/* Function: LogSum2()
*
* Purpose: Returns the log_2 of the sum of two log_2 probabilities.
* log(exp(p1)+exp(p2)) = p1 + log(1 + exp(p2-p1)) for p1 > p2
* Note that this is in log_2 space.
*/
float
LogSum2(float p1, float p2)
{
if (p1 > p2)
return (p1-p2 > 50.) ? p1 : p1 + sreLOG2(1. + pow(2.,(p2-p1)));
else
return (p2-p1 > 50.) ? p2 : p2 + sreLOG2(1. + pow(2.,(p1-p2)));
}
/* yes LogSum2 and FLogsum are identical, this is for backwards compatibility */
float
FLogsum(float s1, float s2)
{
const float max = ESL_MAX(s1, s2);
const float min = ESL_MIN(s1, s2);
#if 0
return (min == -eslINFINITY || (max-min) >= 23.f) ? max : max + sreLOG2(1.0 + sreEXP2(min-max)); /* EPN: While debugging. Replaces logsum table with analytical calculation. Remember to remove! */
#endif
return (min == -eslINFINITY || (max-min) >= 23.f) ? max : max + flogsum_lookup[(int)((max-min)*INTSCALE)];
}
void
init_ilogsum(void)
{
static int firsttime = TRUE;
if (!firsttime) return;
firsttime = FALSE;
int i;
for (i = 0; i < LOGSUM_TBL; i++)
ilogsum_lookup[i] = rint(INTSCALE * (sreLOG2(1.+sreEXP2((double) -i/INTSCALE))));
}
void
FLogsumInit(void)
{
static int firsttime = TRUE;
if (!firsttime) return;
firsttime = FALSE;
int i;
for (i = 0; i < LOGSUM_TBL; i++)
flogsum_lookup[i] = sreLOG2(1. + sreEXP2((double) -i / INTSCALE));
return;
}
/* Function: Prob2Score()
*
* Purpose: Convert a probability to a scaled integer log_2 odds score.
* Round to nearest integer (i.e. note use of +0.5 and floor())
* Return the score.
*/
int
Prob2Score(float p, float null)
{
/*EPN, Fri Feb 8 06:53:02 2008, shouldn't this first line be something like:
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
if (fabs(p - 0.) < eslSMALLX1) return -INFTY; LHS is TRUE if p is virtually 0 (eslSMALLX1 is 5e-9)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
I didn't change it pre-1.0 release, b/c I was afraid of unforeseen downstream consequences,
and it doesn't seem to cause any problems...
*/
if (p == 0.0) return -INFTY;
else return (int) floor(0.5 + INTSCALE * sreLOG2(p/null));
}
/* Function: DegenerateSymbolScore()
*
* Purpose: Given a sequence character x and an hmm emission probability
* vector, calculate the log-odds (base 2) score of
* the symbol.
*
* Easy if x is in the emission alphabet, but not so easy
* is x is a degenerate symbol. The "correct" Bayesian
* philosophy is to calculate score(X) by summing over
* p(x) for all x in the degenerate symbol X to get P(X),
* doing the same sum over the prior to get F(X), and
* doing log_2 (P(X)/F(X)). This gives an X a zero score,
* for instance.
*
* Though this is correct in a formal Bayesian sense --
* we have no information on the sequence, so we can't
* say if it's random or model, so it scores zero --
* it sucks, big time, for scoring biological sequences.
* Sequences with lots of X's score near zero, while
* real sequences have average scores that are negative --
* so the X-laden sequences appear to be lifted out
* of the noise of a full histogram of a database search.
* Correct or not, this is highly undesirable.
*
* So therefore we calculated the expected score of
* the degenerate symbol by summing over all x in X:
* e_x log_2 (p(x)/f(x))
* where the expectation of x, e_x, is calculated from
* the random model.
*
* Empirically, this works; it also has a wooly hand-waving
* probabilistic justification that I'm happy enough about.
*
* Args: p - probabilities of normal symbols
* null - null emission model
* ambig - index of the degenerate character in Alphabet[]
*
* Return: the integer log odds score of x given the emission
* vector and the null model, scaled up by INTSCALE.
*/
int
DegenerateSymbolScore(float *p, float *null, int ambig)
{
int x;
float numer = 0.;
float denom = 0.;
for (x = 0; x < Alphabet_size; x++) {
if (Degenerate[ambig][x]) {
numer += null[x] * sreLOG2(p[x] / null[x]);
denom += null[x];
}
}
return (int) (INTSCALE * numer / denom);
}
/* Function: cm_tr_penalties_Create()
* Date: EPN, Sat Jan 21 12:03:52 2012
*
* Purpose: Allocate and initialize a CM_TR_PENALTIES object.
* A CM and its emit map are required to determine
* truncation penalty scores. This is annoyingly
* complex, see verbose notes within code below.
*
* Some of the code in this function, specifically
* that which calculates the probability of a fragment
* aligning at a given node, is checkable, but only
* if we disallow truncated begins into insert states.
* However, we want to allow truncated begins in reality.
* I've left in a flag for ignoring inserts (<ignore_inserts>)
* I used in testing this function. Set it to TRUE to
* perform the test.
*
* Returns: Newly allocated CM_TR_PENALTIES object. NULL if out
* of memory.
*/
CM_TR_PENALTIES *
cm_tr_penalties_Create(CM_t *cm, int ignore_inserts, char *errbuf)
{
int status;
int v, nd, m, i1, i2;
int lpos, rpos;
int i;
/* variables used for determining ratio of inserts to match at each consensus position */
float *mexpocc = NULL; /* [0..c..clen] probability match state is used to emit at cons posn c */
float *iexpocc = NULL; /* [0..c..clen] probability insert state is used to emit after cons posn c */
double *psi = NULL; /* [0..v..M-1] expected occupancy of state v */
float m_psi, i1_psi, i2_psi; /* temp psi values */
float summed_psi;
CM_TR_PENALTIES *trp = NULL;
/* variables used for calculating global truncation penalties */
float g_5and3; /* fragment probability if 5' and 3' truncation are allowed */
float g_5or3; /* fragment probability if 5' or 3' truncation are allowed */
/* variables used for calculating local truncation penalties */
float *begin = NULL; /* local begin probabilities 0..v..M-1 */
int subtree_clen; /* consensus length of subtree under this node */
float prv53, prv5, prv3; /* previous node's fragment probability, 5'&3', 5' only, 3'only */
float cur53, cur5, cur3; /* current node's fragment probability, 5'&3', 5' only, 3'only */
int nfrag53, nfrag5, nfrag3; /* number of fragments, 5'&3', 5' only, 3'only */
if(cm == NULL || cm->emap == NULL) goto ERROR;
ESL_ALLOC(trp, sizeof(CM_TR_PENALTIES));
trp->M = cm->M;
trp->ignored_inserts = ignore_inserts;
/* Define truncation penalties for each state v. This will be
* the score for doing a truncated begin into state v.
*
* Important note: For this discussion we assume that sequences can
* only be truncated at consensus positions, which means we don't
* have to worry about truncated begins into inserts. This is an
* approximation (also made by Diana and Sean in the 2009 trCYK
* paper) that greatly simplifies the explanation of the calculation
* of the truncation penalties. The examples in my ELN3 notebook
* also use this simplification. However, I need to be able to do
* truncated begins into insert states in some cases (some pass/mode
* combinations see ELN bottom of p.47). I explain first the
* rationale for calculating truncation penalties ignoring inserts
* and then I describe how I adapt those penalties to allow
* for inserts.
*
* This is a lengthy comment. I've divided it into 3 sections:
* Section 1. Global mode truncation penalties, ignoring inserts.
* Section 2. Local mode truncation penalties, ignoring inserts.
* Section 3. Adapting truncation penalties to allow for inserts.
*
**************************************************************
* Section 1. Global mode truncation penalties, ignoring inserts.
*
* We want the truncation penalty to be the log of the probability
* that the particular fragment we're aligning was generated from
* the following generative process. The generative process differs
* between global and local mode.
*
* In global mode:
* o Sample global parsetree which spans consensus positions 1..clen.
* o Randomly choose g and h in range 1..clen, where h >= g and
* truncate sequence from g..h. The first residue will either be
* an insert before position g, or a match at position g of the
* model. The final residue will either be an insert after position
* h or a match at position h of the model.
*
* All g,h fragments are equiprobable, so the probability of any
* particular fragment is 2 / (clen * (clen+1)). So log_2 of this
* value is the truncation penalty for all truncated alignments in
* global mode where both 5' and 3' truncation are allowed.
*
* We store this penalty, per-state in the
* g_ptyAA[TRPENALTY_5P_AND_3P][0..v..M-1]. The penalty is
* identical for all emitting states. The penalty value for
* non-emitters is IMPOSSIBLE because truncated begins are
* not allowed into non-emitters.
*
* If only 5' OR 3' truncation is allowed, we only truncate at g or
* h, which menas there's 1/clen possible fragments and log_2
* (1/clen) is our global truncation penalty.
*
* However, if 5' truncation is allowed we can only do a truncated
* begin into states that with a consensus subtree that spans
* position clen (since we don't allow a truncation at the 3' end).
* Thus any state whose subtree that doesn't span clen gets
* an IMPOSSIBLE value for its truncation score in:
* g_ptyAA[TRPENALTY_5P_ONLY][0..v..M-1].
*
* Likewise, if 3' truncation is allowed we can only do a truncated
* begin into states that with a consensus subtree that spans
* position 1 (since we don't allow a truncation at the 5' end).
*
* There's an example of computing all three types of penalties for
* a simple CM in ELN 3 p43.
*
************************************************************
* Section 2. Local mode truncation penalties, ignoring inserts.
*
* Generative process that generates fragments in local mode:
* o Sample local begin state b with consensus subtree from i..j from
* local begin state distribution.
* o Randomly choose g and h in range i..j, where h >= g and
* truncate sequence from g..h. The first residue will either be
* an insert before position g, or a match at position g of the
* model. The final residue will either be an insert after position
* h or a match at position h of the model.
*
* Unlike in global mode, in local mode all fragments are not
* equiprobable since the local begin state distribution can be
* anything, and each b allows different sets of fragments to be
* generated (because they can only span from i to j).
*
* The truncation penalty should be the log of the probability of
* aligning the current fragment to the model. So we need to know
* the probability of generating each possible fragment.
* We could calculate probability of any fragment g,h with the
* following inefficient algorithm:
*
* For each start fragment point g,
* For each start fragment point h,
* For each state v,
* If lpos[v] <= g && rpos[v] >= h, then
* prob[g][h] += begin[v] * 2. / (st_clen[v] * (st_clen[v]+1));
*
* Where lpos[v]/rpos[v] are the left/right consensus positions in
* consensus subtree rooted at state v. And st_clen[v] is rpos[v] -
* lpos[v] + 1, the consensus length of that subtree.
*
* This gives us prob[g][h], the probability of generating fragment
* g,h. But we want to apply the penalty to a state, not to a
* fragment, to avoid needing to know the fragment boundaries g,h
* during the DP recursion when applying the penalty.
*
* To facilitate this, we need to find state t, the state with
* smallest subtree that contains g,h. State t is relevant because
* it is the state which will root the alignment of the fragment g,h
* by using a truncated begin transition into t. This gives a new
* algorithm:
*
* For each start fragment point g,
* For each start fragment point h,
* Identify state t, the max valued state for which
* lpos[v] <= g && rpos[v] >= h, then {
* prob[t] += prob[g][h]
* fcount[t]++;
* }
*
* prob[t] will be the probability of observing an alignment that
* uses a truncated begin into t to align any fragment. Then we take
* average over all fragments: prob[t] / fcount[t] (since we'll only
* be aligning one of those fragments) and use the log of that
* probability as the penalty for observing a truncated alignment
* rooted at state t. Conveniently, it turns out that all fragments
* that share t are equiprobable (have equal prob[g][h] values), so
* the average probability is the actual probability for each
* fragment, and thus the correct penalty to apply.
*
* Fortunately, we can compute the correct penalty much more
* efficiently than the two algorithms shown above. The
* efficient way is implemented below. A test that the penalties
* are correctly computed is in cm_tr_penalties_Validate().
*
* This discussion assumes we're truncating 5' and 3', but if we're
* only truncating 5' or 3' The situation is a little different.
*
* There's an example of computing all three types of penalties for
* a simple CM in ELN3 p44-45.
*
************************************************************
* Section 3. Adapting truncation penalties to allow for inserts.
*
* We need to be able to do truncated begins into insert states
* because we enforce that the first/final residue of a sequence be
* included in 5'/3' truncated alignments and we want to be able
* to properly align those residues if they're probably emitted
* by insert states.
*
* The methods/logic explained in sections 1 and 2 above I believe
* is correct IF we ignore inserts (assume truncated begins into
* them are impossible). But we need to allow inserts, so I modify
* the truncation penalties as described above to allow for inserts
* as follows. We can calculate the appropriate truncated begin
* penalty for all MATP_MP, MATL_ML, MATR_MR, BIF_B states as with
* the methods described above by ignoring inserts. This gives us a
* probability p of using that state as the root of the truncated
* alignment, i.e. the truncated begin state. (The log_2 of this
* probability is the penalty.) We then partition p amongst the
* MATP_MP, MATL_ML, MATR_MR, BIF_B states and any parent insert
* states, i.e. any insert state that can transition into the
* match/bif state. For each match/bif state there's 0, 1 or 2
* parent inserts. We then partition p based on the relative
* expected occupancy of these inserts versus the match/bif state.
*
* This is certainly 'incorrect' in that it doesn't reflect the
* true probability of a fragment being aligned to each of the
* states, but it should be a close approximation. I think doing
* it correctly is basically impossible in the context of a single
* state-specific penalty (i.e. the penalty would have to be per-fragment
* which would be hard to deal with in the DP functions).
*/
/* allocate and initialize the penalty arrays */
ESL_ALLOC(trp->g_ptyAA, sizeof(float *) * NTRPENALTY);
ESL_ALLOC(trp->l_ptyAA, sizeof(float *) * NTRPENALTY);
ESL_ALLOC(trp->ig_ptyAA, sizeof(int *) * NTRPENALTY);
ESL_ALLOC(trp->il_ptyAA, sizeof(int *) * NTRPENALTY);
for(i = 0; i < NTRPENALTY; i++) {
trp->g_ptyAA[i] = NULL;
trp->l_ptyAA[i] = NULL;
trp->il_ptyAA[i] = NULL;
trp->ig_ptyAA[i] = NULL;
ESL_ALLOC(trp->g_ptyAA[i], sizeof(float) * cm->M);
ESL_ALLOC(trp->l_ptyAA[i], sizeof(float) * cm->M);
ESL_ALLOC(trp->ig_ptyAA[i], sizeof(int) * cm->M);
ESL_ALLOC(trp->il_ptyAA[i], sizeof(int) * cm->M);
esl_vec_FSet(trp->g_ptyAA[i], cm->M, IMPOSSIBLE);
esl_vec_FSet(trp->l_ptyAA[i], cm->M, IMPOSSIBLE);
esl_vec_ISet(trp->ig_ptyAA[i], cm->M, -INFTY);
esl_vec_ISet(trp->il_ptyAA[i], cm->M, -INFTY);
}
/* DumpEmitMap(stdout, cm->emap, cm); */
/* Calculate local begin probabilities and expected occupancy */
ESL_ALLOC(begin, sizeof(float) * cm->M);
cm_CalculateLocalBeginProbs(cm, cm->pbegin, cm->t, begin);
if((status = cm_ExpectedPositionOccupancy(cm, &mexpocc, &iexpocc, &psi, NULL, NULL, NULL)) != eslOK) goto ERROR;
/* Fill global and local truncation penalties in a single loop. We
* step through all nodes and set the truncation penalties for the
* MATP_MP, MATL_ML, MATR_MR, and BIF_B states and any parent
* inserts (i1, i2) of those states.
*/
g_5and3 = 2. / (cm->clen * (cm->clen+1)); /* for global mode: probability of all fragments if we're truncating 5' and 3' */
g_5or3 = 1. / cm->clen; /* for global mode: probability of all fragments if we're only truncating 5' or 3' */
prv5 = prv3 = prv53 = 0.; /* initialize 'previous' probability values used for calc'ing local truncation penalties */
for(nd = 0; nd < cm->nodes; nd++) {
lpos = (cm->ndtype[nd] == MATP_nd || cm->ndtype[nd] == MATL_nd) ? cm->emap->lpos[nd] : cm->emap->lpos[nd] + 1;
rpos = (cm->ndtype[nd] == MATP_nd || cm->ndtype[nd] == MATR_nd) ? cm->emap->rpos[nd] : cm->emap->rpos[nd] - 1;
/* now set penalties for match and insert states m, i1 and maybe i2 (if we're a MATP_MP or BIF_B) */
if(cm->ndtype[nd] == END_nd) {
prv5 = prv3 = prv53 = 0.;
}
else if(cm->ndtype[nd] == BEGL_nd || cm->ndtype[nd] == BEGR_nd) {
prv5 = (cm->ndtype[nd] == BEGL_nd) ? 0. : trp->l_ptyAA[TRPENALTY_5P_ONLY][cm->plast[cm->nodemap[nd]]]; /* parent BIF_B's probability */;
prv3 = (cm->ndtype[nd] == BEGR_nd) ? 0. : trp->l_ptyAA[TRPENALTY_3P_ONLY][cm->plast[cm->nodemap[nd]]]; /* parent BIF_B's probability */;
prv53 = trp->l_ptyAA[TRPENALTY_5P_AND_3P][cm->plast[cm->nodemap[nd]]]; /* parent BIF_B's probability */
}
else if(cm->ndtype[nd] == MATP_nd || cm->ndtype[nd] == MATL_nd || cm->ndtype[nd] == MATR_nd || cm->ndtype[nd] == BIF_nd) {
/* determine match states and insert states that pertain to this node */
m = cm->nodemap[nd]; /* MATP_MP, MATL_ML, MATR_MR, or BIF_B */
InsertsGivenNodeIndex(cm, nd-1, &i1, &i2);
m_psi = psi[m];
if(cm->ndtype[nd] == MATP_MP) { m_psi += (psi[m+1] + psi[m+2]); } /* include MATP_ML and MATP_MR psi */
i1_psi = (i1 == -1) ? 0. : psi[i1];
i2_psi = (i2 == -1) ? 0. : psi[i2];
summed_psi = m_psi + i1_psi + i2_psi;
if(ignore_inserts) {
i1_psi = i2_psi = 0.;
summed_psi = m_psi;
}
/* Global penalties */
/* sanity check, we should only set truncation penalty once per state */
if(NOT_IMPOSSIBLE(trp->g_ptyAA[TRPENALTY_5P_AND_3P][m])) goto ERROR;
if((i1 != -1) && NOT_IMPOSSIBLE(trp->g_ptyAA[TRPENALTY_5P_AND_3P][i1])) goto ERROR;
if((i2 != -1) && NOT_IMPOSSIBLE(trp->g_ptyAA[TRPENALTY_5P_AND_3P][i2])) goto ERROR;
/* divide up the probability g_5and3 amongst relevant states m, i1, i2, weighted by psi */
trp->g_ptyAA[TRPENALTY_5P_AND_3P][m] = (m_psi / summed_psi) * g_5and3;
if(i1 != -1) trp->g_ptyAA[TRPENALTY_5P_AND_3P][i1] = (i1_psi / summed_psi) * g_5and3;
if(i2 != -1) trp->g_ptyAA[TRPENALTY_5P_AND_3P][i2] = (i2_psi / summed_psi) * g_5and3;
/* same thing, for 5P only and 3P only */
if(rpos == cm->clen) { /* else it will remain IMPOSSIBLE */
trp->g_ptyAA[TRPENALTY_5P_ONLY][m] = (m_psi / summed_psi) * g_5or3;
if(i1 != -1) trp->g_ptyAA[TRPENALTY_5P_ONLY][i1] = (i1_psi / summed_psi) * g_5or3;
if(i2 != -1) trp->g_ptyAA[TRPENALTY_5P_ONLY][i2] = (i2_psi / summed_psi) * g_5or3;
}
if(lpos == 1) { /* else it will remain IMPOSSIBLE */
trp->g_ptyAA[TRPENALTY_3P_ONLY][m] = (m_psi / summed_psi) * g_5or3;
if(i1 != -1) trp->g_ptyAA[TRPENALTY_3P_ONLY][i1] = (i1_psi / summed_psi) * g_5or3;
if(i2 != -1) trp->g_ptyAA[TRPENALTY_3P_ONLY][i2] = (i2_psi / summed_psi) * g_5or3;
}
/* Local penalties */
subtree_clen = rpos - lpos + 1;
nfrag5 = subtree_clen;
nfrag3 = subtree_clen;
nfrag53 = (subtree_clen * (subtree_clen+1)) / 2;
/* determine probability of observing a fragment aligned at
* state m (here, m is what I call t above and in notes) and
* partition that probability between m and i1 and/or i2 by
* relative occupancy of match versus inserts
*/
cur5 = begin[m] / (float) nfrag5 + prv5;
cur3 = begin[m] / (float) nfrag3 + prv3;
cur53 = begin[m] / (float) nfrag53 + prv53;
/* sanity check, we should only set truncation penalty once per state */
if(NOT_IMPOSSIBLE(trp->l_ptyAA[TRPENALTY_5P_AND_3P][m])) goto ERROR;
if((i1 != -1) && NOT_IMPOSSIBLE(trp->l_ptyAA[TRPENALTY_5P_AND_3P][i1])) goto ERROR;
if((i2 != -1) && NOT_IMPOSSIBLE(trp->l_ptyAA[TRPENALTY_5P_AND_3P][i2])) goto ERROR;
trp->l_ptyAA[TRPENALTY_5P_AND_3P][m] = (m_psi / summed_psi) * cur53;
if(i1 != -1) trp->l_ptyAA[TRPENALTY_5P_AND_3P][i1] = (i1_psi / summed_psi) * cur53;
if(i2 != -1) trp->l_ptyAA[TRPENALTY_5P_AND_3P][i2] = (i2_psi / summed_psi) * cur53;
trp->l_ptyAA[TRPENALTY_5P_ONLY][m] = (m_psi / summed_psi) * cur5;
if(i1 != -1) trp->l_ptyAA[TRPENALTY_5P_ONLY][i1] = (i1_psi / summed_psi) * cur5;
if(i2 != -1) trp->l_ptyAA[TRPENALTY_5P_ONLY][i2] = (i2_psi / summed_psi) * cur5;
trp->l_ptyAA[TRPENALTY_3P_ONLY][m] = (m_psi / summed_psi) * cur3;
if(i1 != -1) trp->l_ptyAA[TRPENALTY_3P_ONLY][i1] = (i1_psi / summed_psi) * cur3;
if(i2 != -1) trp->l_ptyAA[TRPENALTY_3P_ONLY][i2] = (i2_psi / summed_psi) * cur3;
prv5 = (cm->ndtype[nd] == MATL_nd) ? cur5 : 0.;
prv3 = (cm->ndtype[nd] == MATR_nd) ? cur3 : 0.;
prv53 = cur53;
}
}
/* all penalties are currently probabilities, convert them to log
* probs and set integer penalties (careful, we have to check if
* IMPOSSIBLE first)
*/
for(v = 0; v < cm->M; v++)
{
if((cm->stid[v] == MATP_MP || cm->stid[v] == MATL_ML || cm->stid[v] == MATR_MR || cm->stid[v] == BIF_B) ||
((cm->sttype[v] == IL_st || cm->sttype[v] == IR_st) && (! StateIsDetached(cm, v))))
{
/* Check for rare special case: if we're a MATP_IL and next
* two states are MATP_IR and END_E, then we won't have set
* a trunction penalty. This state will keep an impossible
* truncated begin score, if we did a truncated begin into
* it we'd just emit from the MATP_IL and then go to the
* END_E anyway (the MATP_IR will be detached.
*/
if(cm->stid[v] == MATP_IL && cm->ndtype[cm->ndidx[v]+1] == END_nd) continue;
/* glocal 5P AND 3P: all of these should have been set to a non-IMPOSSIBLE value */
if(! NOT_IMPOSSIBLE(trp->g_ptyAA[TRPENALTY_5P_AND_3P][v])) goto ERROR;
trp->ig_ptyAA[TRPENALTY_5P_AND_3P][v] = Prob2Score(trp->g_ptyAA[TRPENALTY_5P_AND_3P][v], 1.0);
trp->g_ptyAA[TRPENALTY_5P_AND_3P][v] = sreLOG2(trp->g_ptyAA[TRPENALTY_5P_AND_3P][v]);
/* glocal 5P only: some may be IMPOSSIBLE */
if(NOT_IMPOSSIBLE(trp->g_ptyAA[TRPENALTY_5P_ONLY][v])) {
trp->ig_ptyAA[TRPENALTY_5P_ONLY][v] = Prob2Score(trp->g_ptyAA[TRPENALTY_5P_ONLY][v], 1.0);
trp->g_ptyAA[TRPENALTY_5P_ONLY][v] = sreLOG2(trp->g_ptyAA[TRPENALTY_5P_ONLY][v]);
}
/* glocal 5P only: some may be IMPOSSIBLE */
if(NOT_IMPOSSIBLE(trp->g_ptyAA[TRPENALTY_3P_ONLY][v])) {
trp->ig_ptyAA[TRPENALTY_3P_ONLY][v] = Prob2Score(trp->g_ptyAA[TRPENALTY_3P_ONLY][v], 1.0);
trp->g_ptyAA[TRPENALTY_3P_ONLY][v] = sreLOG2(trp->g_ptyAA[TRPENALTY_3P_ONLY][v]);
}
/* local penalties all of these should have been set to a non-IMPOSSIBLE value */
if(! NOT_IMPOSSIBLE(trp->il_ptyAA[TRPENALTY_5P_AND_3P][v])) goto ERROR;
if(! NOT_IMPOSSIBLE(trp->il_ptyAA[TRPENALTY_5P_ONLY][v])) goto ERROR;
if(! NOT_IMPOSSIBLE(trp->il_ptyAA[TRPENALTY_3P_ONLY][v])) goto ERROR;
trp->il_ptyAA[TRPENALTY_5P_AND_3P][v] = Prob2Score(trp->l_ptyAA[TRPENALTY_5P_AND_3P][v], 1.0);
trp->il_ptyAA[TRPENALTY_5P_ONLY][v] = Prob2Score(trp->l_ptyAA[TRPENALTY_5P_ONLY][v], 1.0);
trp->il_ptyAA[TRPENALTY_3P_ONLY][v] = Prob2Score(trp->l_ptyAA[TRPENALTY_3P_ONLY][v], 1.0);
trp->l_ptyAA[TRPENALTY_5P_AND_3P][v] = sreLOG2(trp->l_ptyAA[TRPENALTY_5P_AND_3P][v]);
trp->l_ptyAA[TRPENALTY_5P_ONLY][v] = sreLOG2(trp->l_ptyAA[TRPENALTY_5P_ONLY][v]);
trp->l_ptyAA[TRPENALTY_3P_ONLY][v] = sreLOG2(trp->l_ptyAA[TRPENALTY_3P_ONLY][v]);
}
}
if(ignore_inserts) {
if((status = cm_tr_penalties_Validate(trp, cm, 0.0001, errbuf)) != eslOK) { printf("%s", errbuf); goto ERROR; }
}
/* cm_tr_penalties_Dump(stdout, cm, trp); */
if(mexpocc != NULL) free(mexpocc);
if(iexpocc != NULL) free(iexpocc);
if(psi != NULL) free(psi);
if(begin != NULL) free(begin);
return trp;
ERROR:
if(mexpocc != NULL) free(mexpocc);
if(iexpocc != NULL) free(iexpocc);
if(psi != NULL) free(psi);
if(begin != NULL) free(begin);
if(trp != NULL) cm_tr_penalties_Destroy(trp);
return NULL;
}
