In [5]:
import os, tempfile
import logging
import matplotlib as mpl
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np
In [6]:
from ionchannelABC import theoretical_population_size
from ionchannelABC import IonChannelDistance, EfficientMultivariateNormalTransition, IonChannelAcceptor
from ionchannelABC.experiment import setup
from ionchannelABC.visualization import plot_sim_results, plot_kde_matrix_custom
import myokit
In [7]:
from pyabc import Distribution, RV, History, ABCSMC
from pyabc.epsilon import MedianEpsilon
from pyabc.sampler import MulticoreEvalParallelSampler, SingleCoreSampler
from pyabc.populationstrategy import ConstantPopulationSize
Load experiments used for unified dataset calibration:
In [1]:
from experiments.ito_wang import wang_act, wang_inact
from experiments.ito_courtemanche import courtemanche_kin, courtemanche_rec, courtemanche_deact
In [2]:
modelfile = 'models/nygren_ito.mmt'
Plot steady-state and time constant functions of original model
In [3]:
from ionchannelABC.visualization import plot_variables
In [8]:
sns.set_context('talk')
V = np.arange(-100, 40, 0.01)
nyg_par_map = {'ri': 'ito.r_inf',
'si': 'ito.s_inf',
'rt': 'ito.tau_r',
'st': 'ito.tau_s'}
f, ax = plot_variables(V, nyg_par_map, modelfile, figshape=(2,2))
Combine model and experiments to produce:
In [9]:
observations, model, summary_statistics = setup(modelfile,
wang_act,
wang_inact,
courtemanche_kin,
courtemanche_deact,
courtemanche_rec)
In [10]:
assert len(observations)==len(summary_statistics(model({})))
In [11]:
g = plot_sim_results(modelfile,
wang_act,
wang_inact,
courtemanche_kin,
courtemanche_deact,
courtemanche_rec)
In [12]:
limits = {'ito.p1': (-100, 100),
'ito.p2': (1e-7, 50),
'log_ito.p3': (-7, 0),
'ito.p4': (1e-7, 50),
'log_ito.p5': (-7, 0),
'ito.q1': (-100, 100),
'ito.q2': (1e-7, 50),
'log_ito.q3': (-5, 1),
'ito.q4': (-100, 100),
'ito.q5': (1e-7, 50),
'log_ito.q6': (-7, 0)}
prior = Distribution(**{key: RV("uniform", a, b - a)
for key, (a,b) in limits.items()})
In [13]:
# Test this works correctly with set-up functions
assert len(observations) == len(summary_statistics(model(prior.rvs())))
Set-up path to results database.
In [12]:
db_path = ("sqlite:///" + os.path.join(tempfile.gettempdir(), "nygren_ito_unified.db"))
In [13]:
logging.basicConfig()
abc_logger = logging.getLogger('ABC')
abc_logger.setLevel(logging.DEBUG)
eps_logger = logging.getLogger('Epsilon')
eps_logger.setLevel(logging.DEBUG)
Test theoretical number of particles for approximately 2 particles per dimension in the initial sampling of the parameter hyperspace.
In [14]:
pop_size = theoretical_population_size(2, len(limits))
print("Theoretical minimum population size is {} particles".format(pop_size))
Initialise ABCSMC (see pyABC documentation for further details).
IonChannelDistance calculates the weighting applied to each datapoint based on the experimental variance.
In [15]:
abc = ABCSMC(models=model,
parameter_priors=prior,
distance_function=IonChannelDistance(
exp_id=list(observations.exp_id),
variance=list(observations.variance),
delta=0.05),
population_size=ConstantPopulationSize(2000),
summary_statistics=summary_statistics,
transitions=EfficientMultivariateNormalTransition(),
eps=MedianEpsilon(initial_epsilon=100),
sampler=MulticoreEvalParallelSampler(n_procs=16),
acceptor=IonChannelAcceptor())
In [16]:
obs = observations.to_dict()['y']
obs = {str(k): v for k, v in obs.items()}
In [17]:
abc_id = abc.new(db_path, obs)
Run calibration with stopping criterion of particle 1\% acceptance rate.
In [ ]:
history = abc.run(minimum_epsilon=0., max_nr_populations=100, min_acceptance_rate=0.01)
In [14]:
history = History('sqlite:///results/nygren/ito/unified/nygren_ito_unified.db')
In [15]:
df, w = history.get_distribution()
In [16]:
df.describe()
Out[16]:
In [17]:
sns.set_context('poster')
mpl.rcParams['font.size'] = 14
mpl.rcParams['legend.fontsize'] = 14
g = plot_sim_results(modelfile,
wang_act,
wang_inact,
courtemanche_kin,
courtemanche_deact,
courtemanche_rec,
df=df, w=w)
plt.tight_layout()
In [18]:
import pandas as pd
N = 100
nyg_par_samples = df.sample(n=N, weights=w, replace=True)
nyg_par_samples = nyg_par_samples.set_index([pd.Index(range(N))])
nyg_par_samples = nyg_par_samples.to_dict(orient='records')
In [19]:
sns.set_context('talk')
mpl.rcParams['font.size'] = 14
mpl.rcParams['legend.fontsize'] = 14
f, ax = plot_variables(V, nyg_par_map,
'models/nygren_ito.mmt',
[nyg_par_samples],
figshape=(2,2))
In [20]:
from ionchannelABC.visualization import plot_kde_matrix_custom
import myokit
import numpy as np
In [21]:
m,_,_ = myokit.load(modelfile)
In [22]:
originals = {}
for name in limits.keys():
if name.startswith("log"):
name_ = name[4:]
else:
name_ = name
val = m.value(name_)
if name.startswith("log"):
val_ = np.log10(val)
else:
val_ = val
originals[name] = val_
In [23]:
sns.set_context('paper')
g = plot_kde_matrix_custom(df, w, limits=limits, refval=originals)
plt.tight_layout()
In [ ]: