X-Git-Url: https://git.donarmstrong.com/?p=samtools.git;a=blobdiff_plain;f=bcftools%2Fprob1.c;h=e3d6b5eb42fee41bb3b2e8f392b5ae38d1934bc5;hp=85c2945779f85c9cd050fd72f00cdc78f1deeacb;hb=da45f88bac09817fcd87916d3e104f0e4bc53874;hpb=c219160a2ef3d7e057a86b1e50fa766cb89a5366 diff --git a/bcftools/prob1.c b/bcftools/prob1.c index 85c2945..e3d6b5e 100644 --- a/bcftools/prob1.c +++ b/bcftools/prob1.c @@ -3,14 +3,17 @@ #include #include #include +#include +#include #include "prob1.h" +#include "kstring.h" #include "kseq.h" KSTREAM_INIT(gzFile, gzread, 16384) -#define MC_AVG_ERR 0.007 #define MC_MAX_EM_ITER 16 -#define MC_EM_EPS 1e-4 +#define MC_EM_EPS 1e-5 +#define MC_DEF_INDEL 0.15 unsigned char seq_nt4_table[256] = { 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, @@ -32,17 +35,28 @@ unsigned char seq_nt4_table[256] = { }; struct __bcf_p1aux_t { - int n, M, n1; + int n, M, n1, is_indel; + uint8_t *ploidy; // haploid or diploid ONLY double *q2p, *pdg; // pdg -> P(D|g) - double *phi; + double *phi, *phi_indel; double *z, *zswap; // aux for afs double *z1, *z2, *phi1, *phi2; // only calculated when n1 is set + double **hg; // hypergeometric distribution + double *lf; // log factorial double t, t1, t2; double *afs, *afs1; // afs: accumulative AFS; afs1: site posterior distribution const uint8_t *PL; // point to PL int PL_len; }; +void bcf_p1_indel_prior(bcf_p1aux_t *ma, double x) +{ + int i; + for (i = 0; i < ma->M; ++i) + ma->phi_indel[i] = ma->phi[i] * x; + ma->phi_indel[ma->M] = 1. - ma->phi[ma->M] * x; +} + static void init_prior(int type, double theta, int M, double *phi) { int i; @@ -63,6 +77,7 @@ static void init_prior(int type, double theta, int M, double *phi) void bcf_p1_init_prior(bcf_p1aux_t *ma, int type, double theta) { init_prior(type, theta, ma->M, ma->phi); + bcf_p1_indel_prior(ma, MC_DEF_INDEL); } void bcf_p1_init_subprior(bcf_p1aux_t *ma, int type, double theta) @@ -106,31 +121,46 @@ int bcf_p1_read_prior(bcf_p1aux_t *ma, const char *fn) for (k = 0; k <= ma->M; ++k) ma->phi[k] /= sum; for (k = 0; k <= ma->M; ++k) fprintf(stderr, " %d:%.3lg", k, ma->phi[ma->M - k]); fputc('\n', stderr); - for (sum = 0., k = 1; k <= ma->M; ++k) sum += k * ma->phi[ma->M - k]; - fprintf(stderr, "[heterozygosity] %lf\n", (double)sum / ma->M); + for (sum = 0., k = 1; k < ma->M; ++k) sum += ma->phi[ma->M - k] * (2.* k * (ma->M - k) / ma->M / (ma->M - 1)); + fprintf(stderr, "[%s] heterozygosity=%lf, ", __func__, (double)sum); + for (sum = 0., k = 1; k <= ma->M; ++k) sum += k * ma->phi[ma->M - k] / ma->M; + fprintf(stderr, "theta=%lf\n", (double)sum); + bcf_p1_indel_prior(ma, MC_DEF_INDEL); return 0; } -bcf_p1aux_t *bcf_p1_init(int n) +bcf_p1aux_t *bcf_p1_init(int n, uint8_t *ploidy) { bcf_p1aux_t *ma; int i; ma = calloc(1, sizeof(bcf_p1aux_t)); ma->n1 = -1; ma->n = n; ma->M = 2 * n; + if (ploidy) { + ma->ploidy = malloc(n); + memcpy(ma->ploidy, ploidy, n); + for (i = 0, ma->M = 0; i < n; ++i) ma->M += ploidy[i]; + if (ma->M == 2 * n) { + free(ma->ploidy); + ma->ploidy = 0; + } + } ma->q2p = calloc(256, sizeof(double)); ma->pdg = calloc(3 * ma->n, sizeof(double)); ma->phi = calloc(ma->M + 1, sizeof(double)); + ma->phi_indel = calloc(ma->M + 1, sizeof(double)); ma->phi1 = calloc(ma->M + 1, sizeof(double)); ma->phi2 = calloc(ma->M + 1, sizeof(double)); - ma->z = calloc(2 * ma->n + 1, sizeof(double)); - ma->zswap = calloc(2 * ma->n + 1, sizeof(double)); + ma->z = calloc(ma->M + 1, sizeof(double)); + ma->zswap = calloc(ma->M + 1, sizeof(double)); ma->z1 = calloc(ma->M + 1, sizeof(double)); // actually we do not need this large ma->z2 = calloc(ma->M + 1, sizeof(double)); - ma->afs = calloc(2 * ma->n + 1, sizeof(double)); - ma->afs1 = calloc(2 * ma->n + 1, sizeof(double)); + ma->afs = calloc(ma->M + 1, sizeof(double)); + ma->afs1 = calloc(ma->M + 1, sizeof(double)); + ma->lf = calloc(ma->M + 1, sizeof(double)); for (i = 0; i < 256; ++i) ma->q2p[i] = pow(10., -i / 10.); + for (i = 0; i <= ma->M; ++i) ma->lf[i] = lgamma(i + 1); bcf_p1_init_prior(ma, MC_PTYPE_FULL, 1e-3); // the simplest prior return ma; } @@ -138,68 +168,335 @@ bcf_p1aux_t *bcf_p1_init(int n) int bcf_p1_set_n1(bcf_p1aux_t *b, int n1) { if (n1 == 0 || n1 >= b->n) return -1; + if (b->M != b->n * 2) { + fprintf(stderr, "[%s] unable to set `n1' when there are haploid samples.\n", __func__); + return -1; + } b->n1 = n1; return 0; } +void bcf_p1_set_ploidy(bcf1_t *b, bcf_p1aux_t *ma) +{ + // bcf_p1aux_t fields are not visible outside of prob1.c, hence this wrapper. + // Ideally, this should set ploidy per site to allow pseudo-autosomal regions + b->ploidy = ma->ploidy; +} + void bcf_p1_destroy(bcf_p1aux_t *ma) { if (ma) { - free(ma->q2p); free(ma->pdg); - free(ma->phi); free(ma->phi1); free(ma->phi2); + int k; + free(ma->lf); + if (ma->hg && ma->n1 > 0) { + for (k = 0; k <= 2*ma->n1; ++k) free(ma->hg[k]); + free(ma->hg); + } + free(ma->ploidy); free(ma->q2p); free(ma->pdg); + free(ma->phi); free(ma->phi_indel); free(ma->phi1); free(ma->phi2); free(ma->z); free(ma->zswap); free(ma->z1); free(ma->z2); free(ma->afs); free(ma->afs1); free(ma); } } -#define char2int(s) (((int)s[0])<<8|s[1]) +extern double kf_gammap(double s, double z); +int test16(bcf1_t *b, anno16_t *a); -static int cal_pdg(const bcf1_t *b, bcf_p1aux_t *ma) +int call_multiallelic_gt(bcf1_t *b, bcf_p1aux_t *ma, double threshold) { - int i, j, k; - long *p, tmp; - p = alloca(b->n_alleles * sizeof(long)); - memset(p, 0, sizeof(long) * b->n_alleles); - for (j = 0; j < ma->n; ++j) { - const uint8_t *pi = ma->PL + j * ma->PL_len; - double *pdg = ma->pdg + j * 3; - pdg[0] = ma->q2p[pi[b->n_alleles]]; pdg[1] = ma->q2p[pi[1]]; pdg[2] = ma->q2p[pi[0]]; - for (i = k = 0; i < b->n_alleles; ++i) { - p[i] += (int)pi[k]; - k += b->n_alleles - i; - } - } - for (i = 0; i < b->n_alleles; ++i) p[i] = p[i]<<4 | i; - for (i = 1; i < b->n_alleles; ++i) // insertion sort - for (j = i; j > 0 && p[j] < p[j-1]; --j) - tmp = p[j], p[j] = p[j-1], p[j-1] = tmp; - for (i = b->n_alleles - 1; i >= 0; --i) - if ((p[i]&0xf) == 0) break; - return i; + int nals = 1; + char *p; + for (p=b->alt; *p; p++) + { + if ( *p=='X' || p[0]=='.' ) break; + if ( p[0]==',' ) nals++; + } + if ( b->alt[0] && !*p ) nals++; + + if ( nals==1 ) return 1; + + if ( nals>4 ) + { + if ( *b->ref=='N' ) return 0; + fprintf(stderr,"Not ready for this, more than 4 alleles at %d: %s, %s\n", b->pos+1, b->ref,b->alt); + exit(1); + } + + // find PL and DP FORMAT indexes + uint8_t *pl = NULL; + int npl = 0, idp=-1; + int i; + for (i = 0; i < b->n_gi; ++i) + { + if (b->gi[i].fmt == bcf_str2int("PL", 2)) + { + pl = (uint8_t*)b->gi[i].data; + npl = b->gi[i].len; + } + if (b->gi[i].fmt == bcf_str2int("DP", 2)) idp=i; + } + if ( !pl ) return -1; + + assert(ma->q2p[0] == 1); + + // Init P(D|G) + int npdg = nals*(nals+1)/2; + double *pdg,*_pdg; + _pdg = pdg = malloc(sizeof(double)*ma->n*npdg); + for (i=0; in; i++) + { + int j; + double sum = 0; + for (j=0; jq2p[pl[j]]; + sum += _pdg[j]; + } + if ( sum ) + for (j=0; jinfo, "QS=")) == 0) { fprintf(stderr,"INFO/QS is required with -m, exiting\n"); exit(1); } + double qsum[4]; + if ( sscanf(p+3,"%lf,%lf,%lf,%lf",&qsum[0],&qsum[1],&qsum[2],&qsum[3])!=4 ) { fprintf(stderr,"Could not parse %s\n",p); exit(1); } + + + // Calculate the most likely combination of alleles + int ia,ib,ic, max_als=0, max_als2=0; + double ref_lk = 0, max_lk = INT_MIN, max_lk2 = INT_MIN, lk_sum = INT_MIN; + for (ia=0; ian; isample++) + { + double *p = pdg + isample*npdg; + // assert( log(p[iaa]) <= 0 ); + lk_tot += log(p[iaa]); + } + if ( ia==0 ) ref_lk = lk_tot; + if ( max_lklk_sum ? lk_tot + log(1+exp(lk_sum-lk_tot)) : lk_sum + log(1+exp(lk_tot-lk_sum)); + } + if ( nals>1 ) + { + for (ia=0; ian; isample++) + { + if ( b->ploidy && b->ploidy[isample]==1 ) continue; + double *p = pdg + isample*npdg; + //assert( log(fa*p[iaa] + fb*p[ibb] + fab*p[iab]) <= 0 ); + lk_tot += log(fa*p[iaa] + fb*p[ibb] + fab*p[iab]); + } + if ( max_lklk_sum ? lk_tot + log(1+exp(lk_sum-lk_tot)) : lk_sum + log(1+exp(lk_tot-lk_sum)); + } + } + } + if ( nals>2 ) + { + for (ia=0; ian; isample++) + { + if ( b->ploidy && b->ploidy[isample]==1 ) continue; + double *p = pdg + isample*npdg; + //assert( log(fa*p[iaa] + fb*p[ibb] + fc*p[icc] + fab*p[iab] + fac*p[iac] + fbc*p[ibc]) <= 0 ); + lk_tot += log(fa*p[iaa] + fb*p[ibb] + fc*p[icc] + fab*p[iab] + fac*p[iac] + fbc*p[ibc]); + } + if ( max_lklk_sum ? lk_tot + log(1+exp(lk_sum-lk_tot)) : lk_sum + log(1+exp(lk_tot-lk_sum)); + } + } + } + } + + + // Should we add another allele, does it increase the likelihood significantly? + int n1=0, n2=0; + for (i=0; in_gi; + s.m = b->m_str; s.l = b->l_str - 1; s.s = b->str; + kputs(":GT:GQ", &s); kputc('\0', &s); + b->m_str = s.m; b->l_str = s.l; b->str = s.s; + bcf_sync(b); + + // Call GTs + int isample, gts=0, ac[4] = {0,0,0,0}; + for (isample = 0; isample < b->n_smpl; isample++) + { + int ploidy = b->ploidy ? b->ploidy[isample] : 2; + double *p = pdg + isample*npdg; + int ia, als = 0; + double lk = 0, lk_sum=0; + for (ia=0; ia lk ) { lk = _lk; als = ia<<3 | ia; } + lk_sum += _lk; + } + if ( ploidy==2 ) + { + for (ia=0; ia lk ) { lk = _lk; als = ib<<3 | ia; } + lk_sum += _lk; + } + } + } + lk = -log(1-lk/lk_sum)/0.2302585; + if ( idp>=0 && ((uint16_t*)b->gi[idp].data)[isample]==0 ) + { + ((uint8_t*)b->gi[old_n_gi].data)[isample] = 1<<7; + ((uint8_t*)b->gi[old_n_gi+1].data)[isample] = 0; + continue; + } + ((uint8_t*)b->gi[old_n_gi].data)[isample] = als; + ((uint8_t*)b->gi[old_n_gi+1].data)[isample] = lk<100 ? (int)lk : 99; + + gts |= 1<<(als>>3&7) | 1<<(als&7); + ac[ als>>3&7 ]++; + ac[ als&7 ]++; + } + bcf_fit_alt(b,max_als); + + + // Prepare BCF for output: ref, alt, filter, info, format + memset(&s, 0, sizeof(kstring_t)); kputc('\0', &s); + kputs(b->ref, &s); kputc('\0', &s); + kputs(b->alt, &s); kputc('\0', &s); kputc('\0', &s); + { + int an=0, nalts=0; + for (i=0; i0 && ac[i] ) nalts++; + } + ksprintf(&s, "AN=%d;", an); + if ( nalts ) + { + kputs("AC=", &s); + for (i=1; i0 ) kputc(',', &s); + } + kputc(';', &s); + } + kputs(b->info, &s); + anno16_t a; + int has_I16 = test16(b, &a) >= 0? 1 : 0; + if (has_I16 ) + { + if ( a.is_tested) ksprintf(&s, ";PV4=%.2g,%.2g,%.2g,%.2g", a.p[0], a.p[1], a.p[2], a.p[3]); + ksprintf(&s, ";DP4=%d,%d,%d,%d;MQ=%d", a.d[0], a.d[1], a.d[2], a.d[3], a.mq); + } + kputc('\0', &s); + rm_info(&s, "I16="); + rm_info(&s, "QS="); + } + kputs(b->fmt, &s); kputc('\0', &s); + free(b->str); + b->m_str = s.m; b->l_str = s.l; b->str = s.s; + b->qual = gts>1 ? -4.343*(ref_lk - lk_sum) : -4.343*(max_lk - lk_sum); + if ( b->qual>999 ) b->qual = 999; + bcf_sync(b); + + + free(pdg); + return gts; } -// f0 is the reference allele frequency -static double mc_freq_iter(double f0, const bcf_p1aux_t *ma) + +static int cal_pdg(const bcf1_t *b, bcf_p1aux_t *ma) { - double f, f3[3]; - int i; - f3[0] = (1.-f0)*(1.-f0); f3[1] = 2.*f0*(1.-f0); f3[2] = f0*f0; - for (i = 0, f = 0.; i < ma->n; ++i) { - double *pdg; - pdg = ma->pdg + i * 3; - f += (pdg[1] * f3[1] + 2. * pdg[2] * f3[2]) - / (pdg[0] * f3[0] + pdg[1] * f3[1] + pdg[2] * f3[2]); - } - f /= ma->n * 2.; - return f; + int i, j; + long *p, tmp; + p = alloca(b->n_alleles * sizeof(long)); + memset(p, 0, sizeof(long) * b->n_alleles); + for (j = 0; j < ma->n; ++j) { + const uint8_t *pi = ma->PL + j * ma->PL_len; + double *pdg = ma->pdg + j * 3; + pdg[0] = ma->q2p[pi[2]]; pdg[1] = ma->q2p[pi[1]]; pdg[2] = ma->q2p[pi[0]]; + for (i = 0; i < b->n_alleles; ++i) + p[i] += (int)pi[(i+1)*(i+2)/2-1]; + } + for (i = 0; i < b->n_alleles; ++i) p[i] = p[i]<<4 | i; + for (i = 1; i < b->n_alleles; ++i) // insertion sort + for (j = i; j > 0 && p[j] < p[j-1]; --j) + tmp = p[j], p[j] = p[j-1], p[j-1] = tmp; + for (i = b->n_alleles - 1; i >= 0; --i) + if ((p[i]&0xf) == 0) break; + return i; } + int bcf_p1_call_gt(const bcf_p1aux_t *ma, double f0, int k) { double sum, g[3]; double max, f3[3], *pdg = ma->pdg + k * 3; - int q, i, max_i; - f3[0] = (1.-f0)*(1.-f0); f3[1] = 2.*f0*(1.-f0); f3[2] = f0*f0; + int q, i, max_i, ploidy; + ploidy = ma->ploidy? ma->ploidy[k] : 2; + if (ploidy == 2) { + f3[0] = (1.-f0)*(1.-f0); f3[1] = 2.*f0*(1.-f0); f3[2] = f0*f0; + } else { + f3[0] = 1. - f0; f3[1] = 0; f3[2] = f0; + } for (i = 0, sum = 0.; i < 3; ++i) sum += (g[i] = pdg[i] * f3[i]); for (i = 0, max = -1., max_i = 0; i < 3; ++i) { @@ -219,6 +516,7 @@ static void mc_cal_y_core(bcf_p1aux_t *ma, int beg) { double *z[2], *tmp, *pdg; int _j, last_min, last_max; + assert(beg == 0 || ma->M == ma->n*2); z[0] = ma->z; z[1] = ma->zswap; pdg = ma->pdg; @@ -227,41 +525,81 @@ static void mc_cal_y_core(bcf_p1aux_t *ma, int beg) z[0][0] = 1.; last_min = last_max = 0; ma->t = 0.; - for (_j = beg; _j < ma->n; ++_j) { - int k, j = _j - beg, _min = last_min, _max = last_max; - double p[3], sum; - pdg = ma->pdg + _j * 3; - p[0] = pdg[0]; p[1] = 2. * pdg[1]; p[2] = pdg[2]; - for (; _min < _max && z[0][_min] < TINY; ++_min) z[0][_min] = z[1][_min] = 0.; - for (; _max > _min && z[0][_max] < TINY; --_max) z[0][_max] = z[1][_max] = 0.; - _max += 2; - if (_min == 0) - k = 0, z[1][k] = (2*j+2-k)*(2*j-k+1) * p[0] * z[0][k]; - if (_min <= 1) - k = 1, z[1][k] = (2*j+2-k)*(2*j-k+1) * p[0] * z[0][k] + k*(2*j+2-k) * p[1] * z[0][k-1]; - for (k = _min < 2? 2 : _min; k <= _max; ++k) - z[1][k] = (2*j+2-k)*(2*j-k+1) * p[0] * z[0][k] - + k*(2*j+2-k) * p[1] * z[0][k-1] - + k*(k-1)* p[2] * z[0][k-2]; - for (k = _min, sum = 0.; k <= _max; ++k) sum += z[1][k]; - ma->t += log(sum / ((2. * j + 2) * (2. * j + 1))); - for (k = _min; k <= _max; ++k) z[1][k] /= sum; - if (_min >= 1) z[1][_min-1] = 0.; - if (_min >= 2) z[1][_min-2] = 0.; - if (j < ma->n - 1) z[1][_max+1] = z[1][_max+2] = 0.; - if (_j == ma->n1 - 1) { // set pop1 - ma->t1 = ma->t; - memcpy(ma->z1, z[1], sizeof(double) * (ma->n1 * 2 + 1)); + if (ma->M == ma->n * 2) { + int M = 0; + for (_j = beg; _j < ma->n; ++_j) { + int k, j = _j - beg, _min = last_min, _max = last_max, M0; + double p[3], sum; + M0 = M; M += 2; + pdg = ma->pdg + _j * 3; + p[0] = pdg[0]; p[1] = 2. * pdg[1]; p[2] = pdg[2]; + for (; _min < _max && z[0][_min] < TINY; ++_min) z[0][_min] = z[1][_min] = 0.; + for (; _max > _min && z[0][_max] < TINY; --_max) z[0][_max] = z[1][_max] = 0.; + _max += 2; + if (_min == 0) k = 0, z[1][k] = (M0-k+1) * (M0-k+2) * p[0] * z[0][k]; + if (_min <= 1) k = 1, z[1][k] = (M0-k+1) * (M0-k+2) * p[0] * z[0][k] + k*(M0-k+2) * p[1] * z[0][k-1]; + for (k = _min < 2? 2 : _min; k <= _max; ++k) + z[1][k] = (M0-k+1)*(M0-k+2) * p[0] * z[0][k] + k*(M0-k+2) * p[1] * z[0][k-1] + k*(k-1)* p[2] * z[0][k-2]; + for (k = _min, sum = 0.; k <= _max; ++k) sum += z[1][k]; + ma->t += log(sum / (M * (M - 1.))); + for (k = _min; k <= _max; ++k) z[1][k] /= sum; + if (_min >= 1) z[1][_min-1] = 0.; + if (_min >= 2) z[1][_min-2] = 0.; + if (j < ma->n - 1) z[1][_max+1] = z[1][_max+2] = 0.; + if (_j == ma->n1 - 1) { // set pop1; ma->n1==-1 when unset + ma->t1 = ma->t; + memcpy(ma->z1, z[1], sizeof(double) * (ma->n1 * 2 + 1)); + } + tmp = z[0]; z[0] = z[1]; z[1] = tmp; + last_min = _min; last_max = _max; + } + //for (_j = 0; _j < last_min; ++_j) z[0][_j] = 0.; // TODO: are these necessary? + //for (_j = last_max + 1; _j < ma->M; ++_j) z[0][_j] = 0.; + } else { // this block is very similar to the block above; these two might be merged in future + int j, M = 0; + for (j = 0; j < ma->n; ++j) { + int k, M0, _min = last_min, _max = last_max; + double p[3], sum; + pdg = ma->pdg + j * 3; + for (; _min < _max && z[0][_min] < TINY; ++_min) z[0][_min] = z[1][_min] = 0.; + for (; _max > _min && z[0][_max] < TINY; --_max) z[0][_max] = z[1][_max] = 0.; + M0 = M; + M += ma->ploidy[j]; + if (ma->ploidy[j] == 1) { + p[0] = pdg[0]; p[1] = pdg[2]; + _max++; + if (_min == 0) k = 0, z[1][k] = (M0+1-k) * p[0] * z[0][k]; + for (k = _min < 1? 1 : _min; k <= _max; ++k) + z[1][k] = (M0+1-k) * p[0] * z[0][k] + k * p[1] * z[0][k-1]; + for (k = _min, sum = 0.; k <= _max; ++k) sum += z[1][k]; + ma->t += log(sum / M); + for (k = _min; k <= _max; ++k) z[1][k] /= sum; + if (_min >= 1) z[1][_min-1] = 0.; + if (j < ma->n - 1) z[1][_max+1] = 0.; + } else if (ma->ploidy[j] == 2) { + p[0] = pdg[0]; p[1] = 2 * pdg[1]; p[2] = pdg[2]; + _max += 2; + if (_min == 0) k = 0, z[1][k] = (M0-k+1) * (M0-k+2) * p[0] * z[0][k]; + if (_min <= 1) k = 1, z[1][k] = (M0-k+1) * (M0-k+2) * p[0] * z[0][k] + k*(M0-k+2) * p[1] * z[0][k-1]; + for (k = _min < 2? 2 : _min; k <= _max; ++k) + z[1][k] = (M0-k+1)*(M0-k+2) * p[0] * z[0][k] + k*(M0-k+2) * p[1] * z[0][k-1] + k*(k-1)* p[2] * z[0][k-2]; + for (k = _min, sum = 0.; k <= _max; ++k) sum += z[1][k]; + ma->t += log(sum / (M * (M - 1.))); + for (k = _min; k <= _max; ++k) z[1][k] /= sum; + if (_min >= 1) z[1][_min-1] = 0.; + if (_min >= 2) z[1][_min-2] = 0.; + if (j < ma->n - 1) z[1][_max+1] = z[1][_max+2] = 0.; + } + tmp = z[0]; z[0] = z[1]; z[1] = tmp; + last_min = _min; last_max = _max; } - tmp = z[0]; z[0] = z[1]; z[1] = tmp; - last_min = _min; last_max = _max; } if (z[0] != ma->z) memcpy(ma->z, z[0], sizeof(double) * (ma->M + 1)); } static void mc_cal_y(bcf_p1aux_t *ma) { - if (ma->n1 > 0 && ma->n1 < ma->n) { + if (ma->n1 > 0 && ma->n1 < ma->n && ma->M == ma->n * 2) { // NB: ma->n1 is ineffective when there are haploid samples int k; long double x; memset(ma->z1, 0, sizeof(double) * (2 * ma->n1 + 1)); @@ -277,40 +615,131 @@ static void mc_cal_y(bcf_p1aux_t *ma) } else mc_cal_y_core(ma, 0); } -static void contrast(bcf_p1aux_t *ma, double pc[4]) // mc_cal_y() must be called before hand +#define CONTRAST_TINY 1e-30 + +extern double kf_gammaq(double s, double z); // incomplete gamma function for chi^2 test + +static inline double chi2_test(int a, int b, int c, int d) +{ + double x, z; + x = (double)(a+b) * (c+d) * (b+d) * (a+c); + if (x == 0.) return 1; + z = a * d - b * c; + return kf_gammaq(.5, .5 * z * z * (a+b+c+d) / x); +} + +// chi2=(a+b+c+d)(ad-bc)^2/[(a+b)(c+d)(a+c)(b+d)] +static inline double contrast2_aux(const bcf_p1aux_t *p1, double sum, int k1, int k2, double x[3]) { - int k, n1 = ma->n1, n2 = ma->n - ma->n1; - long double sum1, sum2; - pc[0] = pc[1] = pc[2] = pc[3] = 0.; - if (n1 <= 0 || n2 <= 0) return; - for (k = 0, sum1 = 0.; k <= 2*n1; ++k) sum1 += ma->phi1[k] * ma->z1[k]; - for (k = 0, sum2 = 0.; k <= 2*n2; ++k) sum2 += ma->phi2[k] * ma->z2[k]; - pc[2] = ma->phi1[2*n1] * ma->z1[2*n1] / sum1; - pc[3] = ma->phi2[2*n2] * ma->z2[2*n2] / sum2; - for (k = 2; k < 4; ++k) { - pc[k] = pc[k] > .5? -(-4.343 * log(1. - pc[k] + TINY) + .499) : -4.343 * log(pc[k] + TINY) + .499; - pc[k] = (int)pc[k]; - if (pc[k] > 99) pc[k] = 99; - if (pc[k] < -99) pc[k] = -99; + double p = p1->phi[k1+k2] * p1->z1[k1] * p1->z2[k2] / sum * p1->hg[k1][k2]; + int n1 = p1->n1, n2 = p1->n - p1->n1; + if (p < CONTRAST_TINY) return -1; + if (.5*k1/n1 < .5*k2/n2) x[1] += p; + else if (.5*k1/n1 > .5*k2/n2) x[2] += p; + else x[0] += p; + return p * chi2_test(k1, k2, (n1<<1) - k1, (n2<<1) - k2); +} + +static double contrast2(bcf_p1aux_t *p1, double ret[3]) +{ + int k, k1, k2, k10, k20, n1, n2; + double sum; + // get n1 and n2 + n1 = p1->n1; n2 = p1->n - p1->n1; + if (n1 <= 0 || n2 <= 0) return 0.; + if (p1->hg == 0) { // initialize the hypergeometric distribution + /* NB: the hg matrix may take a lot of memory when there are many samples. There is a way + to avoid precomputing this matrix, but it is slower and quite intricate. The following + computation in this block can be accelerated with a similar strategy, but perhaps this + is not a serious concern for now. */ + double tmp = lgamma(2*(n1+n2)+1) - (lgamma(2*n1+1) + lgamma(2*n2+1)); + p1->hg = calloc(2*n1+1, sizeof(void*)); + for (k1 = 0; k1 <= 2*n1; ++k1) { + p1->hg[k1] = calloc(2*n2+1, sizeof(double)); + for (k2 = 0; k2 <= 2*n2; ++k2) + p1->hg[k1][k2] = exp(lgamma(k1+k2+1) + lgamma(p1->M-k1-k2+1) - (lgamma(k1+1) + lgamma(k2+1) + lgamma(2*n1-k1+1) + lgamma(2*n2-k2+1) + tmp)); + } + } + { // compute + long double suml = 0; + for (k = 0; k <= p1->M; ++k) suml += p1->phi[k] * p1->z[k]; + sum = suml; + } + { // get the max k1 and k2 + double max; + int max_k; + for (k = 0, max = 0, max_k = -1; k <= 2*n1; ++k) { + double x = p1->phi1[k] * p1->z1[k]; + if (x > max) max = x, max_k = k; + } + k10 = max_k; + for (k = 0, max = 0, max_k = -1; k <= 2*n2; ++k) { + double x = p1->phi2[k] * p1->z2[k]; + if (x > max) max = x, max_k = k; + } + k20 = max_k; + } + { // We can do the following with one nested loop, but that is an O(N^2) thing. The following code block is much faster for large N. + double x[3], y; + long double z = 0., L[2]; + x[0] = x[1] = x[2] = 0; L[0] = L[1] = 0; + for (k1 = k10; k1 >= 0; --k1) { + for (k2 = k20; k2 >= 0; --k2) { + if ((y = contrast2_aux(p1, sum, k1, k2, x)) < 0) break; + else z += y; + } + for (k2 = k20 + 1; k2 <= 2*n2; ++k2) { + if ((y = contrast2_aux(p1, sum, k1, k2, x)) < 0) break; + else z += y; + } + } + ret[0] = x[0]; ret[1] = x[1]; ret[2] = x[2]; + x[0] = x[1] = x[2] = 0; + for (k1 = k10 + 1; k1 <= 2*n1; ++k1) { + for (k2 = k20; k2 >= 0; --k2) { + if ((y = contrast2_aux(p1, sum, k1, k2, x)) < 0) break; + else z += y; + } + for (k2 = k20 + 1; k2 <= 2*n2; ++k2) { + if ((y = contrast2_aux(p1, sum, k1, k2, x)) < 0) break; + else z += y; + } + } + ret[0] += x[0]; ret[1] += x[1]; ret[2] += x[2]; + if (ret[0] + ret[1] + ret[2] < 0.95) { // in case of bad things happened + ret[0] = ret[1] = ret[2] = 0; L[0] = L[1] = 0; + for (k1 = 0, z = 0.; k1 <= 2*n1; ++k1) + for (k2 = 0; k2 <= 2*n2; ++k2) + if ((y = contrast2_aux(p1, sum, k1, k2, ret)) >= 0) z += y; + if (ret[0] + ret[1] + ret[2] < 0.95) // It seems that this may be caused by floating point errors. I do not really understand why... + z = 1.0, ret[0] = ret[1] = ret[2] = 1./3; + } + return (double)z; } - pc[0] = ma->phi2[2*n2] * ma->z2[2*n2] / sum2 * (1. - ma->phi1[2*n1] * ma->z1[2*n1] / sum1); - pc[1] = ma->phi1[2*n1] * ma->z1[2*n1] / sum1 * (1. - ma->phi2[2*n2] * ma->z2[2*n2] / sum2); - pc[0] = pc[0] == 1.? 99 : (int)(-4.343 * log(1. - pc[0]) + .499); - pc[1] = pc[1] == 1.? 99 : (int)(-4.343 * log(1. - pc[1]) + .499); } -static double mc_cal_afs(bcf_p1aux_t *ma) +static double mc_cal_afs(bcf_p1aux_t *ma, double *p_ref_folded, double *p_var_folded) { int k; - long double sum = 0.; + long double sum = 0., sum2; + double *phi = ma->is_indel? ma->phi_indel : ma->phi; memset(ma->afs1, 0, sizeof(double) * (ma->M + 1)); mc_cal_y(ma); + // compute AFS for (k = 0, sum = 0.; k <= ma->M; ++k) - sum += (long double)ma->phi[k] * ma->z[k]; + sum += (long double)phi[k] * ma->z[k]; for (k = 0; k <= ma->M; ++k) { - ma->afs1[k] = ma->phi[k] * ma->z[k] / sum; + ma->afs1[k] = phi[k] * ma->z[k] / sum; if (isnan(ma->afs1[k]) || isinf(ma->afs1[k])) return -1.; } + // compute folded variant probability + for (k = 0, sum = 0.; k <= ma->M; ++k) + sum += (long double)(phi[k] + phi[ma->M - k]) / 2. * ma->z[k]; + for (k = 1, sum2 = 0.; k < ma->M; ++k) + sum2 += (long double)(phi[k] + phi[ma->M - k]) / 2. * ma->z[k]; + *p_var_folded = sum2 / sum; + *p_ref_folded = (phi[k] + phi[ma->M - k]) / 2. * (ma->z[ma->M] + ma->z[0]) / sum; + // the expected frequency for (k = 0, sum = 0.; k <= ma->M; ++k) { ma->afs[k] += ma->afs1[k]; sum += k * ma->afs1[k]; @@ -318,49 +747,36 @@ static double mc_cal_afs(bcf_p1aux_t *ma) return sum / ma->M; } -long double bcf_p1_cal_g3(bcf_p1aux_t *p1a, double g[3]) -{ - long double pd = 0., g2[3]; - int i, k; - memset(g2, 0, sizeof(long double) * 3); - for (k = 0; k < p1a->M; ++k) { - double f = (double)k / p1a->M, f3[3], g1[3]; - long double z = 1.; - g1[0] = g1[1] = g1[2] = 0.; - f3[0] = (1. - f) * (1. - f); f3[1] = 2. * f * (1. - f); f3[2] = f * f; - for (i = 0; i < p1a->n; ++i) { - double *pdg = p1a->pdg + i * 3; - double x = pdg[0] * f3[0] + pdg[1] * f3[1] + pdg[2] * f3[2]; - z *= x; - g1[0] += pdg[0] * f3[0] / x; - g1[1] += pdg[1] * f3[1] / x; - g1[2] += pdg[2] * f3[2] / x; - } - pd += p1a->phi[k] * z; - for (i = 0; i < 3; ++i) - g2[i] += p1a->phi[k] * z * g1[i]; - } - for (i = 0; i < 3; ++i) g[i] = g2[i] / pd; - return pd; -} - -int bcf_p1_cal(bcf1_t *b, bcf_p1aux_t *ma, bcf_p1rst_t *rst) +int bcf_p1_cal(const bcf1_t *b, int do_contrast, bcf_p1aux_t *ma, bcf_p1rst_t *rst) { int i, k; long double sum = 0.; + ma->is_indel = bcf_is_indel(b); + rst->perm_rank = -1; // set PL and PL_len for (i = 0; i < b->n_gi; ++i) { - if (b->gi[i].fmt == char2int("PL")) { + if (b->gi[i].fmt == bcf_str2int("PL", 2)) { ma->PL = (uint8_t*)b->gi[i].data; ma->PL_len = b->gi[i].len; break; } } + if (i == b->n_gi) return -1; // no PL if (b->n_alleles < 2) return -1; // FIXME: find a better solution // rst->rank0 = cal_pdg(b, ma); - rst->f_exp = mc_cal_afs(ma); + rst->f_exp = mc_cal_afs(ma, &rst->p_ref_folded, &rst->p_var_folded); rst->p_ref = ma->afs1[ma->M]; + for (k = 0, sum = 0.; k < ma->M; ++k) + sum += ma->afs1[k]; + rst->p_var = (double)sum; + { // compute the allele count + double max = -1; + rst->ac = -1; + for (k = 0; k <= ma->M; ++k) + if (max < ma->z[k]) max = ma->z[k], rst->ac = k; + rst->ac = ma->M - rst->ac; + } // calculate f_flat and f_em for (k = 0, sum = 0.; k <= ma->M; ++k) sum += (long double)ma->z[k]; @@ -370,16 +786,33 @@ int bcf_p1_cal(bcf1_t *b, bcf_p1aux_t *ma, bcf_p1rst_t *rst) rst->f_flat += k * p; } rst->f_flat /= ma->M; - { // calculate f_em - double flast = rst->f_flat; - for (i = 0; i < MC_MAX_EM_ITER; ++i) { - rst->f_em = mc_freq_iter(flast, ma); - if (fabs(rst->f_em - flast) < MC_EM_EPS) break; - flast = rst->f_em; - } + { // estimate equal-tail credible interval (95% level) + int l, h; + double p; + for (i = 0, p = 0.; i <= ma->M; ++i) + if (p + ma->afs1[i] > 0.025) break; + else p += ma->afs1[i]; + l = i; + for (i = ma->M, p = 0.; i >= 0; --i) + if (p + ma->afs1[i] > 0.025) break; + else p += ma->afs1[i]; + h = i; + rst->cil = (double)(ma->M - h) / ma->M; rst->cih = (double)(ma->M - l) / ma->M; } - rst->g[0] = rst->g[1] = rst->g[2] = -1.; - contrast(ma, rst->pc); + if (ma->n1 > 0) { // compute LRT + double max0, max1, max2; + for (k = 0, max0 = -1; k <= ma->M; ++k) + if (max0 < ma->z[k]) max0 = ma->z[k]; + for (k = 0, max1 = -1; k <= ma->n1 * 2; ++k) + if (max1 < ma->z1[k]) max1 = ma->z1[k]; + for (k = 0, max2 = -1; k <= ma->M - ma->n1 * 2; ++k) + if (max2 < ma->z2[k]) max2 = ma->z2[k]; + rst->lrt = log(max1 * max2 / max0); + rst->lrt = rst->lrt < 0? 1 : kf_gammaq(.5, rst->lrt); + } else rst->lrt = -1.0; + rst->cmp[0] = rst->cmp[1] = rst->cmp[2] = rst->p_chi2 = -1.0; + if (do_contrast && rst->p_var > 0.5) // skip contrast2() if the locus is a strong non-variant + rst->p_chi2 = contrast2(ma, rst->cmp); return 0; }