/* The MSV score can be validated against Viterbi (provided we trust
* Viterbi), by creating a multihit local profile in which:
* 1. All t_MM scores = 0
* 2. All other core transitions = -inf
* 3. All t_BMk entries uniformly log 2/(M(M+1))
*/
static void
utest_msv(ESL_GETOPTS *go, ESL_RANDOMNESS *r, ESL_ALPHABET *abc, P7_BG *bg, P7_PROFILE *gm, int nseq, int L)
{
P7_PROFILE *g2 = NULL;
ESL_DSQ *dsq = NULL;
P7_GMX *gx = NULL;
float sc1, sc2;
int k, idx;
if ((dsq = malloc(sizeof(ESL_DSQ) *(L+2))) == NULL) esl_fatal("malloc failed");
if ((gx = p7_gmx_Create(gm->M, L)) == NULL) esl_fatal("matrix creation failed");
if ((g2 = p7_profile_Clone(gm)) == NULL) esl_fatal("profile clone failed");
/* Make g2's scores appropriate for simulating the MSV algorithm in Viterbi */
esl_vec_FSet(g2->tsc, p7P_NTRANS * g2->M, -eslINFINITY);
for (k = 1; k < g2->M; k++) p7P_TSC(g2, k, p7P_MM) = 0.0f;
for (k = 0; k < g2->M; k++) p7P_TSC(g2, k, p7P_BM) = log(2.0f / ((float) g2->M * (float) (g2->M+1)));
for (idx = 0; idx < nseq; idx++)
{
if (esl_rsq_xfIID(r, bg->f, abc->K, L, dsq) != eslOK) esl_fatal("seq generation failed");
if (p7_GMSV (dsq, L, gm, gx, 2.0, &sc1) != eslOK) esl_fatal("MSV failed");
if (p7_GViterbi(dsq, L, g2, gx, &sc2) != eslOK) esl_fatal("viterbi failed");
if (fabs(sc1-sc2) > 0.0001) esl_fatal("MSV score not equal to Viterbi score");
}
p7_gmx_Destroy(gx);
p7_profile_Destroy(g2);
free(dsq);
return;
}
/* Function: p7_profile_SameAsMF()
* Synopsis: Set a generic profile's scores to give MSV scores.
* Incept: MSF Tue Nov 3, 2009 [Janelia]
*
* Purpose: Set a generic profile's scores so that the normal <dp_generic> DP
* algorithms will give the same score as <p7_MSVFilter()>:
* all t_MM scores = 0; all other core transitions = -inf;
* multihit local mode; all <t_BMk> entries uniformly <log 2/(M(M+1))>;
* <tCC, tNN, tJJ> scores 0; total approximated later as -3;
* rounded in the same way as the 8-bit limited precision.
*
* Returns: <eslOK> on success.
*/
int
p7_profile_SameAsMF(const P7_OPROFILE *om, P7_PROFILE *gm)
{
int k;
float tbm = log(2.0f / ((float) gm->M * (float) (gm->M+1)));
/* Transitions */
esl_vec_FSet(gm->tsc, p7P_NTRANS * gm->M, -eslINFINITY);
for (k = 1; k < gm->M; k++) p7P_TSC(gm, k, p7P_MM) = 0.0f;
for (k = 0; k < gm->M; k++) p7P_TSC(gm, k, p7P_BM) = tbm;
return eslOK;
}
/* sample_endpoints()
* Incept: SRE, Mon Jan 22 10:43:20 2007 [Janelia]
*
* Purpose: Given a profile <gm> and random number source <r>, sample
* a begin transition from the implicit probabilistic profile
* model, yielding a sampled start and end node; return these
* via <ret_kstart> and <ret_kend>.
*
* By construction, the entry at node <kstart> is into a
* match state, but the exit from node <kend> might turn
* out to be from either a match or delete state.
*
* We assume that exits j are uniformly distributed for a
* particular entry point i: $a_{ij} =$ constant $\forall
* j$.
*
* Returns: <eslOK> on success.
*
* Throws: <eslEMEM> on allocation error.
*
* Xref: STL11/138
*/
static int
sample_endpoints(ESL_RANDOMNESS *r, const P7_PROFILE *gm, int *ret_kstart, int *ret_kend)
{
float *pstart = NULL;
int k;
int kstart, kend;
int status;
/* We have to backcalculate a probability distribution from the
* lod B->Mk scores in a local model; this is a little time consuming,
* but we don't have to do it often.
*/
ESL_ALLOC(pstart, sizeof(float) * (gm->M+1));
pstart[0] = 0.0f;
for (k = 1; k <= gm->M; k++)
pstart[k] = exp(p7P_TSC(gm, k-1, p7P_BM)) * (gm->M - k + 1); /* multiply p_ij by the number of exits j */
kstart = esl_rnd_FChoose(r, pstart, gm->M+1); /* sample the starting position from that distribution */
kend = kstart + esl_rnd_Roll(r, gm->M-kstart+1); /* and the exit uniformly from possible exits for it */
free(pstart);
*ret_kstart = kstart;
*ret_kend = kend;
return eslOK;
ERROR:
if (pstart != NULL) free(pstart);
*ret_kstart = 0;
*ret_kend = 0;
return status;
}
/* Function: p7_oprofile_GetFwdTransitionArray()
* Synopsis: Retrieve full 32-bit float transition probabilities from a
* profile into a flat array
*
* Purpose: Extract an array of <type> (e.g. p7O_II) transition probabilities
* from the underlying <om> profile. In SIMD implementations,
* these are striped and interleaved, making them difficult to
* directly access. Here, this is trivial.
*
* Args: <om> - optimized profile, containing transition information
* <type> - transition type (e.g. p7O_II)
* <arr> - preallocated array into which floats will be placed
*
* Returns: <eslOK> on success.
*
* Throws: (no abnormal error conditions)
*/
int
p7_oprofile_GetFwdTransitionArray(const P7_OPROFILE *om, int type, float *arr )
{
int i;
for (i=0; i<om->M; i++) {
arr[i] = exp(p7P_TSC(om, i, type));
}
return eslOK;
}
/* Function: p7_ProfileConfig()
* Synopsis: Configure a search profile.
*
* Purpose: Given a model <hmm> with core probabilities, the null1
* model <bg>, a desired search <mode> (one of <p7_LOCAL>,
* <p7_GLOCAL>, <p7_UNILOCAL>, or <p7_UNIGLOCAL>), and an
* expected target sequence length <L>; configure the
* search model in <gm> with lod scores relative to the
* background frequencies in <bg>.
*
* Returns: <eslOK> on success; the profile <gm> now contains
* scores and is ready for searching target sequences.
*
* Throws: <eslEMEM> on allocation error.
*/
int
p7_ProfileConfig(const P7_HMM *hmm, const P7_BG *bg, P7_PROFILE *gm, int L, int mode)
{
int k, x, z; /* counters over states, residues, annotation */
int status;
float *occ = NULL;
float *tp, *rp;
float sc[p7_MAXCODE];
float Z;
/* Contract checks */
if (gm->abc->type != hmm->abc->type) ESL_XEXCEPTION(eslEINVAL, "HMM and profile alphabet don't match");
if (hmm->M > gm->allocM) ESL_XEXCEPTION(eslEINVAL, "profile too small to hold HMM");
if (! (hmm->flags & p7H_CONS)) ESL_XEXCEPTION(eslEINVAL, "HMM must have a consensus to transfer to the profile");
/* Copy some pointer references and other info across from HMM */
gm->M = hmm->M;
gm->max_length = hmm->max_length;
gm->mode = mode;
gm->roff = -1;
gm->eoff = -1;
gm->offs[p7_MOFFSET] = -1;
gm->offs[p7_FOFFSET] = -1;
gm->offs[p7_POFFSET] = -1;
if (gm->name != NULL) free(gm->name);
if (gm->acc != NULL) free(gm->acc);
if (gm->desc != NULL) free(gm->desc);
if ((status = esl_strdup(hmm->name, -1, &(gm->name))) != eslOK) goto ERROR;
if ((status = esl_strdup(hmm->acc, -1, &(gm->acc))) != eslOK) goto ERROR;
if ((status = esl_strdup(hmm->desc, -1, &(gm->desc))) != eslOK) goto ERROR;
if (hmm->flags & p7H_RF) strcpy(gm->rf, hmm->rf);
if (hmm->flags & p7H_MMASK) strcpy(gm->mm, hmm->mm);
if (hmm->flags & p7H_CONS) strcpy(gm->consensus, hmm->consensus); /* must be present, actually, so the flag test is just for symmetry w/ other optional HMM fields */
if (hmm->flags & p7H_CS) strcpy(gm->cs, hmm->cs);
for (z = 0; z < p7_NEVPARAM; z++) gm->evparam[z] = hmm->evparam[z];
for (z = 0; z < p7_NCUTOFFS; z++) gm->cutoff[z] = hmm->cutoff[z];
for (z = 0; z < p7_MAXABET; z++) gm->compo[z] = hmm->compo[z];
/* Entry scores. */
if (p7_profile_IsLocal(gm))
{
/* Local mode entry: occ[k] /( \sum_i occ[i] * (M-i+1))
* (Reduces to uniform 2/(M(M+1)) for occupancies of 1.0) */
Z = 0.;
ESL_ALLOC(occ, sizeof(float) * (hmm->M+1));
if ((status = p7_hmm_CalculateOccupancy(hmm, occ, NULL)) != eslOK) goto ERROR;
for (k = 1; k <= hmm->M; k++)
Z += occ[k] * (float) (hmm->M-k+1);
for (k = 1; k <= hmm->M; k++)
p7P_TSC(gm, k-1, p7P_BM) = log(occ[k] / Z); /* note off-by-one: entry at Mk stored as [k-1][BM] */
free(occ);
}
else /* glocal modes: left wing retraction; must be in log space for precision */
{
Z = log(hmm->t[0][p7H_MD]);
p7P_TSC(gm, 0, p7P_BM) = log(1.0 - hmm->t[0][p7H_MD]);
for (k = 1; k < hmm->M; k++)
{
p7P_TSC(gm, k, p7P_BM) = Z + log(hmm->t[k][p7H_DM]);
Z += log(hmm->t[k][p7H_DD]);
}
}
/* E state loop/move probabilities: nonzero for MOVE allows loops/multihits
* N,C,J transitions are set later by length config
*/
if (p7_profile_IsMultihit(gm)) {
gm->xsc[p7P_E][p7P_MOVE] = -eslCONST_LOG2;
gm->xsc[p7P_E][p7P_LOOP] = -eslCONST_LOG2;
gm->nj = 1.0f;
} else {
gm->xsc[p7P_E][p7P_MOVE] = 0.0f;
gm->xsc[p7P_E][p7P_LOOP] = -eslINFINITY;
gm->nj = 0.0f;
}
/* Transition scores. */
for (k = 1; k < gm->M; k++) {
tp = gm->tsc + k * p7P_NTRANS;
tp[p7P_MM] = log(hmm->t[k][p7H_MM]);
tp[p7P_MI] = log(hmm->t[k][p7H_MI]);
tp[p7P_MD] = log(hmm->t[k][p7H_MD]);
tp[p7P_IM] = log(hmm->t[k][p7H_IM]);
tp[p7P_II] = log(hmm->t[k][p7H_II]);
tp[p7P_DM] = log(hmm->t[k][p7H_DM]);
tp[p7P_DD] = log(hmm->t[k][p7H_DD]);
}
/* Match emission scores. */
sc[hmm->abc->K] = -eslINFINITY; /* gap character */
sc[hmm->abc->Kp-2] = -eslINFINITY; /* nonresidue character */
sc[hmm->abc->Kp-1] = -eslINFINITY; /* missing data character */
for (k = 1; k <= hmm->M; k++) {
for (x = 0; x < hmm->abc->K; x++)
sc[x] = log((double)hmm->mat[k][x] / bg->f[x]);
esl_abc_FExpectScVec(hmm->abc, sc, bg->f);
for (x = 0; x < hmm->abc->Kp; x++) {
rp = gm->rsc[x] + k * p7P_NR;
rp[p7P_MSC] = sc[x];
}
}
/* Insert emission scores */
/* SRE, Fri Dec 5 08:41:08 2008: We currently hardwire insert scores
* to 0, i.e. corresponding to the insertion emission probabilities
* being equal to the background probabilities. Benchmarking shows
* that setting inserts to informative emission distributions causes
* more problems than it's worth: polar biased composition hits
* driven by stretches of "insertion" occur, and are difficult to
* correct for.
*/
for (x = 0; x < gm->abc->Kp; x++)
{
for (k = 1; k < hmm->M; k++) p7P_ISC(gm, k, x) = 0.0f;
p7P_ISC(gm, hmm->M, x) = -eslINFINITY; /* init I_M to impossible. */
}
for (k = 1; k <= hmm->M; k++) p7P_ISC(gm, k, gm->abc->K) = -eslINFINITY; /* gap symbol */
for (k = 1; k <= hmm->M; k++) p7P_ISC(gm, k, gm->abc->Kp-2) = -eslINFINITY; /* nonresidue symbol */
for (k = 1; k <= hmm->M; k++) p7P_ISC(gm, k, gm->abc->Kp-1) = -eslINFINITY; /* missing data symbol */
#if 0
/* original (informative) insert setting: relies on sc[K, Kp-1] initialization to -inf above */
for (k = 1; k < hmm->M; k++) {
for (x = 0; x < hmm->abc->K; x++)
sc[x] = log(hmm->ins[k][x] / bg->f[x]);
esl_abc_FExpectScVec(hmm->abc, sc, bg->f);
for (x = 0; x < hmm->abc->Kp; x++) {
rp = gm->rsc[x] + k*p7P_NR;
rp[p7P_ISC] = sc[x];
}
}
for (x = 0; x < hmm->abc->Kp; x++)
p7P_ISC(gm, hmm->M, x) = -eslINFINITY; /* init I_M to impossible. */
#endif
/* Remaining specials, [NCJ][MOVE | LOOP] are set by ReconfigLength()
*/
gm->L = 0; /* force ReconfigLength to reconfig */
if ((status = p7_ReconfigLength(gm, L)) != eslOK) goto ERROR;
return eslOK;
ERROR:
if (occ != NULL) free(occ);
return status;
}
/* Function: p7_ProfileConfig()
* Synopsis: Configure a search profile.
* Incept: SRE, Sun Sep 25 12:21:25 2005 [St. Louis]
*
* Purpose: Given a model <hmm> with core probabilities, the null1
* model <bg>, a desired search <mode> (one of <p7_LOCAL>,
* <p7_GLOCAL>, <p7_UNILOCAL>, or <p7_UNIGLOCAL>), and an
* expected target sequence length <L>; configure the
* search model in <gm> with lod scores relative to the
* background frequencies in <bg>.
*
* Returns: <eslOK> on success; the profile <gm> now contains
* scores and is ready for searching target sequences.
*
* Throws: <eslEMEM> on allocation error.
*/
int
p7_ProfileConfig(const P7_HMM *hmm, const P7_BG *bg, P7_PROFILE *gm, int L, int mode)
{
int k, x, z; /* counters over states, residues, annotation */
int status;
float *occ = NULL;
float *tp, *rp;
float sc[p7_MAXCODE];
float mthresh;
float Z;
/* Contract checks */
if (gm->abc->type != hmm->abc->type) ESL_XEXCEPTION(eslEINVAL, "HMM and profile alphabet don't match");
if (hmm->M > gm->allocM) ESL_XEXCEPTION(eslEINVAL, "profile too small to hold HMM");
/* Copy some pointer references and other info across from HMM */
gm->M = hmm->M;
gm->mode = mode;
gm->roff = -1;
gm->eoff = -1;
gm->offs[p7_MOFFSET] = -1;
gm->offs[p7_FOFFSET] = -1;
gm->offs[p7_POFFSET] = -1;
if (gm->name != NULL) free(gm->name);
if (gm->acc != NULL) free(gm->acc);
if (gm->desc != NULL) free(gm->desc);
if ((status = esl_strdup(hmm->name, -1, &(gm->name))) != eslOK) goto ERROR;
if ((status = esl_strdup(hmm->acc, -1, &(gm->acc))) != eslOK) goto ERROR;
if ((status = esl_strdup(hmm->desc, -1, &(gm->desc))) != eslOK) goto ERROR;
if (hmm->flags & p7H_RF) strcpy(gm->rf, hmm->rf);
if (hmm->flags & p7H_CS) strcpy(gm->cs, hmm->cs);
for (z = 0; z < p7_NEVPARAM; z++) gm->evparam[z] = hmm->evparam[z];
for (z = 0; z < p7_NCUTOFFS; z++) gm->cutoff[z] = hmm->cutoff[z];
for (z = 0; z < p7_MAXABET; z++) gm->compo[z] = hmm->compo[z];
/* Determine the "consensus" residue for each match position.
* This is only used for alignment displays, not in any calculations.
*/
if (hmm->abc->type == eslAMINO) mthresh = 0.5;
else if (hmm->abc->type == eslDNA) mthresh = 0.9;
else if (hmm->abc->type == eslRNA) mthresh = 0.9;
else mthresh = 0.5;
gm->consensus[0] = ' ';
for (k = 1; k <= hmm->M; k++) {
x = esl_vec_FArgMax(hmm->mat[k], hmm->abc->K);
gm->consensus[k] = ((hmm->mat[k][x] > mthresh) ? toupper(hmm->abc->sym[x]) : tolower(hmm->abc->sym[x]));
}
gm->consensus[hmm->M+1] = '\0';
/* Entry scores. */
if (p7_profile_IsLocal(gm))
{
/* Local mode entry: occ[k] /( \sum_i occ[i] * (M-i+1))
* (Reduces to uniform 2/(M(M+1)) for occupancies of 1.0) */
Z = 0.;
ESL_ALLOC_WITH_TYPE(occ, float*, sizeof(float) * (hmm->M+1));
if ((status = p7_hmm_CalculateOccupancy(hmm, occ, NULL)) != eslOK) goto ERROR;
for (k = 1; k <= hmm->M; k++)
Z += occ[k] * (float) (hmm->M-k+1);
for (k = 1; k <= hmm->M; k++)
p7P_TSC(gm, k-1, p7P_BM) = log((double)(occ[k] / Z)); /* note off-by-one: entry at Mk stored as [k-1][BM] */
free(occ);
}
else /* glocal modes: left wing retraction; must be in log space for precision */
{
