9 #include "bcftools/bcf.h"
11 extern void ks_introsort_uint32_t(size_t n, uint32_t a[]);
13 #define CALL_ETA 0.03f
15 #define CALL_DEFTHETA 0.83f
20 bcf_callaux_t *bcf_call_init(double theta, int min_baseQ)
23 if (theta <= 0.) theta = CALL_DEFTHETA;
24 bca = calloc(1, sizeof(bcf_callaux_t));
26 bca->openQ = 40; bca->extQ = 20; bca->tandemQ = 100;
27 bca->min_baseQ = min_baseQ;
28 bca->e = errmod_init(1. - theta);
29 bca->min_frac = 0.002;
34 void bcf_call_destroy(bcf_callaux_t *bca)
37 errmod_destroy(bca->e);
38 free(bca->bases); free(bca->inscns); free(bca);
40 /* ref_base is the 4-bit representation of the reference base. It is
41 * negative if we are looking at an indel. */
42 int bcf_call_glfgen(int _n, const bam_pileup1_t *pl, int ref_base, bcf_callaux_t *bca, bcf_callret1_t *r)
44 static int *var_pos = NULL, nvar_pos = 0;
45 int i, n, ref4, is_indel, ori_depth = 0;
46 memset(r, 0, sizeof(bcf_callret1_t));
48 ref4 = bam_nt16_nt4_table[ref_base];
50 } else ref4 = 4, is_indel = 1;
51 if (_n == 0) return -1;
52 // enlarge the bases array if necessary
53 if (bca->max_bases < _n) {
55 kroundup32(bca->max_bases);
56 bca->bases = (uint16_t*)realloc(bca->bases, 2 * bca->max_bases);
58 // fill the bases array
59 memset(r, 0, sizeof(bcf_callret1_t));
60 for (i = n = 0; i < _n; ++i) {
61 const bam_pileup1_t *p = pl + i;
62 int q, b, mapQ, baseQ, is_diff, min_dist, seqQ;
64 if (p->is_del || p->is_refskip || (p->b->core.flag&BAM_FUNMAP)) continue;
66 baseQ = q = is_indel? p->aux&0xff : (int)bam1_qual(p->b)[p->qpos]; // base/indel quality
67 seqQ = is_indel? (p->aux>>8&0xff) : 99;
68 if (q < bca->min_baseQ) continue;
69 if (q > seqQ) q = seqQ;
70 mapQ = p->b->core.qual < 255? p->b->core.qual : DEF_MAPQ; // special case for mapQ==255
71 mapQ = mapQ < bca->capQ? mapQ : bca->capQ;
72 if (q > mapQ) q = mapQ;
76 b = bam1_seqi(bam1_seq(p->b), p->qpos); // base
77 b = bam_nt16_nt4_table[b? b : ref_base]; // b is the 2-bit base
78 is_diff = (ref4 < 4 && b == ref4)? 0 : 1;
83 bca->bases[n++] = q<<5 | (int)bam1_strand(p->b)<<4 | b;
84 // collect annotations
85 if (b < 4) r->qsum[b] += q;
86 ++r->anno[0<<2|is_diff<<1|bam1_strand(p->b)];
87 min_dist = p->b->core.l_qseq - 1 - p->qpos;
88 if (min_dist > p->qpos) min_dist = p->qpos;
89 if (min_dist > CAP_DIST) min_dist = CAP_DIST;
90 r->anno[1<<2|is_diff<<1|0] += baseQ;
91 r->anno[1<<2|is_diff<<1|1] += baseQ * baseQ;
92 r->anno[2<<2|is_diff<<1|0] += mapQ;
93 r->anno[2<<2|is_diff<<1|1] += mapQ * mapQ;
94 r->anno[3<<2|is_diff<<1|0] += min_dist;
95 r->anno[3<<2|is_diff<<1|1] += min_dist * min_dist;
97 r->depth = n; r->ori_depth = ori_depth;
99 errmod_cal(bca->e, n, 5, bca->bases, r->p);
101 // Calculate the Variant Distance Bias (make it optional?)
102 if ( nvar_pos < _n ) {
104 var_pos = realloc(var_pos,sizeof(int)*nvar_pos);
106 int alt_dp=0, read_len=0;
107 for (i=0; i<_n; i++) {
108 const bam_pileup1_t *p = pl + i;
109 if ( bam1_seqi(bam1_seq(p->b),p->qpos) == ref_base )
112 var_pos[alt_dp] = p->qpos;
113 if ( (bam1_cigar(p->b)[0]&BAM_CIGAR_MASK)==4 )
114 var_pos[alt_dp] -= bam1_cigar(p->b)[0]>>BAM_CIGAR_SHIFT;
117 read_len += p->b->core.l_qseq;
122 for (i=0; i<alt_dp; i++) {
123 for (j=0; j<i; j++) {
124 mvd += abs(var_pos[i] - var_pos[j]);
128 r->mvd[0] = n ? mvd/n : 0;
130 r->mvd[2] = alt_dp ? read_len/alt_dp : 0;
136 void calc_vdb(int n, const bcf_callret1_t *calls, bcf_call_t *call)
138 // Variant distance bias. Samples merged by means of DP-weighted average.
140 float weight=0, tot_prob=0;
145 int mvd = calls[i].mvd[0];
146 int dp = calls[i].mvd[1];
147 int read_len = calls[i].mvd[2];
149 if ( dp<2 ) continue;
155 prob = (mvd==0) ? 1.0/read_len : (read_len-mvd)*2.0/read_len/read_len;
159 // Sin, quite accurate approximation
160 float mu = read_len/2.9;
161 prob = mvd>2*mu ? 0 : sin(mvd*3.14/2/mu) / (4*mu/3.14);
165 // Scaled gaussian curve, crude approximation, but behaves well. Using fixed depth for bigger depths.
168 float sigma2 = (read_len/1.9/(dp+1)) * (read_len/1.9/(dp+1));
169 float norm = 1.125*sqrt(2*3.14*sigma2);
170 float mu = read_len/2.9;
172 prob = exp(-(mvd-mu)*(mvd-mu)/2/sigma2)/norm;
174 prob = exp(-(mvd-mu)*(mvd-mu)/3.125/sigma2)/norm;
177 //fprintf(pysamerr,"dp=%d mvd=%d read_len=%d -> prob=%f\n", dp,mvd,read_len,prob);
181 tot_prob = weight ? tot_prob/weight : 1;
182 //fprintf(pysamerr,"prob=%f\n", tot_prob);
183 call->vdb = tot_prob;
186 int bcf_call_combine(int n, const bcf_callret1_t *calls, int ref_base /*4-bit*/, bcf_call_t *call)
188 int ref4, i, j, qsum[4];
191 call->ori_ref = ref4 = bam_nt16_nt4_table[ref_base];
192 if (ref4 > 4) ref4 = 4;
193 } else call->ori_ref = -1, ref4 = 0;
195 memset(qsum, 0, 4 * sizeof(int));
196 for (i = 0; i < n; ++i)
197 for (j = 0; j < 4; ++j)
198 qsum[j] += calls[i].qsum[j];
199 for (j = 0; j < 4; ++j) qsum[j] = qsum[j] << 2 | j;
200 // find the top 2 alleles
201 for (i = 1; i < 4; ++i) // insertion sort
202 for (j = i; j > 0 && qsum[j] < qsum[j-1]; --j)
203 tmp = qsum[j], qsum[j] = qsum[j-1], qsum[j-1] = tmp;
204 // set the reference allele and alternative allele(s)
205 for (i = 0; i < 5; ++i) call->a[i] = -1;
208 for (i = 3, j = 1; i >= 0; --i) {
209 if ((qsum[i]&3) != ref4) {
210 if (qsum[i]>>2 != 0) call->a[j++] = qsum[i]&3;
214 if (ref_base >= 0) { // for SNPs, find the "unseen" base
215 if (((ref4 < 4 && j < 4) || (ref4 == 4 && j < 5)) && i >= 0)
216 call->unseen = j, call->a[j++] = qsum[i]&3;
220 if (call->n_alleles == 1) return -1; // no reliable supporting read. stop doing anything
225 call->PL = realloc(call->PL, 15 * n);
230 x = call->n_alleles * (call->n_alleles + 1) / 2;
231 // get the possible genotypes
232 for (i = z = 0; i < call->n_alleles; ++i)
233 for (j = 0; j <= i; ++j)
234 g[z++] = call->a[j] * 5 + call->a[i];
235 for (i = 0; i < n; ++i) {
236 uint8_t *PL = call->PL + x * i;
237 const bcf_callret1_t *r = calls + i;
239 for (j = 0; j < x; ++j)
240 if (min > r->p[g[j]]) min = r->p[g[j]];
242 for (j = 0; j < x; ++j) {
244 y = (int)(r->p[g[j]] - min + .499);
245 if (y > 255) y = 255;
249 // if (ref_base < 0) fprintf(pysamerr, "%d,%d,%f,%d\n", call->n_alleles, x, sum_min, call->unseen);
250 call->shift = (int)(sum_min + .499);
252 // combine annotations
253 memset(call->anno, 0, 16 * sizeof(int));
254 for (i = call->depth = call->ori_depth = 0, tmp = 0; i < n; ++i) {
255 call->depth += calls[i].depth;
256 call->ori_depth += calls[i].ori_depth;
257 for (j = 0; j < 16; ++j) call->anno[j] += calls[i].anno[j];
260 calc_vdb(n, calls, call);
265 int bcf_call2bcf(int tid, int pos, bcf_call_t *bc, bcf1_t *b, bcf_callret1_t *bcr, int is_SP,
266 const bcf_callaux_t *bca, const char *ref)
268 extern double kt_fisher_exact(int n11, int n12, int n21, int n22, double *_left, double *_right, double *two);
272 b->tid = tid; b->pos = pos; b->qual = 0;
273 s.s = b->str; s.m = b->m_str; s.l = 0;
275 if (bc->ori_ref < 0) { // an indel
278 for (j = 0; j < bca->indelreg; ++j) kputc(ref[pos+1+j], &s);
282 for (i = 1; i < 4; ++i) {
283 if (bc->a[i] < 0) break;
285 kputc(',', &s); kputc(ref[pos], &s);
287 if (bca->indel_types[bc->a[i]] < 0) { // deletion
288 for (j = -bca->indel_types[bc->a[i]]; j < bca->indelreg; ++j)
289 kputc(ref[pos+1+j], &s);
290 } else { // insertion; cannot be a reference unless a bug
291 char *inscns = &bca->inscns[bc->a[i] * bca->maxins];
292 for (j = 0; j < bca->indel_types[bc->a[i]]; ++j)
293 kputc("ACGTN"[(int)inscns[j]], &s);
294 for (j = 0; j < bca->indelreg; ++j) kputc(ref[pos+1+j], &s);
299 kputc("ACGTN"[bc->ori_ref], &s); kputc('\0', &s);
300 for (i = 1; i < 5; ++i) {
301 if (bc->a[i] < 0) break;
302 if (i > 1) kputc(',', &s);
303 kputc(bc->unseen == i? 'X' : "ACGT"[bc->a[i]], &s);
309 if (bc->ori_ref < 0) kputs("INDEL;", &s);
310 kputs("DP=", &s); kputw(bc->ori_depth, &s); kputs(";I16=", &s);
311 for (i = 0; i < 16; ++i) {
312 if (i) kputc(',', &s);
313 kputw(bc->anno[i], &s);
317 ksprintf(&s, ";VDB=%.4f", bc->vdb);
324 if (is_SP) kputs(":SP", &s);
327 b->m_str = s.m; b->str = s.s; b->l_str = s.l;
329 memcpy(b->gi[0].data, bc->PL, b->gi[0].len * bc->n);
331 uint16_t *dp = (uint16_t*)b->gi[1].data;
332 int32_t *sp = is_SP? b->gi[2].data : 0;
333 for (i = 0; i < bc->n; ++i) {
334 bcf_callret1_t *p = bcr + i;
335 dp[i] = p->depth < 0xffff? p->depth : 0xffff;
337 if (p->anno[0] + p->anno[1] < 2 || p->anno[2] + p->anno[3] < 2
338 || p->anno[0] + p->anno[2] < 2 || p->anno[1] + p->anno[3] < 2)
342 double left, right, two;
344 kt_fisher_exact(p->anno[0], p->anno[1], p->anno[2], p->anno[3], &left, &right, &two);
345 x = (int)(-4.343 * log(two) + .499);
346 if (x > 255) x = 255;