo, outputDir string) error {
1376 fnm := outputDir + "/samples.csv"
1377 log.Infof("writing sample metadata to %s", fnm)
1378 f, err := os.Create(fnm)
1384 if len(samples) > 0 {
1385 for i := range samples[0].pcaComponents {
1386 pcaLabels += fmt.Sprintf(",PCA%d", i)
1389 _, err = fmt.Fprintf(f, "Index,SampleID,CaseControl,TrainingValidation%s\n", pcaLabels)
1393 for i, si := range samples {
1397 } else if si.isControl {
1402 } else if si.isValidation {
1406 for _, pcaval := range si.pcaComponents {
1407 pcavals += fmt.Sprintf(",%f", pcaval)
1409 _, err = fmt.Fprintf(f, "%d,%s,%s,%s%s\n", i, si.id, cc, tv, pcavals)
1411 return fmt.Errorf("write %s: %w", fnm, err)
1416 return fmt.Errorf("close %s: %w", fnm, err)
1422 func (cmd *sliceNumpy) filterHGVScolpair(colpair [2][]int8) bool {
1423 if cmd.chi2PValue >= 1 {
1426 col0 := make([]bool, 0, len(cmd.chi2Cases))
1427 col1 := make([]bool, 0, len(cmd.chi2Cases))
1428 cases := make([]bool, 0, len(cmd.chi2Cases))
1429 for i, c := range cmd.chi2Cases {
1430 if colpair[0][i] < 0 {
1433 col0 = append(col0, colpair[0][i] != 0)
1434 col1 = append(col1, colpair[1][i] != 0)
1435 cases = append(cases, c)
1437 return len(cases) >= cmd.minCoverage &&
1438 (pvalue(col0, cases) <= cmd.chi2PValue || pvalue(col1, cases) <= cmd.chi2PValue)
1441 func writeNumpyUint32(fnm string, out []uint32, rows, cols int) error {
1442 output, err := os.Create(fnm)
1446 defer output.Close()
1447 bufw := bufio.NewWriterSize(output, 1<<26)
1448 npw, err := gonpy.NewWriter(nopCloser{bufw})
1452 log.WithFields(log.Fields{
1456 "bytes": rows * cols * 4,
1457 }).Infof("writing numpy: %s", fnm)
1458 npw.Shape = []int{rows, cols}
1459 npw.WriteUint32(out)
1464 return output.Close()
1467 func writeNumpyInt32(fnm string, out []int32, rows, cols int) error {
1468 output, err := os.Create(fnm)
1472 defer output.Close()
1473 bufw := bufio.NewWriterSize(output, 1<<26)
1474 npw, err := gonpy.NewWriter(nopCloser{bufw})
1478 log.WithFields(log.Fields{
1482 "bytes": rows * cols * 4,
1483 }).Infof("writing numpy: %s", fnm)
1484 npw.Shape = []int{rows, cols}
1490 return output.Close()
1493 func writeNumpyInt16(fnm string, out []int16, rows, cols int) error {
1494 output, err := os.Create(fnm)
1498 defer output.Close()
1499 bufw := bufio.NewWriterSize(output, 1<<26)
1500 npw, err := gonpy.NewWriter(nopCloser{bufw})
1504 log.WithFields(log.Fields{
1508 "bytes": rows * cols * 2,
1509 }).Infof("writing numpy: %s", fnm)
1510 npw.Shape = []int{rows, cols}
1516 return output.Close()
1519 func writeNumpyInt8(fnm string, out []int8, rows, cols int) error {
1520 output, err := os.Create(fnm)
1524 defer output.Close()
1525 bufw := bufio.NewWriterSize(output, 1<<26)
1526 npw, err := gonpy.NewWriter(nopCloser{bufw})
1530 log.WithFields(log.Fields{
1534 "bytes": rows * cols,
1535 }).Infof("writing numpy: %s", fnm)
1536 npw.Shape = []int{rows, cols}
1542 return output.Close()
1545 func allele2homhet(colpair [2][]int8) {
1546 a, b := colpair[0], colpair[1]
1547 for i, av := range a {
1549 if av < 0 || bv < 0 {
1552 } else if av > 0 && bv > 0 {
1555 } else if av > 0 || bv > 0 {
1559 // ref (or a different variant in same position)
1560 // (this is a no-op) a[i], b[i] = 0, 0
1565 type onehotXref struct {
1567 variant tileVariantID
1573 const onehotXrefSize = unsafe.Sizeof(onehotXref{})
1575 // Build onehot matrix (m[tileVariantIndex][genome] == 0 or 1) for all
1576 // variants of a single tile/tag#.
1578 // Return nil if no tile variant passes Χ² filter.
1579 func (cmd *sliceNumpy) tv2homhet(cgs map[string]CompactGenome, maxv tileVariantID, remap []tileVariantID, tag, chunkstarttag tagID, seq map[tagID][]TileVariant) ([][]int8, []onehotXref) {
1580 if tag == cmd.debugTag {
1581 tv := make([]tileVariantID, len(cmd.cgnames)*2)
1582 for i, name := range cmd.cgnames {
1583 copy(tv[i*2:(i+1)*2], cgs[name].Variants[(tag-chunkstarttag)*2:])
1585 log.WithFields(logrus.Fields{
1586 "cgs[i].Variants[tag*2+j]": tv,
1590 "chunkstarttag": chunkstarttag,
1591 }).Info("tv2homhet()")
1593 if maxv < 1 || (maxv < 2 && !cmd.includeVariant1) {
1594 // everyone has the most common variant (of the variants we don't drop)
1597 tagoffset := tag - chunkstarttag
1599 for _, cg := range cgs {
1601 for _, v := range cg.Variants[tagoffset*2 : tagoffset*2+2] {
1602 if v > 0 && int(v) < len(seq[tag]) && len(seq[tag][v].Sequence) > 0 {
1610 if coverage < cmd.minCoverage {
1613 // "observed" array for p-value calculation (training set
1615 obs := make([][]bool, (maxv+1)*2) // 2 slices (hom + het) for each variant#
1616 // one-hot output (all samples)
1617 outcols := make([][]int8, (maxv+1)*2)
1618 for i := range obs {
1619 obs[i] = make([]bool, cmd.trainingSetSize)
1620 outcols[i] = make([]int8, len(cmd.cgnames))
1622 for cgid, name := range cmd.cgnames {
1623 tsid := cmd.trainingSet[cgid]
1624 cgvars := cgs[name].Variants[tagoffset*2:]
1625 tv0, tv1 := remap[cgvars[0]], remap[cgvars[1]]
1626 for v := tileVariantID(1); v <= maxv; v++ {
1627 if tv0 == v && tv1 == v {
1629 obs[v*2][tsid] = true
1631 outcols[v*2][cgid] = 1
1632 } else if tv0 == v || tv1 == v {
1634 obs[v*2+1][tsid] = true
1636 outcols[v*2+1][cgid] = 1
1641 var xref []onehotXref
1643 for col := 2; col < len(obs); col++ {
1644 // col 0,1 correspond to tile variant 0, i.e.,
1645 // no-call; col 2,3 correspond to the most common
1646 // variant; so we (normally) start at col 4.
1647 if col < 4 && !cmd.includeVariant1 {
1651 maf = homhet2maf(obs[col : col+2])
1652 if maf < cmd.pvalueMinFrequency {
1653 // Skip both columns (hom and het) if
1654 // allele frequency is below threshold
1658 if maf > cmd.maxFrequency {
1659 // Skip both columns if allele
1660 // frequency is above threshold
1665 atomic.AddInt64(&cmd.pvalueCallCount, 1)
1666 p := cmd.pvalue(obs[col])
1667 if cmd.chi2PValue < 1 && !(p < cmd.chi2PValue) {
1670 onehot = append(onehot, outcols[col])
1671 xref = append(xref, onehotXref{
1673 variant: tileVariantID(col >> 1),
1682 func homhet2maf(onehot [][]bool) float64 {
1683 if len(onehot[0]) == 0 {
1687 for i := range onehot[0] {
1691 } else if onehot[1][i] {
1696 return float64(n) / float64(len(onehot[0])*2)
1699 // convert a []onehotXref with length N to a numpy-style []int32
1700 // matrix with N columns, one row per field of onehotXref struct.
1702 // Hom/het row contains hom=0, het=1.
1704 // P-value row contains 1000000x actual p-value.
1705 func onehotXref2int32(xrefs []onehotXref) []int32 {
1707 xdata := make([]int32, 6*xcols)
1708 for i, xref := range xrefs {
1709 xdata[i] = int32(xref.tag)
1710 xdata[xcols+i] = int32(xref.variant)
1712 xdata[xcols*2+i] = 1
1714 xdata[xcols*3+i] = int32(xref.pvalue * 1000000)
1715 xdata[xcols*4+i] = int32(-math.Log10(xref.pvalue) * 1000000)
1716 xdata[xcols*5+i] = int32(xref.maf * 1000000)
1721 // transpose onehot data from in[col][row] to numpy-style
1722 // out[row*cols+col].
1723 func onehotcols2int8(in [][]int8) []int8 {
1729 out := make([]int8, rows*cols)
1730 for row := 0; row < rows; row++ {
1731 outrow := out[row*cols:]
1732 for col, incol := range in {
1733 outrow[col] = incol[row]
1739 // Return [2][]uint32{rowIndices, colIndices} indicating which
1740 // elements of matrixT[c][r] have non-zero values.
1741 func onehotChunk2Indirect(matrixT [][]int8) [2][]uint32 {
1743 for c, col := range matrixT {
1744 for r, val := range col {
1746 nz[0] = append(nz[0], uint32(r))
1747 nz[1] = append(nz[1], uint32(c))