/* Function: p7_oprofile_ReconfigUnihit()
* Synopsis: Quickly reconfig model into unihit mode for target length <L>.
* Incept: MSF Tue Nov 3, 2009 [Janelia]
*
* Purpose: Given a profile <om> that's already been configured once,
* quickly reconfigure it into a unihit mode for target
* length <L>.
*
* This gets called in domain definition, when we need to
* flip the model in and out of unihit <L=0> mode to
* process individual domains.
*/
int
p7_oprofile_ReconfigUnihit(P7_OPROFILE *om, int L)
{
return p7_ReconfigUnihit(om, L);
}
/* Function: p7_domaindef_GlocalByPosteriorHeuristics()
* Synopsis: Define glocal domains in a sequence using posterior probs.
* Incept: EPN, Tue Oct 5 10:02:34 2010
* SRE, Sat Feb 23 08:17:44 2008 [Janelia] (p7_domaindef_ByPosteriorHeuristics())
*
* Purpose: Given a sequence <sq> and model <gm> for which we have
* already calculated a Forward and Backward parsing
* matrices <gxf> and <gxb>; use posterior probability
* heuristics to determine an annotated domain structure;
* and for each domain found, score it (with null2
* calculations) and obtain an optimal accuracy alignment,
* using <fwd> and <bck> matrices as workspace for the
* necessary full-matrix DP calculations. Caller provides a
* new or reused <ddef> object to hold these results.
*
* As a special case, if the profile is in unihit mode
* upon entering, we don't ever modify its configuration.
* This is especially important if this function is
* being used within a search/scan pipeline with a
* specially configured p7 profile in which N->N and/or
* C->C transitions have been set to IMPOSSIBLE. (If
* we were to call ReconfigLength() on such a profile
* we would make those transitions possible.)
*
* One case in which profile reconfiguration is necessary
* is when multiple domains are suspected. However, we
* guard against this if the profile enters in unihit mode
* by no allowing multiple domains (in fact, it should
* never happen because J states are unreachable in unihit
* profiles). If multiple domains are suspected in this case,
* we return eslEINCONCEIVABLE.
*
* Upon return, <ddef> contains the definitions of all the
* domains: their bounds, their null-corrected Forward
* scores, and their optimal posterior accuracy alignments.
*
* <do_null2> is TRUE if we'll eventually apply a null2
* penalty FALSE if not. If FALSE, we can save time by
* skipping Backward calls at some stages.
*
* Returns: <eslOK> on success.
*
* <eslERANGE> on numeric overflow in posterior
* decoding. This should not be possible for multihit
* models.
*
* <eslEINCONCEIVABLE> if profile enters as unihit but
* multiple domains are suspected.
*/
int
p7_domaindef_GlocalByPosteriorHeuristics(const ESL_SQ *sq, P7_PROFILE *gm,
P7_GMX *gxf, P7_GMX *gxb, P7_GMX *fwd, P7_GMX *bck,
P7_DOMAINDEF *ddef, int do_null2)
{
int i, j;
int triggered;
int d;
int i2,j2;
int last_j2;
int nc;
int saveL = gm->L; /* Save the length config of <om>; will restore upon return */
int save_mode = gm->mode; /* Likewise for the mode. */
int status;
int save_mode_is_unihit;
save_mode_is_unihit = (p7_IsMulti(save_mode)) ? FALSE : TRUE; /* if save_mode_is_unihit is TRUE, we never modify profile's configuration (length nor mode) */
if ((status = p7_domaindef_GrowTo(ddef, sq->n)) != eslOK) return status; /* ddef's btot,etot,mocc now ready for seq of length n */
/*printf("GDD P7 mode: %d\n", gm->mode);*/
if ((status = p7_GDomainDecoding(gm, gxf, gxb, ddef)) != eslOK) return status; /* ddef->{btot,etot,mocc} now made. */
/*printf("In p7_domaindef_GlocalByPosteriorHeuristics(): mode: %d rt1: %g rt2: %g rt3: %g nsamples: %d reseed: %d\n", save_mode, ddef->rt1, ddef->rt2, ddef->rt3, ddef->nsamples, ddef->do_reseeding);*/
esl_vec_FSet(ddef->n2sc, sq->n+1, 0.0); /* ddef->n2sc null2 scores are initialized */
ddef->nexpected = ddef->btot[sq->n]; /* posterior expectation for # of domains (same as etot[sq->n]) */
if(! save_mode_is_unihit) p7_ReconfigUnihit(gm, saveL); /* process each domain in unihit mode, regardless of gm->mode */
i = -1;
triggered = FALSE;
for (j = 1; j <= sq->n; j++)
{
/*printf("GDD j: %5d m: %.5f b: %8.3f e: %8.3f bhere: %8.3f ehere: %8.3f\n",
j,
ddef->mocc[j],
ddef->btot[j],
ddef->etot[j],
ddef->btot[j] - ddef->btot[j-1],
ddef->etot[j] - ddef->etot[j-1]);
*/
if (! triggered)
{ /* xref J2/101 for what the logic below is: */
if (ddef->mocc[j] - (ddef->btot[j] - ddef->btot[j-1]) < ddef->rt2) i = j;
else if (i == -1) i = j;
if (ddef->mocc[j] >= ddef->rt1) triggered = TRUE;
}
else if (ddef->mocc[j] - (ddef->etot[j] - ddef->etot[j-1]) < ddef->rt2)
{
/* We have a region i..j to evaluate. */
p7_gmx_GrowTo(fwd, gm->M, j-i+1);
p7_gmx_GrowTo(bck, gm->M, j-i+1);
ddef->nregions++;
if (is_multidomain_region(ddef, i, j))
{
if(save_mode_is_unihit) return eslEINCONCEIVABLE;
/* This region appears to contain more than one domain, so we have to
* resolve it by cluster analysis of posterior trace samples, to define
* one or more domain envelopes.
*/
ddef->nclustered++;
/* Resolve the region into domains by stochastic trace
* clustering; assign position-specific null2 model by
* stochastic trace clustering; there is redundancy
* here; we will consolidate later if null2 strategy
* works
*/
p7_ReconfigMultihit(gm, saveL);
p7_GForward(sq->dsq+i-1, j-i+1, gm, fwd, NULL);
glocal_region_trace_ensemble(ddef, gm, sq->dsq, i, j, fwd, bck, do_null2, &nc);
p7_ReconfigUnihit(gm, saveL);
/* ddef->n2sc is now set on i..j by the traceback-dependent method */
last_j2 = 0;
for (d = 0; d < nc; d++) {
p7_spensemble_GetClusterCoords(ddef->sp, d, &i2, &j2, NULL, NULL, NULL);
if (i2 <= last_j2) ddef->noverlaps++;
/* Note that k..m coords on model are available, but
* we're currently ignoring them. This leads to a
* rare clustering bug that we eventually need to fix
* properly [xref J3/32]: two different regions in one
* profile HMM might have hit same seq domain, and
* when we now go to calculate an OA trace, nothing
* constrains us to find the two different alignments
* to the HMM; in fact, because OA is optimal, we'll
* find one and the *same* alignment, leading to an
* apparent duplicate alignment in the output.
*
* Registered as #h74, Dec 2009, after EBI finds and
* reports it. #h74 is worked around in p7_tophits.c
* by hiding all but one envelope with an identical
* alignment, in the rare event that this
* happens. [xref J5/130].
*/
ddef->nenvelopes++;
if (glocal_rescore_isolated_domain(ddef, gm, sq, fwd, bck, i2, j2, TRUE, do_null2, FALSE) == eslOK)
last_j2 = j2;
}
p7_spensemble_Reuse(ddef->sp);
p7_trace_Reuse(ddef->tr);
}
else
{
/* The region looks simple, single domain; convert the region to an envelope. */
ddef->nenvelopes++;
glocal_rescore_isolated_domain(ddef, gm, sq, fwd, bck, i, j, FALSE, do_null2, FALSE);
}
i = -1;
triggered = FALSE;
}
}
/* If profile was unihit upon entrance, we didn't modify its configuration (length nor mode),
* else restore it to its original multihit mode, and to its original length model */
if (! save_mode_is_unihit) {
p7_ReconfigMultihit(gm, saveL);
}
return eslOK;
}