ruby-minigraph 0.0.20.0

data/ext/minigraph/ggsimple.c

@@ -0,0 +1,570 @@
+#include <assert.h>
+#include "mgpriv.h"
+#include "gfa-priv.h"
+#include "kalloc.h"
+#include "bseq.h"
+#include "algo.h"
+#include "sys.h"
+#include "ggen.h"
+#include "kvec-km.h"
+
+int32_t mg_gc_index(void *km, int min_mapq, int min_map_len, int min_depth_len, const gfa_t *g, int32_t n_seq, mg_gchains_t *const* gcs,
+					double *a_dens, int32_t **soff_, int32_t **qoff_, mg_intv_t **sintv_, mg_intv_t **qintv_)
+{
+	int32_t t, i, j, max_acnt, *scnt, *soff, *qcnt, *qoff;
+	int64_t sum_acnt, sum_alen;
+	mg_intv_t *sintv, *qintv;
+
+	// count the number of intervals on each segment
+	KCALLOC(km, scnt, g->n_seg);
+	KCALLOC(km, qcnt, n_seq);
+	for (t = 0, max_acnt = 0; t < n_seq; ++t) {
+		const mg_gchains_t *gt = gcs[t];
+		for (i = 0; i < gt->n_gc; ++i) {
+			const mg_gchain_t *gc = &gt->gc[i];
+			if (gc->id != gc->parent) continue;
+			if (gc->blen < min_depth_len || gc->mapq < min_mapq) continue;
+			if (gc->n_anchor > max_acnt) max_acnt = gc->n_anchor;
+			++qcnt[t];
+			for (j = 0; j < gc->cnt; ++j)
+				++scnt[gt->lc[gc->off + j].v>>1];
+		}
+	}
+	if (max_acnt == 0) { // no gchain
+		kfree(km, scnt); kfree(km, qcnt);
+		return 0;
+	}
+
+	// compute soff[] and qoff[]
+	KMALLOC(km, soff, g->n_seg + 1);
+	KMALLOC(km, qoff, n_seq + 1);
+	for (soff[0] = 0, i = 1; i <= g->n_seg; ++i)
+		soff[i] = soff[i - 1] + scnt[i - 1];
+	for (qoff[0] = 0, i = 1; i <= n_seq; ++i)
+		qoff[i] = qoff[i - 1] + qcnt[i - 1];
+
+	// populate the interval list
+	memset(scnt, 0, 4 * g->n_seg);
+	memset(qcnt, 0, 4 * n_seq);
+	KMALLOC(km, sintv, soff[g->n_seg]);
+	KMALLOC(km, qintv, qoff[n_seq]);
+	sum_acnt = sum_alen = 0;
+	for (t = 0; t < n_seq; ++t) {
+		const mg_gchains_t *gt = gcs[t];
+		for (i = 0; i < gt->n_gc; ++i) {
+			const mg_gchain_t *gc = &gt->gc[i];
+			mg_intv_t *p;
+			if (gc->id != gc->parent) continue;
+			if (gc->blen < min_depth_len || gc->mapq < min_mapq) continue;
+			p = &qintv[qoff[t] + qcnt[t]];
+			++qcnt[t];
+			p->st = gc->qs, p->en = gc->qe, p->rev = 0, p->far = -1, p->i = -1;
+			for (j = 0; j < gc->cnt; ++j) {
+				const mg_llchain_t *lc = &gt->lc[gc->off + j];
+				int32_t rs, re, tmp;
+				if (lc->cnt > 0) { // compute start and end on the forward strand on the segment
+					const mg128_t *qs = &gt->a[lc->off];
+					const mg128_t *qe = &gt->a[lc->off + lc->cnt - 1];
+					int32_t rs0 = (int32_t)qs->x + 1 - (int32_t)(qs->y>>32&0xff);
+					int32_t re0 = (int32_t)qe->x;
+					assert(rs0 >= 0 && re0 > rs0 && re0 < g->seg[lc->v>>1].len);
+					sum_alen += re0 - rs0, sum_acnt += (qe->x>>32) - (qs->x>>32) + 1;
+					rs = 0, re = g->seg[lc->v>>1].len;
+					if (j == 0) rs = gc->p? gc->p->ss : rs0;
+					if (j == gc->cnt - 1) re = gc->p? gc->p->ee : re0;
+					if (lc->v&1) // swap rs and re
+						tmp = rs, rs = g->seg[lc->v>>1].len - re, re = g->seg[lc->v>>1].len - tmp;
+				} else rs = 0, re = g->seg[lc->v>>1].len;
+				p = &sintv[soff[lc->v>>1] + scnt[lc->v>>1]];
+				++scnt[lc->v>>1];
+				p->st = rs, p->en = re, p->rev = lc->v&1, p->far = -1, p->i = -1;
+			}
+		}
+	}
+	*a_dens = (double)sum_acnt / sum_alen;
+
+	// sort and index intervals
+	for (i = 0; i < g->n_seg; ++i) {
+		assert(soff[i+1] - soff[i] == scnt[i]);
+		mg_intv_index(soff[i+1] - soff[i], &sintv[soff[i]]);
+	}
+	kfree(km, scnt);
+	for (i = 0; i < n_seq; ++i) {
+		assert(qoff[i+1] - qoff[i] == qcnt[i]);
+		mg_intv_index(qoff[i+1] - qoff[i], &qintv[qoff[i]]);
+	}
+	kfree(km, qcnt);
+
+	*sintv_ = sintv, *qintv_ = qintv;
+	*soff_ = soff, *qoff_ = qoff;
+	return max_acnt;
+}
+
+/**********************
+ * Graph augmentation *
+ **********************/
+
+void mg_ggsimple(void *km, const mg_ggopt_t *opt, gfa_t *g, int32_t n_seq, const mg_bseq1_t *seq, mg_gchains_t *const* gcs)
+{
+	int32_t t, i, j, *soff, *qoff, max_acnt, *sc, m_ovlp = 0, *ovlp = 0, n_ins, m_ins, n_inv;
+	int32_t l_pseq, m_pseq;
+	uint64_t *meta;
+	mg_intv_t *sintv, *qintv;
+	double a_dens;
+	gfa_ins_t *ins;
+	char *pseq;
+
+	max_acnt = mg_gc_index(km, opt->min_mapq, opt->min_map_len, opt->min_depth_len, g, n_seq, gcs, &a_dens, &soff, &qoff, &sintv, &qintv);
+	if (max_acnt == 0) return;
+
+	// extract poorly regions
+	m_pseq = l_pseq = 0, pseq = 0;
+	m_ins = n_ins = 0, ins = 0;
+	n_inv = 0;
+	KMALLOC(km, sc, max_acnt);
+	KMALLOC(km, meta, max_acnt);
+	for (t = 0; t < n_seq; ++t) {
+		const mg_gchains_t *gt = gcs[t];
+		for (i = 0; i < gt->n_gc; ++i) {
+			const mg_gchain_t *gc = &gt->gc[i];
+			int32_t off_a, off_l, n_ss, far_q;
+			mg_msseg_t *ss;
+			if (gc->id != gc->parent) continue;
+			if (gc->blen < opt->min_map_len || gc->mapq < opt->min_mapq) continue;
+			assert(gc->cnt > 0);
+
+			// fill sc[]. This part achieves a similar goal to the one in mg_gchain_extra(). It makes more assumptions, but is logically simpler.
+			off_l = gc->off;
+			off_a = gt->lc[off_l].off + 1;
+			far_q = 0;
+			for (j = 1; j < gc->n_anchor; ++j, ++off_a) {
+				const mg128_t *q = &gt->a[off_a - 1], *p = &gt->a[off_a];
+				const mg_llchain_t *lc = &gt->lc[off_l];
+				int32_t s, ed = -1, off_l0 = off_l, pd, qd = (int32_t)p->y - (int32_t)q->y, c = (int32_t)(p->x>>32) - (int32_t)(q->x>>32) - 1;
+				if ((int32_t)q->y > far_q) far_q = (int32_t)q->y; // far_q keeps the rightmost query position seen so far
+				if (off_a == lc->off + lc->cnt) { // we are at the end of the current lchain
+					pd = g->seg[lc->v>>1].len - (int32_t)q->x - 1;
+					for (++off_l; off_l < gc->off + gc->cnt && gt->lc[off_l].cnt == 0; ++off_l)
+						pd += g->seg[gt->lc[off_l].v>>1].len;
+					assert(off_l < gc->off + gc->cnt);
+					if (gt->lc[off_l].ed >= 0) ed = gt->lc[off_l].ed;
+					pd += (int32_t)p->x + 1;
+				} else pd = (int32_t)p->x - (int32_t)q->x;
+				if ((opt->flag&MG_G_NO_QOVLP) && (int32_t)p->y < far_q) s = 1; // query overlap
+				else if (pd == qd && c == 0) s = -opt->match_pen;
+				else if (ed >= 0) {
+					int32_t min_d = pd < qd? pd : qd;
+					double t = 1. / (1.01 - opt->ggs_max_iden);
+					if (t > 10.) t = 10.;
+					s = (int32_t)(ed * t - min_d);
+				} else if (pd > qd) {
+					double x = qd * a_dens;
+					x = x > c? x : c;
+					s = (int32_t)(x + (pd - qd) * a_dens + .499);
+				} else {
+					s = (int32_t)(qd * a_dens + .499);
+					s = s > c? s : c;
+				}
+				sc[j - 1] = s;
+				meta[j-1] = (uint64_t)pd<<32 | off_l0;
+			}
+
+			// get regions to insert
+			ss = mg_mss_all(0, gc->n_anchor - 1, sc, 10, 0, &n_ss);
+			off_a = gt->lc[gc->off].off;
+			for (j = 0; j < n_ss; ++j) {
+				const mg128_t *p, *q;
+				int32_t st, en, ls, le, span, pd, k, n_ovlp, min_len, is_inv = 0;
+				gfa_ins_t I;
+
+				// find the initial positions
+				min_len = opt->ggs_min_end_cnt > 0? opt->ggs_min_end_cnt : 0;
+				if (min_len < ss[j].sc * opt->ggs_min_end_frac) min_len = ss[j].sc * opt->ggs_min_end_frac;
+				if (ss[j].st <= min_len || ss[j].en >= gc->n_anchor - 1 - min_len) continue; // too close to ends
+				st = ss[j].st, en = ss[j].en;
+				q = &gt->a[off_a + st];
+				p = &gt->a[off_a + en];
+				span = p->y>>32&0xff;
+				I.ctg = t;
+				ls = (int32_t)meta[st], le = (int32_t)meta[en]; // first and last lchain; CLOSED interval
+				assert(ls <= le);
+				I.v[0] = gt->lc[ls].v;
+				I.v[1] = gt->lc[le].v;
+				I.voff[0] = (int32_t)q->x + 1 - span;
+				I.voff[1] = (int32_t)p->x + 1;
+				I.coff[0] = (int32_t)q->y + 1 - span;
+				I.coff[1] = (int32_t)p->y + 1;
+				assert(I.voff[0] <= g->seg[I.v[0]>>1].len);
+				assert(I.voff[1] <= g->seg[I.v[1]>>1].len);
+				for (k = st, pd = span; k < en; ++k)
+					pd += meta[k]>>32;
+
+				if (I.coff[0] > I.coff[1]) {
+					if (mg_verbose >= 2 && pd + (I.coff[0] - I.coff[1]) >= opt->min_var_len)
+						fprintf(stderr, "[W::%s] query overlap on gchain %d: [%c%s:%d,%c%s:%d|%d] <=> %s:[%d,%d|%d]\n", __func__, t, "><"[I.v[0]&1], g->seg[I.v[0]>>1].name, I.voff[0], "><"[I.v[1]&1], g->seg[I.v[1]>>1].name, I.voff[1], pd, seq[t].name, I.coff[0], I.coff[1], I.coff[1] - I.coff[0]);
+					continue; // such overlap can't be properly resolved
+				}
+				pd -= gfa_ins_adj(g, opt->ggs_shrink_pen, &I, seq[t].seq);
+
+				min_len = pd > I.coff[1] - I.coff[0]? pd : I.coff[1] - I.coff[0];
+				if (I.coff[0] <= min_len || I.coff[1] >= seq[t].l_seq - min_len) continue; // test if the event is close to ends again
+
+				// filtering
+				if (I.coff[1] - I.coff[0] < opt->min_var_len && pd < opt->min_var_len)
+					continue;
+				for (k = I.coff[0]; k < I.coff[1]; ++k) { // test ambiguous bases
+					int c = seq[t].seq[k];
+					if (c == 'n' || c == 'N') break;
+				}
+				if (k != I.coff[1]) continue; // no ambiguous bases on the insert
+				n_ovlp = mg_intv_overlap(km, qoff[t+1] - qoff[t], &qintv[qoff[t]], I.coff[0], I.coff[1], &ovlp, &m_ovlp); // test overlapping on the query
+				if (n_ovlp == 0) fprintf(stderr, "[W::%s] query interval %s:%d-%d is not covered\n", __func__, seq[t].name, I.coff[0], I.coff[1]);
+				if (n_ovlp != 1) continue;
+				for (k = ls; k <= le; ++k) { // find other mappings overlapping with the insert on the graph
+					uint32_t v = gt->lc[k].v, len = g->seg[v>>1].len;
+					int32_t s = 0, e = len, tmp;
+					if (k == ls) s = (int32_t)gt->a[off_a+st].x + 1 - (int32_t)(gt->a[off_a+st].y>>32&0xff);
+					if (k == le) e = (int32_t)gt->a[off_a+en].x + 1;
+					if (v&1) tmp = s, s = len - e, e = len - tmp;
+					n_ovlp = mg_intv_overlap(km, soff[(v>>1)+1] - soff[v>>1], &sintv[soff[v>>1]], s, e, &ovlp, &m_ovlp);
+					if (n_ovlp == 0) fprintf(stderr, "[W::%s] graph interval %s:%d-%d is not covered by %s:%d-%d\n", __func__, g->seg[v>>1].name, s, e, seq[t].name, I.coff[0], I.coff[1]); // this should be an assert()
+					if (n_ovlp != 1) break;
+				}
+				if (k <= le) continue;
+				if (pd - (I.coff[1] - I.coff[0]) < opt->min_var_len && (I.coff[1] - I.coff[0]) - pd < opt->min_var_len) { // if length difference > min_var_len, just insert
+					int32_t qd = I.coff[1] - I.coff[0], mlen, blen, score;
+					l_pseq = mg_path2seq(km, g, gt, ls, le, I.voff, &pseq, &m_pseq);
+					score = mg_wfa_cmp(km, l_pseq, pseq, qd, &seq[t].seq[I.coff[0]], 5000, &mlen, &blen);
+					if (score > 0) {
+						if (mlen > blen * opt->ggs_max_iden) continue; // make sure k-mer identity is small enough
+						if (blen - mlen < opt->min_var_len * opt->ggs_max_iden) continue;
+					} else if (!(opt->flag & MG_G_NO_INV)) {
+						mg_revcomp_seq(l_pseq, pseq);
+						score = mg_wfa_cmp(km, l_pseq, pseq, qd, &seq[t].seq[I.coff[0]], 5000, &mlen, &blen);
+						if (score > 0 && mlen > blen * opt->ggs_min_inv_iden) is_inv = 1;
+					}
+				}
+				if (mg_dbg_flag & MG_DBG_INSERT) {
+					int32_t mlen, blen, score, qd = I.coff[1] - I.coff[0];
+					l_pseq = mg_path2seq(km, g, gt, ls, le, I.voff, &pseq, &m_pseq);
+					fprintf(stderr, "IN\t[%c%s:%d,%c%s:%d|%d] <=> %s:[%d,%d|%d] inv:%d\n", "><"[I.v[0]&1], g->seg[I.v[0]>>1].name, I.voff[0], "><"[I.v[1]&1], g->seg[I.v[1]>>1].name, I.voff[1], pd, seq[t].name, I.coff[0], I.coff[1], I.coff[1] - I.coff[0], is_inv);
+					fprintf(stderr, "IP\t%s\nIQ\t", pseq);
+					fwrite(&seq[t].seq[I.coff[0]], 1, qd, stderr);
+					if (pd - qd < opt->min_var_len && qd - pd < opt->min_var_len) {
+						score = mg_wfa_cmp(km, l_pseq, pseq, qd, &seq[t].seq[I.coff[0]], 5000, &mlen, &blen);
+					} else score = -1, mlen = 0, blen = pd > qd? pd : qd;
+					fprintf(stderr, "\nIS\t%d==%d\tnwcmp:%d\tmlen:%d\tblen:%d\n", pd, l_pseq, score, mlen, blen);
+				}
+				if (is_inv) { // turn one inversion to two events
+					gfa_ins_t I_inv[2];
+					I_inv[0].ctg = I_inv[1].ctg = I.ctg;
+					// the first event
+					I_inv[0].coff[0] = I_inv[0].coff[1] = I.coff[0];
+					I_inv[0].v[0] = I.v[0];
+					I_inv[0].voff[0] = I.voff[0];
+					I_inv[0].v[1] = I.v[1]^1;
+					I_inv[0].voff[1] = g->seg[I.v[1]>>1].len - I.voff[1];
+					// the second event
+					I_inv[1].coff[0] = I_inv[1].coff[1] = I.coff[1];
+					I_inv[1].v[0] = I.v[0]^1;
+					I_inv[1].voff[0] = g->seg[I.v[0]>>1].len - I.voff[0];
+					I_inv[1].v[1] = I.v[1];
+					I_inv[1].voff[1] = I.voff[1];
+					// insert
+					if (n_ins == m_ins) KEXPAND(km, ins, m_ins);
+					ins[n_ins++] = I_inv[0];
+					if (n_ins == m_ins) KEXPAND(km, ins, m_ins);
+					ins[n_ins++] = I_inv[1];
+					++n_inv;
+				} else {
+					if (n_ins == m_ins) KEXPAND(km, ins, m_ins);
+					ins[n_ins++] = I;
+				}
+			}
+			kfree(0, ss);
+		}
+	}
+	kfree(km, pseq);
+	kfree(km, ovlp);
+	kfree(km, sc);
+	kfree(km, meta);
+	kfree(km, soff); kfree(km, qoff);
+	kfree(km, sintv); kfree(km, qintv);
+
+	if (n_ins > 0) {
+		char **names, **seqs;
+		KMALLOC(km, names, n_seq);
+		KMALLOC(km, seqs, n_seq);
+		for (i = 0; i < n_seq; ++i)
+			names[i] = seq[i].name, seqs[i] = seq[i].seq;
+		n_ins = gfa_ins_filter(g, n_ins, ins);
+		gfa_augment(g, n_ins, ins, n_seq, (const char*const*)names, (const char*const*)seqs);
+		kfree(km, ins);
+		kfree(km, names);
+		kfree(km, seqs);
+	}
+	if (mg_verbose >= 3)
+		fprintf(stderr, "[M::%s::%.3f*%.2f] inserted %d events, including %d inversions\n", __func__,
+				realtime() - mg_realtime0, cputime() / (realtime() - mg_realtime0), n_ins, n_inv);
+}
+
+/**********************
+ * Graph augmentation *
+ **********************/
+
+typedef struct {
+	int32_t lc, vo, qo, po, len, op, sc;
+} ed_intv_t;
+
+static int32_t gg_count_intv(const gfa_t *g, const mg_gchains_t *gt, int32_t i)
+{
+	const mg_gchain_t *gc = &gt->gc[i];
+	int32_t j, l = gc->off, x = gc->ps, n = 0;
+	assert(gc->p);
+	for (j = 0; j < gc->p->n_cigar; ++j) {
+		int32_t op = gc->p->cigar[j]&0xf, len = gc->p->cigar[j]>>4, rl = len;
+		assert(op == 1 || op == 2 || op == 7 || op == 8);
+		if (op == 2 || op == 7 || op == 8) {
+			while (x + rl > g->seg[gt->lc[l].v>>1].len) {
+				rl -= g->seg[gt->lc[l].v>>1].len - x;
+				++n, ++l, x = 0;
+			}
+			x += rl;
+		}
+		++n;
+	}
+	return n;
+}
+
+static void gg_write_intv(const gfa_t *g, const mg_gchains_t *gt, int32_t i, ed_intv_t *intv)
+{
+	const mg_gchain_t *gc = &gt->gc[i];
+	int32_t j, l = gc->off, pl = 0, x = gc->ps, y = gc->qs, n = 0;
+	ed_intv_t *p;
+	assert(gc->p);
+	for (j = 0; j < gc->p->n_cigar; ++j) {
+		int32_t op = gc->p->cigar[j]&0xf, len = gc->p->cigar[j]>>4, rl = len;
+		if (op == 2 || op == 7 || op == 8) {
+			while (x + rl > g->seg[gt->lc[l].v>>1].len) {
+				p = &intv[n++];
+				p->lc = l, p->vo = x, p->qo = y, p->po = pl, p->len = g->seg[gt->lc[l].v>>1].len - x, p->op = op;
+				if (op == 7 || op == 8) y += p->len;
+				rl -= p->len, pl += p->len, ++l, x = 0;
+			}
+		}
+		p = &intv[n++];
+		p->lc = l, p->vo = x, p->qo = y, p->po = pl, p->len = rl, p->op = op;
+		if (op == 7 || op == 8) x += rl, y += rl, pl += rl;
+		else if (op == 1) y += rl;
+		else if (op == 2) x += rl, pl += rl;
+	}
+	assert(y == gc->qe && pl == gc->pe - gc->ps);
+}
+
+static void gg_score_intv(int32_t n_intv, ed_intv_t *intv)
+{
+	int32_t j;
+	for (j = 0; j < n_intv; ++j) {
+		int32_t s;
+		if (intv[j].op == 7)
+			s = intv[j].len >= 10? -intv[j].len : 0;
+		else s = intv[j].len;
+		intv[j].sc = s;
+	}
+}
+
+static void gg_merge_seg(const ed_intv_t *intv, int32_t n_ss, mg_msseg_t *ss)
+{
+	int32_t j0, j;
+	for (j0 = 0, j = 1; j < n_ss; ++j) {
+		mg_msseg_t *s0 = &ss[j0], *s1 = &ss[j];
+		int32_t i, mid = 0;
+		for (i = s0->en + 1; i < s1->st; ++i)
+			mid += intv[i].sc;
+		//fprintf(stderr, "XX\t%d\t%d\t%d\t%d\t%d\t%d\n", j, s0->sc, mid, s1->sc, s0->en+1, s1->st);
+		if (-mid < s0->sc * 0.2 && -mid < s1->sc * 0.2) { // FIXME: mid is sometimes 0
+			s0->en = s1->en, s0->sc += s1->sc + mid;
+			s1->st = s1->en, s1->sc = 0;
+		} else j0 = j;
+	}
+}
+
+void mg_ggsimple_cigar(void *km, const mg_ggopt_t *opt, gfa_t *g, int32_t n_seq, const mg_bseq1_t *seq, mg_gchains_t *const* gcs)
+{
+	int32_t t, i, *soff, *qoff, max_acnt, m_ovlp = 0, *ovlp = 0, n_ins = 0, m_ins, n_inv;
+	int32_t l_pseq, m_pseq;
+	mg_intv_t *sintv, *qintv;
+	double a_dens;
+	gfa_ins_t *ins;
+	char *pseq;
+
+	max_acnt = mg_gc_index(km, opt->min_mapq, opt->min_map_len, opt->min_depth_len, g, n_seq, gcs, &a_dens, &soff, &qoff, &sintv, &qintv);
+	if (max_acnt == 0) return;
+
+	// extract poorly regions
+	m_pseq = l_pseq = 0, pseq = 0;
+	m_ins = n_ins = 0, ins = 0;
+	n_inv = 0;
+	for (t = 0; t < n_seq; ++t) {
+		const mg_gchains_t *gt = gcs[t];
+		for (i = 0; i < gt->n_gc; ++i) {
+			const mg_gchain_t *gc = &gt->gc[i];
+			int32_t j, n_ss, n_intv, *sc;
+			ed_intv_t *intv;
+			mg_msseg_t *ss;
+			if (gc->id != gc->parent) continue;
+			if (gc->p == 0 || gc->blen < opt->min_map_len || gc->mapq < opt->min_mapq) continue;
+			assert(gc->cnt > 0);
+
+			n_intv = gg_count_intv(g, gt, i);
+			KCALLOC(km, intv, n_intv);
+			gg_write_intv(g, gt, i, intv);
+			gg_score_intv(n_intv, intv);
+			KCALLOC(km, sc, n_intv);
+			for (j = 0; j < n_intv; ++j) sc[j] = intv[j].sc;
+			ss = mg_mss_all(0, n_intv, sc, opt->min_var_len, 2 * opt->min_var_len, &n_ss);
+			gg_merge_seg(intv, n_ss, ss);
+
+			// get regions to insert
+			for (j = 0; j < n_ss; ++j) {
+				int32_t st, en, pd, k, n_ovlp, min_len, is_inv = 0, ls, le;
+				gfa_ins_t I;
+				ed_intv_t *is, *ie;
+
+				// find the initial positions
+				st = ss[j].st, en = ss[j].en; // this is a CLOSED interval
+				if (st == en) continue;
+				is = &intv[st], ie = &intv[en - 1];
+				assert(is->op != 7 && ie->op != 7);
+
+				ls = is->lc, le = ie->lc;
+				I.ctg = t;
+				I.v[0] = gt->lc[ls].v;
+				I.v[1] = gt->lc[le].v;
+				I.voff[0] = is->vo;
+				I.voff[1] = ie->vo + (ie->op != 1? ie->len : 0);
+				I.coff[0] = is->qo;
+				I.coff[1] = ie->qo + (ie->op != 2? ie->len : 0);
+				assert(I.voff[0] <= g->seg[I.v[0]>>1].len);
+				assert(I.voff[1] <= g->seg[I.v[1]>>1].len);
+
+				if (I.voff[0] == 0) { // if an insert starts at pos 0, make it start at the end of the previous vertex in the chain
+					assert(ls - 1 >= gc->off);
+					I.v[0] = gt->lc[--ls].v;
+					I.voff[0] = g->seg[I.v[0]>>1].len;
+				}
+				if (I.voff[1] == g->seg[I.v[1]>>1].len) { // if an insert ends at the end of the vertex, make it end at the beginning of the next vertex
+					assert(le + 1 < gc->off + gc->cnt);
+					I.v[1] = gt->lc[++le].v;
+					I.voff[1] = 0;
+				}
+
+				pd = ie->po + (ie->op != 1? ie->len : 0) - is->po;
+				pd -= gfa_ins_adj(g, opt->ggs_shrink_pen, &I, seq[t].seq);
+
+				min_len = pd > I.coff[1] - I.coff[0]? pd : I.coff[1] - I.coff[0];
+				if (I.coff[0] <= min_len || I.coff[1] >= seq[t].l_seq - min_len) continue; // test if the event is close to ends again
+
+				// filtering
+				if (I.coff[1] - I.coff[0] < opt->min_var_len && pd < opt->min_var_len)
+					continue;
+				for (k = I.coff[0]; k < I.coff[1]; ++k) { // test ambiguous bases
+					int c = seq[t].seq[k];
+					if (c == 'n' || c == 'N') break;
+				}
+				if (k != I.coff[1]) continue; // no ambiguous bases on the insert
+				n_ovlp = mg_intv_overlap(km, qoff[t+1] - qoff[t], &qintv[qoff[t]], I.coff[0], I.coff[1], &ovlp, &m_ovlp); // test overlapping on the query
+				if (n_ovlp == 0) fprintf(stderr, "[W::%s] query interval %s:%d-%d is not covered\n", __func__, seq[t].name, I.coff[0], I.coff[1]);
+				if (n_ovlp != 1) continue;
+				for (k = is->lc; k <= ie->lc; ++k) { // find other mappings overlapping with the insert on the graph
+					uint32_t v = gt->lc[k].v, len = g->seg[v>>1].len;
+					int32_t s = 0, e = len, tmp;
+					if (k == is->lc) s = is->vo;
+					if (k == ie->lc) e = ie->vo + (ie->op != 1? ie->len : 0);
+					if (v&1) tmp = s, s = len - e, e = len - tmp;
+					if (s == e) {
+						if (s == 0) ++e;
+						else --s;
+					}
+					n_ovlp = mg_intv_overlap(km, soff[(v>>1)+1] - soff[v>>1], &sintv[soff[v>>1]], s, e, &ovlp, &m_ovlp);
+					if (n_ovlp == 0) fprintf(stderr, "[W::%s] graph interval %c%s:%d-%d is not covered by %s:%d-%d\n", __func__, "><"[v&1], g->seg[v>>1].name, s, e, seq[t].name, I.coff[0], I.coff[1]); // this should be an assert()
+					if (n_ovlp != 1) break;
+				}
+				if (k <= ie->lc) continue;
+				if (pd - (I.coff[1] - I.coff[0]) < opt->min_var_len && (I.coff[1] - I.coff[0]) - pd < opt->min_var_len) { // if length difference > min_var_len, just insert
+					int32_t qd = I.coff[1] - I.coff[0], mlen, blen, score = 0;
+					l_pseq = mg_path2seq(km, g, gt, ls, le, I.voff, &pseq, &m_pseq);
+					score = mg_wfa_cmp(km, l_pseq, pseq, qd, &seq[t].seq[I.coff[0]], 5000, &mlen, &blen);
+					if (score > 0) {
+						if (mlen > blen * opt->ggs_max_iden) continue; // make sure k-mer identity is small enough
+						if (blen - mlen < opt->min_var_len * opt->ggs_max_iden) continue;
+					} else if (!(opt->flag & MG_G_NO_INV)) {
+						mg_revcomp_seq(l_pseq, pseq);
+						score = mg_wfa_cmp(km, l_pseq, pseq, qd, &seq[t].seq[I.coff[0]], 5000, &mlen, &blen);
+						if (score > 0 && mlen > blen * opt->ggs_min_inv_iden) is_inv = 1;
+					}
+				}
+				if (mg_dbg_flag & MG_DBG_INSERT) {
+					int32_t mlen, blen, score, qd = I.coff[1] - I.coff[0];
+					l_pseq = mg_path2seq(km, g, gt, ls, le, I.voff, &pseq, &m_pseq);
+					fprintf(stderr, "IN\t[%c%s:%d,%c%s:%d|%d] <=> %s:[%d,%d|%d] inv:%d\n", "><"[I.v[0]&1], g->seg[I.v[0]>>1].name, I.voff[0], "><"[I.v[1]&1], g->seg[I.v[1]>>1].name, I.voff[1], pd, seq[t].name, I.coff[0], I.coff[1], I.coff[1] - I.coff[0], is_inv);
+					fprintf(stderr, "IP\t%s\nIQ\t", pseq);
+					fwrite(&seq[t].seq[I.coff[0]], 1, qd, stderr);
+					if (pd - qd < opt->min_var_len && qd - pd < opt->min_var_len) {
+						score = mg_wfa_cmp(km, l_pseq, pseq, qd, &seq[t].seq[I.coff[0]], 5000, &mlen, &blen);
+					} else score = -1, mlen = 0, blen = pd > qd? pd : qd;
+					fprintf(stderr, "\nIS\t%d==%d\tnwcmp:%d\tmlen:%d\tblen:%d\n", pd, l_pseq, score, mlen, blen);
+					//if (I.voff[0] == 2305301) { for (k = st; k < en; ++k) fprintf(stderr, "%d%c", intv[k].len, "MIDNSHP=XB"[intv[k].op]); fprintf(stderr, "\n"); }
+				}
+				if (is_inv) { // turn one inversion to two events
+					gfa_ins_t I_inv[2];
+					I_inv[0].ctg = I_inv[1].ctg = I.ctg;
+					// the first event
+					I_inv[0].coff[0] = I_inv[0].coff[1] = I.coff[0];
+					I_inv[0].v[0] = I.v[0];
+					I_inv[0].voff[0] = I.voff[0];
+					I_inv[0].v[1] = I.v[1]^1;
+					I_inv[0].voff[1] = g->seg[I.v[1]>>1].len - I.voff[1];
+					// the second event
+					I_inv[1].coff[0] = I_inv[1].coff[1] = I.coff[1];
+					I_inv[1].v[0] = I.v[0]^1;
+					I_inv[1].voff[0] = g->seg[I.v[0]>>1].len - I.voff[0];
+					I_inv[1].v[1] = I.v[1];
+					I_inv[1].voff[1] = I.voff[1];
+					// insert
+					if (n_ins == m_ins) KEXPAND(km, ins, m_ins);
+					ins[n_ins++] = I_inv[0];
+					if (n_ins == m_ins) KEXPAND(km, ins, m_ins);
+					ins[n_ins++] = I_inv[1];
+					++n_inv;
+				} else {
+					if (n_ins == m_ins) KEXPAND(km, ins, m_ins);
+					ins[n_ins++] = I;
+				}
+			}
+			kfree(0, ss); // this is allocated from malloc() inside mg_mss_all()
+			kfree(km, intv);
+			kfree(km, sc);
+		}
+	}
+	kfree(km, pseq);
+	kfree(km, ovlp);
+	kfree(km, soff); kfree(km, qoff);
+	kfree(km, sintv); kfree(km, qintv);
+
+	if (n_ins > 0) {
+		char **names, **seqs;
+		KMALLOC(km, names, n_seq);
+		KMALLOC(km, seqs, n_seq);
+		for (i = 0; i < n_seq; ++i)
+			names[i] = seq[i].name, seqs[i] = seq[i].seq;
+		n_ins = gfa_ins_filter(g, n_ins, ins);
+		gfa_augment(g, n_ins, ins, n_seq, (const char*const*)names, (const char*const*)seqs);
+		kfree(km, ins);
+		kfree(km, names);
+		kfree(km, seqs);
+	}
+	if (mg_verbose >= 3)
+		fprintf(stderr, "[M::%s::%.3f*%.2f] inserted %d events, including %d inversions\n", __func__,
+				realtime() - mg_realtime0, cputime() / (realtime() - mg_realtime0), n_ins, n_inv);
+}