/* Function: CPlan9SWConfig()
* EPN 05.30.06
* based on SRE's Plan7SWConfig() from HMMER's plan7.c
*
* Purpose: Set the alignment independent parameters of
* a CM Plan 9 model to hmmsw (Smith/Waterman) configuration.
*
* Notes: The desideratum for begin/end probs is that all fragments ij
* (starting at match i, ending at match j) are
* equiprobable -- there is no information in the choice of
* entry/exit. There are M(M+1)/2 possible choices of ij, so
* each must get a probability of 2/M(M+1). This prob is the
* product of a begin, an end, and all the not-end probs in
* the path between i,j.
*
* Thus: entry/exit is asymmetric because of the left/right
* nature of the HMM/profile. Entry probability is distributed
* simply by assigning p_x = pentry / (M-1) to M-1
* internal match states. However, the same approach doesn't
* lead to a flat distribution over exit points. Exit p's
* must be corrected for the probability of a previous exit
* from the model. Requiring a flat distribution over exit
* points leads to an easily solved piece of algebra, giving:
* p_1 = pexit / (M-1)
* p_x = p_1 / (1 - (x-1) p_1)
*
* Modified EPN, Thu Feb 7 15:54:16 2008, as follows:
* To better match a locally configured CM, if <do_match_local_cm>
* we disallow insertions before the first (emitting) match state,
* (from I_0), and after the final (emitting) match state,
* (from I_M). I_0 maps to ROOT_IL and I_M maps to ROOT_IR
* which can never be entered in a locally configured CM
* (b/c the ROOT_S state MUST jump into a local begin state, which
* are always match states>). Also we disallow a M_0->D_1 transition
* because these would be impossible in a locally configured CM.
*
* <do_match_local_cm> is usually TRUE, unless we're configuring
* the CP9 specifically for eventual sub CM alignment, where
* the goal is simply find the most likely start/end point
* of the alignment with this CP9 (in that case we want
* I_0 and I_M reachable).
*
* Args: hmm - the CM Plan 9 model w/ data-dep prob's valid
* pentry - probability of an internal entry somewhere;
* will be evenly distributed over M-1 match states
* pexit - probability of an internal exit somewhere;
* will be distributed over M-1 match states.
* do_match_local_cm - TRUE to make I_0, D_1 and I_M unreachable
* to better match a locally configured CM.
* first_cm_ndtype - only used if do_match_local_cm is TRUE
* if it's MATL or MATP then D_1 should be unreachable (it is in the CM)
* if it's MATR or MATP then D_M should be unreachable (it is in the CM)
*
* Return: (void)
* HMM probabilities are modified.
*/
void
CPlan9SWConfig(CP9_t *hmm, float pentry, float pexit, int do_match_local_cm, int first_cm_ndtype)
{
float basep; /* p1 for exits: the base p */
int k; /* counter over states */
float d;
/* No special (*x* states in Plan 7) states in CM Plan 9 */
/*for (k = 1; k <= hmm->M; k++) printf("before anything: end[%d]: %f\n", k, hmm->end[k]);*/
/* Configure entry.
*/
if(do_match_local_cm) {
hmm->t[0][CTMI] = 0.;
hmm->t[0][CTMM] = 0.; /* already was 0.0, transition from M_0 to M_1 is begin[1] */
hmm->t[0][CTMEL] = 0.; /* already was 0.0, can never do a local end from M_0 */
if((first_cm_ndtype == MATL_nd) || (first_cm_ndtype == MATP_nd)) { /* CM can't possibly reach the CM delete state that maps to D_1, make D_1 unreachable too */
hmm->t[0][CTMD] = 0.;
}
hmm->t[hmm->M][CTMI] = 0.;
hmm->t[hmm->M][CTDI] = 0.;
if((first_cm_ndtype == MATR_nd) || (first_cm_ndtype == MATP_nd)) { /* CM can't possibly reach the CM delete state that maps to D_M, make D_M unreachable too */
hmm->t[hmm->M][CTMD] = 0.;
}
/* renormalize transitions out of M_M */
d = esl_vec_FSum(hmm->t[hmm->M], cp9_TRANS_NMATCH) + hmm->end[hmm->M];
esl_vec_FScale(hmm->t[hmm->M], cp9_TRANS_NMATCH, 1./d);
hmm->end[hmm->M] /= d;
/* renormalize transitions out of D_M */
esl_vec_FNorm(hmm->t[hmm->M] + cp9_TRANS_DELETE_OFFSET, cp9_TRANS_NDELETE); /* delete */
}
hmm->begin[1] = (1. - pentry) * (1. - (hmm->t[0][CTMI] + hmm->t[0][CTMD] + hmm->t[0][CTMEL]));
esl_vec_FSet(hmm->begin+2, hmm->M-1, (pentry * (1.- (hmm->t[0][CTMI] + hmm->t[0][CTMD] + hmm->t[0][CTMEL]))) / (float)(hmm->M-1));
/* note: hmm->t[0][CTMEL] == 0. (can't locally end from begin)
* and if do_match_local_cm, hmm->t[0][CTMI] and hmm->t[0][CTMD] were just set to 0.
*/
/* Configure exit.
* Don't touch hmm->end[hmm->M]
*/
basep = pexit / (float) (hmm->M-1);
for (k = 1; k < hmm->M; k++)
hmm->end[k] = basep / (1. - basep * (float) (k-1));
CPlan9RenormalizeExits(hmm, 1);
/*for (k = 1; k <= hmm->M; k++) printf("after renormalizing: end[%d]: %f\n", k, hmm->end[k]);*/
hmm->flags &= ~CPLAN9_HASBITS; /* reconfig invalidates log-odds scores */
hmm->flags |= CPLAN9_LOCAL_BEGIN; /* local begins now on */
hmm->flags |= CPLAN9_LOCAL_END; /* local ends now on */
CP9Logoddsify(hmm);
}
void
esl_vec_FNorm(float *vec, int n)
{
int x;
float sum;
sum = esl_vec_FSum(vec, n);
if (sum != 0.0) for (x = 0; x < n; x++) vec[x] /= sum;
else for (x = 0; x < n; x++) vec[x] = 1. / (float) n;
}
/* Function: CPlan9Renormalize()
*
* Purpose: Take an HMM in counts form, and renormalize
* all of its probability vectors. Also enforces
* CM Plan9 restrictions on nonexistent transitions.
*
* Args: hmm - the model to renormalize.
*
* Return: (void)
* hmm is changed.
*/
void
CPlan9Renormalize(CP9_t *hmm)
{
int k; /* counter for model position */
float d; /* denominator */
/* match emissions */
esl_vec_FSet(hmm->mat[0], hmm->abc->K, 0.); /*M_0 is B state, non-emitter*/
for (k = 1; k <= hmm->M; k++)
esl_vec_FNorm(hmm->mat[k], hmm->abc->K);
/* insert emissions */
for (k = 0; k <= hmm->M; k++)
esl_vec_FNorm(hmm->ins[k], hmm->abc->K);
/* begin transitions */
d = esl_vec_FSum(hmm->begin+1, hmm->M) + hmm->t[0][CTMI] + hmm->t[0][CTMD] + hmm->t[0][CTMEL];
/* hmm->t[0][CTMEL] should always be 0., can't local end from the M_0 == B state */
esl_vec_FScale(hmm->begin+1, hmm->M, 1./d);
hmm->t[0][CTMI] /= d;
hmm->t[0][CTMD] /= d;
hmm->t[0][CTMEL] /= d;
esl_vec_FNorm(hmm->t[0] + cp9_TRANS_INSERT_OFFSET, cp9_TRANS_NINSERT); /* transitions out of insert for node 0 (state N)*/
esl_vec_FSet (hmm->t[0] + cp9_TRANS_DELETE_OFFSET, cp9_TRANS_NDELETE, 0.);
/* main model transitions */
for (k = 1; k <= hmm->M; k++) /* safe for node M too, hmm->t[hmm->M][CTMM] should be 0.*/
{
d = esl_vec_FSum(hmm->t[k], cp9_TRANS_NMATCH) + hmm->end[k];
esl_vec_FScale(hmm->t[k], cp9_TRANS_NMATCH, 1./d);
hmm->end[k] /= d;
esl_vec_FNorm(hmm->t[k] + cp9_TRANS_INSERT_OFFSET, cp9_TRANS_NINSERT); /* insert */
esl_vec_FNorm(hmm->t[k] + cp9_TRANS_DELETE_OFFSET, cp9_TRANS_NDELETE); /* delete */
}
/* null model emissions */
esl_vec_FNorm(hmm->null, hmm->abc->K);
hmm->flags &= ~CPLAN9_HASBITS; /* clear the log-odds ready flag */
hmm->flags |= CPLAN9_HASPROB; /* set the probabilities OK flag */
}
/* Note that calculate_occupancy has moved to p7_hmm.c, but
* unit tests over there aren't hooked up yet; so leave a copy of the unit test
* here for now.
*/
static void
utest_occupancy(P7_HMM *hmm)
{
char *msg = "modelconfig.c::calculate_occupancy() unit test failed";
float *occ;
float x;
occ = malloc(sizeof(float) * (hmm->M+1));
p7_hmm_CalculateOccupancy(hmm, occ, NULL);
x = esl_vec_FSum(occ+1, hmm->M) / (float) hmm->M;
if (esl_FCompare(x, 0.6, 0.1) != eslOK) esl_fatal(msg);
free(occ);
return;
}
static void
utest_correct_normalization(ESL_RANDOMNESS *r, P7_PROFILE *gm, P7_BG *bg, ESL_DSQ *dsq, int L, P7_GMX *fwd, P7_GMX *bck)
{
char *msg = "normalization unit test failed";
float null2[p7_MAXABET];
float sum;
int x;
esl_rsq_xfIID(r, bg->f, gm->abc->K, L, dsq); /* sample a random digital seq of length L */
p7_GForward (dsq, L, gm, fwd, NULL);
p7_GBackward(dsq, L, gm, bck, NULL);
p7_PosteriorNull2(L, gm, fwd, bck, bck); /* <bck> now contains posterior probs */
p7_Null2Corrections(gm, dsq, L, 0, bck, fwd, null2, NULL, NULL); /* use <fwd> as workspace */
/* Convert null2 from lod score to frequencies f_d */
for (x = 0; x < gm->abc->K; x++)
null2[x] = exp(null2[x]) * bg->f[x];
sum = esl_vec_FSum(null2, gm->abc->K);
if (sum < 0.99 || sum > 1.01) esl_fatal(msg);
}
/* Function: CPlan9RenormalizeExits()
* EPN 05.30.06 based on SRE's Plan7RenormalizeExits() from
* HMMER's plan7.c.
*
* Date: SRE, Fri Aug 14 11:22:19 1998 [St. Louis]
*
* Purpose: Renormalize just the match state transitions;
* for instance, after a Config() function has
* modified the exit distribution.
*
* Args: hmm - hmm to renormalize
* spos - first consensus column modelled by original
* CP9 HMM the sub CP9 HMM models. Often 1.
* Returns: void
*/
void
CPlan9RenormalizeExits(CP9_t *hmm, int spos)
{
int k;
float d;
/* We can't exit from node 0 so we start renormalizing at node 1 */
for (k = 1; k < hmm->M; k++)
{
if(k != (spos-1)) /* we can't exit from the M_spos-1 */
{
d = esl_vec_FSum(hmm->t[k], 4);
/* esl_vec_FScale(hmm->t[k], 4, 1./(d + d*hmm->end[k])); */
esl_vec_FScale(hmm->t[k], 4, (1.-hmm->end[k])/d);
}
}
/* Take care of hmm->M node, which is special */
d = hmm->t[hmm->M][CTMI] + hmm->t[hmm->M][CTMEL]; /* CTMD is IMPOSSIBLE, CTMM is hmm->end[hmm-M] */
if(! (fabs(d-0.) < eslSMALLX1)) { /* don't divide by d if it's zero */
hmm->t[hmm->M][CTMI] *= (1.-hmm->end[hmm->M])/d;
hmm->t[hmm->M][CTMEL] *= (1.-hmm->end[hmm->M])/d;
}
return;
}
/* Function: p7_bg_Read()
* Synopsis: Read background frequencies from a file.
*
* Purpose: Read new background frequencies from file <bgfile>,
* overwriting the frequencies previously in the
* <P7_BG> object <bg>.
*
* Note that <bg> is already created by the caller, not
* created here. Also note that <p7_bg_Read()> only reads
* residue background frequencies used for the "null
* model", whereas a <P7_BG> object contains additional
* information for the bias filter and for the biased
* composition correction.
*
* Args: bgfile - file to read.
* bg - existing <P7_BG> object provided by the caller.
* errbuf - OPTIONAL: space for an error message, upon parse errors; or NULL.
*
* Returns: <eslOK> on success, and background frequencies in <bg>
* are overwritten.
*
* <eslENOTFOUND> if <bgfile> can't be opened for reading.
* <eslEFORMAT> if parsing of <bgfile> fails for some
* reason. In both cases, <errbuf> contains a
* user-directed error message upon return, including (if
* relevant) the file name <bgfile> and the line number on
* which an error was detected. <bg> is unmodified.
*
* Throws: <eslEMEM> on allocation failure; <bg> is unmodified,
* and <errbuf> is empty.
*/
int
p7_bg_Read(char *bgfile, P7_BG *bg, char *errbuf)
{
ESL_FILEPARSER *efp = NULL;
float *fq = NULL;
int n = 0;
char *tok;
int toklen;
int alphatype;
ESL_DSQ x;
int status;
if (errbuf) errbuf[0] = '\0';
status = esl_fileparser_Open(bgfile, NULL, &efp);
if (status == eslENOTFOUND) ESL_XFAIL(eslENOTFOUND, errbuf, "couldn't open bg file %s for reading", bgfile);
else if (status != eslOK) goto ERROR;
esl_fileparser_SetCommentChar(efp, '#');
/* First token is alphabet type: amino | DNA | RNA */
status = esl_fileparser_GetToken(efp, &tok, &toklen);
if (status == eslEOF) ESL_XFAIL(eslEFORMAT, errbuf, "premature end of file [line %d of bgfile %s]", efp->linenumber, bgfile);
else if (status != eslOK) goto ERROR;
alphatype = esl_abc_EncodeType(tok);
if (alphatype == eslUNKNOWN) ESL_XFAIL(eslEFORMAT, errbuf, "expected alphabet type but saw \"%s\" [line %d of bgfile %s]", tok, efp->linenumber, bgfile);
else if (alphatype != bg->abc->type) ESL_XFAIL(eslEFORMAT, errbuf, "bg file's alphabet is %s; expected %s [line %d, %s]", tok, esl_abc_DecodeType(bg->abc->type), efp->linenumber, bgfile);
ESL_ALLOC(fq, sizeof(float) * bg->abc->K);
esl_vec_FSet(fq, bg->abc->K, -1.0);
while ((status = esl_fileparser_NextLine(efp)) == eslOK)
{
status = esl_fileparser_GetTokenOnLine(efp, &tok, &toklen);
if (status == eslEOL) ESL_XFAIL(eslEFORMAT, errbuf, "premature end of file [line %d of bgfile %s", efp->linenumber, bgfile);
else if (status != eslOK) goto ERROR;
if (toklen != 1 || ! esl_abc_CIsCanonical(bg->abc, *tok))
ESL_XFAIL(eslEFORMAT, errbuf, "expected to parse a residue letter; saw %s [line %d of bgfile %s]", tok, efp->linenumber, bgfile);
x = esl_abc_DigitizeSymbol(bg->abc, *tok);
if (fq[x] != -1.0) ESL_XFAIL(eslEFORMAT, errbuf, "already parsed probability of %c [line %d of bgfile %s]", bg->abc->sym[x], efp->linenumber, bgfile);
n++;
status = esl_fileparser_GetTokenOnLine(efp, &tok, &toklen);
if (status == eslEOL) ESL_XFAIL(eslEFORMAT, errbuf, "premature end of file, expected a probability [line %d of bgfile %s]", efp->linenumber, bgfile);
else if (status != eslOK) goto ERROR;
if (! esl_str_IsReal(tok)) ESL_XFAIL(eslEFORMAT, errbuf, "expected a probability, saw %s [line %d of bgfile %s]", tok, efp->linenumber, bgfile);
fq[x] = atof(tok);
status = esl_fileparser_GetTokenOnLine(efp, &tok, &toklen);
if (status == eslOK) ESL_XFAIL(eslEFORMAT, errbuf, "extra unexpected data found [line %d of bgfile %s]", efp->linenumber, bgfile);
else if (status != eslEOL) goto ERROR;
}
if (status != eslEOF) goto ERROR;
if ( n != bg->abc->K)
ESL_XFAIL(eslEFORMAT, errbuf, "expected %d residue frequencies, but found %d in bgfile %s", bg->abc->K, n, bgfile);
if ( esl_FCompare(esl_vec_FSum(fq, bg->abc->K), 1.0, 0.001) != eslOK)
ESL_XFAIL(eslEFORMAT, errbuf, "residue frequencies do not sum to 1.0 in bgfile %s", bgfile);
/* all checking complete. no more error cases. overwrite bg with the new frequencies */
esl_vec_FNorm(fq, bg->abc->K);
esl_vec_FCopy(fq, bg->abc->K, bg->f);
free(fq);
esl_fileparser_Close(efp);
return eslOK;
ERROR:
if (fq) free(fq);
if (efp) esl_fileparser_Close(efp);
return status;
}
