def shift_model(sample_N=None):
shift_model = GP('Nucleosome shift', sample_N=sample_N)
shift_model.design_matrix(incl_shift=True,
incl_prom=False,
incl_gene=False,
incl_antisense=False)
return shift_model
def small_promoter_model(sample_N=None):
small_prom_model = GP('Promoter occupancy', sample_N=sample_N)
small_prom_model.design_matrix(incl_prom=True,
incl_gene=False,
incl_cc=False,
incl_occ=True,
incl_small=True,
incl_nuc=False,
incl_antisense=False)
return small_prom_model
def gene_disorg_model(sample_N=None):
gene_disorg_model = GP('Nucleosome disorganization', sample_N=sample_N)
gene_disorg_model.design_matrix(incl_prom=False,
incl_gene=True,
incl_cc=True,
incl_occ=False,
incl_small=False,
incl_nuc=True,
incl_antisense=False)
return gene_disorg_model
def combined_model(sample_N=None):
combined_model = GP('Combined chromatin', sample_N=sample_N)
sm_model = small_promoter_model(sample_N)
disorg_model = gene_disorg_model(sample_N)
X1 = sm_model.X
X2 = disorg_model.X
same_cols = set(X1.columns).intersection(set(X2.columns))
for col in same_cols:
X1 = X1.drop(col, axis=1)
combined_model.X = X1.join(X2)
return combined_model
def regression_plots():
from src.regression_compare import plot_compare_r2, load_results
from src.gp import plot_res_distribution, plot_res_distribution_time
from src.gp import GP
gp_dir = "%s/gp" % OUTPUT_DIR
mkdirs_safe([gp_dir])
# plot comparison
plot_compare_r2(gp_dir)
plt.savefig('%s/compare_gp_r2.pdf' % gp_dir, transparent=True)
plot_compare_r2(gp_dir, show_legend=True)
plt.savefig('%s/compare_gp_r2_legend.pdf' % gp_dir, transparent=True)
from src.gp import plot_res_distribution_time, plot_res_distribution, GP
results = load_results(gp_dir)
for name in ['Full']:
cur = GP(name, results_path='%s/%s_results.csv' % (gp_dir, name))
plot_res_distribution(cur, selected_genes=selected_genes)
plt.savefig('%s/%s_predictions.pdf' % (gp_dir, name), transparent=True)
for time in [7.5, 30, 120]:
plot_res_distribution_time(cur,
time,
selected_genes=selected_genes)
plt.savefig('%s/%s_%s.pdf' % (gp_dir, name, time),
transparent=True,
dpi=100)
def rna_only_model(sample_N=None):
intercept_model = GP('RNA only', sample_N=sample_N)
intercept_model.design_matrix(incl_prom=False,
incl_gene=False,
incl_antisense=False)
return intercept_model
def prom_model(sample_N=None):
prom_model = GP('Promoter', sample_N=sample_N)
prom_model.design_matrix(incl_prom=True,
incl_gene=False,
incl_antisense=False)
return prom_model
def body_model(sample_N=None):
body_model = GP('Gene body', sample_N=sample_N)
body_model.design_matrix(incl_prom=False,
incl_gene=True,
incl_antisense=False)
return body_model
def full_model(sample_N=None):
full_model = GP('Full', sample_N=sample_N)
full_model.design_matrix(incl_shift=True)
return full_model
def nuc_model(sample_N=None):
nuc_model = GP('Nucleosome fragments', sample_N=sample_N)
nuc_model.design_matrix(incl_nuc=False, incl_antisense=False)
return nuc_model
def sm_model(sample_N=None):
sm_model = GP('Small fragments', sample_N=sample_N)
sm_model.design_matrix(incl_small=False, incl_antisense=False)
return sm_model
def occ_model(sample_N=None):
occ_model = GP('Occupancy', sample_N=sample_N)
occ_model.design_matrix(incl_cc=False, incl_occ=True, incl_antisense=False)
return occ_model
def cc_model(sample_N=None):
cc_model = GP('Cross correlation', sample_N=sample_N)
cc_model.design_matrix(incl_cc=True, incl_occ=False, incl_antisense=False)
return cc_model
def sense_model(sample_N=None):
sense_model = GP('Sense', sample_N=sample_N)
sense_model.design_matrix(incl_antisense=False)
return sense_model
def antisense_model(sample_N=None):
antisense_model = GP('Antisense', sample_N=sample_N)
antisense_model.design_matrix(incl_shift=False,
incl_antisense=True,
incl_sense=False)
return antisense_model
optimizer.zero_grad()
output = model(train_x)
loss = -mll(output, train_y)
loss.backward()
print('Iter %d/%d - Loss: %.3f' % (i + 1, training_iter, loss.item()))
optimizer.step()
return model
train_x = torch.linspace(0, 1, 15)
train_y = torch.sin(train_x * (2 * math.pi))
#model = test_sm(train_x, train_y)
kern = gpytorch.kernels.SpectralMixtureKernel(num_mixtures=4)
gp = GP(train_x, train_y, kern)
"""
likelihood = gpytorch.likelihoods.GaussianLikelihood()
model = SpectralMixtureGPModel(train_x, train_y, likelihood)
model.train()
likelihood.train()
# Use the adam optimizer
optimizer = torch.optim.Adam(model.parameters(), lr=0.1)
# "Loss" for GPs - the marginal log likelihood
mll = gpytorch.mlls.ExactMarginalLogLikelihood(likelihood, model)
#SM works from package but not when wrapped in GP
training_iter = 100
def scatters():
from src.chromatin_summary_plots import plot_distribution
from src.chromatin_metrics_data import ChromatinDataStore
ind_names = [
"Average $\\Delta$ promoter small fragment\noccupancy ",
"Average $\\Delta$ gene body disorganization\n",
"Average $\\Delta$ gene body nucleosome\noccupancy ",
"Combined chromatin score\n",
]
save_names = [
'small_occ_vs_TPM_120',
'disorg_vs_TPM_120',
'nuc_occ_vs_TPM_120',
'combined_vs_TPM_120',
]
xs = [
# mean measures
datastore.promoter_sm_occupancy_delta.mean(axis=1),
datastore.gene_body_disorganization_delta.mean(axis=1),
datastore.gene_body_nuc_occ_delta.mean(axis=1),
datastore.combined_chromatin_score
]
y = datastore.sense_log2_TPM[120]
for i in range(len(xs)):
ind_name = ind_names[i]
x = xs[i]
ind_title_name = ind_name
save_title_name = save_names[i]
xlabel_name = ind_name.replace('\n', ' ')
plot_distribution(x,
y.loc[x.index],
xlabel_name,
'True log$_2$ transcript level, TPM',
highlight=selected_genes,
xlim=(-3, 3),
ylim=(0, 16),
title=('%svs transcript level @ 120 min' %
ind_title_name),
tight_layout=[0.1, 0.075, 0.9, 0.85],
xticks=(-4, 4, 2),
yticks=(0, 16, 5),
plot_aux='cross')
save_path = '%s/%s.pdf' % (save_dir, save_title_name)
plt.savefig(save_path, transparent=True, dpi=scatter_dpi)
print_fl("Wrote %s" % save_path)
save_path = '%s/combined_vs_TPM.pdf' % (save_dir)
plot_combined_vs_TPM(datastore, selected_genes)
plt.savefig(save_path, transparent=True, dpi=scatter_dpi)
save_path = '%s/disorg_vs_TPM.pdf' % (save_dir)
plot_disorg_vs_TPM(datastore, selected_genes)
plt.savefig(save_path, transparent=True, dpi=scatter_dpi)
save_path = '%s/small_occ_vs_TPM.pdf' % (save_dir)
plot_occ_vs_TPM(datastore, selected_genes)
plt.savefig(save_path, transparent=True, dpi=scatter_dpi)
from src.gp import plot_res_distribution_time
from src.gp import GP
from src.regression_compare import plot_compare_r2, load_results
gp_dir = "%s/gp" % OUTPUT_DIR
results = load_results(gp_dir)
name = 'Full'
save_path = '%s/gp_120.pdf' % (save_dir)
time = 120
cur = GP(name, results_path='%s/%s_results.csv' % (gp_dir, name))
plot_res_distribution_time(cur,
time,
selected_genes=selected_genes,
show_pearsonr=True,
plot_aux='none',
show_r2=False,
tight_layout=[0.1, 0.075, 0.9, 0.85])
plt.savefig(save_path, transparent=True, dpi=scatter_dpi)