#include <string.h>
#include "bam.h"
#include "bam2bcf.h"
-#include "ksort.h"
#include "kaln.h"
+#include "kprobaln.h"
#include "khash.h"
KHASH_SET_INIT_STR(rg)
+#include "ksort.h"
+KSORT_INIT_GENERIC(uint32_t)
+
#define MINUS_CONST 0x10000000
#define INDEL_WINDOW_SIZE 50
for (k = 0; k < c->n_cigar; ++k) {
int op = cigar[k] & BAM_CIGAR_MASK;
int l = cigar[k] >> BAM_CIGAR_SHIFT;
- if (op == BAM_CMATCH) {
+ if (op == BAM_CMATCH || op == BAM_CEQUAL || op == BAM_CDIFF) {
if (c->pos > tpos) return y;
if (x + l > tpos) {
*_tpos = tpos;
int bcf_call_gap_prep(int n, int *n_plp, bam_pileup1_t **plp, int pos, bcf_callaux_t *bca, const char *ref,
const void *rghash)
{
- extern void ks_introsort_uint32_t(int, uint32_t*);
- int i, s, j, k, t, n_types, *types, max_rd_len, left, right, max_ins, *score, N, K, l_run, ref_type, n_alt;
- char *inscns = 0, *ref2, *query;
+ int i, s, j, k, t, n_types, *types, max_rd_len, left, right, max_ins, *score1, *score2, max_ref2;
+ int N, K, l_run, ref_type, n_alt;
+ char *inscns = 0, *ref2, *query, **ref_sample;
khash_t(rg) *hash = (khash_t(rg)*)rghash;
if (ref == 0 || bca == 0) return -1;
// mark filtered reads
if (s == n) return -1; // there is no indel at this position.
for (s = N = 0; s < n; ++s) N += n_plp[s]; // N is the total number of reads
{ // find out how many types of indels are present
- int m;
+ int m, n_alt = 0, n_tot = 0;
uint32_t *aux;
aux = calloc(N + 1, 4);
m = max_rd_len = 0;
for (s = 0; s < n; ++s) {
for (i = 0; i < n_plp[s]; ++i) {
const bam_pileup1_t *p = plp[s] + i;
- if (p->indel != 0 && (rghash == 0 || p->aux == 0))
- aux[m++] = MINUS_CONST + p->indel;
+ if (rghash == 0 || p->aux == 0) {
+ ++n_tot;
+ if (p->indel != 0) {
+ ++n_alt;
+ aux[m++] = MINUS_CONST + p->indel;
+ }
+ }
j = bam_cigar2qlen(&p->b->core, bam1_cigar(p->b));
if (j > max_rd_len) max_rd_len = j;
}
// squeeze out identical types
for (i = 1, n_types = 1; i < m; ++i)
if (aux[i] != aux[i-1]) ++n_types;
- if (n_types == 1) { // no indels
+ if (n_types == 1 || (double)n_alt / n_tot < bca->min_frac || n_alt < bca->min_support) { // then skip
free(aux); return -1;
}
+ if (n_types >= 64) {
+ free(aux);
+ if (bam_verbose >= 2)
+ fprintf(stderr, "[%s] excessive INDEL alleles at position %d. Skip the position.\n", __func__, pos + 1);
+ return -1;
+ }
types = (int*)calloc(n_types, sizeof(int));
t = 0;
types[t++] = aux[0] - MINUS_CONST;
for (t = 0; t < n_types; ++t)
if (types[t] == 0) break;
ref_type = t; // the index of the reference type (0)
- assert(n_types < 64);
}
{ // calculate left and right boundary
left = pos > INDEL_WINDOW_SIZE? pos - INDEL_WINDOW_SIZE : 0;
if (ref[i] == 0) break;
right = i;
}
+ /* The following block fixes a long-existing flaw in the INDEL
+ * calling model: the interference of nearby SNPs. However, it also
+ * reduces the power because sometimes, substitutions caused by
+ * indels are not distinguishable from true mutations. Multiple
+ * sequence realignment helps to increase the power.
+ */
+ { // construct per-sample consensus
+ int L = right - left + 1, max_i, max2_i;
+ uint32_t *cns, max, max2;
+ char *ref0, *r;
+ ref_sample = calloc(n, sizeof(void*));
+ cns = calloc(L, 4);
+ ref0 = calloc(L, 1);
+ for (i = 0; i < right - left; ++i)
+ ref0[i] = bam_nt16_table[(int)ref[i+left]];
+ for (s = 0; s < n; ++s) {
+ r = ref_sample[s] = calloc(L, 1);
+ memset(cns, 0, sizeof(int) * L);
+ // collect ref and non-ref counts
+ for (i = 0; i < n_plp[s]; ++i) {
+ bam_pileup1_t *p = plp[s] + i;
+ bam1_t *b = p->b;
+ uint32_t *cigar = bam1_cigar(b);
+ uint8_t *seq = bam1_seq(b);
+ int x = b->core.pos, y = 0;
+ for (k = 0; k < b->core.n_cigar; ++k) {
+ int op = cigar[k]&0xf;
+ int j, l = cigar[k]>>4;
+ if (op == BAM_CMATCH || op == BAM_CEQUAL || op == BAM_CDIFF) {
+ for (j = 0; j < l; ++j)
+ if (x + j >= left && x + j < right)
+ cns[x+j-left] += (bam1_seqi(seq, y+j) == ref0[x+j-left])? 1 : 0x10000;
+ x += l; y += l;
+ } else if (op == BAM_CDEL || op == BAM_CREF_SKIP) x += l;
+ else if (op == BAM_CINS || op == BAM_CSOFT_CLIP) y += l;
+ }
+ }
+ // determine the consensus
+ for (i = 0; i < right - left; ++i) r[i] = ref0[i];
+ max = max2 = 0; max_i = max2_i = -1;
+ for (i = 0; i < right - left; ++i) {
+ if (cns[i]>>16 >= max>>16) max2 = max, max2_i = max_i, max = cns[i], max_i = i;
+ else if (cns[i]>>16 >= max2>>16) max2 = cns[i], max2_i = i;
+ }
+ if ((double)(max&0xffff) / ((max&0xffff) + (max>>16)) >= 0.7) max_i = -1;
+ if ((double)(max2&0xffff) / ((max2&0xffff) + (max2>>16)) >= 0.7) max2_i = -1;
+ if (max_i >= 0) r[max_i] = 15;
+ if (max2_i >= 0) r[max2_i] = 15;
+// for (i = 0; i < right - left; ++i) fputc("=ACMGRSVTWYHKDBN"[(int)r[i]], stderr); fputc('\n', stderr);
+ }
+ free(ref0); free(cns);
+ }
{ // the length of the homopolymer run around the current position
int c = bam_nt16_table[(int)ref[pos + 1]];
if (c == 15) l_run = 1;
free(inscns_aux);
}
// compute the likelihood given each type of indel for each read
- ref2 = calloc(right - left + max_ins + 2, 1);
+ max_ref2 = right - left + 2 + 2 * (max_ins > -types[0]? max_ins : -types[0]);
+ ref2 = calloc(max_ref2, 1);
query = calloc(right - left + max_rd_len + max_ins + 2, 1);
- score = calloc(N * n_types, sizeof(int));
+ score1 = calloc(N * n_types, sizeof(int));
+ score2 = calloc(N * n_types, sizeof(int));
bca->indelreg = 0;
for (t = 0; t < n_types; ++t) {
int l, ir;
- ka_param2_t ap = ka_param2_qual;
- ap.band_width = abs(types[t]) + 3;
+ kpa_par_t apf1 = { 1e-4, 1e-2, 10 }, apf2 = { 1e-6, 1e-3, 10 };
+ apf1.bw = apf2.bw = abs(types[t]) + 3;
// compute indelreg
if (types[t] == 0) ir = 0;
else if (types[t] > 0) ir = est_indelreg(pos, ref, types[t], &inscns[t*max_ins]);
else ir = est_indelreg(pos, ref, -types[t], 0);
if (ir > bca->indelreg) bca->indelreg = ir;
// fprintf(stderr, "%d, %d, %d\n", pos, types[t], ir);
- // write ref2
- for (k = 0, j = left; j <= pos; ++j)
- ref2[k++] = bam_nt16_nt4_table[bam_nt16_table[(int)ref[j]]];
- if (types[t] <= 0) j += -types[t];
- else for (l = 0; l < types[t]; ++l)
- ref2[k++] = inscns[t*max_ins + l];
- if (types[0] < 0) { // mask deleted sequences to avoid a particular error in the model.
- int jj, tmp = types[t] >= 0? -types[0] : -types[0] + types[t];
- for (jj = 0; jj < tmp && j < right && ref[j]; ++jj, ++j)
- ref2[k++] = 4;
- }
- for (; j < right && ref[j]; ++j)
- ref2[k++] = bam_nt16_nt4_table[bam_nt16_table[(int)ref[j]]];
- if (j < right) right = j;
- // align each read to ref2
+ // realignment
for (s = K = 0; s < n; ++s) {
+ // write ref2
+ for (k = 0, j = left; j <= pos; ++j)
+ ref2[k++] = bam_nt16_nt4_table[(int)ref_sample[s][j-left]];
+ if (types[t] <= 0) j += -types[t];
+ else for (l = 0; l < types[t]; ++l)
+ ref2[k++] = inscns[t*max_ins + l];
+ for (; j < right && ref[j]; ++j)
+ ref2[k++] = bam_nt16_nt4_table[(int)ref_sample[s][j-left]];
+ for (; k < max_ref2; ++k) ref2[k] = 4;
+ if (j < right) right = j;
+ // align each read to ref2
for (i = 0; i < n_plp[s]; ++i, ++K) {
bam_pileup1_t *p = plp[s] + i;
- int qbeg, qend, tbeg, tend, sc;
+ int qbeg, qend, tbeg, tend, sc, kk;
uint8_t *seq = bam1_seq(p->b);
+ uint32_t *cigar = bam1_cigar(p->b);
+ if (p->b->core.flag&4) continue; // unmapped reads
+ // FIXME: the following loop should be better moved outside; nonetheless, realignment should be much slower anyway.
+ for (kk = 0; kk < p->b->core.n_cigar; ++kk)
+ if ((cigar[kk]&BAM_CIGAR_MASK) == BAM_CREF_SKIP) break;
+ if (kk < p->b->core.n_cigar) continue;
+ // FIXME: the following skips soft clips, but using them may be more sensitive.
// determine the start and end of sequences for alignment
qbeg = tpos2qpos(&p->b->core, bam1_cigar(p->b), left, 0, &tbeg);
qend = tpos2qpos(&p->b->core, bam1_cigar(p->b), right, 1, &tend);
// write the query sequence
for (l = qbeg; l < qend; ++l)
query[l - qbeg] = bam_nt16_nt4_table[bam1_seqi(seq, l)];
- // do alignment; this takes most of computing time for indel calling
- sc = ka_global_score((uint8_t*)ref2 + tbeg - left, tend - tbeg + abs(types[t]),
- (uint8_t*)query, qend - qbeg, &ap);
- score[K*n_types + t] = -sc;
+ { // do realignment; this is the bottleneck
+ const uint8_t *qual = bam1_qual(p->b), *bq;
+ uint8_t *qq;
+ qq = calloc(qend - qbeg, 1);
+ bq = (uint8_t*)bam_aux_get(p->b, "ZQ");
+ if (bq) ++bq; // skip type
+ for (l = qbeg; l < qend; ++l) {
+ qq[l - qbeg] = bq? qual[l] + (bq[l] - 64) : qual[l];
+ if (qq[l - qbeg] > 30) qq[l - qbeg] = 30;
+ if (qq[l - qbeg] < 7) qq[l - qbeg] = 7;
+ }
+ sc = kpa_glocal((uint8_t*)ref2 + tbeg - left, tend - tbeg + abs(types[t]),
+ (uint8_t*)query, qend - qbeg, qq, &apf1, 0, 0);
+ l = (int)(100. * sc / (qend - qbeg) + .499); // used for adjusting indelQ below
+ if (l > 255) l = 255;
+ score1[K*n_types + t] = score2[K*n_types + t] = sc<<8 | l;
+ if (sc > 5) {
+ sc = kpa_glocal((uint8_t*)ref2 + tbeg - left, tend - tbeg + abs(types[t]),
+ (uint8_t*)query, qend - qbeg, qq, &apf2, 0, 0);
+ l = (int)(100. * sc / (qend - qbeg) + .499);
+ if (l > 255) l = 255;
+ score2[K*n_types + t] = sc<<8 | l;
+ }
+ free(qq);
+ }
/*
for (l = 0; l < tend - tbeg + abs(types[t]); ++l)
fputc("ACGTN"[(int)ref2[tbeg-left+l]], stderr);
for (s = K = 0; s < n; ++s) {
for (i = 0; i < n_plp[s]; ++i, ++K) {
bam_pileup1_t *p = plp[s] + i;
- int *sct = &score[K*n_types], indelQ, seqQ;
+ int *sct = &score1[K*n_types], indelQ1, indelQ2, seqQ, indelQ;
for (t = 0; t < n_types; ++t) sc[t] = sct[t]<<6 | t;
for (t = 1; t < n_types; ++t) // insertion sort
for (j = t; j > 0 && sc[j] < sc[j-1]; --j)
* compromise for multi-allelic indels.
*/
if ((sc[0]&0x3f) == ref_type) {
- indelQ = (sc[1]>>6) - (sc[0]>>6);
+ indelQ1 = (sc[1]>>14) - (sc[0]>>14);
seqQ = est_seqQ(bca, types[sc[1]&0x3f], l_run);
} else {
for (t = 0; t < n_types; ++t) // look for the reference type
if ((sc[t]&0x3f) == ref_type) break;
- indelQ = (sc[t]>>6) - (sc[0]>>6);
+ indelQ1 = (sc[t]>>14) - (sc[0]>>14);
seqQ = est_seqQ(bca, types[sc[0]&0x3f], l_run);
}
- p->aux = (sc[0]&0x3f)<<16 | seqQ<<8 | indelQ;
+ tmp = sc[0]>>6 & 0xff;
+ indelQ1 = tmp > 111? 0 : (int)((1. - tmp/111.) * indelQ1 + .499); // reduce indelQ
+ sct = &score2[K*n_types];
+ for (t = 0; t < n_types; ++t) sc[t] = sct[t]<<6 | t;
+ for (t = 1; t < n_types; ++t) // insertion sort
+ for (j = t; j > 0 && sc[j] < sc[j-1]; --j)
+ tmp = sc[j], sc[j] = sc[j-1], sc[j-1] = tmp;
+ if ((sc[0]&0x3f) == ref_type) {
+ indelQ2 = (sc[1]>>14) - (sc[0]>>14);
+ } else {
+ for (t = 0; t < n_types; ++t) // look for the reference type
+ if ((sc[t]&0x3f) == ref_type) break;
+ indelQ2 = (sc[t]>>14) - (sc[0]>>14);
+ }
+ tmp = sc[0]>>6 & 0xff;
+ indelQ2 = tmp > 111? 0 : (int)((1. - tmp/111.) * indelQ2 + .499);
+ // pick the smaller between indelQ1 and indelQ2
+ indelQ = indelQ1 < indelQ2? indelQ1 : indelQ2;
+ if (indelQ > 255) indelQ = 255;
+ if (seqQ > 255) seqQ = 255;
+ p->aux = (sc[0]&0x3f)<<16 | seqQ<<8 | indelQ; // use 22 bits in total
sumq[sc[0]&0x3f] += indelQ < seqQ? indelQ : seqQ;
-// fprintf(stderr, "pos=%d read=%d:%d name=%s call=%d q=%d\n", pos, s, i, bam1_qname(p->b), types[sc[0]&0x3f], indelQ);
+// fprintf(stderr, "pos=%d read=%d:%d name=%s call=%d indelQ=%d seqQ=%d\n", pos, s, i, bam1_qname(p->b), types[sc[0]&0x3f], indelQ, seqQ);
}
}
// determine bca->indel_types[] and bca->inscns
for (t = 0; t < n_types; ++t)
sumq[t] = sumq[t]<<6 | t;
for (t = 1; t < n_types; ++t) // insertion sort
- for (j = t; j > 0 && sumq[j] < sumq[j-1]; --j)
+ for (j = t; j > 0 && sumq[j] > sumq[j-1]; --j)
tmp = sumq[j], sumq[j] = sumq[j-1], sumq[j-1] = tmp;
for (t = 0; t < n_types; ++t) // look for the reference type
if ((sumq[t]&0x3f) == ref_type) break;
if (x == bca->indel_types[j]) break;
p->aux = j<<16 | (j == 4? 0 : (p->aux&0xffff));
if ((p->aux>>16&0x3f) > 0) ++n_alt;
-// fprintf(stderr, "pos=%d read=%d:%d name=%s call=%d q=%d\n", pos, s, i, bam1_qname(p->b), p->aux>>16&63, p->aux&0xff);
+// fprintf(stderr, "X pos=%d read=%d:%d name=%s call=%d type=%d q=%d seqQ=%d\n", pos, s, i, bam1_qname(p->b), p->aux>>16&63, bca->indel_types[p->aux>>16&63], p->aux&0xff, p->aux>>8&0xff);
}
}
}
- free(score);
+ free(score1); free(score2);
// free
+ for (i = 0; i < n; ++i) free(ref_sample[i]);
+ free(ref_sample);
free(types); free(inscns);
return n_alt > 0? 0 : -1;
}