int main(int argc, char const *argv[], char const *envp[]) {
PRE();
DCL();
EXP();
INT();
FLP();
ARR();
STR();
MEM();
ENV(envp);
SIG();
ERR();
MSC();
POS();
return 0;
}
/* Function: p7_GMSV()
* Synopsis: The MSV score algorithm (slow, correct version)
* Incept: SRE, Thu Dec 27 08:33:39 2007 [Janelia]
*
* Purpose: Calculates the maximal score of ungapped local segment
* pair alignments, taking advantage of the fact that this
* is simply equivalent to setting all MM transitions to 1.0
* in a multihit local profile.
*
* Args: dsq - sequence in digitized form, 1..L
* L - length of dsq
* gm - profile (can be in any mode)
* gx - DP matrix with room for an MxL alignment
* nu - configuration: expected number of hits (use 2.0 as a default)
* opt_sc - optRETURN: MSV lod score in nats.
*
* Returns: <eslOK> on success.
*
* Note: This is written deliberately as a modified p7_GViterbi
* routine. It could be faster -- we don't need the
* interleaved dp matrix or residue scores, since we aren't
* calculating D or I states, for example, and we could do
* without some of the special states -- but speed is the
* job of the optimized implementations. Rather, the goal
* here is to establish a stable, probabilistically correct
* reference calculation. (Thus, the CC, NN, JJ transitions
* are real scores here, not fixed to 0 as in the optimized
* versions.)
*/
int
p7_GMSV(const ESL_DSQ *dsq, int L, const P7_PROFILE *gm, P7_GMX *gx, float nu, float *opt_sc)
{
float **dp = gx->dp;
float *xmx = gx->xmx;
float tloop = logf((float) L / (float) (L+3));
float tmove = logf( 3.0f / (float) (L+3));
float tbmk = logf( 2.0f / ((float) gm->M * (float) (gm->M+1)));
float tej = logf((nu - 1.0f) / nu);
float tec = logf(1.0f / nu);
int i,k;
XMX(0,p7G_N) = 0;
XMX(0,p7G_B) = tmove; /* 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) = -eslINFINITY; /* need seq to get here */
for (i = 1; i <= L; i++)
{
float const *rsc = gm->rsc[dsq[i]];
MMX(i,0) = -eslINFINITY;
XMX(i,p7G_E) = -eslINFINITY;
for (k = 1; k <= gm->M; k++)
{
MMX(i,k) = MSC(k) + ESL_MAX(MMX(i-1,k-1), XMX(i-1,p7G_B) + tbmk);
XMX(i,p7G_E) = ESL_MAX(XMX(i,p7G_E), MMX(i,k));
}
XMX(i,p7G_J) = ESL_MAX( XMX(i-1,p7G_J) + tloop, XMX(i, p7G_E) + tej);
XMX(i,p7G_C) = ESL_MAX( XMX(i-1,p7G_C) + tloop, XMX(i, p7G_E) + tec);
XMX(i,p7G_N) = XMX(i-1,p7G_N) + tloop;
XMX(i,p7G_B) = ESL_MAX( XMX(i, p7G_N) + tmove, XMX(i, p7G_J) + tmove);
}
gx->M = gm->M;
gx->L = L;
if (opt_sc != NULL) *opt_sc = XMX(L,p7G_C) + tmove;
return eslOK;
}
/* 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_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_SparseViterbi()
* Synopsis: Viterbi optimal path algorithm, in sparse DP.
*
* Purpose: Compare profile <gm> to digital sequence <dsq> of length <L>,
* by the Viterbi algorithm, using sparse dynamic programming,
* as constrained by the sparse mask <sm>.
* Fill in the sparse Viterbi matrix <sx>; (optionally) trace
* back the optimal path and return it in the trace structure <opt_tr>
* if the caller provides one; and (optionally) return the
* Viterbi raw score in nats in <*opt_sc>.
*
* <sx> can be reused from previous calculations, even
* smaller ones; see <p7_sparsemx_Reuse()>. If necessary,
* it will be reallocated here, to be large enough for the
* <gm->M> by <L> calculation restricted to masked cells
* <sm>.
*
* Args: dsq - digital target sequence, 1..L
* L - length of <dsq>
* gm - profile
* sm - sparse mask
* sx - Viterbi matrix to fill; (may be reallocated here)
* opt_tr - optRESULT: trace structure with optimal traceback; or NULL if caller doesn't want it
* opt_sc - optRETURN: raw Viterbi score in nats; or NULL if result unwanted
*
* Returns: <eslOK> on success; <opt_tr>, if non-NULL, contains the optimal traceback;
* and <*opt_sc> optionally contains the raw Viterbi score.
*/
int
p7_SparseViterbi(const ESL_DSQ *dsq, int L, const P7_PROFILE *gm, const P7_SPARSEMASK *sm, P7_SPARSEMX *sx, P7_TRACE *opt_tr, float *opt_sc)
{
float const *tsc = gm->tsc; /* sets up TSC() macro, access to profile's transitions */
float const *rsc; /* will be set up for MSC(), ISC() macros for residue scores */
float *xpc; /* ptr that steps through current special cells */
float *dpc; /* ptr to step thru current row i main DP cells */
float *dpp; /* ptr to step thru previous row i-1 main DP cells */
float *last_dpc; /* used to reinit dpp after each sparse row computation */
int ng;
float xE, xN, xJ, xB, xL, xG, xC; /* tmp scores on special states. only stored when in row bands, and on ia-1 before a seg */
float mlc, mgc; /* temporary score calculations M(i,k) */
float dlc, dgc; /* precalculated D(i,k+1) value on current row */
int *kc = sm->k[0]; /* <kc> points to the list of sparse cell indices k for current row i */
int *kp; /* <kp> points to the previous row's sparse cell index list */
int i,k; /* i,k row,col (seq position, profile position) cell coords */
int y,z; /* indices in lists of k coords on prev, current row */
int status;
/* Contract checks on arguments */
ESL_DASSERT1( (sm->L == L) );
ESL_DASSERT1( (sm->M == gm->M) );
/* Assure that <sx> is allocated large enough (we might be reusing it).
* Set its type now, so we can Dump/Validate/etc. during debugging this routine, if needed.
*/
if ( (status = p7_sparsemx_Reinit(sx, sm)) != eslOK) return status;
sx->type = p7S_VITERBI;
xN = 0.0f;
xJ = -eslINFINITY;
xC = -eslINFINITY;
ng = 0;
xpc = sx->xmx;
dpc = sx->dp;
for (i = 1; i <= L; i++)
{
if (! sm->n[i]) { ng++; continue; } /* skip rows that have no included cells */
/* Reinitialize and store specials for row ia-1 just outside sparsified segment */
if (i == 1 || ng) {
*xpc++ = xE = -eslINFINITY;
*xpc++ = xN = xN + ( ng ? ng * gm->xsc[p7P_N][p7P_LOOP] : 0.0); /* test ng, because we must watch out for 0*-inf special case */
*xpc++ = xJ = xJ + ( ng ? ng * gm->xsc[p7P_J][p7P_LOOP] : 0.0);
*xpc++ = xB = ESL_MAX( xN + gm->xsc[p7P_N][p7P_MOVE], xJ + gm->xsc[p7P_J][p7P_MOVE]);
*xpc++ = xL = xB + gm->xsc[p7P_B][0]; /* B->L */
*xpc++ = xG = xB + gm->xsc[p7P_B][1]; /* B->G */
*xpc++ = xC = xC + ( ng ? ng * gm->xsc[p7P_C][p7P_LOOP] : 0.0);
*xpc++ = -eslINFINITY; /* JJ: this space only used in a Decoding matrix. */
*xpc++ = -eslINFINITY; /* CC: this space only used in a Decoding matrix. */
ng = 0;
}
rsc = gm->rsc[dsq[i]]; /* now MSC(k), ISC(k) residue score macros work */
last_dpc = dpc; /* remember where dpc started; dpp will be set here after we finish each row calculation */
kp = kc; /* last row we did becomes prev row now; ready to step through k indices of previous row's sparse cells */
kc = sm->k[i]; /* ditto for current row i */
dlc = dgc = xE = -eslINFINITY;
for (z=0, y=0; z < sm->n[i]; z++) /* Iterate over the one or more sparse cells (i,k) that we calculate on this row. */
{
k = kc[z]; /* next sparse cell to calculate: (i,k) */
/* Try to find cell i-1,k-1; then compute M(i,k) from it */
mlc = xL + TSC(p7P_LM, k-1);
mgc = xG + TSC(p7P_GM, k-1);
while (y < sm->n[i-1] && kp[y] < k-1) { y++; dpp+=p7S_NSCELLS; }
if (y < sm->n[i-1] && kp[y] == k-1) {
mlc = ESL_MAX( ESL_MAX( dpp[p7R_ML] + TSC(p7P_MM, k-1),
dpp[p7R_IL] + TSC(p7P_IM, k-1)),
ESL_MAX( dpp[p7R_DL] + TSC(p7P_DM, k-1),
mlc));
mgc = ESL_MAX( ESL_MAX( dpp[p7R_MG] + TSC(p7P_MM, k-1),
dpp[p7R_IG] + TSC(p7P_IM, k-1)),
ESL_MAX( dpp[p7R_DG] + TSC(p7P_DM, k-1),
mgc));
}
*dpc++ = mlc = MSC(k) + mlc;
*dpc++ = mgc = MSC(k) + mgc;
/* Try to find cell i-1,k; then compute I(i,k) from it */
while (y < sm->n[i-1] && kp[y] < k) { y++; dpp+=p7S_NSCELLS; }
if (y < sm->n[i-1] && kp[y] == k) {
*dpc++ = ESL_MAX( dpp[p7R_ML] + TSC(p7P_MI,k), dpp[p7R_IL] + TSC(p7P_II, k)); // +ISC(k), if we weren't enforcing it to zero
*dpc++ = ESL_MAX( dpp[p7R_MG] + TSC(p7P_MI,k), dpp[p7R_IG] + TSC(p7P_II, k)); // ditto
} else {
*dpc++ = -eslINFINITY;
*dpc++ = -eslINFINITY;
}
/* local exit paths (a F/V difference here: in V, no Dk->E path can win */
xE = ESL_MAX(xE, mlc);
/* delayed store of Dk; advance calculation of next D_k+1 */
*dpc++ = dlc;
*dpc++ = dgc;
if (z < sm->n[i]-1 && kc[z+1] == k+1) { /* is there a (i,k+1) cell to our right? */
dlc = ESL_MAX( mlc + TSC(p7P_MD, k), dlc + TSC(p7P_DD, k));
dgc = ESL_MAX( mgc + TSC(p7P_MD, k), dgc + TSC(p7P_DD, k));
} else { /* if not, we MUST consider {MD}Gk->Dk+1..E glocal exit path, even from internal sparse cells - not just last cell! */
xE = ESL_MAX( xE, TSC(p7P_DGE, k) + ESL_MAX( mgc + TSC(p7P_MD, k), dgc + TSC(p7P_DD, k))); // yes, the D path can contribute; we only use wing-retraction on sparse cells k where k+1 is unmarked; if k=M, for example, we must check D->E
dlc = dgc = -eslINFINITY;
}
}
*xpc++ = xE; // we already max'ed over all Mk->E exits, both local and glocal
*xpc++ = xN = xN + gm->xsc[p7P_N][p7P_LOOP];
*xpc++ = xJ = ESL_MAX( xJ + gm->xsc[p7P_J][p7P_LOOP], xE + gm->xsc[p7P_E][p7P_LOOP]);
*xpc++ = xB = ESL_MAX( xJ + gm->xsc[p7P_J][p7P_MOVE], xN + gm->xsc[p7P_N][p7P_MOVE]);
*xpc++ = xL = xB + gm->xsc[p7P_B][0]; /* B->L */
*xpc++ = xG = xB + gm->xsc[p7P_B][1]; /* B->G */
*xpc++ = xC = ESL_MAX( xE + gm->xsc[p7P_E][p7P_MOVE], xC + gm->xsc[p7P_C][p7P_LOOP]);
*xpc++ = -eslINFINITY; /* JJ: this space only used in a Decoding matrix. */
*xpc++ = -eslINFINITY; /* CC: this space only used in a Decoding matrix. */
/* now dpc is on the start of the next sparsified row */
dpp = last_dpc;
}
xC += ( ng ? ng * gm->xsc[p7P_C][p7P_LOOP] : 0.0f) + gm->xsc[p7P_C][p7P_MOVE];
if (opt_sc) *opt_sc = xC;
if (opt_tr && xC != -eslINFINITY) return p7_sparse_trace_Viterbi(gm, sx, opt_tr);
else 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;
}