KSTREAM_INIT(gzFile, gzread, 16384)
#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] = {
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
ma->z2 = calloc(ma->M + 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;
}
void bcf_p1_destroy(bcf_p1aux_t *ma)
{
if (ma) {
+ 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);
if ((p[i]&0xf) == 0) break;
return i;
}
-// f0 is the reference allele frequency
-static double mc_freq_iter(double f0, const 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 bcf_p1_call_gt(const bcf_p1aux_t *ma, double f0, int k)
{
}
for (i = 0, sum = 0.; i < 3; ++i)
sum += (g[i] = pdg[i] * f3[i]);
- for (i = 0, max = -1., max_i = 0; i <= ploidy; ++i) {
+ for (i = 0, max = -1., max_i = 0; i < 3; ++i) {
g[i] /= sum;
if (g[i] > max) max = g[i], max_i = i;
}
last_min = last_max = 0;
ma->t = 0.;
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;
+ 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] = (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];
+ 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] = (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];
+ 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 / ((2. * j + 2) * (2. * j + 1)));
+ 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.;
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) {
} 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)
{
- int k, n1 = ma->n1, n2 = ma->n - ma->n1;
- long double sum1, sum2;
- pc[0] = pc[1] = pc[2] = pc[3] = -1.;
- 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 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])
+{
+ 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, double *p_ref_folded, double *p_var_folded)
return sum / ma->M;
}
-int bcf_p1_cal(const 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 == bcf_str2int("PL", 2)) {
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);
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];
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)
+ 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-1, p = 0.; i >= 0; --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;
}