Compress binary packages with xz.
[samtools.git] / bam2bcf.c
1 #include <math.h>
2 #include <stdint.h>
3 #include "bam.h"
4 #include "kstring.h"
5 #include "bam2bcf.h"
6 #include "errmod.h"
7 #include "bcftools/bcf.h"
8
9 extern  void ks_introsort_uint32_t(size_t n, uint32_t a[]);
10
11 #define CALL_ETA 0.03f
12 #define CALL_MAX 256
13 #define CALL_DEFTHETA 0.83f
14 #define DEF_MAPQ 20
15
16 #define CAP_DIST 25
17
18 bcf_callaux_t *bcf_call_init(double theta, int min_baseQ)
19 {
20         bcf_callaux_t *bca;
21         if (theta <= 0.) theta = CALL_DEFTHETA;
22         bca = calloc(1, sizeof(bcf_callaux_t));
23         bca->capQ = 60;
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;
28         bca->min_support = 1;
29         return bca;
30 }
31
32 void bcf_call_destroy(bcf_callaux_t *bca)
33 {
34         if (bca == 0) return;
35         errmod_destroy(bca->e);
36         free(bca->bases); free(bca->inscns); free(bca);
37 }
38 /* ref_base is the 4-bit representation of the reference base. It is
39  * negative if we are looking at an indel. */
40 int bcf_call_glfgen(int _n, const bam_pileup1_t *pl, int ref_base, bcf_callaux_t *bca, bcf_callret1_t *r)
41 {
42     static int *var_pos = NULL, nvar_pos = 0;
43         int i, n, ref4, is_indel, ori_depth = 0;
44         memset(r, 0, sizeof(bcf_callret1_t));
45         if (ref_base >= 0) {
46                 ref4 = bam_nt16_nt4_table[ref_base];
47                 is_indel = 0;
48         } else ref4 = 4, is_indel = 1;
49         if (_n == 0) return -1;
50         // enlarge the bases array if necessary
51         if (bca->max_bases < _n) {
52                 bca->max_bases = _n;
53                 kroundup32(bca->max_bases);
54                 bca->bases = (uint16_t*)realloc(bca->bases, 2 * bca->max_bases);
55         }
56         // fill the bases array
57         memset(r, 0, sizeof(bcf_callret1_t));
58         for (i = n = 0; i < _n; ++i) {
59                 const bam_pileup1_t *p = pl + i;
60                 int q, b, mapQ, baseQ, is_diff, min_dist, seqQ;
61                 // set base
62                 if (p->is_del || p->is_refskip || (p->b->core.flag&BAM_FUNMAP)) continue;
63                 ++ori_depth;
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;
71                 if (q > 63) q = 63;
72                 if (q < 4) q = 4;
73                 if (!is_indel) {
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;
77                 } else {
78                         b = p->aux>>16&0x3f;
79                         is_diff = (b != 0);
80                 }
81                 bca->bases[n++] = q<<5 | (int)bam1_strand(p->b)<<4 | b;
82                 // collect annotations
83                 if (b < 4) r->qsum[b] += q;
84                 ++r->anno[0<<2|is_diff<<1|bam1_strand(p->b)];
85                 min_dist = p->b->core.l_qseq - 1 - p->qpos;
86                 if (min_dist > p->qpos) min_dist = p->qpos;
87                 if (min_dist > CAP_DIST) min_dist = CAP_DIST;
88                 r->anno[1<<2|is_diff<<1|0] += baseQ;
89                 r->anno[1<<2|is_diff<<1|1] += baseQ * baseQ;
90                 r->anno[2<<2|is_diff<<1|0] += mapQ;
91                 r->anno[2<<2|is_diff<<1|1] += mapQ * mapQ;
92                 r->anno[3<<2|is_diff<<1|0] += min_dist;
93                 r->anno[3<<2|is_diff<<1|1] += min_dist * min_dist;
94         }
95         r->depth = n; r->ori_depth = ori_depth;
96         // glfgen
97         errmod_cal(bca->e, n, 5, bca->bases, r->p);
98
99     // Calculate the Variant Distance Bias (make it optional?)
100     if ( nvar_pos < _n ) {
101         nvar_pos = _n;
102         var_pos = realloc(var_pos,sizeof(int)*nvar_pos);
103     }
104     int alt_dp=0, read_len=0;
105     for (i=0; i<_n; i++) {
106         const bam_pileup1_t *p = pl + i;
107         if ( bam1_seqi(bam1_seq(p->b),p->qpos) == ref_base ) 
108             continue;
109
110         var_pos[alt_dp] = p->qpos;
111         if ( (bam1_cigar(p->b)[0]&BAM_CIGAR_MASK)==4 )
112             var_pos[alt_dp] -= bam1_cigar(p->b)[0]>>BAM_CIGAR_SHIFT;
113
114         alt_dp++;
115         read_len += p->b->core.l_qseq;
116     }
117     float mvd=0;
118     int j;
119     n=0;
120     for (i=0; i<alt_dp; i++) {
121         for (j=0; j<i; j++) {
122             mvd += abs(var_pos[i] - var_pos[j]);
123             n++;
124         }
125     }
126     r->mvd[0] = n ? mvd/n : 0;
127     r->mvd[1] = alt_dp;
128     r->mvd[2] = alt_dp ? read_len/alt_dp : 0;
129
130         return r->depth;
131 }
132
133
134 void calc_vdb(int n, const bcf_callret1_t *calls, bcf_call_t *call)
135 {
136     // Variant distance bias. Samples merged by means of DP-weighted average.
137
138     float weight=0, tot_prob=0;
139
140     int i;
141     for (i=0; i<n; i++)
142     {
143         int mvd      = calls[i].mvd[0];
144         int dp       = calls[i].mvd[1];
145         int read_len = calls[i].mvd[2];
146
147         if ( dp<2 ) continue;
148
149         float prob = 0;
150         if ( dp==2 )
151         {
152             // Exact formula
153             prob = (mvd==0) ? 1.0/read_len : (read_len-mvd)*2.0/read_len/read_len;
154         }
155         else if ( dp==3 )
156         {
157             // Sin, quite accurate approximation
158             float mu = read_len/2.9;
159             prob = mvd>2*mu ? 0 : sin(mvd*3.14/2/mu) / (4*mu/3.14);
160         }
161         else
162         {
163             // Scaled gaussian curve, crude approximation, but behaves well. Using fixed depth for bigger depths.
164             if ( dp>5 )
165                 dp = 5;
166             float sigma2 = (read_len/1.9/(dp+1)) * (read_len/1.9/(dp+1));
167             float norm   = 1.125*sqrt(2*3.14*sigma2);
168             float mu     = read_len/2.9;
169             if ( mvd < mu )
170                 prob = exp(-(mvd-mu)*(mvd-mu)/2/sigma2)/norm;
171             else
172                 prob = exp(-(mvd-mu)*(mvd-mu)/3.125/sigma2)/norm;
173         }
174
175         //fprintf(stderr,"dp=%d mvd=%d read_len=%d -> prob=%f\n", dp,mvd,read_len,prob);
176         tot_prob += prob*dp;
177         weight += dp;
178     }
179     tot_prob = weight ? tot_prob/weight : 1; 
180     //fprintf(stderr,"prob=%f\n", tot_prob);
181     call->vdb = tot_prob;
182 }
183
184 int bcf_call_combine(int n, const bcf_callret1_t *calls, int ref_base /*4-bit*/, bcf_call_t *call)
185 {
186         int ref4, i, j, qsum[4];
187         int64_t tmp;
188         if (ref_base >= 0) {
189                 call->ori_ref = ref4 = bam_nt16_nt4_table[ref_base];
190                 if (ref4 > 4) ref4 = 4;
191         } else call->ori_ref = -1, ref4 = 0;
192         // calculate qsum
193         memset(qsum, 0, 4 * sizeof(int));
194         for (i = 0; i < n; ++i)
195                 for (j = 0; j < 4; ++j)
196                         qsum[j] += calls[i].qsum[j];
197         for (j = 0; j < 4; ++j) qsum[j] = qsum[j] << 2 | j;
198         // find the top 2 alleles
199         for (i = 1; i < 4; ++i) // insertion sort
200                 for (j = i; j > 0 && qsum[j] < qsum[j-1]; --j)
201                         tmp = qsum[j], qsum[j] = qsum[j-1], qsum[j-1] = tmp;
202         // set the reference allele and alternative allele(s)
203         for (i = 0; i < 5; ++i) call->a[i] = -1;
204         call->unseen = -1;
205         call->a[0] = ref4;
206         for (i = 3, j = 1; i >= 0; --i) {
207                 if ((qsum[i]&3) != ref4) {
208                         if (qsum[i]>>2 != 0) call->a[j++] = qsum[i]&3;
209                         else break;
210                 }
211         }
212         if (ref_base >= 0) { // for SNPs, find the "unseen" base
213                 if (((ref4 < 4 && j < 4) || (ref4 == 4 && j < 5)) && i >= 0)
214                         call->unseen = j, call->a[j++] = qsum[i]&3;
215                 call->n_alleles = j;
216         } else {
217                 call->n_alleles = j;
218                 if (call->n_alleles == 1) return -1; // no reliable supporting read. stop doing anything
219         }
220         // set the PL array
221         if (call->n < n) {
222                 call->n = n;
223                 call->PL = realloc(call->PL, 15 * n);
224         }
225         {
226                 int x, g[15], z;
227                 double sum_min = 0.;
228                 x = call->n_alleles * (call->n_alleles + 1) / 2;
229                 // get the possible genotypes
230                 for (i = z = 0; i < call->n_alleles; ++i)
231                         for (j = 0; j <= i; ++j)
232                                 g[z++] = call->a[j] * 5 + call->a[i];
233                 for (i = 0; i < n; ++i) {
234                         uint8_t *PL = call->PL + x * i;
235                         const bcf_callret1_t *r = calls + i;
236                         float min = 1e37;
237                         for (j = 0; j < x; ++j)
238                                 if (min > r->p[g[j]]) min = r->p[g[j]];
239                         sum_min += min;
240                         for (j = 0; j < x; ++j) {
241                                 int y;
242                                 y = (int)(r->p[g[j]] - min + .499);
243                                 if (y > 255) y = 255;
244                                 PL[j] = y;
245                         }
246                 }
247 //              if (ref_base < 0) fprintf(stderr, "%d,%d,%f,%d\n", call->n_alleles, x, sum_min, call->unseen);
248                 call->shift = (int)(sum_min + .499);
249         }
250         // combine annotations
251         memset(call->anno, 0, 16 * sizeof(int));
252         for (i = call->depth = call->ori_depth = 0, tmp = 0; i < n; ++i) {
253                 call->depth += calls[i].depth;
254                 call->ori_depth += calls[i].ori_depth;
255                 for (j = 0; j < 16; ++j) call->anno[j] += calls[i].anno[j];
256         }
257
258     calc_vdb(n, calls, call);
259
260         return 0;
261 }
262
263 int bcf_call2bcf(int tid, int pos, bcf_call_t *bc, bcf1_t *b, bcf_callret1_t *bcr, int is_SP,
264                                  const bcf_callaux_t *bca, const char *ref)
265 {
266         extern double kt_fisher_exact(int n11, int n12, int n21, int n22, double *_left, double *_right, double *two);
267         kstring_t s;
268         int i, j;
269         b->n_smpl = bc->n;
270         b->tid = tid; b->pos = pos; b->qual = 0;
271         s.s = b->str; s.m = b->m_str; s.l = 0;
272         kputc('\0', &s);
273         if (bc->ori_ref < 0) { // an indel
274                 // write REF
275                 kputc(ref[pos], &s);
276                 for (j = 0; j < bca->indelreg; ++j) kputc(ref[pos+1+j], &s);
277                 kputc('\0', &s);
278                 // write ALT
279                 kputc(ref[pos], &s);
280                 for (i = 1; i < 4; ++i) {
281                         if (bc->a[i] < 0) break;
282                         if (i > 1) {
283                                 kputc(',', &s); kputc(ref[pos], &s);
284                         }
285                         if (bca->indel_types[bc->a[i]] < 0) { // deletion
286                                 for (j = -bca->indel_types[bc->a[i]]; j < bca->indelreg; ++j)
287                                         kputc(ref[pos+1+j], &s);
288                         } else { // insertion; cannot be a reference unless a bug
289                                 char *inscns = &bca->inscns[bc->a[i] * bca->maxins];
290                                 for (j = 0; j < bca->indel_types[bc->a[i]]; ++j)
291                                         kputc("ACGTN"[(int)inscns[j]], &s);
292                                 for (j = 0; j < bca->indelreg; ++j) kputc(ref[pos+1+j], &s);
293                         }
294                 }
295                 kputc('\0', &s);
296         } else { // a SNP
297                 kputc("ACGTN"[bc->ori_ref], &s); kputc('\0', &s);
298                 for (i = 1; i < 5; ++i) {
299                         if (bc->a[i] < 0) break;
300                         if (i > 1) kputc(',', &s);
301                         kputc(bc->unseen == i? 'X' : "ACGT"[bc->a[i]], &s);
302                 }
303                 kputc('\0', &s);
304         }
305         kputc('\0', &s);
306         // INFO
307         if (bc->ori_ref < 0) kputs("INDEL;", &s);
308         kputs("DP=", &s); kputw(bc->ori_depth, &s); kputs(";I16=", &s);
309         for (i = 0; i < 16; ++i) {
310                 if (i) kputc(',', &s);
311                 kputw(bc->anno[i], &s);
312         }
313     if ( bc->vdb!=1 )
314     {
315         ksprintf(&s, ";VDB=%.4f", bc->vdb);
316     }
317         kputc('\0', &s);
318         // FMT
319         kputs("PL", &s);
320         if (bcr) {
321                 kputs(":DP", &s);
322                 if (is_SP) kputs(":SP", &s);
323         }
324         kputc('\0', &s);
325         b->m_str = s.m; b->str = s.s; b->l_str = s.l;
326         bcf_sync(b);
327         memcpy(b->gi[0].data, bc->PL, b->gi[0].len * bc->n);
328         if (bcr) {
329                 uint16_t *dp = (uint16_t*)b->gi[1].data;
330                 int32_t *sp = is_SP? b->gi[2].data : 0;
331                 for (i = 0; i < bc->n; ++i) {
332                         bcf_callret1_t *p = bcr + i;
333                         dp[i] = p->depth < 0xffff? p->depth : 0xffff;
334                         if (is_SP) {
335                                 if (p->anno[0] + p->anno[1] < 2 || p->anno[2] + p->anno[3] < 2
336                                         || p->anno[0] + p->anno[2] < 2 || p->anno[1] + p->anno[3] < 2)
337                                 {
338                                         sp[i] = 0;
339                                 } else {
340                                         double left, right, two;
341                                         int x;
342                                         kt_fisher_exact(p->anno[0], p->anno[1], p->anno[2], p->anno[3], &left, &right, &two);
343                                         x = (int)(-4.343 * log(two) + .499);
344                                         if (x > 255) x = 255;
345                                         sp[i] = x;
346                                 }
347                         }
348                 }
349         }
350         return 0;
351 }