7 #include "bcftools/bcf.h"
9 extern void ks_introsort_uint32_t(size_t n, uint32_t a[]);
11 #define CALL_ETA 0.03f
13 #define CALL_DEFTHETA 0.83f
18 bcf_callaux_t *bcf_call_init(double theta, int min_baseQ)
21 if (theta <= 0.) theta = CALL_DEFTHETA;
22 bca = calloc(1, sizeof(bcf_callaux_t));
24 bca->openQ = 40; bca->extQ = 20; bca->tandemQ = 100;
25 bca->min_baseQ = min_baseQ;
26 bca->e = errmod_init(1. - theta);
27 bca->min_frac = 0.002;
29 bca->per_sample_flt = 0;
33 void bcf_call_destroy(bcf_callaux_t *bca)
36 errmod_destroy(bca->e);
37 free(bca->bases); free(bca->inscns); free(bca);
39 /* ref_base is the 4-bit representation of the reference base. It is
40 * negative if we are looking at an indel. */
41 int bcf_call_glfgen(int _n, const bam_pileup1_t *pl, int ref_base, bcf_callaux_t *bca, bcf_callret1_t *r)
43 static int *var_pos = NULL, nvar_pos = 0;
44 int i, n, ref4, is_indel, ori_depth = 0;
45 memset(r, 0, sizeof(bcf_callret1_t));
47 ref4 = bam_nt16_nt4_table[ref_base];
49 } else ref4 = 4, is_indel = 1;
50 if (_n == 0) return -1;
51 // enlarge the bases array if necessary
52 if (bca->max_bases < _n) {
54 kroundup32(bca->max_bases);
55 bca->bases = (uint16_t*)realloc(bca->bases, 2 * bca->max_bases);
57 // fill the bases array
58 for (i = n = r->n_supp = 0; i < _n; ++i) {
59 const bam_pileup1_t *p = pl + i;
60 int q, b, mapQ, baseQ, is_diff, min_dist, seqQ;
62 if (p->is_del || p->is_refskip || (p->b->core.flag&BAM_FUNMAP)) continue;
64 baseQ = q = is_indel? p->aux&0xff : (int)bam1_qual(p->b)[p->qpos]; // base/indel quality
65 seqQ = is_indel? (p->aux>>8&0xff) : 99;
66 if (q < bca->min_baseQ) continue;
67 if (q > seqQ) q = seqQ;
68 mapQ = p->b->core.qual < 255? p->b->core.qual : DEF_MAPQ; // special case for mapQ==255
69 mapQ = mapQ < bca->capQ? mapQ : bca->capQ;
70 if (q > mapQ) q = mapQ;
74 b = bam1_seqi(bam1_seq(p->b), p->qpos); // base
75 b = bam_nt16_nt4_table[b? b : ref_base]; // b is the 2-bit base
76 is_diff = (ref4 < 4 && b == ref4)? 0 : 1;
81 if (is_diff) ++r->n_supp;
82 bca->bases[n++] = q<<5 | (int)bam1_strand(p->b)<<4 | b;
83 // collect annotations
84 if (b < 4) r->qsum[b] += q;
85 ++r->anno[0<<2|is_diff<<1|bam1_strand(p->b)];
86 min_dist = p->b->core.l_qseq - 1 - p->qpos;
87 if (min_dist > p->qpos) min_dist = p->qpos;
88 if (min_dist > CAP_DIST) min_dist = CAP_DIST;
89 r->anno[1<<2|is_diff<<1|0] += baseQ;
90 r->anno[1<<2|is_diff<<1|1] += baseQ * baseQ; // FIXME: signed int is not enough for thousands of samples
91 r->anno[2<<2|is_diff<<1|0] += mapQ;
92 r->anno[2<<2|is_diff<<1|1] += mapQ * mapQ; // FIXME: signed int is not enough for thousands of samples
93 r->anno[3<<2|is_diff<<1|0] += min_dist;
94 r->anno[3<<2|is_diff<<1|1] += min_dist * min_dist;
96 r->depth = n; r->ori_depth = ori_depth;
98 errmod_cal(bca->e, n, 5, bca->bases, r->p);
100 // Calculate the Variant Distance Bias (make it optional?)
101 if ( nvar_pos < _n ) {
103 var_pos = realloc(var_pos,sizeof(int)*nvar_pos);
105 int alt_dp=0, read_len=0;
106 for (i=0; i<_n; i++) {
107 const bam_pileup1_t *p = pl + i;
108 if ( bam1_seqi(bam1_seq(p->b),p->qpos) == ref_base )
111 var_pos[alt_dp] = p->qpos;
112 if ( (bam1_cigar(p->b)[0]&BAM_CIGAR_MASK)==4 )
113 var_pos[alt_dp] -= bam1_cigar(p->b)[0]>>BAM_CIGAR_SHIFT;
116 read_len += p->b->core.l_qseq;
121 for (i=0; i<alt_dp; i++) {
122 for (j=0; j<i; j++) {
123 mvd += abs(var_pos[i] - var_pos[j]);
127 r->mvd[0] = n ? mvd/n : 0;
129 r->mvd[2] = alt_dp ? read_len/alt_dp : 0;
135 void calc_vdb(int n, const bcf_callret1_t *calls, bcf_call_t *call)
137 // Variant distance bias. Samples merged by means of DP-weighted average.
139 float weight=0, tot_prob=0;
144 int mvd = calls[i].mvd[0];
145 int dp = calls[i].mvd[1];
146 int read_len = calls[i].mvd[2];
148 if ( dp<2 ) continue;
154 prob = (mvd==0) ? 1.0/read_len : (read_len-mvd)*2.0/read_len/read_len;
158 // Sin, quite accurate approximation
159 float mu = read_len/2.9;
160 prob = mvd>2*mu ? 0 : sin(mvd*3.14/2/mu) / (4*mu/3.14);
164 // Scaled gaussian curve, crude approximation, but behaves well. Using fixed depth for bigger depths.
167 float sigma2 = (read_len/1.9/(dp+1)) * (read_len/1.9/(dp+1));
168 float norm = 1.125*sqrt(2*3.14*sigma2);
169 float mu = read_len/2.9;
171 prob = exp(-(mvd-mu)*(mvd-mu)/2/sigma2)/norm;
173 prob = exp(-(mvd-mu)*(mvd-mu)/3.125/sigma2)/norm;
176 //fprintf(stderr,"dp=%d mvd=%d read_len=%d -> prob=%f\n", dp,mvd,read_len,prob);
180 tot_prob = weight ? tot_prob/weight : 1;
181 //fprintf(stderr,"prob=%f\n", tot_prob);
182 call->vdb = tot_prob;
185 int bcf_call_combine(int n, const bcf_callret1_t *calls, int ref_base /*4-bit*/, bcf_call_t *call)
187 int ref4, i, j, qsum[4];
190 call->ori_ref = ref4 = bam_nt16_nt4_table[ref_base];
191 if (ref4 > 4) ref4 = 4;
192 } else call->ori_ref = -1, ref4 = 0;
194 memset(qsum, 0, 4 * sizeof(int));
195 for (i = 0; i < n; ++i)
196 for (j = 0; j < 4; ++j)
197 qsum[j] += calls[i].qsum[j];
199 for (j=0; j<4; j++) { qsum_tot += qsum[j]; call->qsum[j] = 0; }
200 for (j = 0; j < 4; ++j) qsum[j] = qsum[j] << 2 | j;
201 // find the top 2 alleles
202 for (i = 1; i < 4; ++i) // insertion sort
203 for (j = i; j > 0 && qsum[j] < qsum[j-1]; --j)
204 tmp = qsum[j], qsum[j] = qsum[j-1], qsum[j-1] = tmp;
205 // set the reference allele and alternative allele(s)
206 for (i = 0; i < 5; ++i) call->a[i] = -1;
209 for (i = 3, j = 1; i >= 0; --i) {
210 if ((qsum[i]&3) != ref4) {
213 if ( j<4 ) call->qsum[j] = (float)(qsum[i]>>2)/qsum_tot; // ref N can make j>=4
214 call->a[j++] = qsum[i]&3;
219 call->qsum[0] = (float)(qsum[i]>>2)/qsum_tot;
221 if (ref_base >= 0) { // for SNPs, find the "unseen" base
222 if (((ref4 < 4 && j < 4) || (ref4 == 4 && j < 5)) && i >= 0)
223 call->unseen = j, call->a[j++] = qsum[i]&3;
227 if (call->n_alleles == 1) return -1; // no reliable supporting read. stop doing anything
232 call->PL = realloc(call->PL, 15 * n);
237 x = call->n_alleles * (call->n_alleles + 1) / 2;
238 // get the possible genotypes
239 for (i = z = 0; i < call->n_alleles; ++i)
240 for (j = 0; j <= i; ++j)
241 g[z++] = call->a[j] * 5 + call->a[i];
242 for (i = 0; i < n; ++i) {
243 uint8_t *PL = call->PL + x * i;
244 const bcf_callret1_t *r = calls + i;
246 for (j = 0; j < x; ++j)
247 if (min > r->p[g[j]]) min = r->p[g[j]];
249 for (j = 0; j < x; ++j) {
251 y = (int)(r->p[g[j]] - min + .499);
252 if (y > 255) y = 255;
256 // if (ref_base < 0) fprintf(stderr, "%d,%d,%f,%d\n", call->n_alleles, x, sum_min, call->unseen);
257 call->shift = (int)(sum_min + .499);
259 // combine annotations
260 memset(call->anno, 0, 16 * sizeof(int));
261 for (i = call->depth = call->ori_depth = 0, tmp = 0; i < n; ++i) {
262 call->depth += calls[i].depth;
263 call->ori_depth += calls[i].ori_depth;
264 for (j = 0; j < 16; ++j) call->anno[j] += calls[i].anno[j];
267 calc_vdb(n, calls, call);
272 int bcf_call2bcf(int tid, int pos, bcf_call_t *bc, bcf1_t *b, bcf_callret1_t *bcr, int fmt_flag,
273 const bcf_callaux_t *bca, const char *ref)
275 extern double kt_fisher_exact(int n11, int n12, int n21, int n22, double *_left, double *_right, double *two);
279 b->tid = tid; b->pos = pos; b->qual = 0;
280 s.s = b->str; s.m = b->m_str; s.l = 0;
282 if (bc->ori_ref < 0) { // an indel
285 for (j = 0; j < bca->indelreg; ++j) kputc(ref[pos+1+j], &s);
289 for (i = 1; i < 4; ++i) {
290 if (bc->a[i] < 0) break;
292 kputc(',', &s); kputc(ref[pos], &s);
294 if (bca->indel_types[bc->a[i]] < 0) { // deletion
295 for (j = -bca->indel_types[bc->a[i]]; j < bca->indelreg; ++j)
296 kputc(ref[pos+1+j], &s);
297 } else { // insertion; cannot be a reference unless a bug
298 char *inscns = &bca->inscns[bc->a[i] * bca->maxins];
299 for (j = 0; j < bca->indel_types[bc->a[i]]; ++j)
300 kputc("ACGTN"[(int)inscns[j]], &s);
301 for (j = 0; j < bca->indelreg; ++j) kputc(ref[pos+1+j], &s);
306 kputc("ACGTN"[bc->ori_ref], &s); kputc('\0', &s);
307 for (i = 1; i < 5; ++i) {
308 if (bc->a[i] < 0) break;
309 if (i > 1) kputc(',', &s);
310 kputc(bc->unseen == i? 'X' : "ACGT"[bc->a[i]], &s);
316 if (bc->ori_ref < 0) ksprintf(&s,"INDEL;IS=%d,%f;", bca->max_support, bca->max_frac);
317 kputs("DP=", &s); kputw(bc->ori_depth, &s); kputs(";I16=", &s);
318 for (i = 0; i < 16; ++i) {
319 if (i) kputc(',', &s);
320 kputw(bc->anno[i], &s);
322 ksprintf(&s,";QS=%f,%f,%f,%f", bc->qsum[0],bc->qsum[1],bc->qsum[2],bc->qsum[3]);
324 ksprintf(&s, ";VDB=%.4f", bc->vdb);
328 if (bcr && fmt_flag) {
329 if (fmt_flag & B2B_FMT_DP) kputs(":DP", &s);
330 if (fmt_flag & B2B_FMT_DV) kputs(":DV", &s);
331 if (fmt_flag & B2B_FMT_SP) kputs(":SP", &s);
334 b->m_str = s.m; b->str = s.s; b->l_str = s.l;
336 memcpy(b->gi[0].data, bc->PL, b->gi[0].len * bc->n);
337 if (bcr && fmt_flag) {
338 uint16_t *dp = (fmt_flag & B2B_FMT_DP)? b->gi[1].data : 0;
339 uint16_t *dv = (fmt_flag & B2B_FMT_DV)? b->gi[1 + ((fmt_flag & B2B_FMT_DP) != 0)].data : 0;
340 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;
341 for (i = 0; i < bc->n; ++i) {
342 bcf_callret1_t *p = bcr + i;
343 if (dp) dp[i] = p->depth < 0xffff? p->depth : 0xffff;
344 if (dv) dv[i] = p->n_supp < 0xffff? p->n_supp : 0xffff;
346 if (p->anno[0] + p->anno[1] < 2 || p->anno[2] + p->anno[3] < 2
347 || p->anno[0] + p->anno[2] < 2 || p->anno[1] + p->anno[3] < 2)
351 double left, right, two;
353 kt_fisher_exact(p->anno[0], p->anno[1], p->anno[2], p->anno[3], &left, &right, &two);
354 x = (int)(-4.343 * log(two) + .499);
355 if (x > 255) x = 255;