def calc_lik_with_clonal(combo, si, di, phi_i, pi, ni):
# calculate with given phi
# (lls currently uses precision fudge factor to get around 0 probability errors when pv = 1)
pvs = np.array([get_pv(phi_i, c, pi, ni) for c in combo])
lls = np.array([pm.binomial_like(si, di, pvs[i]) for i,c in enumerate(combo)])-0.00000001
# calculate with clonal phi
pvs_cl = np.array([get_pv(np.array(1), c, pi, ni) for c in combo])
lls_cl = np.array([pm.binomial_like(si, di, pvs_cl[i]) for i,c in enumerate(combo)])-0.00000001
return np.array([[pvs, lls], [pvs_cl, lls_cl]])
def test_withzeros(self):
"Makes sure no NaN's happen when some probabilities are zero."
q = np.zeros(npix)
q[0]=.99
# Log-distribution of number of pixels positive.
lpf = anopheles.utils.ubl(q)
lpp = unequal_binomial_lp(q)
pp = np.exp(lpp)
pf = np.exp(lpf)
assert_equal(pp,pf)
lpb = anopheles.utils.bin_ubl(3,8,.1,q)
assert(not np.isnan(lpb))
assert(not np.any(np.isnan(pf)))
# Binomial mixture from Fortran.
pbf = np.exp([anopheles.utils.bin_ubl(x,n_obs,prob_detect,q) for x in xrange(n_obs+1)])
# Do binomial mixture by hand.
pbp = np.zeros(n_obs+1)
for i in xrange(npix+1):
pbp+=np.exp([pm.binomial_like(x,8,prob_detect*float(i)/npix)+lpf[i] for x in xrange(n_obs+1)])
assert_almost_equal(pbf,pbp)
def obs(pi=pi, phi=phi):
logp = pl.log(1 - phi) * num_nonzeros + mc.binomial_like(
r[nonzeros] * n[nonzeros], n[nonzeros], pi[nonzeros])
for n_i in n[~nonzeros]:
logp += pl.log(phi + (1 - phi) *
pl.exp(pl.log(1 - pi[~nonzeros]) * n[~nonzeros])).sum()
return logp
def N_pos_now(
value=pm.utils.round_array(pos[this_slice]),
n=pm.utils.round_array(pos[this_slice] + neg[this_slice]),
eps_p_f=eps_p_f_now,
a1=a1,
a2=a2,
):
return pm.binomial_like(value, n=n, p=pm.flib.stukel_invlogit(eps_p_f, a1, a2))
def test_binomial_case(self):
"""Checks for correspondence with the binomial distribution in the case of equal
presence probabilities."""
q = np.ones(5)*.2
lpf, lpp, pbf, pbp = standard_things(q)
# In this case you can compute the log-p of number of pixels positive directly.
lpo = np.array([pm.binomial_like(x,npix,q[0]) for x in range(npix+1)])
assert_almost_equal(lpf, lpo)
assert_almost_equal(lpp, lpo)
def standard_things(q,prob_detect=prob_detect,n_obs=n_obs):
npix=len(q)
# Log-distribution of number of pixels positive.
lpf = anopheles.utils.ubl(q)
lpp = unequal_binomial_lp(q)
assert_equal(lpf, lpp)
# Binomial mixture from Fortran.
pbf = np.exp([anopheles.utils.bin_ubl(x,n_obs,prob_detect,q) for x in xrange(n_obs+1)])
# Do binomial mixture by hand.
pbp = np.zeros(n_obs+1)
for i in xrange(npix+1):
pbp+=np.exp([pm.binomial_like(x,n_obs,prob_detect*float(i)/npix)+lpf[i] for x in xrange(n_obs+1)])
assert_almost_equal(pbf,pbp)
return lpf, lpp, pbf, pbp
def like(self,
fdia=256,
D=14.0**2 * pi,
r=0.078,
x0=29.7,
y0=91.7,
oo=-9.0,
pe=0.414,
dinvasion=296.4,
Kbaldio=0.0029,
Dbaldio=14.0**2 * pi,
per=0.0,
mxr=1.0,
vparams=False,
hilosgpu=32,
v=0.0024,
vermapa=False):
#print fdia, D, r, x0, y0, oo, pe, dinvasion, Kbaldio, Dbaldio, per, mxr, hilosgpu, v,
self.simula(fdia,
D,
r,
x0,
y0,
oo,
pe,
dinvasion,
Kbaldio,
Dbaldio,
per,
mxr,
hilosgpu=hilosgpu,
v=v,
vermapa=vermapa)
l0 = binomial_like(self.mosquitos[:, 4], 1.,
clip(self.esperados, 1e-6, 1 - 1e-6))
if isnan(l0): l0 = -1e9
elif isinf(l0): l0 = -1e9
self.likelihood = l0
return l0
def calc_lik(combo, si, di, phi_i, pi, ni):
pvs = np.array([get_pv(phi_i, c, pi, ni) for c in combo])
lls = np.array([pm.binomial_like(si, di, pvs[i]) for i,c in enumerate(combo)])-0.00000001
return np.array([pvs, lls])
def data_vivax(value = vivax_pos[where_vivax], splrep = None, p = p_vivax, n = np.sum(cur_obs,axis=1)):
return pm.binomial_like(x=value, n=n, p=p)
def obs(pi=pi):
return mc.binomial_like(r*n, n, pi)
pop_C_prev = pop_C_k.stats()["quantiles"][50] / float(pop_C_N)
pop_C_prev_per_1000 = "%.0f" % (pop_C_prev * 1000)
print pop_C_prev_per_1000
pop_C_ui = pop_C_k.stats()["95% HPD interval"] / float(pop_C_N)
pop_C_ui_per_1000 = "[%.0f, %.0f]" % tuple(pop_C_ui * 1000)
print pop_C_ui_per_1000
### @export 'binomial-model-ppc'
r = pl.array(schiz["r"])
n = pl.array(schiz["n"], dtype=int)
k = r * n
pi = mc.Uninformative("pi", value=0.5)
mc.binomial_like(k, n, pi)
@mc.potential
def obs(pi=pi):
return mc.binomial_like(k, n, pi)
@mc.deterministic
def pred(pi=pi):
return mc.rbinomial(n, pi)
mc.MCMC([pi, obs, pred]).sample(20000, 10000, 10, verbose=False, progress_bar=False)
pl.figure(**book_graphics.quarter_page_params)
def survey_likelihood(sp_sub, survey_plan, data, i, a1, a2):
data_ = np.ones_like(sp_sub)*data[i]
return pm.binomial_like(data_, survey_plan.n[i], pm.stukel_invlogit(sp_sub, a1, a2))
pop_C_prev = pop_C_k.stats()['quantiles'][50] / float(pop_C_N)
pop_C_prev_per_1000 = '%.0f' % (pop_C_prev * 1000)
print pop_C_prev_per_1000
pop_C_ui = pop_C_k.stats()['95% HPD interval'] / float(pop_C_N)
pop_C_ui_per_1000 = '[%.0f, %.0f]' % tuple(pop_C_ui * 1000)
print pop_C_ui_per_1000
### @export 'binomial-model-ppc'
r = pl.array(schiz['r'])
n = pl.array(schiz['n'], dtype=int)
k = r * n
pi = mc.Uninformative('pi', value=.5)
mc.binomial_like(k, n, pi)
@mc.potential
def obs(pi=pi):
return mc.binomial_like(k, n, pi)
@mc.deterministic
def pred(pi=pi):
return mc.rbinomial(n, pi)
mc.MCMC([pi, obs, pred]).sample(20000,
10000,
10,
def p_obs(value=p, pi=pi, n=n):
return mc.binomial_like(value * n, n, pi + 1.0e-9)
def p_obs(value=p, pi=pi, n=n):
return mc.binomial_like(value * n, n, pi + 1.e-9)
def d_now(value = vivax_pos[i], splrep = splreps[i_vivax], p = p, n = np.sum(cur_obs)):
return pm.binomial_like(x=value, n=n, p=p)
r = k/n
iter = 20000
burn = 10000
thin = 10
results = {}
xmax = .07
### @export 'distribution-comparison'
pl.figure(**book_graphics.quarter_page_params)
ax = pl.axes([.1, .3, .85, .65])
x = pl.arange(0, n_small*pi_true*4, .1)
# plot binomial distribution
y1 = [pl.exp(mc.binomial_like(x_i, n_small, pi_true)) for x_i in x]
pl.step(x, y1, 'k',
linewidth=1, linestyle='step:', alpha=.8,
label='Binomial')
# plot poisson distribution
y2 = [pl.exp(mc.poisson_like(x_i, n_small*pi_true)) for x_i in x]
pl.plot(x, y2, 'k',
linewidth=1, linestyle='steps--', alpha=.8,
label='Poisson')
pl.legend(loc='upper right', fancybox=True, shadow=True)
pl.yticks([0, .05])
pl.xticks([25, 50, 75], ['','',''])
pl.axis([-.1, n_small*pi_true*4, -.02, 1.1*max(y1)])
pl.xlabel('Count')
def data_vivax(value=vivax_pos[where_vivax], splrep=None, p=p_vivax, n=np.sum(cur_obs, axis=1)):
return pm.binomial_like(x=value, n=n, p=p)
r = k / n
iter = 20000
burn = 10000
thin = 10
results = {}
xmax = .07
### @export 'distribution-comparison'
pl.figure(**book_graphics.quarter_page_params)
ax = pl.axes([.1, .3, .85, .65])
x = pl.arange(0, n_small * pi_true * 4, .1)
# plot binomial distribution
y1 = [pl.exp(mc.binomial_like(x_i, n_small, pi_true)) for x_i in x]
pl.step(x, y1, 'k', linewidth=1, linestyle='step:', alpha=.8, label='Binomial')
# plot poisson distribution
y2 = [pl.exp(mc.poisson_like(x_i, n_small * pi_true)) for x_i in x]
pl.plot(x,
y2,
'k',
linewidth=1,
linestyle='steps--',
alpha=.8,
label='Poisson')
pl.legend(loc='upper right', fancybox=True, shadow=True)
pl.yticks([0, .05])
pl.xticks([25, 50, 75], ['', '', ''])
def survey_likelihood(x, survey_plan, data, i):
data_ = np.ones_like(x)*data[i]
return pm.binomial_like(data_, survey_plan.n[i], pm.invlogit(x))
def p_obs(value=p, pi=pi_latent, n=n):
pi_flat = pl.array(pi)
return mc.binomial_like((value * n)[i_nonzero], n[i_nonzero],
pi_flat[i_nonzero])
def obs(value=r, n=n, logit_p=logit_p):
return mc.binomial_like(r, n, mc.invlogit(logit_p))
def p_obs(value=p, pi=pi_latent, n=n):
pi_flat = pl.array(pi)
return mc.binomial_like((value * n)[i_nonzero], n[i_nonzero], pi_flat[i_nonzero])
def post_process_clusters(mcmc, sv_df, snv_df, clus_out_dir, sup, dep, norm,
cn_states, sparams, cparams, output_params, map_):
merge_clusts = cparams['merge_clusts']
subclone_diff = cparams['subclone_diff']
phi_limit = cparams['phi_limit']
merge_clusts = cparams['merge_clusts']
cnv_pval = cparams['clonal_cnv_pval']
hpd_alpha = cparams['hpd_alpha']
adjust_phis = cparams['adjust_phis']
clus_penalty = output_params['cluster_penalty']
smc_het = output_params['smc_het']
plot = output_params['plot']
try:
sv_df = sv_df[sv_df.classification.values != 'SIMU_SV']
except AttributeError:
pass
npoints = len(snv_df) + len(sv_df)
sup, dep, norm, cn_states = sup[:
npoints], dep[:
npoints], norm[:
npoints], cn_states[:
npoints]
z_trace = mcmc.trace('z')[:]
# assign points to highest probability cluster
clus_counts = [np.bincount(z_trace[:, i]) for i in range(npoints)]
clus_max_prob = [index_max(c) for c in clus_counts]
clus_mp_counts = np.bincount(clus_max_prob)
clus_idx = np.nonzero(clus_mp_counts)[0]
clus_mp_counts = clus_mp_counts[clus_idx]
# cluster distribution
clus_info = pd.DataFrame(clus_idx, columns=['clus_id'])
clus_info['size'] = clus_mp_counts
if len(clus_info) < 1:
print(
"Warning! Could not converge on any major SV clusters. Skipping.\n"
)
return None
center_trace = mcmc.trace("phi_k")[:]
phis = np.array([
mean_confidence_interval(center_trace[:, cid], hpd_alpha)
for cid in clus_idx
])
original_phis = phis.copy()
adjusted_phis = get_adjusted_phis(clus_info, center_trace, cparams)
hpd_lo = '_'.join([str(int(100 - (100 * hpd_alpha))), 'HPD', 'lo'])
hpd_hi = '_'.join([str(int(100 - (100 * hpd_alpha))), 'HPD', 'hi'])
phis = adjusted_phis if adjust_phis else phis
clus_info['phi'] = phis[:, 0]
clus_info[hpd_lo] = phis[:, 1]
clus_info[hpd_hi] = phis[:, 2]
if adjust_phis:
clus_info['phi_unadjusted'] = original_phis[:, 0]
clus_info['%s_unadjusted' % hpd_lo] = original_phis[:, 1]
clus_info['%s_unadjusted' % hpd_hi] = original_phis[:, 2]
else:
clus_info['phi_adjusted'] = adjusted_phis[:, 0]
clus_info['%s_adjusted' % hpd_lo] = adjusted_phis[:, 1]
clus_info['%s_adjusted' % hpd_hi] = adjusted_phis[:, 2]
clus_ids = clus_info.clus_id.values
clus_members = np.array(
[np.where(np.array(clus_max_prob) == i)[0] for i in clus_ids])
col_names = map(lambda x: 'cluster' + str(x), clus_ids)
df_probs = pd.DataFrame(clus_counts, dtype=float)[clus_ids].fillna(0)
df_probs = df_probs.apply(lambda x: x / sum(x), axis=1)
df_probs.columns = col_names
# cluster certainty
clus_max_df = pd.DataFrame(clus_max_prob,
columns=['most_likely_assignment'])
phi_cols = ["average_ccf", hpd_lo, hpd_hi]
phi_matrix = pd.DataFrame(phis[:], index=clus_ids,
columns=phi_cols).loc[clus_max_prob]
phi_matrix.index = range(len(phi_matrix))
ccert = clus_max_df.join(phi_matrix)
clus_info.index = range(len(clus_info))
print('\n\n')
print(clus_info[['clus_id', 'size', 'phi']])
print('Compiling and writing output...')
dump_out_dir = clus_out_dir
if len(snv_df) > 0 and len(sv_df) == 0:
# snvs only trace output
dump_out_dir = '%s/snvs' % clus_out_dir
trace_out = '%s/' % (dump_out_dir)
write_output.dump_trace(center_trace, trace_out + 'phi_trace.txt')
write_output.dump_trace(z_trace, trace_out + 'z_trace.txt')
try:
alpha_trace = mcmc.trace('alpha')[:]
write_output.dump_trace(alpha_trace, trace_out + 'alpha_trace.txt')
except KeyError:
pass
# cluster plotting
if plot:
plot_clusters(mcmc.trace, clus_idx, clus_max_prob, sup, dep,
clus_out_dir, cparams)
# merge clusters
if len(clus_info) > 1 and merge_clusts:
clus_merged = pd.DataFrame(columns=clus_info.columns,
index=clus_info.index)
clus_merged, clus_members, merged_ids = merge_clusters(clus_out_dir,clus_info,clus_merged,\
clus_members,[],sup,dep,norm,cn_states,sparams,cparams)
if len(clus_merged) != len(clus_info):
clus_info = clus_merged
df_probs, ccert = merge_results(clus_merged, merged_ids, df_probs,
ccert)
snv_probs = pd.DataFrame()
snv_ccert = pd.DataFrame()
snv_members = np.empty(0)
z_phi = get_per_variant_phi(z_trace, center_trace)
# compile run fit statistics
run_fit = pd.DataFrame()
if map_ is not None:
nclus = len(clus_info)
# bic = -2 * map_.lnL + (1 + npoints + nclus * 2) + (nclus * clus_penalty) * np.log(npoints)
phis = ccert.average_ccf.values
cns, pvs = cluster.get_most_likely_cn_states(cn_states, sup, dep, phis,
sparams['pi'], cnv_pval,
norm)
lls = []
for si, di, pvi in zip(sup, dep, pvs):
lls.append(pm.binomial_like(si, di, pvi))
svc_ic = -2 * np.sum(lls) + (npoints +
nclus * clus_penalty) * np.log(npoints)
run_fit = pd.DataFrame([['svc_IC', svc_ic], ['BIC', map_.BIC],
['AIC', map_.AIC], ['AICc', map_.AICc],
['lnL', map_.lnL], ['logp', map_.logp],
['logp_at_max', map_.logp_at_max],
['param_len', map_.len],
['data_len', map_.data_len]])
if len(snv_df) > 0:
snv_pos = ['chrom', 'pos']
snv_probs = df_probs.loc[:len(snv_df) - 1]
snv_probs = snv_df[snv_pos].join(snv_probs)
snv_ccert = ccert.loc[:len(snv_df) - 1]
snv_ccert = snv_df[snv_pos].join(snv_ccert)
snv_max_probs = np.array(clus_max_prob)[:len(snv_df)]
snv_members = np.array(
[np.where(np.array(snv_max_probs) == i)[0] for i in clus_ids])
snv_sup = sup[:len(snv_df)]
snv_dep = dep[:len(snv_df)]
snv_norm = norm[:len(snv_df)]
snv_cn_states = cn_states[:len(snv_df)]
snv_z_phi = z_phi[:len(snv_df)]
write_output.write_out_files(snv_df,
clus_info.copy(),
snv_members,
snv_probs,
snv_ccert,
clus_out_dir,
sparams['sample'],
sparams['pi'],
snv_sup,
snv_dep,
snv_norm,
snv_cn_states,
run_fit,
smc_het,
cnv_pval,
snv_z_phi,
are_snvs=True)
sv_probs = pd.DataFrame()
sv_ccert = pd.DataFrame()
sv_members = np.empty(0)
if len(sv_df) > 0:
lb = len(snv_df) if len(snv_df) > 0 else 0
sv_pos = ['chr1', 'pos1', 'dir1', 'chr2', 'pos2', 'dir2']
sv_probs = df_probs.loc[lb:lb + len(sv_df) - 1]
sv_probs.index = sv_df.index
sv_probs = sv_df[sv_pos].join(sv_probs)
sv_ccert = ccert.loc[lb:lb + len(sv_df) - 1]
sv_ccert.index = sv_df.index
sv_ccert = sv_df[sv_pos].join(sv_ccert)
sv_max_probs = np.array(clus_max_prob)[:len(sv_df)]
sv_members = np.array(
[np.where(np.array(sv_max_probs) == i)[0] for i in clus_ids])
sv_sup = sup[lb:lb + len(sv_df)]
sv_dep = dep[lb:lb + len(sv_df)]
sv_norm = norm[lb:lb + len(sv_df)]
sv_cn_states = cn_states[lb:lb + len(sv_df)]
sv_z_phi = z_phi[lb:lb + len(sv_df)]
write_output.write_out_files(sv_df, clus_info.copy(), sv_members,
sv_probs, sv_ccert, clus_out_dir,
sparams['sample'], sparams['pi'], sv_sup,
sv_dep, sv_norm, sv_cn_states, run_fit,
smc_het, cnv_pval, sv_z_phi)
def obs(pi=pi):
return mc.binomial_like(k, n, pi)
def obs(pi=pi, phi=phi):
logp = pl.log(1-phi)*num_nonzeros + mc.binomial_like(r[nonzeros]*n[nonzeros], n[nonzeros], pi[nonzeros])
for n_i in n[~nonzeros]:
logp += pl.log(phi + (1-phi) * pl.exp(pl.log(1-pi[~nonzeros]) * n[~nonzeros])).sum()
return logp
