/* Function: p7_GNull2_ByTrace()
* Synopsis: Assign null2 scores to an envelope by the sampling method.
* Incept: SRE, Thu May 1 10:00:43 2008 [Janelia]
*
* Purpose: Given a traceback <tr> for an alignment of model <gm> to
* some target sequence; calculate null2 odds ratios $\frac{f'{x}}{f{x}}$
* as the state-usage-weighted emission probabilities,
* with state usages calculated by counting emissions used
* at positions <zstart..zend> in the trace.
*
* Because we only need to collect state usages from the
* trace <tr>, the target sequence is irrelevant. Because
* we are only averaging emission odds ratios from model
* <gm>, the configuration of <gm> is irrelevant (uni
* vs. multihit, or length config).
*
* Args: gm - model, in any configuration; only emission odds are used
* tr - traceback for any region (or all) of a target sequence
* zstart - first elem in <tr> to collect from; use 0 for complete
* zend - last elem in <tr> to collect from; use tr->N-1 for complete
* wrk - DP matrix w/ at least one row, for workspace
* null2 - RESULT: odds ratios f'(x)/f(x) for all Kp residues
*
* Returns: <eslOK> on success, and the <ddef->n2sc> scores are set
* for region <i..j>.
*
* Throws: <eslEMEM> on allocation error.
*/
int
p7_GNull2_ByTrace(const P7_PROFILE *gm, const P7_TRACE *tr, int zstart, int zend, P7_GMX *wrk, float *null2)
{
float **dp = wrk->dp; /* so that {MDI}MX() macros work */
float *xmx = wrk->xmx; /* so that XMX() macro works */
int Ld = 0;
int M = gm->M;
int k; /* index over model position */
int x; /* index over residues */
int z; /* index over trace position */
float xfactor;
/* We'll use the i=0 row in wrk for working space: dp[0][] and xmx[0..4]. */
esl_vec_FSet(wrk->dp[0], (M+1)*p7G_NSCELLS, 0.0);
esl_vec_FSet(wrk->xmx, p7G_NXCELLS, 0.0);
/* Calculate emitting state usage in this particular trace segment: */
for (z = zstart; z <= zend; z++)
{
switch (tr->st[z]) {
case p7T_M: Ld++; MMX(0,tr->k[z]) += 1.0; break;
case p7T_I: Ld++; IMX(0,tr->k[z]) += 1.0; break;
case p7T_N: if (tr->st[z-1] == p7T_N) { Ld++; XMX(0,p7G_N) += 1.0; } break;
case p7T_C: if (tr->st[z-1] == p7T_C) { Ld++; XMX(0,p7G_C) += 1.0; } break;
case p7T_J: if (tr->st[z-1] == p7T_J) { Ld++; XMX(0,p7G_J) += 1.0; } break;
}
}
esl_vec_FScale(wrk->dp[0], (M+1)*p7G_NSCELLS, (1.0 / (float) Ld));
esl_vec_FScale(wrk->xmx, p7G_NXCELLS, (1.0 / (float) Ld));
/* Calculate null2's odds ratio emission probabilities, by taking
* posterior weighted sum over all emission vectors used in paths
* explaining the domain.
*/
esl_vec_FSet(null2, gm->abc->K, 0.0);
xfactor = XMX(0,p7G_N) + XMX(0,p7G_C) + XMX(0,p7G_J);
for (x = 0; x < gm->abc->K; x++)
{
for (k = 1; k < M; k++)
{
null2[x] += MMX(0,k) * expf(p7P_MSC(gm, k, x));
null2[x] += IMX(0,k) * expf(p7P_ISC(gm, k, x));
}
null2[x] += MMX(0,M) * expf(p7P_MSC(gm, M, x));
null2[x] += xfactor;
}
/* now null2[x] = \frac{f_d(x)}{f_0(x)} odds ratios for all x in alphabet,
* 0..K-1, where f_d(x) are the ad hoc "null2" residue frequencies
* for this envelope.
*/
/* make valid scores for all degeneracies, by averaging the odds ratios. */
esl_abc_FAvgScVec(gm->abc, null2);
null2[gm->abc->K] = 1.0; /* gap character */
null2[gm->abc->Kp-2] = 1.0; /* nonresidue "*" */
null2[gm->abc->Kp-1] = 1.0; /* missing data "~" */
return eslOK;
}
/* Function: p7_GDecoding()
* Synopsis: Posterior decoding of residue assignments.
* Incept: SRE, Fri Feb 29 10:16:21 2008 [Janelia]
*
* Purpose: Calculates a posterior decoding of the residues in a
* target sequence, given profile <gm> and filled Forward
* and Backward matrices <fwd>, <bck> for the profile
* aligned to that target sequence. The resulting posterior
* decoding is stored in a DP matrix <pp>, provided by the
* caller.
*
* Each residue <i> must have been emitted by match state
* <1..M>, insert state <1..M-1>, or an NN, CC, or JJ loop
* transition. For <dp = pp->dp>, <xmx = pp->xmx>,
* <MMX(i,k)> is the probability that match <k> emitted
* residue <i>; <IMX(i,k)> is the probability that insert
* <k> emitted residue <i>; <XMX(i,N)>,<XMX(i,C)>,
* <XMX(i,J)> are the probabilities that residue <i> was
* emitted on an NN, CC, or JJ transition. The sum over all
* these possibilities for a given residue <i> is 1.0.
*
* Thus the only nonzero entries in a posterior decoding matrix
* <pp> are <M_{1..M}>, <I_{1..M-1}>, <N_{1..L-1}> (residue L
* can't be emitted by N), <C_{2..L}> (residue 1 can't be
* emitted by C), and <J_{2..L-1}> (residues 1,L can't be
* emitted by J).
*
* In particular, row i=0 is unused (all zeros) in a pp
* matrix; the null2 calculation will take advantage of
* this by using the zero row for workspace.
*
* The caller may pass the Backward matrix <bck> as <pp>,
* in which case <bck> will be overwritten with
* <pp>. However, the caller may \emph{not} overwrite <fwd>
* this way; an <(i-1)> dependency in the calculation of
* NN, CC, JJ transitions prevents this.
*
* Args: gm - profile (must be the same that was used to fill <fwd>, <bck>).
* fwd - filled Forward matrix
* bck - filled Backward matrix
* pp - RESULT: posterior decoding matrix.
*
* Returns: <eslOK> on success.
*
* Throws: (no abnormal error conditions)
*
* Note: Burns time renormalizing each row. If you don't do this,
* probabilities will have an error of +/- 0.001 or so, creeping
* in from error in FLogsum()'s table approximation and even
* in log() and exp() themselves; including "probabilities"
* up to ~1.001. Though this isn't going to break anything
* in normal use, it does drive the unit tests wild; the SSE
* implementation is more accurate, and unit tests that try
* to compare SSE and generic results will see differences,
* some sufficient to alter the choice of OA traceback.
*
*/
int
p7_GDecoding(const P7_PROFILE *gm, const P7_GMX *fwd, P7_GMX *bck, P7_GMX *pp)
{
float **dp = pp->dp;
float *xmx = pp->xmx;
int L = fwd->L;
int M = gm->M;
int i,k;
float overall_sc = fwd->xmx[p7G_NXCELLS*L + p7G_C] + gm->xsc[p7P_C][p7P_MOVE];
float denom;
pp->M = M;
pp->L = L;
XMX(0, p7G_E) = 0.0;
XMX(0, p7G_N) = 0.0;
XMX(0, p7G_J) = 0.0;
XMX(0, p7G_B) = 0.0;
XMX(0, p7G_C) = 0.0;
for (k = 0; k <= M; k++)
MMX(0,k) = IMX(0,k) = DMX(0,k) = 0.0;
for (i = 1; i <= L; i++)
{
denom = 0.0;
MMX(i,0) = IMX(i,0) = DMX(i,0) = 0.0;
for (k = 1; k < M; k++)
{
MMX(i,k) = expf(fwd->dp[i][k*p7G_NSCELLS + p7G_M] + bck->dp[i][k*p7G_NSCELLS + p7G_M] - overall_sc); denom += MMX(i,k);
IMX(i,k) = expf(fwd->dp[i][k*p7G_NSCELLS + p7G_I] + bck->dp[i][k*p7G_NSCELLS + p7G_I] - overall_sc); denom += IMX(i,k);
DMX(i,k) = 0.;
}
MMX(i,M) = expf(fwd->dp[i][M*p7G_NSCELLS + p7G_M] + bck->dp[i][M*p7G_NSCELLS + p7G_M] - overall_sc); denom += MMX(i,M);
IMX(i,M) = 0.;
DMX(i,M) = 0.;
/* order doesn't matter. note that this whole function is trivially simd parallel */
XMX(i,p7G_E) = 0.;
XMX(i,p7G_N) = expf(fwd->xmx[p7G_NXCELLS*(i-1) + p7G_N] + bck->xmx[p7G_NXCELLS*i + p7G_N] + gm->xsc[p7P_N][p7P_LOOP] - overall_sc);
XMX(i,p7G_J) = expf(fwd->xmx[p7G_NXCELLS*(i-1) + p7G_J] + bck->xmx[p7G_NXCELLS*i + p7G_J] + gm->xsc[p7P_J][p7P_LOOP] - overall_sc);
XMX(i,p7G_B) = 0.;
XMX(i,p7G_C) = expf(fwd->xmx[p7G_NXCELLS*(i-1) + p7G_C] + bck->xmx[p7G_NXCELLS*i + p7G_C] + gm->xsc[p7P_C][p7P_LOOP] - overall_sc);
denom += XMX(i,p7G_N) + XMX(i,p7G_J) + XMX(i,p7G_C);
denom = 1.0 / denom;
for (k = 1; k < M; k++) { MMX(i,k) *= denom; IMX(i,k) *= denom; }
MMX(i,M) *= denom;
XMX(i,p7G_N) *= denom;
XMX(i,p7G_J) *= denom;
XMX(i,p7G_C) *= denom;
}
return eslOK;
}
/* Function: p7_omx_FDeconvert()
* Synopsis: Convert an optimized DP matrix to generic one.
* Incept: SRE, Tue Aug 19 17:58:13 2008 [Janelia]
*
* Purpose: Convert the 32-bit float values in optimized DP matrix
* <ox> to a generic one <gx>. Caller provides <gx> with sufficient
* space to hold the <ox->M> by <ox->L> matrix.
*
* This function is used to gain access to the
* somewhat more powerful debugging and display
* tools available for generic DP matrices.
*/
int
p7_omx_FDeconvert(P7_OMX *ox, P7_GMX *gx)
{
int Q = p7O_NQF(ox->M);
int i, q, r, k;
union { __m128 v; float p[4]; } u;
float **dp = gx->dp;
float *xmx = gx->xmx;
for (i = 0; i <= ox->L; i++)
{
MMX(i,0) = DMX(i,0) = IMX(i,0) = -eslINFINITY;
for (q = 0; q < Q; q++)
{
u.v = MMO(ox->dpf[i],q); for (r = 0; r < 4; r++) { k = (Q*r)+q+1; if (k <= ox->M) MMX(i, (Q*r)+q+1) = u.p[r]; }
u.v = DMO(ox->dpf[i],q); for (r = 0; r < 4; r++) { k = (Q*r)+q+1; if (k <= ox->M) DMX(i, (Q*r)+q+1) = u.p[r]; }
u.v = IMO(ox->dpf[i],q); for (r = 0; r < 4; r++) { k = (Q*r)+q+1; if (k <= ox->M) IMX(i, (Q*r)+q+1) = u.p[r]; }
}
XMX(i,p7G_E) = ox->xmx[i*p7X_NXCELLS+p7X_E];
XMX(i,p7G_N) = ox->xmx[i*p7X_NXCELLS+p7X_N];
XMX(i,p7G_J) = ox->xmx[i*p7X_NXCELLS+p7X_J];
XMX(i,p7G_B) = ox->xmx[i*p7X_NXCELLS+p7X_B];
XMX(i,p7G_C) = ox->xmx[i*p7X_NXCELLS+p7X_C];
}
gx->L = ox->L;
gx->M = ox->M;
return eslOK;
}
static inline int
select_i(const P7_PROFILE *gm, const P7_GMX *gx, int i, int k)
{
float **dp = gx->dp; /* so {MDI}MX() macros work */
float const *tsc = gm->tsc; /* so TSCDELTA() macro works */
float path[2];
path[0] = TSCDELTA(p7P_MI, k) * MMX(i-1,k);
path[1] = TSCDELTA(p7P_II, k) * IMX(i-1,k);
return ((path[0] >= path[1]) ? p7T_M : p7T_I);
}
static inline int
select_m(const P7_PROFILE *gm, const P7_GMX *gx, int i, int k)
{
float **dp = gx->dp; /* so {MDI}MX() macros work */
float *xmx = gx->xmx; /* so XMX() macro works */
float const *tsc = gm->tsc; /* so TSCDELTA() macro works */
float path[4];
int state[4] = { p7T_M, p7T_I, p7T_D, p7T_B };
path[0] = TSCDELTA(p7P_MM, k-1) * MMX(i-1,k-1);
path[1] = TSCDELTA(p7P_IM, k-1) * IMX(i-1,k-1);
path[2] = TSCDELTA(p7P_DM, k-1) * DMX(i-1,k-1);
path[3] = TSCDELTA(p7P_BM, k-1) * XMX(i-1,p7G_B);
return state[esl_vec_FArgMax(path, 4)];
}
static inline float
get_postprob(const P7_GMX *pp, int scur, int sprv, int k, int i)
{
float **dp = pp->dp;
float *xmx = pp->xmx;
switch (scur) {
case p7T_M: return MMX(i,k);
case p7T_I: return IMX(i,k);
case p7T_N: if (sprv == scur) return XMX(i,p7G_N);
case p7T_C: if (sprv == scur) return XMX(i,p7G_C);
case p7T_J: if (sprv == scur) return XMX(i,p7G_J);
default: return 0.0;
}
}
/* Function: p7_GViterbi()
* Synopsis: The Viterbi algorithm.
* Incept: SRE, Tue Jan 30 10:50:53 2007 [Einstein's, St. Louis]
*
* Purpose: The standard Viterbi dynamic programming algorithm.
*
* Given a digital sequence <dsq> of length <L>, a profile
* <gm>, and DP matrix <gx> allocated for at least <L>
* by <gm->M> cells; calculate the maximum scoring path by
* Viterbi; return the Viterbi score in <ret_sc>, and the
* Viterbi matrix is in <gx>.
*
* The caller may then retrieve the Viterbi path by calling
* <p7_GTrace()>.
*
* The Viterbi lod score is returned in nats. The caller
* needs to subtract a null model lod score, then convert
* to bits.
*
* Args: dsq - sequence in digitized form, 1..L
* L - length of dsq
* gm - profile.
* gx - DP matrix with room for an MxL alignment
* opt_sc - optRETURN: Viterbi lod score in nats
*
* Return: <eslOK> on success.
*/
int
p7_GViterbi(const ESL_DSQ *dsq, int L, const P7_PROFILE *gm, P7_GMX *gx, float *opt_sc)
{
float const *tsc = gm->tsc;
float **dp = gx->dp;
float *xmx = gx->xmx;
int M = gm->M;
int i,k;
float esc = p7_profile_IsLocal(gm) ? 0 : -eslINFINITY;
/* Initialization of the zero row. */
XMX(0,p7G_N) = 0; /* S->N, p=1 */
XMX(0,p7G_B) = gm->xsc[p7P_N][p7P_MOVE]; /* S->N->B, no N-tail */
XMX(0,p7G_E) = XMX(0,p7G_C) = XMX(0,p7G_J) = -eslINFINITY; /* need seq to get here */
for (k = 0; k <= gm->M; k++)
MMX(0,k) = IMX(0,k) = DMX(0,k) = -eslINFINITY; /* need seq to get here */
/* DP recursion */
for (i = 1; i <= L; i++)
{
float const *rsc = gm->rsc[dsq[i]];
float sc;
MMX(i,0) = IMX(i,0) = DMX(i,0) = -eslINFINITY;
XMX(i,p7G_E) = -eslINFINITY;
for (k = 1; k < gm->M; k++)
{
/* match state */
sc = ESL_MAX( MMX(i-1,k-1) + TSC(p7P_MM,k-1),
IMX(i-1,k-1) + TSC(p7P_IM,k-1));
sc = ESL_MAX(sc, DMX(i-1,k-1) + TSC(p7P_DM,k-1));
sc = ESL_MAX(sc, XMX(i-1,p7G_B) + TSC(p7P_BM,k-1));
MMX(i,k) = sc + MSC(k);
/* E state update */
XMX(i,p7G_E) = ESL_MAX(XMX(i,p7G_E), MMX(i,k) + esc);
/* in Viterbi alignments, Dk->E can't win in local mode (and
* isn't possible in glocal mode), so don't bother
* looking. */
/* insert state */
sc = ESL_MAX(MMX(i-1,k) + TSC(p7P_MI,k),
IMX(i-1,k) + TSC(p7P_II,k));
IMX(i,k) = sc + ISC(k);
/* delete state */
DMX(i,k) = ESL_MAX(MMX(i,k-1) + TSC(p7P_MD,k-1),
DMX(i,k-1) + TSC(p7P_DD,k-1));
}
/* Unrolled match state M. */
sc = ESL_MAX( MMX(i-1,M-1) + TSC(p7P_MM,M-1),
IMX(i-1,M-1) + TSC(p7P_IM,M-1));
sc = ESL_MAX(sc, DMX(i-1,M-1 ) + TSC(p7P_DM,M-1));
sc = ESL_MAX(sc, XMX(i-1,p7G_B) + TSC(p7P_BM,M-1));
MMX(i,M) = sc + MSC(M);
/* Unrolled delete state D_M
* (Unlike internal Dk->E transitions that can never appear in
* Viterbi alignments, D_M->E is possible in glocal mode.)
*/
DMX(i,M) = ESL_MAX(MMX(i,M-1) + TSC(p7P_MD,M-1),
DMX(i,M-1) + TSC(p7P_DD,M-1));
/* E state update; transition from M_M scores 0 by def'n */
sc = ESL_MAX(XMX(i,p7G_E), MMX(i,M));
XMX(i,p7G_E) = ESL_MAX(sc, DMX(i,M));
/* Now the special states. E must already be done, and B must follow N,J.
* remember, N, C and J emissions are zero score by definition.
*/
/* J state */
sc = XMX(i-1,p7G_J) + gm->xsc[p7P_J][p7P_LOOP]; /* J->J */
XMX(i,p7G_J) = ESL_MAX(sc, XMX(i, p7G_E) + gm->xsc[p7P_E][p7P_LOOP]); /* E->J is E's "loop" */
/* C state */
sc = XMX(i-1,p7G_C) + gm->xsc[p7P_C][p7P_LOOP];
XMX(i,p7G_C) = ESL_MAX(sc, XMX(i, p7G_E) + gm->xsc[p7P_E][p7P_MOVE]);
/* N state */
XMX(i,p7G_N) = XMX(i-1,p7G_N) + gm->xsc[p7P_N][p7P_LOOP];
/* B state */
sc = XMX(i,p7G_N) + gm->xsc[p7P_N][p7P_MOVE]; /* N->B is N's move */
XMX(i,p7G_B) = ESL_MAX(sc, XMX(i,p7G_J) + gm->xsc[p7P_J][p7P_MOVE]); /* J->B is J's move */
}
/* T state (not stored) */
if (opt_sc != NULL) *opt_sc = XMX(L,p7G_C) + gm->xsc[p7P_C][p7P_MOVE];
gx->M = gm->M;
gx->L = L;
return eslOK;
}
/* Function: p7_GOptimalAccuracy()
* Synopsis: Optimal accuracy decoding: fill.
* Incept: SRE, Fri Feb 29 11:56:49 2008 [Janelia]
*
* Purpose: Calculates the fill step of the optimal accuracy decoding
* algorithm \citep{Kall05}.
*
* Caller provides the posterior decoding matrix <pp>,
* which was calculated by Forward/Backward on a target sequence
* of length <L> using the query model <gm>.
*
* Caller also provides a DP matrix <gx>, allocated for the
* <gm->M> by <pp->L> comparison. The routine fills this in
* with OA scores.
*
* Args: gm - query profile
* pp - posterior decoding matrix created by <p7_GPosteriorDecoding()>
* gx - RESULT: caller provided DP matrix for <gm->M> by <L>
* ret_e - RETURN: expected number of correctly decoded positions
*
* Returns: <eslOK> on success, and <*ret_e> contains the final OA
* score, which is the expected number of correctly decoded
* positions in the target sequence (up to <L>).
*
* Throws: (no abnormal error conditions)
*/
int
p7_GOptimalAccuracy(const P7_PROFILE *gm, const P7_GMX *pp, P7_GMX *gx, float *ret_e)
{
int L = pp->L;
float **dp = gx->dp;
float *xmx = gx->xmx;
float const *tsc = gm->tsc;
int i,k;
int M = gm->M;
float esc = p7_profile_IsLocal(gm) ? 1.0 : 0.0;
float t1, t2;
/* Initialization of the zero row (i=0; no residues to account for. */
XMX(0,p7G_N) = 0.; /* S->N, p=1 */
XMX(0,p7G_B) = 0.; /* S->N->B, no N-tail */
XMX(0,p7G_E) = XMX(0,p7G_C) = XMX(0,p7G_J) = -eslINFINITY; /* need seq to get here */
for (k = 0; k <= M; k++)
MMX(0,k) = IMX(0,k) = DMX(0,k) = -eslINFINITY; /* need seq to get here */
for (i = 1; i <= L; i++)
{
MMX(i,0) = IMX(i,0) = DMX(i,0) = XMX(i,p7G_E) = -eslINFINITY;
for (k = 1; k < M; k++)
{
MMX(i,k) = ESL_MAX(ESL_MAX(TSCDELTA(p7P_MM, k-1) * (MMX(i-1,k-1) + pp->dp[i][k*p7G_NSCELLS + p7G_M]),
TSCDELTA(p7P_IM, k-1) * (IMX(i-1,k-1) + pp->dp[i][k*p7G_NSCELLS + p7G_M])),
ESL_MAX(TSCDELTA(p7P_DM, k-1) * (DMX(i-1,k-1) + pp->dp[i][k*p7G_NSCELLS + p7G_M]),
TSCDELTA(p7P_BM, k-1) * (XMX(i-1,p7G_B)+ pp->dp[i][k*p7G_NSCELLS + p7G_M])));
XMX(i,p7G_E) = ESL_MAX(XMX(i,p7G_E),
esc * MMX(i,k));
IMX(i,k) = ESL_MAX(TSCDELTA(p7P_MI, k) * (MMX(i-1,k) + pp->dp[i][k*p7G_NSCELLS + p7G_I]),
TSCDELTA(p7P_II, k) * (IMX(i-1,k) + pp->dp[i][k*p7G_NSCELLS + p7G_I]));
DMX(i,k) = ESL_MAX(TSCDELTA(p7P_MD, k-1) * MMX(i,k-1),
TSCDELTA(p7P_DD, k-1) * DMX(i,k-1));
}
/* last node (k=M) is unrolled; it has no I state, and it has a p=1.0 {MD}->E transition even in local mode */
MMX(i,M) = ESL_MAX(ESL_MAX(TSCDELTA(p7P_MM, M-1) * (MMX(i-1,M-1) + pp->dp[i][M*p7G_NSCELLS + p7G_M]),
TSCDELTA(p7P_IM, M-1) * (IMX(i-1,M-1) + pp->dp[i][M*p7G_NSCELLS + p7G_M])),
ESL_MAX(TSCDELTA(p7P_DM, M-1) * (DMX(i-1,M-1) + pp->dp[i][M*p7G_NSCELLS + p7G_M]),
TSCDELTA(p7P_BM, M-1) * (XMX(i-1,p7G_B)+ pp->dp[i][M*p7G_NSCELLS + p7G_M])));
DMX(i,M) = ESL_MAX(TSCDELTA(p7P_MD, M-1) * MMX(i,M-1),
TSCDELTA(p7P_DD, M-1) * DMX(i,M-1));
/* note: we calculated XMX before DMX in the loop, because we probably had MMX(i,k) in a register.
* but now we can't do that, because XMX depends on DMX
*/
XMX(i,p7G_E) = ESL_MAX(XMX(i,p7G_E), ESL_MAX(MMX(i,M), DMX(i, M)));
/* now the special states; it's important that E is already done, and B is done after N,J */
t1 = ( (gm->xsc[p7P_J][p7P_LOOP] == -eslINFINITY) ? FLT_MIN : 1.0);
t2 = ( (gm->xsc[p7P_E][p7P_LOOP] == -eslINFINITY) ? FLT_MIN : 1.0);
XMX(i, p7G_J) = ESL_MAX( t1 * (XMX(i-1,p7G_J) + pp->xmx[i*p7G_NXCELLS + p7G_J]),
t2 * XMX(i, p7G_E));
t1 = ( (gm->xsc[p7P_C][p7P_LOOP] == -eslINFINITY) ? FLT_MIN : 1.0);
t2 = ( (gm->xsc[p7P_E][p7P_MOVE] == -eslINFINITY) ? FLT_MIN : 1.0);
XMX(i,p7G_C) = ESL_MAX( t1 * (XMX(i-1,p7G_C) + pp->xmx[i*p7G_NXCELLS + p7G_C]),
t2 * XMX(i, p7G_E));
t1 = ( (gm->xsc[p7P_N][p7P_LOOP] == -eslINFINITY) ? FLT_MIN : 1.0);
XMX(i,p7G_N) = t1 * (XMX(i-1,p7G_N) + pp->xmx[i*p7G_NXCELLS + p7G_N]);
t1 = ( (gm->xsc[p7P_N][p7P_MOVE] == -eslINFINITY) ? FLT_MIN : 1.0);
t2 = ( (gm->xsc[p7P_J][p7P_MOVE] == -eslINFINITY) ? FLT_MIN : 1.0);
XMX(i,p7G_B) = ESL_MAX( t1 * XMX(i, p7G_N),
t2 * XMX(i, p7G_J));
}
*ret_e = XMX(L,p7G_C);
return eslOK;
}
/* Function: p7_GStochasticTrace()
* Synopsis: Stochastic traceback of a Forward matrix.
* Incept: SRE, Thu Jan 3 15:39:20 2008 [Janelia]
*
* Purpose: Stochastic traceback of Forward matrix <gx> to
* sample an alignment of digital sequence <dsq>
* (of length <L>) to the profile <gm>.
*
* The sampled traceback is returned in <tr>, which the
* caller must have at least made an initial allocation of
* (the <tr> will be grown as needed here).
*
* Args: r - source of random numbers
* dsq - digital sequence aligned to, 1..L
* L - length of dsq
* gm - profile
* mx - Forward matrix to trace, L x M
* tr - storage for the recovered traceback.
*
* Returns: <eslOK> on success.
*/
int
p7_GStochasticTrace(ESL_RANDOMNESS *r, const ESL_DSQ *dsq, int L, const P7_PROFILE *gm, const P7_GMX *gx, P7_TRACE *tr)
{
int status;
int i; /* position in seq (1..L) */
int k; /* position in model (1..M) */
int M = gm->M;
float **dp = gx->dp;
float *xmx = gx->xmx;
float const *tsc = gm->tsc;
float *sc; /* scores of possible choices: up to 2M-1, in the case of exits to E */
int scur, sprv;
/* we'll index M states as 1..M, and D states as 2..M = M+2..2M: M0, D1 are impossibles. */
ESL_ALLOC(sc, sizeof(float) * (2*M+1));
k = 0;
i = L;
if ((status = p7_trace_Append(tr, p7T_T, k, i)) != eslOK) goto ERROR;
if ((status = p7_trace_Append(tr, p7T_C, k, i)) != eslOK) goto ERROR;
sprv = p7T_C;
while (sprv != p7T_S)
{
switch (tr->st[tr->N-1]) {
/* C(i) comes from C(i-1) or E(i) */
case p7T_C:
if (XMX(i,p7G_C) == -eslINFINITY) ESL_XEXCEPTION(eslFAIL, "impossible C reached at i=%d", i);
sc[0] = XMX(i-1, p7G_C) + gm->xsc[p7P_C][p7P_LOOP];
sc[1] = XMX(i, p7G_E) + gm->xsc[p7P_E][p7P_MOVE];
esl_vec_FLogNorm(sc, 2);
scur = (esl_rnd_FChoose(r, sc, 2) == 0) ? p7T_C : p7T_E;
break;
/* E connects from any M or D state. k set here */
case p7T_E:
if (XMX(i, p7G_E) == -eslINFINITY) ESL_XEXCEPTION(eslFAIL, "impossible E reached at i=%d", i);
if (p7_profile_IsLocal(gm)) { /* local models come from any M, D */
sc[0] = sc[M+1] = -eslINFINITY;
for (k = 1; k <= M; k++) sc[k] = MMX(i,k);
for (k = 2; k <= M; k++) sc[k+M] = DMX(i,k);
esl_vec_FLogNorm(sc, 2*M+1); /* now sc is a prob vector */
k = esl_rnd_FChoose(r, sc, 2*M+1);
if (k <= M) scur = p7T_M;
else { k -= M; scur = p7T_D; }
} else { /* glocal models come from M_M or D_M */
k = M;
sc[0] = MMX(i,M);
sc[1] = DMX(i,M);
esl_vec_FLogNorm(sc, 2); /* now sc is a prob vector */
scur = (esl_rnd_FChoose(r, sc, 2) == 0) ? p7T_M : p7T_D;
}
break;
/* M connects from {MDI} i-1,k-1, or B */
case p7T_M:
if (MMX(i,k) == -eslINFINITY) ESL_XEXCEPTION(eslFAIL, "impossible M reached at k=%d,i=%d", k,i);
sc[0] = XMX(i-1,p7G_B) + TSC(p7P_BM, k-1);
sc[1] = MMX(i-1,k-1) + TSC(p7P_MM, k-1);
sc[2] = IMX(i-1,k-1) + TSC(p7P_IM, k-1);
sc[3] = DMX(i-1,k-1) + TSC(p7P_DM, k-1);
esl_vec_FLogNorm(sc, 4);
switch (esl_rnd_FChoose(r, sc, 4)) {
case 0: scur = p7T_B; break;
case 1: scur = p7T_M; break;
case 2: scur = p7T_I; break;
case 3: scur = p7T_D; break;
}
k--;
i--;
break;
/* D connects from M,D at i,k-1 */
case p7T_D:
if (DMX(i, k) == -eslINFINITY) ESL_XEXCEPTION(eslFAIL, "impossible D reached at k=%d,i=%d", k,i);
sc[0] = MMX(i, k-1) + TSC(p7P_MD, k-1);
sc[1] = DMX(i, k-1) + TSC(p7P_DD, k-1);
esl_vec_FLogNorm(sc, 2);
scur = (esl_rnd_FChoose(r, sc, 2) == 0) ? p7T_M : p7T_D;
k--;
break;
/* I connects from M,I at i-1,k */
case p7T_I:
if (IMX(i,k) == -eslINFINITY) ESL_XEXCEPTION(eslFAIL, "impossible I reached at k=%d,i=%d", k,i);
sc[0] = MMX(i-1,k) + TSC(p7P_MI, k);
sc[1] = IMX(i-1,k) + TSC(p7P_II, k);
esl_vec_FLogNorm(sc, 2);
scur = (esl_rnd_FChoose(r, sc, 2) == 0) ? p7T_M : p7T_I;
i--;
break;
/* N connects from S, N */
case p7T_N:
if (XMX(i, p7G_N) == -eslINFINITY) ESL_XEXCEPTION(eslFAIL, "impossible N reached at i=%d", i);
scur = (i == 0) ? p7T_S : p7T_N;
break;
/* B connects from N, J */
case p7T_B:
if (XMX(i,p7G_B) == -eslINFINITY) ESL_XEXCEPTION(eslFAIL, "impossible B reached at i=%d", i);
sc[0] = XMX(i, p7G_N) + gm->xsc[p7P_N][p7P_MOVE];
sc[1] = XMX(i, p7G_J) + gm->xsc[p7P_J][p7P_MOVE];
esl_vec_FLogNorm(sc, 2);
scur = (esl_rnd_FChoose(r, sc, 2) == 0) ? p7T_N : p7T_J;
break;
/* J connects from E(i) or J(i-1) */
case p7T_J:
if (XMX(i,p7G_J) == -eslINFINITY) ESL_XEXCEPTION(eslFAIL, "impossible J reached at i=%d", i);
sc[0] = XMX(i-1,p7G_J) + gm->xsc[p7P_J][p7P_LOOP];
sc[1] = XMX(i, p7G_E) + gm->xsc[p7P_E][p7P_LOOP];
esl_vec_FLogNorm(sc, 2);
scur = (esl_rnd_FChoose(r, sc, 2) == 0) ? p7T_J : p7T_E;
break;
default: ESL_XEXCEPTION(eslFAIL, "bogus state in traceback");
} /* end switch over statetype[tpos-1] */
/* Append this state and the current i,k to be explained to the growing trace */
if ((status = p7_trace_Append(tr, scur, k, i)) != eslOK) goto ERROR;
/* For NCJ, we had to defer i decrement. */
if ( (scur == p7T_N || scur == p7T_J || scur == p7T_C) && scur == sprv) i--;
sprv = scur;
} /* end traceback, at S state */
if ((status = p7_trace_Reverse(tr)) != eslOK) goto ERROR;
tr->M = gm->M;
tr->L = L;
free(sc);
return eslOK;
ERROR:
if (sc != NULL) free(sc);
return status;
}
/* Function: p7_GViterbi_longtarget()
* Synopsis: The Viterbi algorithm.
* Incept: SRE, Tue Jan 30 10:50:53 2007 [Einstein's, St. Louis]
*
* Purpose: Given a digital sequence <dsq> of length <L>, a profile
* <gm>, and DP matrix <gx> allocated for at least <L>
* by <gm->M> cells; calculates the Viterbi score for
* regions of <dsq>, and captures the positions at which
* such regions exceed the score required to be
* significant in the eyes of the calling function
* (usually p=0.001).
*
* Args: dsq - sequence in digitized form, 1..L
* L - length of dsq
* gm - profile.
* gx - DP matrix with room for an MxL alignment
* filtersc - null or bias correction, required for translating a P-value threshold into a score threshold
* P - p-value below which a region is captured as being above threshold
* windowlist - RETURN: array of hit windows (start and end of diagonal) for the above-threshold areas
*
* Return: <eslOK> on success.
*/
int
p7_GViterbi_longtarget(const ESL_DSQ *dsq, int L, const P7_PROFILE *gm, P7_GMX *gx,
float filtersc, double P, P7_HMM_WINDOWLIST *windowlist)
{
float const *tsc = gm->tsc;
float **dp = gx->dp;
float *xmx = gx->xmx;
int M = gm->M;
int i,k;
float esc = p7_profile_IsLocal(gm) ? 0 : -eslINFINITY;
int16_t sc_thresh;
float invP;
/* Initialization of the zero row. */
XMX(0,p7G_N) = 0; /* S->N, p=1 */
XMX(0,p7G_B) = gm->xsc[p7P_N][p7P_MOVE]; /* S->N->B, no N-tail */
XMX(0,p7G_E) = XMX(0,p7G_C) = XMX(0,p7G_J) = -eslINFINITY; /* need seq to get here */
for (k = 0; k <= gm->M; k++)
MMX(0,k) = IMX(0,k) = DMX(0,k) = -eslINFINITY; /* need seq to get here */
/*
* In p7_ViterbiFilter, converting from a scaled int Viterbi score
* S (aka xE the score getting to state E) to a probability
* goes like this:
* S = XMX(i,p7G_E)
* vsc = S + gm->xsc[p7P_E][p7P_MOVE] + gm->xsc[p7P_C][p7P_MOVE];
* P = esl_gumbel_surv((vfsc - filtersc) / eslCONST_LOG2 , gm->evparam[p7_VMU], gm->evparam[p7_VLAMBDA]);
* and we're computing the threshold vsc, so invert it:
* (vsc - filtersc) / eslCONST_LOG2 = esl_gumbel_invsurv( P, gm->evparam[p7_VMU], gm->evparam[p7_VLAMBDA])
* vsc = filtersc + eslCONST_LOG2 * esl_gumbel_invsurv( P, gm->evparam[p7_VMU], gm->evparam[p7_VLAMBDA])
* S = vsc - gm->xsc[p7P_E][p7P_MOVE] - gm->xsc[p7P_C][p7P_MOVE]
*/
invP = esl_gumbel_invsurv(P, gm->evparam[p7_VMU], gm->evparam[p7_VLAMBDA]);
sc_thresh = (int) ceil (filtersc + (eslCONST_LOG2 * invP)
- gm->xsc[p7P_E][p7P_MOVE] - gm->xsc[p7P_C][p7P_MOVE] );
/* DP recursion */
for (i = 1; i <= L; i++)
{
float const *rsc = gm->rsc[dsq[i]];
float sc;
MMX(i,0) = IMX(i,0) = DMX(i,0) = -eslINFINITY;
XMX(i,p7G_E) = -eslINFINITY;
for (k = 1; k < gm->M; k++)
{
/* match state */
sc = ESL_MAX( MMX(i-1,k-1) + TSC(p7P_MM,k-1),
IMX(i-1,k-1) + TSC(p7P_IM,k-1));
sc = ESL_MAX(sc, DMX(i-1,k-1) + TSC(p7P_DM,k-1));
sc = ESL_MAX(sc, XMX(i-1,p7G_B) + TSC(p7P_BM,k-1));
MMX(i,k) = sc + MSC(k);
/* E state update */
XMX(i,p7G_E) = ESL_MAX(XMX(i,p7G_E), MMX(i,k) + esc);
/* in Viterbi alignments, Dk->E can't win in local mode (and
* isn't possible in glocal mode), so don't bother
* looking. */
/* insert state */
sc = ESL_MAX(MMX(i-1,k) + TSC(p7P_MI,k),
IMX(i-1,k) + TSC(p7P_II,k));
IMX(i,k) = sc + ISC(k);
/* delete state */
DMX(i,k) = ESL_MAX(MMX(i,k-1) + TSC(p7P_MD,k-1),
DMX(i,k-1) + TSC(p7P_DD,k-1));
}
/* Unrolled match state M. */
sc = ESL_MAX( MMX(i-1,M-1) + TSC(p7P_MM,M-1),
IMX(i-1,M-1) + TSC(p7P_IM,M-1));
sc = ESL_MAX(sc, DMX(i-1,M-1 ) + TSC(p7P_DM,M-1));
sc = ESL_MAX(sc, XMX(i-1,p7G_B) + TSC(p7P_BM,M-1));
MMX(i,M) = sc + MSC(M);
/* Unrolled delete state D_M
* (Unlike internal Dk->E transitions that can never appear in
* Viterbi alignments, D_M->E is possible in glocal mode.)
*/
DMX(i,M) = ESL_MAX(MMX(i,M-1) + TSC(p7P_MD,M-1),
DMX(i,M-1) + TSC(p7P_DD,M-1));
/* E state update; transition from M_M scores 0 by def'n */
sc = ESL_MAX(XMX(i,p7G_E), MMX(i,M));
XMX(i,p7G_E) = ESL_MAX(sc, DMX(i,M));
if (XMX(i,p7G_E) >= sc_thresh) {
//hit score threshold. Add a window to the list, then reset scores.
for (k = 1; k <= gm->M; k++) {
if (MMX(i,k) == XMX(i,p7G_E)) {
p7_hmmwindow_new(windowlist, 0, i, 0, k, 1, 0.0, p7_NOCOMPLEMENT );
}
MMX(i,0) = IMX(i,0) = DMX(i,0) = -eslINFINITY;
}
} else {
/* Now the special states. E must already be done, and B must follow N,J.
* remember, N, C and J emissions are zero score by definition.
*/
/* J state */
sc = XMX(i-1,p7G_J) + gm->xsc[p7P_J][p7P_LOOP]; /* J->J */
XMX(i,p7G_J) = ESL_MAX(sc, XMX(i, p7G_E) + gm->xsc[p7P_E][p7P_LOOP]); /* E->J is E's "loop" */
/* C state */
sc = XMX(i-1,p7G_C) + gm->xsc[p7P_C][p7P_LOOP];
XMX(i,p7G_C) = ESL_MAX(sc, XMX(i, p7G_E) + gm->xsc[p7P_E][p7P_MOVE]);
/* N state */
XMX(i,p7G_N) = XMX(i-1,p7G_N) + gm->xsc[p7P_N][p7P_LOOP];
/* B state */
sc = XMX(i,p7G_N) + gm->xsc[p7P_N][p7P_MOVE]; /* N->B is N's move */
XMX(i,p7G_B) = ESL_MAX(sc, XMX(i,p7G_J) + gm->xsc[p7P_J][p7P_MOVE]); /* J->B is J's move */
}
}
/* T state (not stored) */
gx->M = gm->M;
gx->L = L;
return eslOK;
}
/* Function: p7_GTrace()
* Incept: SRE, Thu Feb 1 10:25:56 2007 [UA 8018 St. Louis to Dulles]
*
* Purpose: Traceback of a Viterbi matrix: retrieval
* of optimum alignment.
*
* This function is currently implemented as a
* reconstruction traceback, rather than using a shadow
* matrix. Because H3 uses floating point scores, and we
* can't compare floats for equality, we have to compare
* floats for near-equality and therefore, formally, we can
* only guarantee a near-optimal traceback. However, even in
* the unlikely event that a suboptimal is returned, the
* score difference from true optimal will be negligible.
*
* Args: dsq - digital sequence aligned to, 1..L
* L - length of <dsq>
* gm - profile
* mx - Viterbi matrix to trace, L x M
* tr - storage for the recovered traceback.
*
* Return: <eslOK> on success.
* <eslFAIL> if even the optimal path has zero probability;
* in this case, the trace is set blank (<tr->N = 0>).
*
* Note: Care is taken to evaluate the prev+tsc+emission
* calculations in exactly the same order that Viterbi did
* them, lest you get numerical problems with
* a+b+c = d; d-c != a+b because d,c are nearly equal.
* (This bug appeared in dev: xref J1/121.)
*/
int
p7_GTrace(const ESL_DSQ *dsq, int L, const P7_PROFILE *gm, const P7_GMX *gx, P7_TRACE *tr)
{
int i = L; /* position in seq (1..L) */
int k = 0; /* position in model (1..M) */
int M = gm->M;
float **dp = gx->dp; /* so {MDI}MX() macros work */
float *xmx = gx->xmx; /* so XMX() macro works */
float tol = 1e-5; /* floating point "equality" test */
float const *tsc = gm->tsc;
int sprv, scur; /* previous, current state in trace */
int status;
#ifdef p7_DEBUGGING
if (tr->N != 0) ESL_EXCEPTION(eslEINVAL, "trace isn't empty: forgot to Reuse()?");
#endif
if ((status = p7_trace_Append(tr, p7T_T, k, i)) != eslOK) return status;
if ((status = p7_trace_Append(tr, p7T_C, k, i)) != eslOK) return status;
sprv = p7T_C;
while (sprv != p7T_S) {
float const *rsc = (i>0 ? gm->rsc[dsq[i]] : NULL);
switch (sprv) {
case p7T_C: /* C(i) comes from C(i-1) or E(i) */
if (XMX(i,p7G_C) == -eslINFINITY) ESL_EXCEPTION(eslFAIL, "impossible C reached at i=%d", i);
if (esl_FCompare(XMX(i, p7G_C), XMX(i-1, p7G_C) + gm->xsc[p7P_C][p7P_LOOP], tol) == eslOK) scur = p7T_C;
else if (esl_FCompare(XMX(i, p7G_C), XMX(i, p7G_E) + gm->xsc[p7P_E][p7P_MOVE], tol) == eslOK) scur = p7T_E;
else ESL_EXCEPTION(eslFAIL, "C at i=%d couldn't be traced", i);
break;
case p7T_E: /* E connects from any M state. k set here */
if (XMX(i, p7G_E) == -eslINFINITY) ESL_EXCEPTION(eslFAIL, "impossible E reached at i=%d", i);
if (p7_profile_IsLocal(gm))
{
scur = p7T_M; /* can't come from D, in a *local* Viterbi trace. */
for (k = M; k >= 1; k--) if (esl_FCompare(XMX(i, p7G_E), MMX(i,k), tol) == eslOK) break;
if (k == 0) ESL_EXCEPTION(eslFAIL, "E at i=%d couldn't be traced", i);
}
else /* glocal mode: we either come from D_M or M_M */
{
if (esl_FCompare(XMX(i, p7G_E), MMX(i,M), tol) == eslOK) { scur = p7T_M; k = M; }
else if (esl_FCompare(XMX(i, p7G_E), DMX(i,M), tol) == eslOK) { scur = p7T_D; k = M; }
else ESL_EXCEPTION(eslFAIL, "E at i=%d couldn't be traced", i);
}
break;
case p7T_M: /* M connects from i-1,k-1, or B */
if (MMX(i,k) == -eslINFINITY) ESL_EXCEPTION(eslFAIL, "impossible M reached at k=%d,i=%d", k,i);
if (esl_FCompare(MMX(i,k), XMX(i-1,p7G_B) + TSC(p7P_BM, k-1) + MSC(k), tol) == eslOK) scur = p7T_B;
else if (esl_FCompare(MMX(i,k), MMX(i-1,k-1) + TSC(p7P_MM, k-1) + MSC(k), tol) == eslOK) scur = p7T_M;
else if (esl_FCompare(MMX(i,k), IMX(i-1,k-1) + TSC(p7P_IM, k-1) + MSC(k), tol) == eslOK) scur = p7T_I;
else if (esl_FCompare(MMX(i,k), DMX(i-1,k-1) + TSC(p7P_DM, k-1) + MSC(k), tol) == eslOK) scur = p7T_D;
else ESL_EXCEPTION(eslFAIL, "M at k=%d,i=%d couldn't be traced", k,i);
k--; i--;
break;
case p7T_D: /* D connects from M,D at i,k-1 */
if (DMX(i, k) == -eslINFINITY) ESL_EXCEPTION(eslFAIL, "impossible D reached at k=%d,i=%d", k,i);
if (esl_FCompare(DMX(i,k), MMX(i, k-1) + TSC(p7P_MD, k-1), tol) == eslOK) scur = p7T_M;
else if (esl_FCompare(DMX(i,k), DMX(i, k-1) + TSC(p7P_DD, k-1), tol) == eslOK) scur = p7T_D;
else ESL_EXCEPTION(eslFAIL, "D at k=%d,i=%d couldn't be traced", k,i);
k--;
break;
case p7T_I: /* I connects from M,I at i-1,k*/
if (IMX(i,k) == -eslINFINITY) ESL_EXCEPTION(eslFAIL, "impossible I reached at k=%d,i=%d", k,i);
if (esl_FCompare(IMX(i,k), MMX(i-1,k) + TSC(p7P_MI, k) + ISC(k), tol) == eslOK) scur = p7T_M;
else if (esl_FCompare(IMX(i,k), IMX(i-1,k) + TSC(p7P_II, k) + ISC(k), tol) == eslOK) scur = p7T_I;
else ESL_EXCEPTION(eslFAIL, "I at k=%d,i=%d couldn't be traced", k,i);
i--;
break;
case p7T_N: /* N connects from S, N */
if (XMX(i, p7G_N) == -eslINFINITY) ESL_EXCEPTION(eslFAIL, "impossible N reached at i=%d", i);
scur = ( (i == 0) ? p7T_S : p7T_N);
break;
case p7T_B: /* B connects from N, J */
if (XMX(i,p7G_B) == -eslINFINITY) ESL_EXCEPTION(eslFAIL, "impossible B reached at i=%d", i);
if (esl_FCompare(XMX(i,p7G_B), XMX(i, p7G_N) + gm->xsc[p7P_N][p7P_MOVE], tol) == eslOK) scur = p7T_N;
else if (esl_FCompare(XMX(i,p7G_B), XMX(i, p7G_J) + gm->xsc[p7P_J][p7P_MOVE], tol) == eslOK) scur = p7T_J;
else ESL_EXCEPTION(eslFAIL, "B at i=%d couldn't be traced", i);
break;
case p7T_J: /* J connects from E(i) or J(i-1) */
if (XMX(i,p7G_J) == -eslINFINITY) ESL_EXCEPTION(eslFAIL, "impossible J reached at i=%d", i);
if (esl_FCompare(XMX(i,p7G_J), XMX(i-1,p7G_J) + gm->xsc[p7P_J][p7P_LOOP], tol) == eslOK) scur = p7T_J;
else if (esl_FCompare(XMX(i,p7G_J), XMX(i, p7G_E) + gm->xsc[p7P_E][p7P_LOOP], tol) == eslOK) scur = p7T_E;
else ESL_EXCEPTION(eslFAIL, "J at i=%d couldn't be traced", i);
break;
default: ESL_EXCEPTION(eslFAIL, "bogus state in traceback");
} /* end switch over statetype[tpos-1] */
/* Append this state and the current i,k to be explained to the growing trace */
if ((status = p7_trace_Append(tr, scur, k, i)) != eslOK) return status;
/* For NCJ, we had to defer i decrement. */
if ( (scur == p7T_N || scur == p7T_J || scur == p7T_C) && scur == sprv) i--;
sprv = scur;
} /* end traceback, at S state */
tr->M = gm->M;
tr->L = L;
return p7_trace_Reverse(tr);
}
/* Function: p7_GNull2_ByExpectation()
* Synopsis: Calculate null2 model from posterior probabilities.
* Incept: SRE, Thu Feb 28 09:52:28 2008 [Janelia]
*
* Purpose: Calculate the "null2" model for the envelope encompassed
* by a posterior probability calculation <pp> for model
* <gm>. Return the null2 odds emission probabilities
* $\frac{f'{x}}{f{x}}$ in <null2>, which caller
* provides as space for at least <alphabet->Kp> residues.
*
* The expectation method is applied to envelopes in
* simple, well resolved regions (regions containing just a
* single envelope, where no stochastic traceback
* clustering was required).
*
* Make sure that the posterior probability matrix <pp> has
* been calculated by the caller for only the envelope; thus
* its rows are numbered <1..Ld>, for envelope <ienv..jenv>
* of length <Ld=jenv-ienv+1>.
*
* Args: gm - profile, in any mode, target length model set to <L>
* pp - posterior prob matrix, for <gm> against domain envelope <dsq+i-1> (offset)
* null2 - RETURN: null2 odds ratios per residue; <0..Kp-1>; caller allocated space
*
* Returns: <eslOK> on success; <null2> contains the null2 scores. The 0
* row of <pp> has been used as temp space, and happens to contain
* the expected frequency that each M,I,N,C,J state is used in this
* <pp> matrix to generate residues.
*
* Throws: (no abnormal error conditions)
*/
int
p7_GNull2_ByExpectation(const P7_PROFILE *gm, P7_GMX *pp, float *null2)
{
int M = gm->M;
int Ld = pp->L;
float **dp = pp->dp;
float *xmx = pp->xmx;
float xfactor;
int x; /* over symbols 0..K-1 */
int i; /* over offset envelope dsq positions 1..Ld */
int k; /* over model M states 1..M, I states 1..M-1 */
/* Calculate expected # of times that each emitting state was used
* in generating the Ld residues in this domain.
* The 0 row in <wrk> is used to hold these numbers.
*/
esl_vec_FCopy(pp->dp[1], (M+1)*p7G_NSCELLS, pp->dp[0]);
esl_vec_FCopy(pp->xmx+p7G_NXCELLS, p7G_NXCELLS, pp->xmx);
for (i = 2; i <= Ld; i++)
{
esl_vec_FAdd(pp->dp[0], pp->dp[i], (M+1)*p7G_NSCELLS);
esl_vec_FAdd(pp->xmx, pp->xmx+i*p7G_NXCELLS, p7G_NXCELLS);
}
/* Convert those expected #'s to log frequencies; these we'll use as
* the log posterior weights.
*/
esl_vec_FLog(pp->dp[0], (M+1)*p7G_NSCELLS);
esl_vec_FLog(pp->xmx, p7G_NXCELLS);
esl_vec_FIncrement(pp->dp[0], (M+1)*p7G_NSCELLS, -log((float)Ld));
esl_vec_FIncrement(pp->xmx, p7G_NXCELLS, -log((float)Ld));
/* Calculate null2's log odds emission probabilities, by taking
* posterior weighted sum over all emission vectors used in paths
* explaining the domain.
* This is dog-slow; a point for future optimization.
*/
xfactor = XMX(0,p7G_N);
xfactor = p7_FLogsum(xfactor, XMX(0,p7G_C));
xfactor = p7_FLogsum(xfactor, XMX(0,p7G_J));
esl_vec_FSet(null2, gm->abc->K, -eslINFINITY);
for (x = 0; x < gm->abc->K; x++)
{
for (k = 1; k < M; k++)
{
null2[x] = p7_FLogsum(null2[x], MMX(0,k) + p7P_MSC(gm, k, x));
null2[x] = p7_FLogsum(null2[x], IMX(0,k) + p7P_ISC(gm, k, x));
}
null2[x] = p7_FLogsum(null2[x], MMX(0,M) + p7P_MSC(gm, k, x));
null2[x] = p7_FLogsum(null2[x], xfactor);
}
esl_vec_FExp (null2, gm->abc->K);
/* now null2[x] = \frac{f_d(x)}{f_0(x)} for all x in alphabet,
* 0..K-1, where f_d(x) are the ad hoc "null2" residue frequencies
* for this envelope.
*/
/* make valid scores for all degeneracies, by averaging the odds ratios. */
esl_abc_FAvgScVec(gm->abc, null2); /* does not set gap, nonres, missing */
null2[gm->abc->K] = 1.0; /* gap character */
null2[gm->abc->Kp-2] = 1.0; /* nonresidue "*" */
null2[gm->abc->Kp-1] = 1.0; /* missing data "~" */
return eslOK;
}
/* Function: p7_GForward()
* Synopsis: The Forward algorithm.
* Incept: SRE, Mon Apr 16 13:57:35 2007 [Janelia]
*
* Purpose: The Forward dynamic programming algorithm.
*
* Given a digital sequence <dsq> of length <L>, a profile
* <gm>, and DP matrix <gx> allocated for at least <gm->M>
* by <L> cells; calculate the probability of the sequence
* given the model using the Forward algorithm; return the
* Forward matrix in <gx>, and the Forward score in <ret_sc>.
*
* The Forward score is in lod score form. To convert to a
* bitscore, the caller needs to subtract a null model lod
* score, then convert to bits.
*
* Args: dsq - sequence in digitized form, 1..L
* L - length of dsq
* gm - profile.
* gx - DP matrix with room for an MxL alignment
* opt_sc - optRETURN: Forward lod score in nats
*
* Return: <eslOK> on success.
*/
int
p7_GForward(const ESL_DSQ *dsq, int L, const P7_PROFILE *gm, P7_GMX *gx, float *opt_sc)
{
float const *tsc = gm->tsc;
float **dp = gx->dp;
float *xmx = gx->xmx;
int M = gm->M;
int i, k;
float esc = p7_profile_IsLocal(gm) ? 0 : -eslINFINITY;
/* Initialization of the zero row, and the lookup table of the log
* sum routine.
*/
XMX(0,p7G_N) = 0; /* S->N, p=1 */
XMX(0,p7G_B) = gm->xsc[p7P_N][p7P_MOVE]; /* S->N->B, no N-tail */
XMX(0,p7G_E) = XMX(0,p7G_C) = XMX(0,p7G_J) = -eslINFINITY; /* need seq to get here */
for (k = 0; k <= M; k++)
MMX(0,k) = IMX(0,k) = DMX(0,k) = -eslINFINITY; /* need seq to get here */
p7_FLogsumInit();
/* Recursion. Done as a pull.
* Note some slightly wasteful boundary conditions:
* tsc[0] = impossible for all eight transitions (no node 0)
* D_1 is wastefully calculated (doesn't exist)
*/
for (i = 1; i <= L; i++)
{
float const *rsc = gm->rsc[dsq[i]];
float sc;
MMX(i,0) = IMX(i,0) = DMX(i,0) = -eslINFINITY;
XMX(i, p7G_E) = -eslINFINITY;
for (k = 1; k < M; k++)
{
/* match state */
sc = p7_FLogsum(p7_FLogsum(MMX(i-1,k-1) + TSC(p7P_MM,k-1),
IMX(i-1,k-1) + TSC(p7P_IM,k-1)),
p7_FLogsum(XMX(i-1,p7G_B) + TSC(p7P_BM,k-1),
DMX(i-1,k-1) + TSC(p7P_DM,k-1)));
MMX(i,k) = sc + MSC(k);
/* insert state */
sc = p7_FLogsum(MMX(i-1,k) + TSC(p7P_MI,k),
IMX(i-1,k) + TSC(p7P_II,k));
IMX(i,k) = sc + ISC(k);
/* delete state */
DMX(i,k) = p7_FLogsum(MMX(i,k-1) + TSC(p7P_MD,k-1),
DMX(i,k-1) + TSC(p7P_DD,k-1));
/* E state update */
XMX(i,p7G_E) = p7_FLogsum(p7_FLogsum(MMX(i,k) + esc,
DMX(i,k) + esc),
XMX(i,p7G_E));
}
/* unrolled match state M_M */
sc = p7_FLogsum(p7_FLogsum(MMX(i-1,M-1) + TSC(p7P_MM,M-1),
IMX(i-1,M-1) + TSC(p7P_IM,M-1)),
p7_FLogsum(XMX(i-1,p7G_B) + TSC(p7P_BM,M-1),
DMX(i-1,M-1) + TSC(p7P_DM,M-1)));
MMX(i,M) = sc + MSC(M);
IMX(i,M) = -eslINFINITY;
/* unrolled delete state D_M */
DMX(i,M) = p7_FLogsum(MMX(i,M-1) + TSC(p7P_MD,M-1),
DMX(i,M-1) + TSC(p7P_DD,M-1));
/* unrolled E state update */
XMX(i,p7G_E) = p7_FLogsum(p7_FLogsum(MMX(i,M),
DMX(i,M)),
XMX(i,p7G_E));
/* J state */
XMX(i,p7G_J) = p7_FLogsum(XMX(i-1,p7G_J) + gm->xsc[p7P_J][p7P_LOOP],
XMX(i, p7G_E) + gm->xsc[p7P_E][p7P_LOOP]);
/* C state */
XMX(i,p7G_C) = p7_FLogsum(XMX(i-1,p7G_C) + gm->xsc[p7P_C][p7P_LOOP],
XMX(i, p7G_E) + gm->xsc[p7P_E][p7P_MOVE]);
/* N state */
XMX(i,p7G_N) = XMX(i-1,p7G_N) + gm->xsc[p7P_N][p7P_LOOP];
/* B state */
XMX(i,p7G_B) = p7_FLogsum(XMX(i, p7G_N) + gm->xsc[p7P_N][p7P_MOVE],
XMX(i, p7G_J) + gm->xsc[p7P_J][p7P_MOVE]);
}
if (opt_sc != NULL) *opt_sc = XMX(L,p7G_C) + gm->xsc[p7P_C][p7P_MOVE];
gx->M = M;
gx->L = L;
return eslOK;
}
/* Function: p7_GBackward()
* Synopsis: The Backward algorithm.
* Incept: SRE, Fri Dec 28 14:31:58 2007 [Janelia]
*
* Purpose: The Backward dynamic programming algorithm.
*
* Given a digital sequence <dsq> of length <L>, a profile
* <gm>, and DP matrix <gx> allocated for at least <gm->M>
* by <L> cells; calculate the probability of the sequence
* given the model using the Backward algorithm; return the
* Backward matrix in <gx>, and the Backward score in <ret_sc>.
*
* The Backward score is in lod score form. To convert to a
* bitscore, the caller needs to subtract a null model lod
* score, then convert to bits.
*
* Args: dsq - sequence in digitized form, 1..L
* L - length of dsq
* gm - profile
* gx - DP matrix with room for an MxL alignment
* opt_sc - optRETURN: Backward lod score in nats
*
* Return: <eslOK> on success.
*/
int
p7_GBackward(const ESL_DSQ *dsq, int L, const P7_PROFILE *gm, P7_GMX *gx, float *opt_sc)
{
float const *tsc = gm->tsc;
float const *rsc = NULL;
float **dp = gx->dp;
float *xmx = gx->xmx;
int M = gm->M;
int i, k;
float esc = p7_profile_IsLocal(gm) ? 0 : -eslINFINITY;
/* Note: backward calculates the probability we can get *out* of
* cell i,k; exclusive of emitting residue x_i.
*/
p7_FLogsumInit();
/* Initialize the L row. */
XMX(L,p7G_J) = XMX(L,p7G_B) = XMX(L,p7G_N) = -eslINFINITY;
XMX(L,p7G_C) = gm->xsc[p7P_C][p7P_MOVE]; /* C<-T */
XMX(L,p7G_E) = XMX(L,p7G_C) + gm->xsc[p7P_E][p7P_MOVE]; /* E<-C, no tail */
MMX(L,M) = DMX(L,M) = XMX(L,p7G_E); /* {MD}_M <- E (prob 1.0) */
IMX(L,M) = -eslINFINITY; /* no I_M state */
for (k = M-1; k >= 1; k--) {
MMX(L,k) = p7_FLogsum( XMX(L,p7G_E) + esc,
DMX(L, k+1) + TSC(p7P_MD,k));
DMX(L,k) = p7_FLogsum( XMX(L,p7G_E) + esc,
DMX(L, k+1) + TSC(p7P_DD,k));
IMX(L,k) = -eslINFINITY;
}
/* Main recursion */
for (i = L-1; i >= 1; i--)
{
rsc = gm->rsc[dsq[i+1]];
XMX(i,p7G_B) = MMX(i+1,1) + TSC(p7P_BM,0) + MSC(1); /* t_BM index is 0 because it's stored off-by-one. */
for (k = 2; k <= M; k++)
XMX(i,p7G_B) = p7_FLogsum(XMX(i, p7G_B), MMX(i+1,k) + TSC(p7P_BM,k-1) + MSC(k));
XMX(i,p7G_J) = p7_FLogsum( XMX(i+1,p7G_J) + gm->xsc[p7P_J][p7P_LOOP],
XMX(i, p7G_B) + gm->xsc[p7P_J][p7P_MOVE]);
XMX(i,p7G_C) = XMX(i+1,p7G_C) + gm->xsc[p7P_C][p7P_LOOP];
XMX(i,p7G_E) = p7_FLogsum( XMX(i, p7G_J) + gm->xsc[p7P_E][p7P_LOOP],
XMX(i, p7G_C) + gm->xsc[p7P_E][p7P_MOVE]);
XMX(i,p7G_N) = p7_FLogsum( XMX(i+1,p7G_N) + gm->xsc[p7P_N][p7P_LOOP],
XMX(i, p7G_B) + gm->xsc[p7P_N][p7P_MOVE]);
MMX(i,M) = DMX(i,M) = XMX(i,p7G_E);
IMX(i,M) = -eslINFINITY;
for (k = M-1; k >= 1; k--)
{
MMX(i,k) = p7_FLogsum( p7_FLogsum(MMX(i+1,k+1) + TSC(p7P_MM,k) + MSC(k+1),
IMX(i+1,k) + TSC(p7P_MI,k) + ISC(k)),
p7_FLogsum(XMX(i,p7G_E) + esc,
DMX(i, k+1) + TSC(p7P_MD,k)));
IMX(i,k) = p7_FLogsum( MMX(i+1,k+1) + TSC(p7P_IM,k) + MSC(k+1),
IMX(i+1,k) + TSC(p7P_II,k) + ISC(k));
DMX(i,k) = p7_FLogsum( MMX(i+1,k+1) + TSC(p7P_DM,k) + MSC(k+1),
p7_FLogsum( DMX(i, k+1) + TSC(p7P_DD,k),
XMX(i, p7G_E) + esc));
}
}
/* At i=0, only N,B states are reachable. */
rsc = gm->rsc[dsq[1]];
XMX(0,p7G_B) = MMX(1,1) + TSC(p7P_BM,0) + MSC(1); /* t_BM index is 0 because it's stored off-by-one. */
for (k = 2; k <= M; k++)
XMX(0,p7G_B) = p7_FLogsum(XMX(0, p7G_B), MMX(1,k) + TSC(p7P_BM,k-1) + MSC(k));
XMX(i,p7G_J) = -eslINFINITY;
XMX(i,p7G_C) = -eslINFINITY;
XMX(i,p7G_E) = -eslINFINITY;
XMX(i,p7G_N) = p7_FLogsum( XMX(1, p7G_N) + gm->xsc[p7P_N][p7P_LOOP],
XMX(0, p7G_B) + gm->xsc[p7P_N][p7P_MOVE]);
for (k = M; k >= 1; k--)
MMX(i,M) = IMX(i,M) = DMX(i,M) = -eslINFINITY;
if (opt_sc != NULL) *opt_sc = XMX(0,p7G_N);
gx->M = M;
gx->L = L;
return eslOK;
}