static int
make_occasionally_dishonest_casino(ESL_HMM **ret_hmm, ESL_ALPHABET **ret_abc)
{
ESL_ALPHABET *abc = esl_alphabet_Create(eslDICE);
ESL_HMM *hmm = esl_hmm_Create(abc, 2);
int x;
/* State 0 = fair die */
hmm->pi[0] = 1.0;
hmm->pi[1] = 0.0;
hmm->pi[2] = 0.0; /* no L=0 seqs */
hmm->t[0][0] = 0.96;
hmm->t[0][1] = 0.03;
hmm->t[0][2] = 0.01; /* end from state 0; mean length 100 */
for (x = 0; x < abc->K; x++)
hmm->e[0][x] = 1.0 / (float) abc->K;
/* State 1 = loaded die */
hmm->t[1][0] = 0.05;
hmm->t[1][1] = 0.95;
hmm->t[1][2] = 0.0; /* no end from state 1 */
for (x = 0; x < abc->K-1; x++) hmm->e[1][x] = 0.5 / ((float) abc->K-1);
hmm->e[1][abc->K-1] = 0.5;
esl_hmm_Configure(hmm, NULL);
*ret_hmm = hmm;
*ret_abc = abc;
return eslOK;
}
/* Function: p7_bg_SetFilter()
* Synopsis: Configure filter HMM with new model composition.
* Incept: SRE, Fri Dec 5 09:08:15 2008 [Janelia]
*
* Purpose: The "filter HMM" is an experimental filter in the
* acceleration pipeline for avoiding biased composition
* sequences. It has no effect on final scoring, if a
* sequence passes all steps of the pipeline; it is only
* used to eliminate biased sequences from further
* consideration early in the pipeline, before the big guns
* of domain postprocessing are applied.
*
* At least at present, it doesn't actually work as well as
* one would hope. This will be an area of future work.
* What we really want to do is make a better null model of
* real protein sequences (and their biases), and incorporate
* that model into the flanks (NCJ states) of the profile.
*
* <compo> is the average model residue composition, from
* either the HMM or the copy in a profile or optimized
* profile. <M> is the length of the model in nodes.
*
* Returns: <eslOK> on success.
*
* Throws: (no abnormal error conditions)
*
* Xref: J4/25: generalized to use composition vector, not
* specifically an HMM.
*
* Note: This looks like a two-state HMM, but if you start thinking
* about its length distribution ("oh my god, L0 assumes a
* fixed L=400 expectation, it's all wrong, it's not conditional
* on the target sequence length and length modeling's messed
* up!"), don't panic. It's set up as a conditional-on-L model;
* the P(L) term is added in p7_bg_FilterScore() below.
*/
int
p7_bg_SetFilter(P7_BG *bg, int M, const float *compo)
{
float L0 = 400.0; /* mean length in state 0 of filter HMM (normal background) */
float L1 = (float) M / 8.0; /* mean length in state 1 of filter HMM (biased segment) */
/* State 0 is the normal iid model. */
bg->fhmm->t[0][0] = L0 / (L0+1.0f);
bg->fhmm->t[0][1] = 1.0f / (L0+1.0f);
bg->fhmm->t[0][2] = 1.0f; /* 1.0 transition to E means we'll set length distribution externally. */
esl_vec_FCopy(bg->f, bg->abc->K, bg->fhmm->e[0]);
/* State 1 is the potentially biased model composition. */
bg->fhmm->t[1][0] = 1.0f / (L1+1.0f);
bg->fhmm->t[1][1] = L1 / (L1+1.0f);
bg->fhmm->t[1][2] = 1.0f; /* 1.0 transition to E means we'll set length distribution externally. */
esl_vec_FCopy(compo, bg->abc->K, bg->fhmm->e[1]);
bg->fhmm->pi[0] = 0.999;
bg->fhmm->pi[1] = 0.001;
esl_hmm_Configure(bg->fhmm, bg->f);
return eslOK;
}
ESL_HMM *
create_null_hmm(ESL_ALPHABET *abc)
{
ESL_HMM *hmm;
hmm = esl_hmm_Create(abc, 1);
/* state 0 = normal iid model.*/
hmm->t[0][0] = 1.0f;
hmm->t[0][1] = 1.0f; /* external length distribution */
/* SW50 iid frequencies: H3 default background */
hmm->e[0][0] = 0.0787945; /* A */
hmm->e[0][1] = 0.0151600; /* C */
hmm->e[0][2] = 0.0535222; /* D */
hmm->e[0][3] = 0.0668298; /* E */
hmm->e[0][4] = 0.0397062; /* F */
hmm->e[0][5] = 0.0695071; /* G */
hmm->e[0][6] = 0.0229198; /* H */
hmm->e[0][7] = 0.0590092; /* I */
hmm->e[0][8] = 0.0594422; /* K */
hmm->e[0][9] = 0.0963728; /* L */
hmm->e[0][10] = 0.0237718; /* M */
hmm->e[0][11] = 0.0414386; /* N */
hmm->e[0][12] = 0.0482904; /* P */
hmm->e[0][13] = 0.0395639; /* Q */
hmm->e[0][14] = 0.0540978; /* R */
hmm->e[0][15] = 0.0683364; /* S */
hmm->e[0][16] = 0.0540687; /* T */
hmm->e[0][17] = 0.0673417; /* V */
hmm->e[0][18] = 0.0114135; /* W */
hmm->e[0][19] = 0.0304133; /* Y */
hmm->pi[0] = 1.0;
esl_hmm_Configure(hmm, NULL);
return hmm;
}
ESL_HMM *
create_test_hmm(ESL_ALPHABET *abc)
{
ESL_HMM *hmm;
int L = 400;
int M = 200;
hmm = esl_hmm_Create(abc, 2);
/* state 0 = normal iid model. state 1 = biased state */
hmm->t[0][0] = (float) L / (float) (L+1);
hmm->t[0][1] = 1.0f / (float) (L+1);
hmm->t[0][2] = 1.0; /* external length distribution */
hmm->t[1][0] = (float) 2.0f / (float) (M+2);
hmm->t[1][1] = (float) M / (float) (M+2);
hmm->t[1][2] = 1.0;
/* SW50 iid frequencies: H3 default background */
hmm->e[0][0] = 0.0787945; /* A */
hmm->e[0][1] = 0.0151600; /* C */
hmm->e[0][2] = 0.0535222; /* D */
hmm->e[0][3] = 0.0668298; /* E */
hmm->e[0][4] = 0.0397062; /* F */
hmm->e[0][5] = 0.0695071; /* G */
hmm->e[0][6] = 0.0229198; /* H */
hmm->e[0][7] = 0.0590092; /* I */
hmm->e[0][8] = 0.0594422; /* K */
hmm->e[0][9] = 0.0963728; /* L */
hmm->e[0][10] = 0.0237718; /* M */
hmm->e[0][11] = 0.0414386; /* N */
hmm->e[0][12] = 0.0482904; /* P */
hmm->e[0][13] = 0.0395639; /* Q */
hmm->e[0][14] = 0.0540978; /* R */
hmm->e[0][15] = 0.0683364; /* S */
hmm->e[0][16] = 0.0540687; /* T */
hmm->e[0][17] = 0.0673417; /* V */
hmm->e[0][18] = 0.0114135; /* W */
hmm->e[0][19] = 0.0304133; /* Y */
/* average of MFS_1 core emissions */
hmm->e[1][0] = 0.1068; /* A */
hmm->e[1][1] = 0.0110; /* C */
hmm->e[1][2] = 0.0242; /* D */
hmm->e[1][3] = 0.0293; /* E */
hmm->e[1][4] = 0.0621; /* F */
hmm->e[1][5] = 0.0899; /* G */
hmm->e[1][6] = 0.0139; /* H */
hmm->e[1][7] = 0.0762; /* I */
hmm->e[1][8] = 0.0319; /* K */
hmm->e[1][9] = 0.1274; /* L */
hmm->e[1][10] = 0.0338; /* M */
hmm->e[1][11] = 0.0285; /* N */
hmm->e[1][12] = 0.0414; /* P */
hmm->e[1][13] = 0.0266; /* Q */
hmm->e[1][14] = 0.0375; /* R */
hmm->e[1][15] = 0.0747; /* S */
hmm->e[1][16] = 0.0568; /* T */
hmm->e[1][17] = 0.0815; /* V */
hmm->e[1][18] = 0.0161; /* W */
hmm->e[1][19] = 0.0303; /* Y */
hmm->pi[0] = 0.99;
hmm->pi[1] = 0.01;
esl_hmm_Configure(hmm, NULL);
return hmm;
}
