#include "bam.h"
#include "kstring.h"
#include "bam2bcf.h"
+#include "errmod.h"
#include "bcftools/bcf.h"
extern void ks_introsort_uint32_t(size_t n, uint32_t a[]);
#define CALL_ETA 0.03f
#define CALL_MAX 256
-#define CALL_DEFTHETA 0.85f
+#define CALL_DEFTHETA 0.83f
+#define DEF_MAPQ 20
-struct __bcf_callaux_t {
- int max_info, capQ;
- double *fk;
- uint32_t *info;
-};
+#define CAP_DIST 25
-bcf_callaux_t *bcf_call_init(double theta)
+bcf_callaux_t *bcf_call_init(double theta, int min_baseQ)
{
bcf_callaux_t *bca;
- int n;
if (theta <= 0.) theta = CALL_DEFTHETA;
bca = calloc(1, sizeof(bcf_callaux_t));
bca->capQ = 60;
- bca->fk = calloc(CALL_MAX, sizeof(double));
- bca->fk[0] = 1.;
- for (n = 1; n < CALL_MAX; ++n)
- bca->fk[n] = theta >= 1.? 1. : pow(theta, n) * (1.0 - CALL_ETA) + CALL_ETA;
- bca->fk[CALL_MAX-1] = 0.;
+ bca->openQ = 40; bca->extQ = 20; bca->tandemQ = 100;
+ bca->min_baseQ = min_baseQ;
+ bca->e = errmod_init(1. - theta);
+ bca->min_frac = 0.002;
+ bca->min_support = 1;
+ bca->per_sample_flt = 0;
return bca;
}
void bcf_call_destroy(bcf_callaux_t *bca)
{
- if (bca) {
- free(bca->info); free(bca->fk); free(bca);
- }
+ if (bca == 0) return;
+ errmod_destroy(bca->e);
+ free(bca->bases); free(bca->inscns); free(bca);
}
+/* ref_base is the 4-bit representation of the reference base. It is
+ * negative if we are looking at an indel. */
+int bcf_call_glfgen(int _n, const bam_pileup1_t *pl, int ref_base, bcf_callaux_t *bca, bcf_callret1_t *r)
+{
+ static int *var_pos = NULL, nvar_pos = 0;
+ int i, n, ref4, is_indel, ori_depth = 0;
+ memset(r, 0, sizeof(bcf_callret1_t));
+ if (ref_base >= 0) {
+ ref4 = bam_nt16_nt4_table[ref_base];
+ is_indel = 0;
+ } else ref4 = 4, is_indel = 1;
+ if (_n == 0) return -1;
+ // enlarge the bases array if necessary
+ if (bca->max_bases < _n) {
+ bca->max_bases = _n;
+ kroundup32(bca->max_bases);
+ bca->bases = (uint16_t*)realloc(bca->bases, 2 * bca->max_bases);
+ }
+ // fill the bases array
+ for (i = n = r->n_supp = 0; i < _n; ++i) {
+ const bam_pileup1_t *p = pl + i;
+ int q, b, mapQ, baseQ, is_diff, min_dist, seqQ;
+ // set base
+ if (p->is_del || p->is_refskip || (p->b->core.flag&BAM_FUNMAP)) continue;
+ ++ori_depth;
+ baseQ = q = is_indel? p->aux&0xff : (int)bam1_qual(p->b)[p->qpos]; // base/indel quality
+ seqQ = is_indel? (p->aux>>8&0xff) : 99;
+ if (q < bca->min_baseQ) continue;
+ if (q > seqQ) q = seqQ;
+ mapQ = p->b->core.qual < 255? p->b->core.qual : DEF_MAPQ; // special case for mapQ==255
+ mapQ = mapQ < bca->capQ? mapQ : bca->capQ;
+ if (q > mapQ) q = mapQ;
+ if (q > 63) q = 63;
+ if (q < 4) q = 4;
+ if (!is_indel) {
+ b = bam1_seqi(bam1_seq(p->b), p->qpos); // base
+ b = bam_nt16_nt4_table[b? b : ref_base]; // b is the 2-bit base
+ is_diff = (ref4 < 4 && b == ref4)? 0 : 1;
+ } else {
+ b = p->aux>>16&0x3f;
+ is_diff = (b != 0);
+ }
+ if (is_diff) ++r->n_supp;
+ bca->bases[n++] = q<<5 | (int)bam1_strand(p->b)<<4 | b;
+ // collect annotations
+ if (b < 4) r->qsum[b] += q;
+ ++r->anno[0<<2|is_diff<<1|bam1_strand(p->b)];
+ min_dist = p->b->core.l_qseq - 1 - p->qpos;
+ if (min_dist > p->qpos) min_dist = p->qpos;
+ if (min_dist > CAP_DIST) min_dist = CAP_DIST;
+ r->anno[1<<2|is_diff<<1|0] += baseQ;
+ r->anno[1<<2|is_diff<<1|1] += baseQ * baseQ; // FIXME: signed int is not enough for thousands of samples
+ r->anno[2<<2|is_diff<<1|0] += mapQ;
+ r->anno[2<<2|is_diff<<1|1] += mapQ * mapQ; // FIXME: signed int is not enough for thousands of samples
+ r->anno[3<<2|is_diff<<1|0] += min_dist;
+ r->anno[3<<2|is_diff<<1|1] += min_dist * min_dist;
+ }
+ r->depth = n; r->ori_depth = ori_depth;
+ // glfgen
+ errmod_cal(bca->e, n, 5, bca->bases, r->p);
-typedef struct {
- float esum[5], fsum[5];
- uint32_t c[5];
- int w[8];
-} auxaux_t;
+ // Calculate the Variant Distance Bias (make it optional?)
+ if ( nvar_pos < _n ) {
+ nvar_pos = _n;
+ var_pos = realloc(var_pos,sizeof(int)*nvar_pos);
+ }
+ int alt_dp=0, read_len=0;
+ for (i=0; i<_n; i++) {
+ const bam_pileup1_t *p = pl + i;
+ if ( bam1_seqi(bam1_seq(p->b),p->qpos) == ref_base )
+ continue;
-/*
- The following code is nearly identical to bam_maqcns_glfgen() under
- the simplified SOAPsnp model. It does the following:
+ var_pos[alt_dp] = p->qpos;
+ if ( (bam1_cigar(p->b)[0]&BAM_CIGAR_MASK)==4 )
+ var_pos[alt_dp] -= bam1_cigar(p->b)[0]>>BAM_CIGAR_SHIFT;
- 1) Collect strand, base, quality and mapping quality information for
- each base and combine them in an integer:
+ alt_dp++;
+ read_len += p->b->core.l_qseq;
+ }
+ float mvd=0;
+ int j;
+ n=0;
+ for (i=0; i<alt_dp; i++) {
+ for (j=0; j<i; j++) {
+ mvd += abs(var_pos[i] - var_pos[j]);
+ n++;
+ }
+ }
+ r->mvd[0] = n ? mvd/n : 0;
+ r->mvd[1] = alt_dp;
+ r->mvd[2] = alt_dp ? read_len/alt_dp : 0;
- x = min{baseQ,mapQ}<<24 | 1<<21 | strand<<18 | base<<16 | baseQ<<8 | mapQ
+ return r->depth;
+}
- 2) Sort the list of integers for the next step.
- 3) For each base, calculate e_b, the sum of weighted qualities. For
- each type of base on each strand, the best quality has the highest
- weight. Only the top 255 bases on each strand are used (different
- from maqcns).
+void calc_vdb(int n, const bcf_callret1_t *calls, bcf_call_t *call)
+{
+ // Variant distance bias. Samples merged by means of DP-weighted average.
- 4) Rescale the total read depth to 255.
+ float weight=0, tot_prob=0;
- 5) Calculate Q(D|g) = -10\log_{10}P(D|g) (d_a is the allele count):
+ int i;
+ for (i=0; i<n; i++)
+ {
+ int mvd = calls[i].mvd[0];
+ int dp = calls[i].mvd[1];
+ int read_len = calls[i].mvd[2];
- Q(D|<aa>)=\sum_{b\not=a}e_b
- Q(D|<aA>)=3*(d_a+d_A)+\sum_{b\not=a,b\not=A}e_b
- */
-int bcf_call_glfgen(int _n, const bam_pileup1_t *pl, int ref_base /*4-bit*/, bcf_callaux_t *bca, bcf_callret1_t *r)
-{
- int i, j, k, c, n;
- float *p = r->p;
- auxaux_t aux;
+ if ( dp<2 ) continue;
- memset(r, 0, sizeof(bcf_callret1_t));
- if (_n == 0) return -1;
+ float prob = 0;
+ if ( dp==2 )
+ {
+ // Exact formula
+ prob = (mvd==0) ? 1.0/read_len : (read_len-mvd)*2.0/read_len/read_len;
+ }
+ else if ( dp==3 )
+ {
+ // Sin, quite accurate approximation
+ float mu = read_len/2.9;
+ prob = mvd>2*mu ? 0 : sin(mvd*3.14/2/mu) / (4*mu/3.14);
+ }
+ else
+ {
+ // Scaled gaussian curve, crude approximation, but behaves well. Using fixed depth for bigger depths.
+ if ( dp>5 )
+ dp = 5;
+ float sigma2 = (read_len/1.9/(dp+1)) * (read_len/1.9/(dp+1));
+ float norm = 1.125*sqrt(2*3.14*sigma2);
+ float mu = read_len/2.9;
+ if ( mvd < mu )
+ prob = exp(-(mvd-mu)*(mvd-mu)/2/sigma2)/norm;
+ else
+ prob = exp(-(mvd-mu)*(mvd-mu)/3.125/sigma2)/norm;
+ }
- // enlarge the aux array if necessary
- if (bca->max_info < _n) {
- bca->max_info = _n;
- kroundup32(bca->max_info);
- bca->info = (uint32_t*)realloc(bca->info, 4 * bca->max_info);
- }
- // fill the aux array
- for (i = n = 0; i < _n; ++i) {
- const bam_pileup1_t *p = pl + i;
- uint32_t q, x = 0, qq;
- if (p->is_del || (p->b->core.flag&BAM_FUNMAP)) continue; // skip unmapped reads and deleted bases
- q = (uint32_t)bam1_qual(p->b)[p->qpos]; // base quality
- x |= (uint32_t)bam1_strand(p->b) << 18 | q << 8 | p->b->core.qual;
- if (p->b->core.qual < q) q = p->b->core.qual; // cap the overall quality at mapping quality
- x |= q << 24;
- qq = bam1_seqi(bam1_seq(p->b), p->qpos); // base
- q = bam_nt16_nt4_table[qq? qq : ref_base]; // q is the 2-bit base
- if (q < 4) x |= 1 << 21 | q << 16;
-
- bca->info[n++] = x;
- }
- ks_introsort_uint32_t(n, bca->info);
- r->depth = n;
- // generate esum and fsum
- memset(&aux, 0, sizeof(auxaux_t));
- for (j = n - 1, r->sum_Q2 = 0; j >= 0; --j) { // calculate esum and fsum
- uint32_t info = bca->info[j];
- int tmp;
- if (info>>24 < 4 && (info>>8&0x3f) != 0) info = 4<<24 | (info&0xffffff);
- k = info>>16&7;
- if (info>>24 > 0) {
- aux.esum[k&3] += bca->fk[aux.w[k]] * (info>>24);
- aux.fsum[k&3] += bca->fk[aux.w[k]];
- if (aux.w[k] + 1 < CALL_MAX) ++aux.w[k];
- ++aux.c[k&3];
- }
- tmp = (int)(info&0xff) < bca->capQ? (int)(info&0xff) : bca->capQ;
- r->sum_Q2 += tmp * tmp;
- }
- memcpy(r->esum, aux.esum, 5 * sizeof(float));
- // rescale ->c[]
- for (j = c = 0; j != 4; ++j) c += aux.c[j];
- if (c > 255) {
- for (j = 0; j != 4; ++j) aux.c[j] = (int)(254.0 * aux.c[j] / c + 0.499);
- for (j = c = 0; j != 4; ++j) c += aux.c[j];
- }
- // generate likelihood
- for (j = 0; j != 5; ++j) {
- float tmp;
- // homozygous
- for (k = 0, tmp = 0.0; k != 5; ++k)
- if (j != k) tmp += aux.esum[k];
- p[j*5+j] = tmp; // anything that is not j
- // heterozygous
- for (k = j + 1; k < 5; ++k) {
- for (i = 0, tmp = 0.0; i != 5; ++i)
- if (i != j && i != k) tmp += aux.esum[i];
- p[j*5+k] = p[k*5+j] = 3.01 * (aux.c[j] + aux.c[k]) + tmp;
- }
- }
- return 0;
+ //fprintf(stderr,"dp=%d mvd=%d read_len=%d -> prob=%f\n", dp,mvd,read_len,prob);
+ tot_prob += prob*dp;
+ weight += dp;
+ }
+ tot_prob = weight ? tot_prob/weight : 1;
+ //fprintf(stderr,"prob=%f\n", tot_prob);
+ call->vdb = tot_prob;
}
-/*
- 1) Find the top 2 bases (from esum[4]).
-
- 2) If the reference base is among the top 2, consider the locus is
- potentially biallelic and set call->a[2] as -1; otherwise, the
- locus is potentially triallelic. If the reference is ambiguous,
- take the weakest call as the pseudo-reference.
- */
int bcf_call_combine(int n, const bcf_callret1_t *calls, int ref_base /*4-bit*/, bcf_call_t *call)
{
- int ref4, i, j;
- int64_t sum[5], tmp;
- call->ori_ref = ref4 = bam_nt16_nt4_table[ref_base];
- if (ref4 > 4) ref4 = 4;
- { // calculate esum
- double esum[5];
- memset(esum, 0, sizeof(double) * 4);
- for (i = 0; i < n; ++i) {
- for (j = 0; j < 4; ++j)
- esum[j] += calls[i].esum[j];
- }
+ int ref4, i, j, qsum[4];
+ int64_t tmp;
+ if (ref_base >= 0) {
+ call->ori_ref = ref4 = bam_nt16_nt4_table[ref_base];
+ if (ref4 > 4) ref4 = 4;
+ } else call->ori_ref = -1, ref4 = 0;
+ // calculate qsum
+ memset(qsum, 0, 4 * sizeof(int));
+ for (i = 0; i < n; ++i)
for (j = 0; j < 4; ++j)
- sum[j] = (int)(esum[j] * 100. + .499) << 2 | j;
- }
+ qsum[j] += calls[i].qsum[j];
+ int qsum_tot=0;
+ for (j=0; j<4; j++) { qsum_tot += qsum[j]; call->qsum[j] = 0; }
+ for (j = 0; j < 4; ++j) qsum[j] = qsum[j] << 2 | j;
// find the top 2 alleles
for (i = 1; i < 4; ++i) // insertion sort
- for (j = i; j > 0 && sum[j] < sum[j-1]; --j)
- tmp = sum[j], sum[j] = sum[j-1], sum[j-1] = tmp;
+ for (j = i; j > 0 && qsum[j] < qsum[j-1]; --j)
+ tmp = qsum[j], qsum[j] = qsum[j-1], qsum[j-1] = tmp;
// set the reference allele and alternative allele(s)
for (i = 0; i < 5; ++i) call->a[i] = -1;
call->unseen = -1;
call->a[0] = ref4;
for (i = 3, j = 1; i >= 0; --i) {
- if ((sum[i]&3) != ref4) {
- if (sum[i]>>2 != 0) call->a[j++] = sum[i]&3;
+ if ((qsum[i]&3) != ref4) {
+ if (qsum[i]>>2 != 0)
+ {
+ if ( j<4 ) call->qsum[j] = (float)(qsum[i]>>2)/qsum_tot; // ref N can make j>=4
+ call->a[j++] = qsum[i]&3;
+ }
else break;
}
+ else
+ call->qsum[0] = (float)(qsum[i]>>2)/qsum_tot;
+ }
+ if (ref_base >= 0) { // for SNPs, find the "unseen" base
+ if (((ref4 < 4 && j < 4) || (ref4 == 4 && j < 5)) && i >= 0)
+ call->unseen = j, call->a[j++] = qsum[i]&3;
+ call->n_alleles = j;
+ } else {
+ call->n_alleles = j;
+ if (call->n_alleles == 1) return -1; // no reliable supporting read. stop doing anything
}
- if (((ref4 < 4 && j < 4) || (ref4 == 4 && j < 5)) && i >= 0)
- call->unseen = j, call->a[j++] = sum[i]&3;
- call->n_alleles = j;
// set the PL array
if (call->n < n) {
call->n = n;
x = call->n_alleles * (call->n_alleles + 1) / 2;
// get the possible genotypes
for (i = z = 0; i < call->n_alleles; ++i)
- for (j = i; j < call->n_alleles; ++j)
- g[z++] = call->a[i] * 5 + call->a[j];
+ for (j = 0; j <= i; ++j)
+ g[z++] = call->a[j] * 5 + call->a[i];
for (i = 0; i < n; ++i) {
uint8_t *PL = call->PL + x * i;
const bcf_callret1_t *r = calls + i;
PL[j] = y;
}
}
+// if (ref_base < 0) fprintf(stderr, "%d,%d,%f,%d\n", call->n_alleles, x, sum_min, call->unseen);
call->shift = (int)(sum_min + .499);
}
- for (i = call->depth = 0, tmp = 0; i < n; ++i) {
+ // combine annotations
+ memset(call->anno, 0, 16 * sizeof(int));
+ for (i = call->depth = call->ori_depth = 0, tmp = 0; i < n; ++i) {
call->depth += calls[i].depth;
- tmp += calls[i].sum_Q2;
+ call->ori_depth += calls[i].ori_depth;
+ for (j = 0; j < 16; ++j) call->anno[j] += calls[i].anno[j];
}
- call->rmsQ = (int)(sqrt((double)tmp / call->depth) + .499);
+
+ calc_vdb(n, calls, call);
+
return 0;
}
-int bcf_call2bcf(int tid, int pos, bcf_call_t *bc, bcf1_t *b)
+int bcf_call2bcf(int tid, int pos, bcf_call_t *bc, bcf1_t *b, bcf_callret1_t *bcr, int fmt_flag,
+ const bcf_callaux_t *bca, const char *ref)
{
+ extern double kt_fisher_exact(int n11, int n12, int n21, int n22, double *_left, double *_right, double *two);
kstring_t s;
- int i;
+ int i, j;
+ b->n_smpl = bc->n;
b->tid = tid; b->pos = pos; b->qual = 0;
s.s = b->str; s.m = b->m_str; s.l = 0;
kputc('\0', &s);
- kputc("ACGTN"[bc->ori_ref], &s); kputc('\0', &s);
- for (i = 1; i < 5; ++i) {
- if (bc->a[i] < 0) break;
- if (i > 1) kputc(',', &s);
-// kputc(bc->unseen == i && i != 3? 'X' : "ACGT"[bc->a[i]], &s);
- kputc(bc->unseen == i? 'X' : "ACGT"[bc->a[i]], &s);
+ if (bc->ori_ref < 0) { // an indel
+ // write REF
+ kputc(ref[pos], &s);
+ for (j = 0; j < bca->indelreg; ++j) kputc(ref[pos+1+j], &s);
+ kputc('\0', &s);
+ // write ALT
+ kputc(ref[pos], &s);
+ for (i = 1; i < 4; ++i) {
+ if (bc->a[i] < 0) break;
+ if (i > 1) {
+ kputc(',', &s); kputc(ref[pos], &s);
+ }
+ if (bca->indel_types[bc->a[i]] < 0) { // deletion
+ for (j = -bca->indel_types[bc->a[i]]; j < bca->indelreg; ++j)
+ kputc(ref[pos+1+j], &s);
+ } else { // insertion; cannot be a reference unless a bug
+ char *inscns = &bca->inscns[bc->a[i] * bca->maxins];
+ for (j = 0; j < bca->indel_types[bc->a[i]]; ++j)
+ kputc("ACGTN"[(int)inscns[j]], &s);
+ for (j = 0; j < bca->indelreg; ++j) kputc(ref[pos+1+j], &s);
+ }
+ }
+ kputc('\0', &s);
+ } else { // a SNP
+ kputc("ACGTN"[bc->ori_ref], &s); kputc('\0', &s);
+ for (i = 1; i < 5; ++i) {
+ if (bc->a[i] < 0) break;
+ if (i > 1) kputc(',', &s);
+ kputc(bc->unseen == i? 'X' : "ACGT"[bc->a[i]], &s);
+ }
+ kputc('\0', &s);
}
kputc('\0', &s);
+ // INFO
+ if (bc->ori_ref < 0) ksprintf(&s,"INDEL;IS=%d,%f;", bca->max_support, bca->max_frac);
+ kputs("DP=", &s); kputw(bc->ori_depth, &s); kputs(";I16=", &s);
+ for (i = 0; i < 16; ++i) {
+ if (i) kputc(',', &s);
+ kputw(bc->anno[i], &s);
+ }
+ ksprintf(&s,";QS=%f,%f,%f,%f", bc->qsum[0],bc->qsum[1],bc->qsum[2],bc->qsum[3]);
+ if (bc->vdb != 1)
+ ksprintf(&s, ";VDB=%.4f", bc->vdb);
+ kputc('\0', &s);
+ // FMT
+ kputs("PL", &s);
+ if (bcr && fmt_flag) {
+ if (fmt_flag & B2B_FMT_DP) kputs(":DP", &s);
+ if (fmt_flag & B2B_FMT_DV) kputs(":DV", &s);
+ if (fmt_flag & B2B_FMT_SP) kputs(":SP", &s);
+ }
kputc('\0', &s);
- kputs("MQ=", &s); kputw(bc->rmsQ, &s); kputs(";DP=", &s); kputw(bc->depth, &s); kputc('\0', &s);
- kputs("PL", &s); kputc('\0', &s);
b->m_str = s.m; b->str = s.s; b->l_str = s.l;
- bcf_sync(bc->n, b);
+ bcf_sync(b);
memcpy(b->gi[0].data, bc->PL, b->gi[0].len * bc->n);
+ if (bcr && fmt_flag) {
+ uint16_t *dp = (fmt_flag & B2B_FMT_DP)? b->gi[1].data : 0;
+ uint16_t *dv = (fmt_flag & B2B_FMT_DV)? b->gi[1 + ((fmt_flag & B2B_FMT_DP) != 0)].data : 0;
+ int32_t *sp = (fmt_flag & B2B_FMT_SP)? b->gi[1 + ((fmt_flag & B2B_FMT_DP) != 0) + ((fmt_flag & B2B_FMT_DV) != 0)].data : 0;
+ for (i = 0; i < bc->n; ++i) {
+ bcf_callret1_t *p = bcr + i;
+ if (dp) dp[i] = p->depth < 0xffff? p->depth : 0xffff;
+ if (dv) dv[i] = p->n_supp < 0xffff? p->n_supp : 0xffff;
+ if (sp) {
+ if (p->anno[0] + p->anno[1] < 2 || p->anno[2] + p->anno[3] < 2
+ || p->anno[0] + p->anno[2] < 2 || p->anno[1] + p->anno[3] < 2)
+ {
+ sp[i] = 0;
+ } else {
+ double left, right, two;
+ int x;
+ kt_fisher_exact(p->anno[0], p->anno[1], p->anno[2], p->anno[3], &left, &right, &two);
+ x = (int)(-4.343 * log(two) + .499);
+ if (x > 255) x = 255;
+ sp[i] = x;
+ }
+ }
+ }
+ }
return 0;
}
projects / samtools.git / blobdiff