#include <math.h>
#include <stdint.h>
+#include <assert.h>
#include "bam.h"
#include "kstring.h"
#include "bam2bcf.h"
bca->e = errmod_init(1. - theta);
bca->min_frac = 0.002;
bca->min_support = 1;
+ bca->per_sample_flt = 0;
return bca;
}
+
+int dist_from_end(const bam_pileup1_t *p)
+{
+ int icig, n_tot_bases = 0, iread = 0, edist = p->qpos + 1;
+ for (icig=0; icig<p->b->core.n_cigar; icig++)
+ {
+ // Conversion from uint32_t to MIDNSHP
+ // 0123456
+ // MIDNSHP
+ int cig = bam1_cigar(p->b)[icig] & BAM_CIGAR_MASK;
+ int ncig = bam1_cigar(p->b)[icig] >> BAM_CIGAR_SHIFT;
+ if ( cig==0 )
+ {
+ n_tot_bases += ncig;
+ iread += ncig;
+ }
+ else if ( cig==1 )
+ {
+ n_tot_bases += ncig;
+ iread += ncig;
+ }
+ else if ( cig==4 )
+ {
+ iread += ncig;
+ if ( iread<=p->qpos ) edist -= ncig;
+ }
+ }
+ // distance from either end
+ if ( edist > n_tot_bases/2. ) edist = n_tot_bases - edist + 1;
+ if ( edist<0 ) { fprintf(stderr,"uh: %d %d -> %d (%d)\n", p->qpos,n_tot_bases,edist,p->b->core.pos+1); exit(1); }
+ return edist;
+}
+
void bcf_call_destroy(bcf_callaux_t *bca)
{
if (bca == 0) return;
errmod_destroy(bca->e);
+ if (bca->npos) { free(bca->ref_pos); free(bca->alt_pos); bca->npos = 0; }
free(bca->bases); free(bca->inscns); free(bca);
}
/* ref_base is the 4-bit representation of the reference base. It is
bca->bases = (uint16_t*)realloc(bca->bases, 2 * bca->max_bases);
}
// fill the bases array
- memset(r, 0, sizeof(bcf_callret1_t));
- for (i = n = 0; i < _n; ++i) {
+ bca->read_len = 0;
+ 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
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;
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;
+ 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;
+ 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;
+
+ // collect read positions for ReadPosBias
+ int epos = dist_from_end(p);
+ int npos = p->b->core.l_qseq;
+ if ( bca->npos < npos )
+ {
+ bca->ref_pos = realloc(bca->ref_pos, sizeof(int)*npos);
+ bca->alt_pos = realloc(bca->alt_pos, sizeof(int)*npos);
+ int j;
+ for (j=bca->npos; j<npos; j++) bca->ref_pos[j] = 0;
+ for (j=bca->npos; j<npos; j++) bca->alt_pos[j] = 0;
+ bca->npos = npos;
+ }
+ if ( bam1_seqi(bam1_seq(p->b),p->qpos) == ref_base )
+ bca->ref_pos[epos]++;
+ else
+ bca->alt_pos[epos]++;
+ bca->read_len += npos;
+ //fprintf(stderr,"qpos=%d epos=%d fwd=%d var=%d len=%d %s\n", p->qpos, epos, p->b->core.flag&BAM_FREVERSE ? 0 : 1, bam1_seqi(bam1_seq(p->b),p->qpos) == ref_base ? 1 : 0, p->b->core.l_qseq, bam1_qname(p->b));
}
+ if ( n ) bca->read_len /= n;
r->depth = n; r->ori_depth = ori_depth;
// glfgen
errmod_cal(bca->e, n, 5, bca->bases, r->p);
return r->depth;
}
-int bcf_call_combine(int n, const bcf_callret1_t *calls, int ref_base /*4-bit*/, bcf_call_t *call)
+double mann_whitney_1947(int n, int m, int U)
+{
+ if (U<0) return 0;
+ if (n==0||m==0) return U==0 ? 1 : 0;
+ return (double)n/(n+m)*mann_whitney_1947(n-1,m,U-m) + (double)m/(n+m)*mann_whitney_1947(n,m-1,U);
+}
+
+void calc_ReadPosBias(bcf_callaux_t *bca, bcf_call_t *call)
+{
+ int i, nref = 0, nalt = 0;
+ unsigned long int U = 0;
+ for (i=0; i<bca->npos; i++)
+ {
+ nref += bca->ref_pos[i];
+ nalt += bca->alt_pos[i];
+ U += nref*bca->alt_pos[i];
+ bca->ref_pos[i] = 0;
+ bca->alt_pos[i] = 0;
+ }
+ if ( !nref || !nalt )
+ {
+ call->read_pos_bias = -1;
+ return;
+ }
+
+ if ( nref>=8 || nalt>=8 )
+ {
+ // normal approximation
+ double mean = ((double)nref*nalt+1.0)/2.0;
+ double var2 = (double)nref*nalt*(nref+nalt+1.0)/12.0;
+ double z = (U-mean)/sqrt(var2);
+ call->read_pos_bias = z;
+ //fprintf(stderr,"nref=%d nalt=%d U=%ld mean=%e var=%e zval=%e\n", nref,nalt,U,mean,sqrt(var2),call->read_pos_bias);
+ }
+ else
+ {
+ double p = mann_whitney_1947(nalt,nref,U);
+ // biased form claimed by GATK to behave better empirically
+ // double var2 = (1.0+1.0/(nref+nalt+1.0))*(double)nref*nalt*(nref+nalt+1.0)/12.0;
+ double var2 = (double)nref*nalt*(nref+nalt+1.0)/12.0;
+ double z;
+ if ( p >= 1./sqrt(var2*2*M_PI) ) z = 0; // equal to mean
+ else
+ {
+ if ( U >= nref*nalt/2. ) z = sqrt(-2*log(sqrt(var2*2*M_PI)*p));
+ else z = -sqrt(-2*log(sqrt(var2*2*M_PI)*p));
+ }
+ call->read_pos_bias = z;
+ //fprintf(stderr,"nref=%d nalt=%d U=%ld p=%e var2=%e zval=%e\n", nref,nalt,U, p,var2,call->read_pos_bias);
+ }
+}
+
+float mean_diff_to_prob(float mdiff, int dp, int readlen)
+{
+ if ( dp==2 )
+ {
+ if ( mdiff==0 )
+ return (2.0*readlen + 4.0*(readlen-1.0))/((float)readlen*readlen);
+ else
+ return 8.0*(readlen - 4.0*mdiff)/((float)readlen*readlen);
+ }
+
+ // This is crude empirical approximation and is not very accurate for
+ // shorter read lengths (<100bp). There certainly is a room for
+ // improvement.
+ const float mv[24][2] = { {0,0}, {0,0}, {0,0},
+ { 9.108, 4.934}, { 9.999, 3.991}, {10.273, 3.485}, {10.579, 3.160},
+ {10.828, 2.889}, {11.014, 2.703}, {11.028, 2.546}, {11.244, 2.391},
+ {11.231, 2.320}, {11.323, 2.138}, {11.403, 2.123}, {11.394, 1.994},
+ {11.451, 1.928}, {11.445, 1.862}, {11.516, 1.815}, {11.560, 1.761},
+ {11.544, 1.728}, {11.605, 1.674}, {11.592, 1.652}, {11.674, 1.613},
+ {11.641, 1.570} };
+
+ float m, v;
+ if ( dp>=24 )
+ {
+ m = (int)readlen/8.;
+ if (dp>100) dp = 100;
+ v = 1.476/(0.182*pow(dp,0.514));
+ v = v*(readlen/100.);
+ }
+ else
+ {
+ m = mv[dp][0];
+ v = mv[dp][1];
+ m = (int)m*readlen/100.;
+ v = v*readlen/100.;
+ v *= 1.2; // allow more variability
+ }
+ return 1.0/(v*sqrt(2*M_PI)) * exp(-0.5*((mdiff-m)/v)*((mdiff-m)/v));
+}
+
+void calc_vdb(bcf_callaux_t *bca, bcf_call_t *call)
+{
+ int i, dp = 0;
+ float mean_pos = 0, mean_diff = 0;
+ for (i=0; i<bca->npos; i++)
+ {
+ if ( !bca->alt_pos[i] ) continue;
+ dp += bca->alt_pos[i];
+ mean_pos += bca->alt_pos[i]*i;
+ }
+ if ( !dp )
+ {
+ call->vdb = -1;
+ return;
+ }
+ mean_pos /= dp;
+ for (i=0; i<bca->npos; i++)
+ {
+ if ( !bca->alt_pos[i] ) continue;
+ mean_diff += bca->alt_pos[i] * fabs(i - mean_pos);
+ }
+ mean_diff /= dp;
+ call->vdb = mean_diff_to_prob(mean_diff, dp, bca->read_len);
+}
+
+/**
+ * bcf_call_combine() - sets the PL array and VDB, RPB annotations, finds the top two alleles
+ * @n: number of samples
+ * @calls: each sample's calls
+ * @bca: auxiliary data structure for holding temporary values
+ * @ref_base: the reference base
+ * @call: filled with the annotations
+ */
+int bcf_call_combine(int n, const bcf_callret1_t *calls, bcf_callaux_t *bca, int ref_base /*4-bit*/, bcf_call_t *call)
{
int ref4, i, j, qsum[4];
int64_t tmp;
for (i = 0; i < n; ++i)
for (j = 0; j < 4; ++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
call->a[0] = ref4;
for (i = 3, j = 1; i >= 0; --i) {
if ((qsum[i]&3) != ref4) {
- if (qsum[i]>>2 != 0) call->a[j++] = qsum[i]&3;
+ 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->ori_depth += calls[i].ori_depth;
for (j = 0; j < 16; ++j) call->anno[j] += calls[i].anno[j];
}
+
+ calc_vdb(bca, call);
+ calc_ReadPosBias(bca, call);
+
return 0;
}
-int bcf_call2bcf(int tid, int pos, bcf_call_t *bc, bcf1_t *b, bcf_callret1_t *bcr, int is_SP,
+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);
}
kputc('\0', &s);
// INFO
- if (bc->ori_ref < 0) kputs("INDEL;", &s);
+ 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);
+ if (bc->read_pos_bias != -1 )
+ ksprintf(&s, ";RPB=%.4f", bc->read_pos_bias);
kputc('\0', &s);
// FMT
kputs("PL", &s);
- if (bcr) {
- kputs(":DP", &s);
- if (is_SP) kputs(":SP", &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);
b->m_str = s.m; b->str = s.s; b->l_str = s.l;
bcf_sync(b);
memcpy(b->gi[0].data, bc->PL, b->gi[0].len * bc->n);
- if (bcr) {
- uint16_t *dp = (uint16_t*)b->gi[1].data;
- int32_t *sp = is_SP? b->gi[2].data : 0;
+ 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;
- dp[i] = p->depth < 0xffff? p->depth : 0xffff;
- if (is_SP) {
+ 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)
{