import (
"bufio"
"bytes"
+ "encoding/gob"
"flag"
"fmt"
"io"
"os"
"regexp"
"runtime"
+ "runtime/debug"
"sort"
"strconv"
"strings"
regionsFilename := flags.String("regions", "", "only output columns/annotations that intersect regions in specified bed `file`")
expandRegions := flags.Int("expand-regions", 0, "expand specified regions by `N` base pairs on each side`")
mergeOutput := flags.Bool("merge-output", false, "merge output into one matrix.npy and one matrix.annotations.csv")
+ hgvsSingle := flags.Bool("single-hgvs-matrix", false, "also generate hgvs-based matrix")
+ hgvsChunked := flags.Bool("chunked-hgvs-matrix", false, "also generate hgvs-based matrix per chromosome")
flags.IntVar(&cmd.threads, "threads", 16, "number of memory-hungry assembly threads")
cmd.filter.Flags(flags)
err = flags.Parse(args)
Name: "lightning slice-numpy",
Client: arvados.NewClientFromEnv(),
ProjectUUID: *projectUUID,
- RAM: 650000000000,
+ RAM: 750000000000,
VCPUs: 96,
Priority: *priority,
KeepCache: 2,
"-regions=" + *regionsFilename,
"-expand-regions=" + fmt.Sprintf("%d", *expandRegions),
"-merge-output=" + fmt.Sprintf("%v", *mergeOutput),
+ "-single-hgvs-matrix=" + fmt.Sprintf("%v", *hgvsSingle),
+ "-chunked-hgvs-matrix=" + fmt.Sprintf("%v", *hgvsChunked),
}
runner.Args = append(runner.Args, cmd.filter.Args()...)
var output string
log.Printf("after applying mask, len(reftile) == %d", len(reftile))
}
+ type hgvsColSet map[hgvs.Variant][2][]int8
+ encodeHGVS := throttle{Max: len(refseq)}
+ encodeHGVSTodo := map[string]chan hgvsColSet{}
+ tmpHGVSCols := map[string]*os.File{}
+ if *hgvsChunked {
+ for seqname := range refseq {
+ var f *os.File
+ f, err = os.Create(*outputDir + "/tmp." + seqname + ".gob")
+ if err != nil {
+ return 1
+ }
+ defer os.Remove(f.Name())
+ bufw := bufio.NewWriterSize(f, 1<<24)
+ enc := gob.NewEncoder(bufw)
+ tmpHGVSCols[seqname] = f
+ todo := make(chan hgvsColSet, 128)
+ encodeHGVSTodo[seqname] = todo
+ encodeHGVS.Go(func() error {
+ for colset := range todo {
+ err := enc.Encode(colset)
+ if err != nil {
+ encodeHGVS.Report(err)
+ for range todo {
+ }
+ return err
+ }
+ }
+ return bufw.Flush()
+ })
+ }
+ }
+
var toMerge [][]int16
- if *mergeOutput {
+ if *mergeOutput || *hgvsSingle {
toMerge = make([][]int16, len(infiles))
}
seq[tv.Tag] = variants
}
for _, cg := range ent.CompactGenomes {
- if matchGenome.MatchString(cg.Name) {
- cgs[cg.Name] = cg
+ if !matchGenome.MatchString(cg.Name) {
+ continue
+ }
+ // pad to full slice size
+ // to avoid out-of-bounds
+ // checks later
+ if sliceSize := 2 * int(cg.EndTag-cg.StartTag); len(cg.Variants) < sliceSize {
+ cg.Variants = append(cg.Variants, make([]tileVariantID, sliceSize-len(cg.Variants))...)
}
+ cgs[cg.Name] = cg
}
return nil
})
go func() {
defer throttleCPU.Release()
count := make(map[[blake2b.Size256]byte]int, len(variants))
+
+ rt := reftile[tag]
+ if rt != nil {
+ count[blake2b.Sum256(rt.tiledata)] = 0
+ }
+
for _, cg := range cgs {
idx := int(tag-tagstart) * 2
- if idx < len(cg.Variants) {
- for allele := 0; allele < 2; allele++ {
- v := cg.Variants[idx+allele]
- if v > 0 && len(variants[v].Sequence) > 0 {
- count[variants[v].Blake2b]++
- }
+ for allele := 0; allele < 2; allele++ {
+ v := cg.Variants[idx+allele]
+ if v > 0 && len(variants[v].Sequence) > 0 {
+ count[variants[v].Blake2b]++
}
}
}
remap[i] = rank[tv.Blake2b]
}
variantRemap[tag-tagstart] = remap
+ if rt != nil {
+ rt.variant = rank[blake2b.Sum256(rt.tiledata)]
+ }
}()
}
throttleCPU.Wait()
// mention it in annotations?)
continue
}
- variants, ok := seq[tag]
- if !ok {
- // how could we even have a reftile if there is no sequence data??
- return fmt.Errorf("bug: have no variants for tag %d but reftile is %+v", tag, rt)
- }
+ fmt.Fprintf(annow, "%d,%d,%d,=,%s,%d,,,\n", tag, outcol, rt.variant, rt.seqname, rt.pos)
+ variants := seq[tag]
reftilestr := strings.ToUpper(string(rt.tiledata))
remap := variantRemap[tag-tagstart]
- done := make([]bool, len(variants))
+ maxv := tileVariantID(0)
+ for _, v := range remap {
+ if maxv < v {
+ maxv = v
+ }
+ }
+ done := make([]bool, maxv+1)
+ variantDiffs := make([][]hgvs.Variant, maxv+1)
for v, tv := range variants {
v := remap[v]
- if done[v] {
+ if v == rt.variant || done[v] {
continue
} else {
done[v] = true
}
if len(tv.Sequence) < taglen || !bytes.HasSuffix(rt.tiledata, tv.Sequence[len(tv.Sequence)-taglen:]) {
- fmt.Fprintf(annow, "%d,%d,%d,.,%s,%d,.,,\n", tag, outcol, v, rt.seqname, rt.pos)
+ fmt.Fprintf(annow, "%d,%d,%d,,%s,%d,,,\n", tag, outcol, v, rt.seqname, rt.pos)
continue
}
if lendiff := len(rt.tiledata) - len(tv.Sequence); lendiff < -1000 || lendiff > 1000 {
continue
}
diffs, _ := hgvs.Diff(reftilestr, strings.ToUpper(string(tv.Sequence)), 0)
+ for i := range diffs {
+ diffs[i].Position += rt.pos
+ }
for _, diff := range diffs {
- diff.Position += rt.pos
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)
}
+ if *hgvsChunked {
+ variantDiffs[v] = diffs
+ }
+ }
+ if *hgvsChunked {
+ // We can now determine, for each HGVS
+ // variant (diff) in this reftile
+ // region, whether a given genome
+ // phase/allele (1) has the variant, (0) has
+ // =ref or a different variant in that
+ // position, or (-1) is lacking
+ // coverage / couldn't be diffed.
+ hgvsCol := hgvsColSet{}
+ for _, diffs := range variantDiffs {
+ for _, diff := range diffs {
+ if _, ok := hgvsCol[diff]; ok {
+ continue
+ }
+ hgvsCol[diff] = [2][]int8{
+ make([]int8, len(cgnames)),
+ make([]int8, len(cgnames)),
+ }
+ }
+ }
+ for row, name := range cgnames {
+ variants := cgs[name].Variants[(tag-tagstart)*2:]
+ for ph := 0; ph < 2; ph++ {
+ v := variants[ph]
+ if int(v) >= len(remap) {
+ v = 0
+ } else {
+ v = remap[v]
+ }
+ if v == rt.variant {
+ // hgvsCol[*][ph][row] is already 0
+ } else if len(variantDiffs[v]) == 0 {
+ // lacking coverage / couldn't be diffed
+ for _, col := range hgvsCol {
+ col[ph][row] = -1
+ }
+ } else {
+ for _, diff := range variantDiffs[v] {
+ hgvsCol[diff][ph][row] = 1
+ }
+ }
+ }
+ }
+ encodeHGVSTodo[rt.seqname] <- hgvsCol
}
outcol++
}
}
}
seq = nil
+ cgs = nil
+ debug.FreeOSMemory()
throttleNumpyMem.Release()
- if *mergeOutput {
+ if *mergeOutput || *hgvsSingle {
log.Infof("%04d: matrix fragment %d rows x %d cols", infileIdx, rows, cols)
toMerge[infileIdx] = out
- } else {
+ }
+ if !*mergeOutput {
fnm := fmt.Sprintf("%s/matrix.%04d.npy", *outputDir, infileIdx)
err = writeNumpyInt16(fnm, out, rows, cols)
if err != nil {
return err
}
+ debug.FreeOSMemory()
}
log.Infof("%s: done (%d/%d)", infile, int(atomic.AddInt64(&done, 1)), len(infiles))
return nil
if err = throttleMem.Wait(); err != nil {
return 1
}
- if *mergeOutput {
- log.Info("merging output matrix and annotations")
- annoFilename := fmt.Sprintf("%s/matrix.annotations.csv", *outputDir)
- annof, err := os.Create(annoFilename)
+ if *hgvsChunked {
+ log.Info("flushing hgvsCols temp files")
+ for seqname := range refseq {
+ close(encodeHGVSTodo[seqname])
+ }
+ err = encodeHGVS.Wait()
if err != nil {
return 1
}
- annow := bufio.NewWriterSize(annof, 1<<20)
+ for seqname := range refseq {
+ log.Infof("%s: reading hgvsCols from temp file", seqname)
+ f := tmpHGVSCols[seqname]
+ _, err = f.Seek(0, io.SeekStart)
+ if err != nil {
+ return 1
+ }
+ var hgvsCols hgvsColSet
+ dec := gob.NewDecoder(bufio.NewReaderSize(f, 1<<24))
+ for err == nil {
+ err = dec.Decode(&hgvsCols)
+ }
+ if err != io.EOF {
+ return 1
+ }
+ log.Infof("%s: sorting %d hgvs variants", seqname, len(hgvsCols))
+ variants := make([]hgvs.Variant, 0, len(hgvsCols))
+ for v := range hgvsCols {
+ variants = append(variants, v)
+ }
+ sort.Slice(variants, func(i, j int) bool {
+ vi, vj := &variants[i], &variants[j]
+ if vi.Position != vj.Position {
+ return vi.Position < vj.Position
+ } else if vi.Ref != vj.Ref {
+ return vi.Ref < vj.Ref
+ } else {
+ return vi.New < vj.New
+ }
+ })
+ rows := len(cgnames)
+ cols := len(variants) * 2
+ log.Infof("%s: building hgvs matrix (rows=%d, cols=%d, mem=%d)", seqname, rows, cols, rows*cols)
+ out := make([]int8, rows*cols)
+ for varIdx, variant := range variants {
+ hgvsCols := hgvsCols[variant]
+ for row := range cgnames {
+ for ph := 0; ph < 2; ph++ {
+ out[row*cols+varIdx+ph] = hgvsCols[ph][row]
+ }
+ }
+ }
+ err = writeNumpyInt8(fmt.Sprintf("%s/hgvs.%s.npy", *outputDir, seqname), out, rows, cols)
+ if err != nil {
+ return 1
+ }
+ out = nil
+
+ fnm := fmt.Sprintf("%s/hgvs.%s.annotations.csv", *outputDir, seqname)
+ log.Infof("%s: writing hgvs column labels to %s", seqname, fnm)
+ var hgvsLabels bytes.Buffer
+ for varIdx, variant := range variants {
+ fmt.Fprintf(&hgvsLabels, "%d,%s:g.%s\n", varIdx, seqname, variant.String())
+ }
+ err = ioutil.WriteFile(fnm, hgvsLabels.Bytes(), 0666)
+ if err != nil {
+ return 1
+ }
+ }
+ }
+
+ if *mergeOutput || *hgvsSingle {
+ var annow *bufio.Writer
+ var annof *os.File
+ if *mergeOutput {
+ annoFilename := fmt.Sprintf("%s/matrix.annotations.csv", *outputDir)
+ annof, err = os.Create(annoFilename)
+ if err != nil {
+ return 1
+ }
+ annow = bufio.NewWriterSize(annof, 1<<20)
+ }
rows := len(cgnames)
cols := 0
for _, chunk := range toMerge {
cols += len(chunk) / rows
}
- out := make([]int16, rows*cols)
- hgvs := map[string][2][]int16{} // hgvs -> [[g0,g1,g2,...], [g0,g1,g2,...]] (slice of genomes for each phase)
+ log.Infof("merging output matrix (rows=%d, cols=%d, mem=%d) and annotations", rows, cols, rows*cols*2)
+ var out []int16
+ if *mergeOutput {
+ out = make([]int16, rows*cols)
+ }
+ hgvsCols := map[string][2][]int16{} // hgvs -> [[g0,g1,g2,...], [g0,g1,g2,...]] (slice of genomes for each phase)
startcol := 0
for outIdx, chunk := range toMerge {
chunkcols := len(chunk) / rows
- for row := 0; row < rows; row++ {
- copy(out[row*cols+startcol:], chunk[row*chunkcols:(row+1)*chunkcols])
+ if *mergeOutput {
+ for row := 0; row < rows; row++ {
+ copy(out[row*cols+startcol:], chunk[row*chunkcols:(row+1)*chunkcols])
+ }
}
toMerge[outIdx] = nil
if err != nil {
return 1
}
- err = os.Remove(annotationsFilename)
- if err != nil {
- return 1
+ if *mergeOutput {
+ err = os.Remove(annotationsFilename)
+ if err != nil {
+ return 1
+ }
}
for _, line := range bytes.Split(buf, []byte{'\n'}) {
if len(line) == 0 {
continue
}
- fields := bytes.SplitN(line, []byte{','}, 8)
+ fields := bytes.SplitN(line, []byte{','}, 9)
+ tag, _ := strconv.Atoi(string(fields[0]))
incol, _ := strconv.Atoi(string(fields[1]))
tileVariant, _ := strconv.Atoi(string(fields[2]))
hgvsID := string(fields[3])
seqname := string(fields[4])
pos, _ := strconv.Atoi(string(fields[5]))
refseq := fields[6]
+ if hgvsID == "" {
+ // Null entry for un-diffable
+ // tile variant
+ continue
+ }
+ if hgvsID == "=" {
+ // Null entry for ref tile
+ continue
+ }
if mask != nil && !mask.Check(strings.TrimPrefix(seqname, "chr"), pos, pos+len(refseq)) {
// The tile intersects one of
// the selected regions, but
// variant does not.
continue
}
- fmt.Fprintf(annow, "%s,%d,%d,%s,%s,%d,%s,%s\n", fields[0], incol+startcol/2, tileVariant, hgvsID, seqname, pos, refseq, fields[7])
- hgvscols := hgvs[hgvsID]
- if hgvscols[0] == nil {
- hgvscols = [2][]int16{make([]int16, len(cgnames)), make([]int16, len(cgnames))}
- hgvs[hgvsID] = hgvscols
+ hgvsColPair := hgvsCols[hgvsID]
+ if hgvsColPair[0] == nil {
+ // values in new columns start
+ // out as -1 ("no data yet")
+ // or 0 ("=ref") here, may
+ // change to 1 ("hgvs variant
+ // present") below, either on
+ // this line or a future line.
+ hgvsColPair = [2][]int16{make([]int16, len(cgnames)), make([]int16, len(cgnames))}
+ rt, ok := reftile[tagID(tag)]
+ if !ok {
+ err = fmt.Errorf("bug: seeing annotations for tag %d, but it has no reftile entry", tag)
+ return 1
+ }
+ for ph := 0; ph < 2; ph++ {
+ for row := 0; row < rows; row++ {
+ v := chunk[row*chunkcols+incol*2+ph]
+ if tileVariantID(v) == rt.variant {
+ hgvsColPair[ph][row] = 0
+ } else {
+ hgvsColPair[ph][row] = -1
+ }
+ }
+ }
+ hgvsCols[hgvsID] = hgvsColPair
+ if annow != nil {
+ hgvsref := hgvs.Variant{
+ Position: pos,
+ Ref: string(refseq),
+ New: string(refseq),
+ }
+ 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])
+ }
+ }
+ if annow != nil {
+ 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])
}
for ph := 0; ph < 2; ph++ {
for row := 0; row < rows; row++ {
v := chunk[row*chunkcols+incol*2+ph]
if int(v) == tileVariant {
- if len(hgvsID) == 0 {
- // we have the tile variant sequence, but the diff against ref didn't work out (see lendiff above)
- hgvscols[ph][row] = -2
- } else {
- hgvscols[ph][row] = 1
- }
- } else if v < 0 {
- hgvscols[ph][row] = -1
+ hgvsColPair[ph][row] = 1
}
}
}
startcol += chunkcols
}
- err = annow.Flush()
- if err != nil {
- return 1
- }
- err = annof.Close()
- if err != nil {
- return 1
- }
- err = writeNumpyInt16(fmt.Sprintf("%s/matrix.npy", *outputDir), out, rows, cols)
- if err != nil {
- return 1
+ if *mergeOutput {
+ err = annow.Flush()
+ if err != nil {
+ return 1
+ }
+ err = annof.Close()
+ if err != nil {
+ return 1
+ }
+ err = writeNumpyInt16(fmt.Sprintf("%s/matrix.npy", *outputDir), out, rows, cols)
+ if err != nil {
+ return 1
+ }
}
out = nil
- cols = len(hgvs) * 2
- log.Printf("building hgvs-based matrix: %d rows x %d cols", rows, cols)
- out = make([]int16, rows*cols)
- hgvsIDs := make([]string, 0, len(hgvs))
- for hgvsID := range hgvs {
- hgvsIDs = append(hgvsIDs, hgvsID)
- }
- sort.Strings(hgvsIDs)
- var hgvsLabels bytes.Buffer
- for idx, hgvsID := range hgvsIDs {
- fmt.Fprintf(&hgvsLabels, "%d,%s\n", idx, hgvsID)
- for ph := 0; ph < 2; ph++ {
- hgvscol := hgvs[hgvsID][ph]
- for row, val := range hgvscol {
- out[row*cols+idx*2+ph] = val
+ if *hgvsSingle {
+ cols = len(hgvsCols) * 2
+ log.Printf("building hgvs-based matrix: %d rows x %d cols", rows, cols)
+ out = make([]int16, rows*cols)
+ hgvsIDs := make([]string, 0, cols/2)
+ for hgvsID := range hgvsCols {
+ hgvsIDs = append(hgvsIDs, hgvsID)
+ }
+ sort.Strings(hgvsIDs)
+ var hgvsLabels bytes.Buffer
+ for idx, hgvsID := range hgvsIDs {
+ fmt.Fprintf(&hgvsLabels, "%d,%s\n", idx, hgvsID)
+ for ph := 0; ph < 2; ph++ {
+ hgvscol := hgvsCols[hgvsID][ph]
+ for row, val := range hgvscol {
+ out[row*cols+idx*2+ph] = val
+ }
}
}
- }
- err = writeNumpyInt16(fmt.Sprintf("%s/hgvs.npy", *outputDir), out, rows, cols)
- if err != nil {
- return 1
- }
+ err = writeNumpyInt16(fmt.Sprintf("%s/hgvs.npy", *outputDir), out, rows, cols)
+ if err != nil {
+ return 1
+ }
- fnm := fmt.Sprintf("%s/hgvs.annotations.csv", *outputDir)
- log.Printf("writing hgvs labels: %s", fnm)
- err = ioutil.WriteFile(fnm, hgvsLabels.Bytes(), 0777)
- if err != nil {
- return 1
+ fnm := fmt.Sprintf("%s/hgvs.annotations.csv", *outputDir)
+ log.Printf("writing hgvs labels: %s", fnm)
+ err = ioutil.WriteFile(fnm, hgvsLabels.Bytes(), 0777)
+ if err != nil {
+ return 1
+ }
}
}
return 0
}
return output.Close()
}
+
+func writeNumpyInt8(fnm string, out []int8, rows, cols int) error {
+ output, err := os.Create(fnm)
+ if err != nil {
+ return err
+ }
+ defer output.Close()
+ bufw := bufio.NewWriterSize(output, 1<<26)
+ npw, err := gonpy.NewWriter(nopCloser{bufw})
+ if err != nil {
+ return err
+ }
+ log.WithFields(log.Fields{
+ "filename": fnm,
+ "rows": rows,
+ "cols": cols,
+ }).Infof("writing numpy: %s", fnm)
+ npw.Shape = []int{rows, cols}
+ npw.WriteInt8(out)
+ err = bufw.Flush()
+ if err != nil {
+ return err
+ }
+ return output.Close()
+}
projects / lightning.git / blobdiff