- // 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
- memset(r, 0, sizeof(bcf_callret1_t));
- r->indelreg = 10000;
- for (i = n = 0; i < _n; ++i) {
- const bam_pileup1_t *p = pl + i;
- int q, b, mapQ, indelQ, is_diff, min_dist;
- if (p->is_del || p->is_refskip || (p->b->core.flag&BAM_FUNMAP)) continue;
- { // compute indel (base) quality
- // this can be made more efficient, but realignment is the bottleneck anyway
- int j, k, x, y, op, len = 0, max_left, max_rght, seqQ, indelreg;
- bam1_core_t *c = &p->b->core;
- uint32_t *cigar = bam1_cigar(p->b);
- uint8_t *qual = bam1_qual(p->b);
- for (k = y = 0, x = c->pos; k < c->n_cigar && y <= p->qpos; ++k) {
- op = cigar[k]&0xf;
- len = cigar[k]>>4;
- if (op == BAM_CMATCH) {
- if (pos > x && pos < x + len) break;
- x += len; y += len;
- } else if (op == BAM_CINS || op == BAM_CSOFT_CLIP) y += len;
- else if (op == BAM_CDEL || op == BAM_CREF_SKIP) x += len;
- }
- if (k == c->n_cigar) continue; // this actually should not happen
- max_left = max_rght = 0; indelreg = 0;
- if (pos == x + len - 1 && k+2 < c->n_cigar && ((cigar[k+1]&0xf) == BAM_CINS || (cigar[k+1]&0xf) == BAM_CDEL)
- && (cigar[k+2]&0xf) == BAM_CMATCH)
- {
- for (j = y; j < y + len; ++j)
- if (max_left < qual[j]) max_left = qual[j];
- if ((cigar[k+1]&0xf) == BAM_CINS) y += cigar[k+1]>>4;
- else x += cigar[k+1]>>4;
- op = cigar[k+2]&0xf; len = cigar[k+2]>>4;
- for (j = y; j < y + len; ++j) {
- if (max_rght < qual[j]) max_rght = qual[j];
- if (qual[j] > BAM2BCF_INDELREG_THRES && indelreg == 0)
- indelreg = j - y + 1;
- }
- } else {
- for (j = y; j <= p->qpos; ++j)
- if (max_left < qual[j]) max_left = qual[j];
- for (j = p->qpos + 1; j < y + len; ++j)
- if (max_rght < qual[j]) max_rght = qual[j];
-
- }
- indelQ = max_left < max_rght? max_left : max_rght;
- // estimate the sequencing error rate
- seqQ = bca->openQ;
- if (p->indel != 0) seqQ += bca->extQ * (abs(p->indel) - 1); // FIXME: better to model homopolymer
- if (p->indel != 0) { // a different model for tandem repeats
- uint8_t *seq = bam1_seq(p->b);
- int tandemQ, qb = bam1_seqi(seq, p->qpos), l;
- for (j = p->qpos + 1; j < c->l_qseq; ++j)
- if (qb != bam1_seqi(seq, j)) break;
- l = j;
- for (j = (int)p->qpos - 1; j >= 0; --j)
- if (qb != bam1_seqi(seq, j)) break;
- l = l - (j + 1);
- tandemQ = (int)((double)(abs(p->indel)) / l * bca->tandemQ + .499);
- if (seqQ > tandemQ) seqQ = tandemQ;
- }
-// fprintf(stderr, "%s\t%d\t%d\t%d\t%d\t%d\t%d\n", bam1_qname(p->b), pos+1, p->indel, indelQ, seqQ, max_left, max_rght);
- if (indelQ > seqQ) indelQ = seqQ;
- q = indelQ;
- }
- if (q < bca->min_baseQ) continue;
- mapQ = p->b->core.qual < bca->capQ? p->b->core.qual : bca->capQ;
- if (q > mapQ) q = mapQ;
- if (q > 63) q = 63;
- if (q < 4) q = 4;
- b = p->indel? 1 : 0;
- bca->bases[n++] = q<<5 | (int)bam1_strand(p->b)<<4 | b;
- // collect annotations
- r->qsum[b] += q;
- is_diff = b;
- ++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] += indelQ;
- r->anno[1<<2|is_diff<<1|1] += indelQ * indelQ;
- r->anno[2<<2|is_diff<<1|0] += mapQ;
- r->anno[2<<2|is_diff<<1|1] += mapQ * mapQ;
- 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;
- // glfgen
- errmod_cal(bca->e, n, 2, bca->bases, r->p);
- return r->depth;
+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 0
+//todo
+ double var = 0, avg = (double)(nref+nalt)/bca->npos;
+ for (i=0; i<bca->npos; i++)
+ {
+ double ediff = bca->ref_pos[i] + bca->alt_pos[i] - avg;
+ var += ediff*ediff;
+ bca->ref_pos[i] = 0;
+ bca->alt_pos[i] = 0;
+ }
+ call->read_pos.avg = avg;
+ call->read_pos.var = sqrt(var/bca->npos);
+ call->read_pos.dp = nref+nalt;
+#endif
+ 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 = 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 = 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));